Search

Cultivated Cistanche deserticola (known as Rou Cong-Rong) is one kind of herbaceous crop in the arid lands and warm deserts of northwestern China, such as Xinjiang, Inner Mongolia, and Qinghai province. The cash crop is also characterized by its high medicinal and economic value. However, manual picking cannot fully meet the large-scale production at present, due to the time and labor cost. Intellectualized picking has already been the major trend to improve efficiency and labor saving in industrial planting. However, it is difficult to accurately identify and locate cistanche tubulosa during picking, due to the complex environmental factors, such as light, shelter and dense targets. In this study, an improved YOLOv5s model (called YOLOv5s-ESTC) was proposed to increase the recognition accuracy of planted Cistanche deserticola in a complex environment. The lightweight model of object detection was more easily deployed into the mobile terminal and embedded device. Firstly, the backbone network of the YOLOv5s was replaced with the lightweight Eifficientnetv2 using the Fused MBConv module. The performance of the model was optimized to minimize the parameter count, computational complexity, and memory usage. Secondly, the C3STR block was blended in the backbone, due to the integrated C3 and Swin Transformer block. The output feature maps from C3STR were fused in the neck network. The receptive field of Cistanche deserticola was expanded within the Swin transformer layer by layer, in order to leverage the shifted window based on the multi-head selfattention. Therefore, the improved model had effectively captured the subtle features in the image. There was some interaction of texture features between pixel neighborhoods, particularly those associated with the smaller-sized targets. As such, the performance of detection was significantly enhanced for Cistanche deserticola. Lastly, the coordinate attention mechanism was integrated into the network. Meanwhile, the relevant feature of the target channel was strengthened to suppress the invalid background. The ablation experiment was designed to validate the performance of an improved model. The results show that the detection accuracy and precision of the improved YOLOv5-ESTC reached 89.8%, and 92.3%, respectively, which were improved by 1.6 and 2.9 percentage, compared with the original YOLOv5s. The parameters and computational complexity were 5.69× 106 MB and 6.8 GB, respectively, which were reduced by 1.33 MB and 9.1 GB, compared with the original model. The weight of the improved model was 11.9 MB, and the inferred time per image was 8.9 ms, which fully met the real-time detection. A comparative test was designed to verify the detection performance of the improved model. The experimental results show that the detection accuracy of the YOLOv5s-ESTC model was improved by 1.1, 3.0, 10.8, 10.3, 8.5, 1.6, 1.2, 1.4, and 0.5 percent, compared with the mainstream SSD, Faster R-CNN, YOLOv3, YOLOv4, YOLOX, YOLOv5s, YOLOv6s, YOLOv7s, and YOLOv8s models in the same environment, respectively. The improved model accurately detected Cistanche deserticola targets without missing any mistakes, particularly in the complex environment, including small targets, the soil background similar to cistanche tubulosa, and the densely distributed samples. The higher recognition accuracy was achieved in the improved model, compared with the rest using deep learning. The model before and after improvement was also deployed to the mobile terminal device. The YOLOv5s-ESTC model exhibited excellent detection capability on the mobile platform. The detection time was 110.81 ms, indicating a notable improvement of 37.8%, compared with the original model. Moreover, the detection accuracy of the YOLOv5s-ESTC model remained at a high level. In conclusion, the improved model can be expected to effectively detect the Cistanche deserticola targets in natural environments. The finding can provide a valuable reference to develop intelligent harvesting equipment for cistanche cultivation. [ABSTRACT FROM AUTHOR]

Manual picking cannot fully meet the large-scale production in China at present. Robotic picking has been an inevitable trend, particularly with the shortage of labor resources and the rapid development of mechanical automation. It is very necessary to accurately identify and position the apples in the complex environments. Fruit attitude fusion acquisition can be synchronously realized and then classified the apple information. Sometimes, only a small portion of target fruit is covered from the orchard environment, including the leaves, branches, and fruits. There are the small differences among the fruit growth patterns. The convolutional neural network is easy to cause the deep feature map, and then lose the key information of fruit covering parts after multiple convolution operations, resulting in the misrecognition of the fruit growth pattern. At the same time, the detection network can easily identify two apples as one for the overlapping fruits in the natural environment, thus causing the omission of the occluded fruits. In this study, an improved YOLOv7 model was proposed to recognize the apple posture from the growth morphologies. Firstly, the multi-scale feature fusion network was improved to add a 160×160 feature scale layer in the backbone network. The sensitivity of the model was enhanced to identify the tiny local features; Secondly, CBAM attention mechanism was introduced to improve the target region of interest; Finally, the Soft-NMS was used to effectively avoid the high-density overlapping targets being suppressed at one time, thus reducing the missed detection. The experimental results show that the recognition accuracy, recall and average recognition precision of DCS-YOLOv7 were 86.9%, 80.5% and 87.1%, respectively, which were 4.2%, 2.2% and 3.7% higher than the original YOLOv7 model. The average accuracy and speed were greatly improved to fully meet the requirements of picking robot. In addition, an apple gesture recognition was proposed using semantic segmentation and the minimum outer join features. Firstly, comparison tests showed that the Unet model exhibited the best performance in apple image segmentation. The average pixel accuracies were 0.7 and 0.2 percentage points higher than those of DeepLabv3+ and PSPNet. The average intersection and merger ratios were 1.6 and 1.1 percentage points higher as well. The average speed of segmentation also outperformed the rest. As such, the UNet instance segmentation network was chosen as the apple segmentation model. The apple image was segmented using UNet semantic segmentation network. The apple and calyx contour features were obtained by the contour extraction, and then the pose of unobstructed apple was obtained using the apple minimum external feature. The accuracy was 94% to detect the apple pose. The average processing time for each image was 15.7ms, indicating the better acquisition for the pose of apple target. The validity and correctness of recognition model were verified with the high detection accuracy to integrate the recognition of fruit growth pattern and posture. The recognition of fruit posture was considered to classify the growth pattern of apples. The end effector can rapidly and accurately pick the fruits in a suitable way. The finding can lay the foundation for the non-destructive and efficient picking of apples. [ABSTRACT FROM AUTHOR]

To solve the problems of the existing depth acquisition methods with missing data and low resolution, the novel depth enhancement method based on soft clustering was proposed and named the soft clustering solver. By the strong edge-preservation of the soft clustering method, the accuracy of depth enhancement could be improved. The affinity matrix derived from the soft clustering was combined with the weighted least square model to establish the confidence-weighted least square model in the solver, and the iteratively based solution method was proposed. To evaluate the proposed method, the experiments on several depth enhancement tasks of depth inpainting, depth super-resolution and depth rectification were conducted. Various evaluation metrics were used, including peak signal to noise ratio (PSNR), structural similarity index measure (SSIM), rooted mean squared error (RMSE) and running time. The results show that for depth inpainting, the average PSNR reaches 42. 28 with average SSIM of 98.83%. For depth super-resolution and depth rectification, the average RMSE values are 8.96 and 2.36, respectively. By the proposed method, the image with resolution of 2048 x 1024 pixels can be processed with only 5.03 s. [ABSTRACT FROM AUTHOR]

Defects are required to be removed in many rows of dried Hami jujubes. In this study, an online detection and counting system was designed for dried Hami jujubes that were randomly distributed on the conveyor belt. The system was divided into four parts, including a target detector, multi-target tracker, counter, and software interface. The target detector was used to detect the defect categories of dried Hami jujube and then mark the detection box. The multi-target tracker was used to distribute the identity information of dried Hami jujube in the video. The counter was used to count the number of dried Hami jujube in each defect category. The software interface was used to control the program execution. The accuracy of multi-target tracking depended mainly on target detection. The target detector was first designed to compare several models of YOLO series, in order to better detect the defects with high counting accuracy. The better detection of defects was achieved in the YOLOv7 model. After that, the 3×3 CBS convolution layer in the YOLOv7 model was replaced with DCNv2, DWconv, PConv, GSconv, and DSconv. A comparison was made to explore the influence of the improved model on the defect detection of dried Hami jujube. It was found that the DSconv convolution of the improved YOLOv7 model shared fewer parameters, the higher detection speed, and the better accuracy, compared with the original. Four attention mechanisms were compared, including CBAM, ECA, SE, and SimAM. ELAN and SPP layers were also added to the backbone network of the YOLOv7 model. SimAM was performed the best to increase the accuracy of target detection. The ByteTrack multi-target tracker was selected to deal with the high similarity in the different kinds of dried Hami jujubes under the actual environment. The reason was that the appearance information dominated the performance of the multi-target detector for data association. A line-drawing counting was also proposed for statistical counting. The coordinates and ID information were focused at both ends of the counting line to count. Finally, the counting performance was achieved for dried Hami jujube. The software interface was designed using PyQt5 in the practical application. The model was then verified on the detection system of dried Hami jujube. The improved model was also deployed into the detection and counting system for online detection and counting of dried Hami jujube. The experimental results show that: 1) The floating-point computation of the improved YOLOv7 model was 64.6 G, which was 38.6% lower than that of the YOLOv7 model. The mAP, Recall, and F1-score on the test set reached 98.03%, 93.43%, and 95.00%, respectively, which increased by 4.40, 6.88, and 7.00 percent points, respectively, compared with the original. 2) The counting model with ByteTrack as the target detector showed an accuracy rate of 90.12% for the multi-category counting. The finding can also provide technical support for the detection, counting, sorting, and grading of dried Hami jujube. [ABSTRACT FROM AUTHOR]

China is the world’s largest producer and exporter of silk products, as well as a major consumer of silk products. Luster, as one of the main components of the visual style and appearance of silk products, is an important indicator for evaluating the appearance quality of silk fabrics. Therefore, the testing and evaluation of the luster of silk fabrics is of great significance for better style of products and enhancing the visual appearance of the fabrics. The evaluation methods for the luster of fabrics are mainly divided into subjective evaluation methods and instrumental testing methods. Subjective evaluation methods rely on human subjective perceptions with unstable consequences. The instrumental methods can only measure the luster of part surface of the fabrics. When the luster distribution of the fabrics is uneven, the test results are inaccurate. Comparison with other evaluation methods, there are many advantages on the technology of image processing. However, in most studies of luster evaluation based on image processing, the models established can only perform well for specific sample sets with poor generalization abilities. The reason is that the effective features of fabric luster images remain to be further studied in the construction of the key luster features of fabric images. In this paper, white, blue, black, pink and brown silk crepe satin plain fabrics were selected as samples, and the samples were washed by a programmable standard washing machine. The spectral data of the samples were collected by a spectrophotometer, and the subjective scores of the samples were obtained through visual experiments. Exploratory cluster analysis was carried out on the collected data set to obtain the structural characteristics of the data set, and correlation analysis and factor analysis were carried out to explore the influencing factors of the subjective evaluation of fabric luster. The results show that in the color space CIE L*a*b*, a* and b* have a correlation of-0. 45 with the subjective evaluation of samples, while the correlation between color brightness and subjective evaluation is low. In the general linear model analysis of the effects of a* and b*, it is found that the change of fabric color significantly affects the subjective cores of fabric luster. The results of factor analysis show that the change of spectral reflectance of samples has the highest impact on the subjective score of fabric luster in the wavelength range of 530 nm to 560 nm. On this basis, the image analysis of fabric luster was carried out in this paper, and features such as non-uniformity, luster grade and highlight ratio of the image were constructed. These three features can better reflect the contrast characteristics of the highlights and background brightness of the fabric, and it is concluded that the contrast characteristics of the highlights and background brightness of the fabric are the key factors to evaluate the luster of the fabric based on image processing technology. In the human visual system, there is a phenomenon that enhances higher gray level and reduces lower gray level when human eyes are gazing at different grayscale edges. It is the high contrast between the bright spots and the background of silk fabrics that makes silk fabrics have a shiny luster. Compared with neural network or other integrated learning models, this method has more intuitive indexes, and is different from other fabric luster evaluation technologies based on image processing. It does not require learning and training on a specific sample set, so it will have a more promising application prospect.   [ABSTRACT FROM AUTHOR]

Large size laser crystals are a key component of high-power all solid-state lasers. However, during the crystal growth process, defects such as bubbles and inclusions are inevitably generated, which affect the performance of the laser. In order to quickly and effectively evaluate crystal defects, a three-dimensional imaging system for crystal defects based on sheet beam scanning was established in this paper. In the system, a Powell lens is used to shape a 532 nm point laser to obtain a sheetlike laser. The scattered light generated by crystal defects is obtained through a telecentric lens and CMOS detector to obtain the surface distribution of crystal defects. At the same time, high-precision displacement platforms are used to scan the surface array one by one, completing the three-dimensional distribution scanning of the crystal. Furthermore, combined with digital image processing technology, characterize the surface distribution characteristics of crystal defects and reconstruct the volume distribution characteristics of crystal defects. Based on the above system, intrinsic defects in Yb:YAG crystals with dimensions of about 50 mm x 50 mm x 100 mm were measured with a minimum defect detection resolution of 38. 69 μm. This article provides a new method for accurately characterizing the internal defects of crystals, and also provides visual data support for high-precision processing of crystal billets in the later stage. [ABSTRACT FROM AUTHOR]

According to the problem of the current RGB-D feature extraction network used for pose estimation was too large, this paper proposed a lightweight improved XYZNet RGB-D feature extraction network. First, this paper designed a lightweight sub-network BaseNet to replace ResNet18 in XYZNet, which made the network scale significantly reduced and obtained more powerful performance. Then, this paper proposed a re-parameterized multiscale convolutional attention (Rep-MSCA) sub-module based on depth separable convolution, which enhanced the ability of BaseNet to extract contextual information of different scales, and constrained the amount of parameters in the model. Finally, in order to improve the geometric feature extraction ability of PointNet in XYZNet with a small parameter cost, this paper designed a re-parameterized residual multi-layer perceptron (Rep-ResP) module. The floating point operations (FLOPs) and parameters of the improved network are 60.8% and 64.8% lower, the inference speed is 21.2% higher, and the accuracy of the mainstream datasets LineMOD and YCB-Video is 0.5% and 0.6% higher. The proposed model is more suitable for deployment in scenarios where hardware resources are tight. [ABSTRACT FROM AUTHOR]

In the large-scale target detection of remote sensing images based on deep learning, it is difficult to obtain some ground objects, and show poor performance in training results. Therefore, the morphological building index and the generative adversarial network were used for sample augmentation to reduce the problem of model overfitting caused by insufficient detection samples. By extracting the morphological building index related to the texture structure information, superimposing it with the original sample, the texture and spatial characteristics of the building could be strengthened. The existing samples were used to train the generative adversarial network to augment some targets categories. After compositing them with the samples enhanced by the morphological building index, the original sample set was expanded. Compared with the augmentation strategies of flipping, cropping, and changing the color, the detection accuracy of this method got 2%~5% improvement on YLOLOv5, EfficientDet and other models. Experiments have proved that the sample augmentation method combining building index and generative adversarial network can significantly improve the detection accuracy of small-sample remote sensing image targets of special interest categories such as power stations. [ABSTRACT FROM AUTHOR]

To solve the problem of insufficient suppression of halo artifacts in the classic methods, the bilateral weights were embedded into the guided filter to propose a new local method of bilateral weighted guided image filtering. The objective function was constructed as bilateral weighted ridge regression model, which could be solved by the iterative execution of bilateral filter. The fast approximation of bilateral filter was adopted to improve computational efficiency. Through the analysis on the filter kernel, it was shown that the more flexible edge-preserving and filtering control were provided by the proposed bilateral weighted guided filtering than traditional methods. The proposed method was extended for the color image filtering. The experiments on two image processing applications of edge-preserving smoothing and portrait skinning were conducted to evaluate the proposed method. The running efficiencies were compared and analyzed. The results show that the outputs produced by bilateral weighted guided filtering have better visual quality compared with those of most classic edge-preserving filtering methods, and the BIQI, IL-NIQE and SSEQ of the proposed method can reach respective 20.082, 47.026 and 53.103. [ABSTRACT FROM AUTHOR]

Medical image processing technology has developed rapidly with the rise of deep learning. The medical image segmentation technology based on deep learning has become the mainstream method in the segmentation field, which solves the shortcomings of the traditional segmentation method's insufficient segmentation accuracy. This technology has been maturely applied to the segmentation of some pathological images. This study introduces and compares the segmentation methods based on deep learning in recent years, and focuses on the major contributions of U-Net and its improved models in the segmentation field, and summarizes the common medical image modalities and evaluation indicators of segmentation algorithms and commonly used segmentation data sets. Finally, the future development of medical image segmentation technology is prospected. [ABSTRACT FROM AUTHOR]

In recent years, the construction industry has faced the urgent need to replace labor with robots due to rising labor costs and the disappearance of demographic dividends. To address the challenges of feature extraction failure and erroneous extraction in wall slotting robots under complex and variable lighting conditions, an improved multi-scale Retinex algorithm is proposed for image preprocessing. By applying the original multi-scale Retinex to the image's brightness component, the algorithm obtains the brightness increment at each pixel position. The attenuation function is subsequently applied to suppress the brightness increment in high-lighted regions, while simultaneously converting the brightness increment into RGB channel variations. The experimental results demonstrated that the preprocessing algorithm achieved relatively consistent color images under various lighting conditions, effectively suppressed color deviation, and overcame the issues of graying in the single-scale Retinex stage and color distortion in highlight conditions. By employing the target detection and positioning method based on this preprocessing algorithm, the wall slotting robot can accurately detect and position targets even under complex and variable lighting conditions, providing an experimental foundation for the development of wall slotting robots. [ABSTRACT FROM AUTHOR]

Fractional vegetation cover (FVC) is an essential agronomic index. Quick and accurate acquisition of coverage is very important for the real-time monitoring of crop growth status in breeding and precision agricultural management. Image background segmentation is the key step in canopy cover extraction, and accurately segmenting the image from the background can effectively reduce the error of canopy cover extraction. The performance of traditional segmentation methods largely depends on the quality of the training data set, and is easily affected by the changes in segmentation thresholds and lightintensity during the different growth periods of crops, resulting in image segmentation method with low accuracy anduniversality, which ultimately leads to the problem of unsatisfactory vegetation cover extraction. In order to solve the above issues, in this study, Gaussian Mixture Model clustering was proposed using the Lab color space features enhanced by CLAHESV (contrast limited adaptive histogram equalization-saturation value). Taking rice at the late tillering stage as the object, the visible images of rice at 2, 3, 4, and 5 m height were collected by unmanned aerial vehicle (UAV). The saturation (S) and Value (V) components in HSV color space were enhanced by contrast limited adaptive histogram equalization algorithm (CLAHE). Gaussian Mixture Model (GMM) combined with the a-component of Lab color space was applied to segment the image background and extract the rice coverage, and then compared with the GMM-RGB, GMM-HSV, GMM-Lab, and GMM-a. The results show that the two GMM models with the a-component shared a better performance of segmentation than RGB, HSV, and Lab at different heights, where the accuracy of GMM-CLAHE-SV-a was the best. Compared with the GMM-a, the average overall accuracy of GMM-CLAHE-SV-a with image segmentation increased by 2.16, 1.01, 1.03, and 1.26 percentage points, respectively, while the average Kappa coefficient increased by 0.0414, 0.0173, 0.0190, and 0.0221, respectively, at heights of 2, 3, 4, and 5 m; The average extraction error of coverage decreased by 8.75, 7.01, 5.93, and 5.34 percentage points, respectively, whereas, the fitting accuracies were improved by 0.0960, 0.0502, 0.0622, and 0.1906, respectively, at heights of 2, 3, 4, and 5 m. The image segmentation and coverage extraction performance of GMM-CLAHE-SV-a were superior to GMM-a, thus effectively reducing the influence of light intensity and reflection. UAV images can be directly processed without labeling the training set or thresholds. The high universality of the improved model can also be expected to quickly segment the rice pixels and extract the fractional vegetation cover information in complex field environments. [ABSTRACT FROM AUTHOR]

This study proposes a lightweight underwater image restoration method to obtain high-quality biological images and address issues such as color cast and low contrast. The process compensates for color distortion in underwater images through global averaging and max pooling,reducing the influence of irrelevant features in the input images and enhancing the model′s focus on key features. It introduces ghost convolution to reduce redundancy in feature maps and model parameters. Additionally,dynamic convolution is used to adjust the brightness and contrast of images in three color spaces adaptively. The results show that the proposed model outperforms contrast methods based on non-physical and physical models in terms of peak signal-to-noise ratio (PSNR) and structural similarity index ( SSIM), with values of 24. 298 and 0. 891, respectively, on the Enhanced Underwater Vision Perception (EUVP) underwater scene dataset. Compared to the deep learning models Shallow-Uwnet and Water-Net on the EUVP underwater biological image dataset,the proposed model achieves computational efficiency of only 0. 27% and 0. 04%,respectively. The improved images show an SSIM improvement of 3. 77% and 6. 72%,compared to the two aforementioned models. By directly adjusting pixel values to restore underwater image colors,this model occupies only 41 KB of storage space,making it suitable for deployment on underwater robots. This study demonstrates that the proposed model effectively enhances underwater biological images,and due to its low computational load and storage requirements,it contributes to the accurate support of ocean data and promotes the development of marine resources. [ABSTRACT FROM AUTHOR]

Bridge cracks are one of the key factor for the service performance of bridges. Crack detection is crucial to bridge maintenance. At present, most of the bridge crack detection is manual detection. The cracks are usually inspected by personnel visually inspection and manually recording. It is costly in time. Automated algorithms of crack identification based on machine vision are adopted in bridge crack detection by bridge surface images. Recently, the speed and accuracy of automatic crack identification is greatly improved by artificial intelligence technology. The bridge crack detection system based on machine vision and its application were introduced in detail. The advantages of this technology in bridge crack detection and the need for improvement are summarized. [ABSTRACT FROM AUTHOR]

Fish quantity is one of the most meaningful tasks in aquaculture. Precise quantity can greatly contribute to the proper fish density in ponds, thus offering fair fish trade to sales and customers. However, fry (the juvenile fish) counting is still longterm hard work, due to the tiny size and huge batch in practice. This work aims to explore an efficient way for a precise number of fries. A self-built image acquisition device was designed and then applied by image segmentation using edge curvature. A whole closed space was constructed underneath with a white light background in the device. A camera inside was equipped downward to collect images of fries. The fries were in the size of 1-3 cm length and then placed in a 40 cm×30 cm×6 cm transparent box inside 0.9-1.1 cm depth water. Totally 18 000 fry images were acquired, which were of 6 planar density classes, including 0.25, 0.33, 0.42, 0.50, 0.58 and 0.67 fry/cm2 . The preprocessing was carried out to reduce the negative and redundant information in the collected images, such as the noises and the background. Therefore, the images were transferred to the binary images, and then the morphology operations like dilation and erosion were completed next. As the least self-contained unit in a binary image, the individual connected domains were extracted and then saved for the following deeper analysis, as well as their skeleton maps and edge maps. Among them, an individual connected domain was distinguished as a single fry or an adhesion of multiple fries. This identification was determined by two keys-one was the number of branch points of the skeleton map, and another was the distance from the points of the skeleton to the edge. Those domains that were recognized as single fry were directly counted as 1 into the amount. The number of fries needed to be reconfirmed in the domain with an adhesion of multiple fries. The detailed steps were listed as follow: 1) To calculate and filter, the curvature of the edge was purposed to search the points whose quantitative value always were negative, those points were named after ‘pit point’ and thought of as the places that overlaps appeared; 2) If the pit points of the cluster were adjacent and concentrated in a community, those pit points were marked as the pre-segmenting locations, while those locations were named as ‘pit point group’, and all of pit point group on an edge was found and numbered; 3) According to the different pre-segmenting locations, all forms of overlap were classified as three types, including crossed-fries, cascaded-fries, and single-pit cluster. Then potential overlaps were determined; 4) In each overlap, two highlighted segmentation points were picked out from the potential pit point groups, and then linked as a segmentation line to depart the original connected domain; 5) All segmentation lines were tested by presupposed rules to decide whether being saved or not. 6) The number of fries was set as the number of separate domains until all segmentation was realized in a connected domain. Finally, the fry number was calculated as the sum of the number from all domains in a complete image. The results show that the mean absolute percentage error (MAPE) was less than 1.37% after detection when the fry plane density was below 0.58 fry/cm2 . Meantime, both accuracy and efficiency were considered concurrently, if the plane density was kept within 0.33-0.58 fry/cm2 . The overlapping images of high-density fry can be expected to accurately and rapidly count over 1 cm length of fry. The finding can provide algorithmic support for the development of fry counters. [ABSTRACT FROM AUTHOR]

Efficient and accurate identification and position of compound buds can greatly contribute to the orchard robots for automatic bud removal operations, thereby improving the efficiency of flower and fruit thinning. This study aims to accurately identify and position the small objects in complex backgrounds for grape bud removal in the field. (1) A visual recognition system was proposed for the concurrent bud and secondary bud using a combination of far and near visual. Euclidean distance was applied on the visual mapping plane. The concurrent bud was then identified in the far view, whereas, the secondary bud was identified in the near view within the concurrent bud mapping area, in order to improve the recognition accuracy of small target objects in the field. The original images (2 132 images) were randomly divided into the training set (1 735 images) and validation set (397 images) in the ratio of 8:2. The training set and validation set were uniformly cropped to 2944×1656 pixel size, and then scaled to 1280×720 pixel size with 2.3 times of equal proportion. Data augmentation was carried out by three random combinations of augmentation through scaling, flipping, and color gamut transformation for the concurrent and secondary bud datasets. Once the augmentation scale was too large for the target features, the images were manually excluded from the dataset. As such, the training set was enlarged to 15 840 images. YOLOv5m and YOLOv5s were selected as the network models for juxtaposed bud and parabudding detection. The sizes of the trained models were 42.1 and 14.4 MB, respectively, with AP of 0.702 and 0.773, F1 scores of 0.685 and 0.765 on the test set of concurrent and secondary bud images, respectively, where the average inference time per image was 10.62 and 7.01 ms. 60 compound bud images showed that the overall average accuracy of this improved model was 0.905, and the average detection time was 18.1 ms for each process; (2) An online detection and position were then proposed for the compound bud in the field. The online detection and position were introduced in the prediction stage to improve the light-resistant ability of the depth camera when positioning in a complex environment, in order to reduce the fluctuation and error of the depth information. The same compound bud was online detected and positioned for 5 consecutive frames. The initial frame and the latest frame image data were deleted and updated in real time, until the relative error of positioning coordinates in 5 consecutive frames was less than the error threshold, in order to obtain the average positioning coordinates of the compound bud. Better performance was achieved in the positioning accuracy of compound bud in natural background and the resistance to light coupling in complex light environment. The test results showed that the positioning accuracy of the improved model was ±0.916 and ±0.654 mm for the concurrent and secondary bud, respectively. The repeatability and accuracy can fully meet the demand for the secondary bud wiping, indicating a reliable and effective system. The finding can provide high-quality technical support for the online positioning of the concurrent and secondary buds for sprout wiping operation in sprouting robots, in order to realize the automatic sprout wiping in orchards. [ABSTRACT FROM AUTHOR]

An method of the line-structured light stripe center extraction based on the matching of adjacent pixels of skeleton is proposed. The traditional method is improved in the phase of image preprocessing and light stripe center extraction. In the phase of image preprocessing, Markov's random field theory is applied to denoise binary images, and a method of extracting region of interest (ROI) based on the area characteristics of connected regions is proposed. In the phase of light stripe center extraction, firstly, a light stripe skeleton pruning method is proposed to prune and smooth the stripe skeleton obtained by the refined ROI. Secondly, considering the geometric and gray distribution characteristics of the light stripe image, the pixels in the ROI are divided based on an analysis of adjacent areas, and then the gray barycenter is obtained. Finally, Savitzky-Golay filter is used to extract the center of light stripe with sub-pixel precision. Experimental results show that the proposed method has strong applicability for different types of light stripe extraction and achieves higher precision compared with Steger method. And it's worth noting that the speed of the proposed method is about 6.98 times higher than that of the Steger method. [ABSTRACT FROM AUTHOR]

An accurate and rapid detection of eggshell damage has been one of the most important steps in the field of egg processing. Eggshells are prone to mechanical impact and damage during laying and processing, leading to the leakage of egg contents. It has also posed serious bacterial infections and food safety risks on the rest of the intact eggs in the processing line, particularly for the cleaning work and production costs. Furthermore, the current manual detection relies mainly on the subjective experience with slow speed and low accuracy. Therefore, it is of great practical significance to rapidly, accurately, and low-costly detect broken shell eggs in egg processing. In this study, an online real-time detection system was proposed for the broken shell eggs using the improved YOLOv7 (you only look once) model. Significant differences were also presented in the defects between broken shell eggs. Specifically, the YOLOv7 network was selected to replace the loss function CIoU (complete-IoU) with WIoUv2 (wise-IoU). The coordinate attention (CA) modules were added into the deformable convolutional DCNv2 (deformable convnet) in the backbone network. At the same time, the detection head (IDetect) module was replaced in the YOLOv7 network with a decoupled detection head (IDetect-Decoupled) with implicit knowledge learning. Among them, the CA module was used to effectively identify and locate targets of interest for optimal model performance, where the deformable convolution was combined to decouple the detection heads. WIoU loss functions were used to learn the local and global information while improving the accuracy and speed of shell-breaking egg detection with the less missed and false detections. The generalization and robustness of the model were further improved using data augmentation. Various operations were performed on the training set, such as random cropping, 50% probability of random horizontal and vertical flipping, 50% probability of random rotation and scaling, random noise, brightness and color adjustments. The dataset was screened to remove the unqualified data after data augmentation. Finally, a final dataset was obtained with 1 729 images, including 3 782 intact and 1 574 cracked eggs. The experimental results on the PC side show that better performance was achieved in the improved model, with an average precision (mAP) of 94.0% on the test set. The single image detection time of 13.1 ms was 2.9 percentage points higher than before, where the detection time was only extended by 1.0 ms. The parameter quantity of the improved model was 3.64×107, indicating a decrease of 2.1 percentage points, compared with the original model. In addition, the improved YOLOv7 model enhanced the mAP of broken shell eggs by 5.4, 3.4, and 4.5 percentage points, compared with the SSD, Faster R-CNN, and YOLOv5 models, respectively. The best detection accuracy was achieved for the broken shell eggs, indicating the high effectiveness of the improved model. Finally, the model was also deployed to the device side through the format conversion. The ONNXRuntime deep learning was utilized to conduct the online detection and verification under its inference framework. The improved model was reduced by 3.8 percentage points on the false detection rate of broken eggs, whereas, the missed detection rate remained unchanged. The average frame rate was about 54 frames/s for online detection. The strong robustness of the improved model fully met the requirements of online real-time detection. This finding can provide the technical reference for the online detection of broken shell eggs. [ABSTRACT FROM AUTHOR]

Potatos are one of the most important food crops to maintain food security and stability in the world. However, the manual cutting of seed preparation has seriously restricted the development of the potato planting industry, due to the labor)intensive and high labor costs. It is in high demand to rapidly and accurately detect the seed buds for the intelligent cutting of potato seeds. The performance of detection can also directly dominate the quality of potato seed after cutting, which in turn determines the total yield and economic benefits. In this study, a target detection model was proposed for the potato seed buds using improved YOLOv7. The large-scale preparation was then realized to detect the small area occupied by the seed buds and a few extractable features under the complex background of the seed surface in the potato seed-cutting machinery. Firstly, a Contextual Transformer of self-attention mechanism was added into the Backbone part, in order to enhance the target objects and the removal of redundant background. Different weights were also given to the target objects and background region. Secondly, the InceptionNeXt module was selected to replace the original ELAN-H module in the Head part. The loss of high dimensional features was reduced, particularly for the potato seed buds with the increase of network depth. Multi-scale fusion was better performed to improve the detection of potato seed buds. Finally, the bounding box loss function was changed to normalized wasserstein distance(NWD), in order to reduce the loss value and speed up the convergence of the network model. The samples were selected as well-stored surfaces free of insect pests, dry rot and disease spots. The reason was that the quality of training samples depended mainly on the detection of the potato seed buds. Meanwhile, the additional samples with surface damage and soil were utilized for the diversity of the dataset. The total number of 500 samples was obtained in this case. The front and back sides of potato seed were captured at a distance of 30 cm from the samples using a CCD camera. A total of 1 000 JPEG format images were randomly divided into the training set (800 images), validation set (100 images), and test set (100 images), according to the ratio of 8:1:1. At the same time, data augmentation was performed on the dataset, including the mirror, rotate, cropping, brightness adjustment, and the noise addition into the images. Model generalization was then improved beyond the insufficient number of samples. The experimental results indicated that the mean average precision (mAP) of the improved YOLOv7 model reached 95.40%, which was 4.2 percentage points higher than the original. The detection accuracies of improved model were 34.90, 26.32, 27.25, 22.88, 35.92, 17.23, 15.70 percentage points higher than those of similar target detection models, such as Faster-RCNN(ResNet50), Faster-RCNN(VGG), SSD, YOLOv3, YOLOv4, YOLOv5, YOLOX, respectively. Therefore, the missed detection rates of the improved model were 4% and 11% for the potato seed with smooth and soil or damaged surface, respectively, indicating better detection than the rest in the actual application. The finding can provide strong support to the recognition of potato seed buds for the intelligent cutting of potato seed. [ABSTRACT FROM AUTHOR]

In view of the problem that it is difficult to ensure high efficiency and high accuracy in circle fitting algorithm in the same time, a circle fitting algorithm based on multi-level optimization is proposed in this study. The 3σ criterion is used to remove the coarse error points, and the random sampling consistency is improved by reducing the randomness of subset selection, the descending operation of the circle model and the threshold transformation of the number of adaptive iterations, so as to extract the high quality internal group points. The iterative weighted least square method with iteration termination condition is applied to achieve the great processing of point groups. In this study, the effectiveness of the algorithm is verified from the three aspects of defect circle, impurity interference and other noise, and the proposed method is compared with other mainstream circle fitting algorithms. The results show that the fitting accuracy of the proposed algorithm is less than 0.7 pixels under different degrees of circular defect and impurity interference, and the algorithm performs better than other algorithms in fitting effect. In the case of the noise interference of about 20%~265%, the fitting accuracy of the algorithm does not exceed 1 pixel, and the running time is less than 0.7 s. These results indicate that the proposed algorithm can resist a large number of salt-and-pepper noise interference and maintain high accuracy and detection efficiency. [ABSTRACT FROM AUTHOR]

Grape has been one of the most popular fruits with great nutritional value and economic benefits. Manual picking of mature grapes cannot fully meet the large-scale production in recent years, particularly with the expansion of planting areas. A picking robot can be expected to monitor the growth of grapes in orchards in real time. Automatic grape picking can also be promoted to realize intelligent agricultural production. In this study, an improved YOLOv5s model (MRW-YOLOv5s) was proposed to rapidly and accurately identify the grapes in orchards. Firstly, the lightweight network MobileNetv3 was used as the feature extraction network, in order to reduce the amount of model parameters. A coordinate attention module (CA) was also embedded into the bneck structure of MobileNetv3 to strengthen the feature extraction capability of the network. Secondly, RepVGG Block was introduced into the neck network, where the multi-branch features were integrated to improve the detection accuracy of the model. Moreover, the structural reparameterization of the RepVGG Block was implemented to further accelerate the inference speed of the model. Finally, Wise Intersection over Union Loss (WIoU Loss) with the dynamic nonmonotonic focusing mechanism was taken as the bounding box regression loss function, in order to accelerate the network convergence for the better detection accuracy of the model. Gradient-weighted class activation mapping (Grad-CAM) was also selected to capture the grape targets when the backbone network of the improved model was embedded with the CA module. A better performance was then achieved, compared with the model embedded with Efficient Channel Attention (ECA) and Convolutional Block Attention Module (CBAM). In addition, there was the lowest speed of bounding box loss regression in the convergence curve of the loss function, while the highest loss value after convergence, where the EIoU was the bounding box loss function. Once the CIoU and Wise-IoU v1 were taken as the bounding box loss functions, there were similar convergence speeds and loss values, indicating a slightly lower value than that of EIoU. Moreover, there was the highest convergence speed of the model and the lowest loss value, when the Wise-IoU v3 was used as the bounding box loss function. Therefore, the Wise-IoU v3 can be expected to accelerate the convergence of the model for better accuracy of the model. The results showed that the number of parameters of the improved MRW-YOLOv5s model was only 7.56 M. The mean Average Precision (mAP) on the test set reached 97.74%, and the average detection time per image was 10.03 ms, which were 2.32 percentage points and 6ms higher than those of the original YOLOv5s model, respectively. The mAP of the MRW-YOLOv5s model was 9.89, 7.53, 2.12, 0.91, and 2.42 percentage points higher, respectively, compared with the mainstream object detection models, such as the SSD, RetinaNet, YOLOv4, YOLOv7, and YOLOX. In terms of the number of model parameters, the improved model was only 7.56 M, which was 68.2%, 79.2%, 88.2%, 79.7%, and 15.4% less than the above five models, respectively. The average detection speed of the improved model was only 10.03 ms, which was 2.64, 13.19, 10.59, 4.14, and 5.46 ms higher than the above five models, respectively. Furthermore, the weight size of the improved model was only 26.97 MB, which was more conducive to model deployment. Therefore, the MRW-YOLOv5s model can greatly contribute to the detection accuracy, parameter size, and detection speed. The finding can also provide technical support for the intelligent orchards and mechanization of picking. [ABSTRACT FROM AUTHOR]

The root is one of the most crucial parts of the plant to affect the overall healthy plant. The leaf can be the corresponding indicator at various stages of growth, even in the ultimate crop. However, some significant challenges still remain in the in-situ detection of soil roots. In this study, an in-situ measurement approach was presented for the tomato root phenotypes using the array ESP32-camera (CAM). Tomato root images were also captured for the in-situ measurement of root phenotypic parameters. A 4×4 array ESP32-CAM combined with 4×4 OV2640 lens module was used for the in-situ automatic wireless acquisition of soil root images. The camera was calibrated and corrected for the aberrations using the Zhang Zhengyou calibration. The image alignment was achieved using scale-invariant feature transform and K-nearest neighbor feature detection matching. The image stitching was obtained for the inter-camera transformation matrix using offline calibration. The semantic segmentation of root images was improved to introduce the efficient multi-head self-attention mechanism. The U-Net model was improved to mix the dice loss and cross-entropy loss. In-situ image acquisition experiments of tomato root systems were also conducted to obtain the images of periodic root changes. The root system was measured manually, i.e., the root length was measured in sections by a soft ruler, the root diameter was obtained by averaging the points, the root area was approximated by the product of length and average diameter, and the root depth and width were obtained by measuring the vertical longitudinal depth and horizontal longitudinal width. Morphological processing and skeleton extraction were used to measure the root length, root area, and root mean diameter using pixel point scanning. By contrast, the root phenotypic parameters (such as the root depth and root width) were measured using the root convexity package. The results showed that the root mean square error (RMSE) of standard shape image stitching iterations of 4×4 camera array was less than 1.11 mm, and the subjective quality scores of the structural similarity index and difference of edge map of global stitched images were above 0.85, the peak signal-to-noise ratio (PSNR) was greater than 32dB. The improved U-Net model shared the greater improvement in the search for root system completeness and accuracy. Specifically, the precision, recall, intersection over the union, and F1 value of the improved U-Net model on the tomato root segmentation were 86.06%, 78.98%, 71.41%, and 82.37%, respectively, which were 18.97 percentage points, 13.21 percentage points, 21.67 percentage points, and 16.30 percentage points higher than those of the original U-Net model training, respectively. Furthermore, the mean absolute percentage errors of calculated area, length, mean diameter, root depth, and root width of root systems were 7.78%, 5.66%, 8.48%, 2.40%, and 2.23%, respectively, and the coefficients of determination were 0.91, 0.93, 0.84, 0.98, and 0.99, respectively, compared with the manual measurements. Insitu measurement of the root phenotype of tomato plants can also be extended to the other plants or fruit trees. [ABSTRACT FROM AUTHOR]

An intelligent monitoring system was developed to continuously, rapidly and accurately identify the grassland vegetation cover and the key phenological periods. The fixed monitoring, mobile monitoring and cloud platforms were also combined to provide a solution. The innovative system consisted of three main components: the fixed and mobile monitoring subsystem, as well as the cloud platform of grassland phenology intelligent monitoring. The fixed monitoring subsystem included various components, such as phenology cameras, power modules, communication modules, edge computing controllers, and supporting poles. The mobile monitoring subsystem consisted of handheld devices with capabilities, such as dustproof, waterproof, and shockproof, along with the necessary applications. The survey plot frames of the grassland field were required for the on-site image collection. The cloud-based grassland phenology intelligent monitoring platform was designed with a browser/server architecture, and then deployed on the Alibaba Cloud’s lightweight application server. As such, the platform was connected seamlessly to a MySQL database, with a user-friendly front-end built on the LayUI framework and a robust back-end using the Spring Boot framework. The efficient storage of collected images and periodic processing was obtained from the comprehensive data, such as the grassland vegetation index, vegetation cover, and phenological parameters. The cloud platform offered a wide range of functionalities, including data querying, processing, displaying, and sharing. Both the fixed and mobile monitoring subsystem were utilized to capture the grassland vegetation images. The fixed and mobile monitoring subsystem were operated and uploaded independently to the cloud platform using a cloud server. A series of processing steps were used to extract the regions of interest, and automatically calculate the grassland vegetation index using advanced techniques of image processing. The vegetation index was then converted into a grayscale space for the singlechannel grayscale image. Binarization segmentation was performed to accurately extract all vegetation pixels using the maximum interclass variance method, according to the grayscale value differences between vegetation and non-vegetation areas. This process enabled the system to identify the grassland vegetation cover, and the obtained data was stored in a dedicated database. A double logistic function was used to fit the time series of the vegetation index for the vegetation growth pattern. A precise curve of vegetation growth was formed after fitting. Furthermore, a dynamic threshold method was selected to extract the important phenological parameters suitable for the growth pattern of black oatgrass using TIMESAT software. The innovative system demonstrated that an impressive accuracy rate of 90% was achieved for the grassland vegetation cover identification using the combination of the excess green index (EXG) with the maximum interclass variance method. This performance fully met the high demands of automated and bulk extraction of grassland vegetation cover. Additionally, the extraction of vegetation indices with GCC, EXG, and NGRDI, followed by fitting the vegetation growth curve with a double logistic function, can be expected to accurately identify the crucial phenological key periods, in accordance with the growth patterns of black oatgrass. The system can provide reliable technological and data support for the intelligent monitoring and management of grassland vegetation. [ABSTRACT FROM AUTHOR]

An accurate detection has been a great challenge at the varying maturity stages of lotus seedpods within the dynamic field environments. In this study, a groundbreaking approach was proposed to robustly detect and distinguish the maturity levels of lotus seedpods. The improved YOLOv5s model was also selected for computer vision and precision agriculture. The Bottleneck Transformer (BoT) self-attention mechanism module was first incorporated into the core feature network. The global and local features were then integrated into the mapping structure. This fusion also amplified the discernibility between different maturity stages of lotus seedpods under the ever-changing visual conditions within the lotus fields. Intersection over Union (EIoU) loss function was used to enhance the precision of bounding box regression. The improved model was then optimized to precisely delineate the contours and boundaries of lotus seedpods. A substantial elevation of detection was found in precision agriculture. K-means++ clustering was introduced to enhance the network convergence, in order to recalibrate the calculation of initial anchor box sizes. A more accurate and efficient learning process was then achieved to ultimately enhance the performance of the improved model more accessible for practical applications. The results show that outstanding performance was achieved in the improved YOLOv5s model on the test dataset, with the boasting precision (P) rates of 98.95%, recall (R) rates of 97.00%, and an impressive mean average precision (mAP) score of 98.30%. Notably, the average detection time remained at 6.4 ms, with a compact model size of 13.4 megabytes. The improved YOLOv5s model performed better than most detection models, including YOLOv3, YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, and YOLOv7. There were substantial improvements of 0.2%, 1.8%, 1.5%, 0.5%, and 0.9% in mAP, respectively, indicating better detection of the lotus seedpod maturity. Furthermore, a real-time detection system of lotus seedpod maturity was established to integrate the improved YOLOv5s model into a Raspberry Pi 4B mobile controller. Additionally, four distance ranges were tested to verify the improved model. The better performance was obtained in the 0.5-meter to 1.0-meter detection ranges, compared with the YOLOv5 and YOLOv3-tiny models. This YOLOv5s-based approach can be expected to detect the maturity of lotus seedpods for the development of intelligent harvesting equipment. The applicability of the improved model can also be extended to the defying gravity during growth for sustainable and efficient crop management in the frontiers of precision agriculture and industry practices. In summary, modern computer vision can greatly contribute to agricultural needs, such as the improved YOLOv5s model. The accurate and rapid detection of lotus seedpod maturity was achieved for better efficiency, sustainability, and productivity in precision agriculture and smart farming, as well as the crop maturity detection in gravity-defying plants. [ABSTRACT FROM AUTHOR]

The number of trees is essential information for modern forestry management, which affects managers' formulation of development strategies. At present, tree inventory mainly relies on manual counting, which is costly, time-consuming and labor-intensive. Using unmanned aerial vehicle (UAV) to count targets has become increasingly popular in agriculture and forestry due to its low cost, ease of operation, and flexibility of use. This study focused on using UAV images to count spruce numbers. A total of 603 images with an average canopy density of 81% and an average plant density of 6 667 plants/hm² were selected for the dataset, each containing an average of 354 spruce trees. Among these, 205 images had interference factors such as weeds. The images were divided into a training set and a test set at a ratio of 7:3. The training set was expanded by randomly flipping the images to improve the robustness to different flight attitudes of UAV. After data augmentation, 844 training images and 181 test images were obtained. Aiming at the problem of dense spruces in natural environments including severe adhesion and background interference such as weeds with similar characteristics to spruces, we selected IntegrateNet, a model known for its strong performance in dense target counting tasks as baseline model for spruces. This study then worked to improve the IntegrateNet model to achieve a more accurate counting of dense spruce that is closer to real-world conditions. First of all, this paper used the self-calibrated convolutions (SCConv) to replace the ordinary convolutional layers at 1/8, 1/16 feature maps and density maps in the baseline model to expand the convolution receptive field. Secondly, in order to deal with background interference problems such as weeds and the serious adhesion of the target, this paper added the criss-cross attention mechanism (CCA) to the feature fusion of the IntegrateNet model. It can consider the horizontal and vertical context information of each pixel to generate richer semantic features to improve the contextual information extraction ability of the model. The mean counting accuracy(MCA), mean absolute error(MAE), root mean square error(RMSE), and the coefficient of determination R² are used as evaluation indicators. The ablation experiments and comparative experiments are designed to verify the performance of the proposed model. The ablation experiments show that the improved methods proposed in this paper can effectively improve the counting accuracy of the model. The comparative experiments show that the improved IntegrateNet model proposed in this study has MCA, MAE and RMSE reached 98.32%, 8.99 plants and 13.79 plants respectively, and the R² was 0.99.Compared with TasselNetv3_lite, TasselNetv3_seg, and IntegrateNet models, the MCA of the improved IntegrateNet model increased by 16.44, 10.55, and 9.26 percentage points, the MAE of the improved IntegrateNet model decreased by 25.62, 10.45, and 6.99 plants, and the RMSE of the improved IntegrateNet model decreased by 48.25, 13.84, and 12.52 plants. In summary, the improved IntegrateNet model proposed in this study demonstrated a significant increase in accuracy for counting dense spruce in natural environments. Moreover, our research is not limited to spruce, as the methodology can be applied to other tree species as well. It can providing technical support for subsequent forestry intelligent statistics. [ABSTRACT FROM AUTHOR]

[Objective] Wheat, as one of the major global food crops, plays a key role in food production and food supply. Different influencing factors can lead to different types of wheat lodging, e.g., root lodging may be due to improper use of fertilizers. While stem lodging is mostly due to harsh environments, different types of wheat lodging can have different impacts on yield and quality. The aim of this study was to categorize the types of wheat lodging by unmanned aerial vehicle (UAV) image detection and to investigate the effect of UAV flight altitude on the classification performance. [Methods] Three UAV flight altitudes (15, 45, and 91 m) were set to acquire images of wheat test fields. The main research methods contained three parts: an automatic segmentation algorithm, wheat classification model selection, and an improved classification model based on EfficientNetV2-C. In the first part, the automatic segmentation algorithm was used to segment the UAV to acquire the wheat test field at three different heights and made it into the training dataset needed for the classification model. The main steps were first to preprocess the original wheat test field images acquired by the UAV through scaling, skew correction, and other methods to save computation time and improve segmentation accuracy. Subsequently, the pre-processed image information was analyzed, and the green part of the image was extracted using the super green algorithm, which was binarized and combined with the edge contour ex‐ traction algorithm to remove the redundant part of the image to extract the region of interest, so that the image was segmented for the first time. Finally, the idea of accumulating pixels to find sudden value added was used to find the segmentation coordinates of two dif‐ ferent sizes of wheat test field in the image, and the region of interest of the wheat test field was segmented into a long rectangle and a short rectangle test field twice, so as to obtain the structural parameters of different sizes of wheat test field and then to generate the da‐ taset of different heights. In the second part, four machine learning classification models of support vector machine (SVM), K nearest neighbor (KNN), decision tree (DT), and naive bayes (NB), and two deep learning classification models (ResNet101 and Efficient‐ NetV2) were selected. Under the unimproved condition, six classification models were utilized to classify the images collected from three UAVs at different flight altitudes, respectively, and the optimal classification model was selected for improvement. In the third part, an improved model, EfficientNetV2-C, with EfficientNetV2 as the base model, was proposed to classify and recognized the lodg‐ ing type of wheat in test field images. The main improvement points were attention mechanism improvement and loss function im‐ provement. The attention mechanism was to replace the original model squeeze and excitation (SE) with coordinate attention (CA), which was able to embed the position information into the channel attention, aggregate the features along the width and height direc‐ tions, respectively, during feature extraction, and capture the long-distance correlation in the width direction while retaining the longdistance correlation in the length direction, accurate location information, enhancing the feature extraction capability of the network in space. The loss function was replaced by class-balanced focal loss (CB-Focal Loss), which could assign different loss weights accord‐ ing to the number of valid samples in each class when targeting unbalanced datasets, effectively solving the impact of data imbalance on the classification accuracy of the model. [Results and Discussions] Four machine learning classification results: SVM average classification accuracy was 81.95%, DT aver‐ age classification accuracy was 79.56%, KNN average classification accuracy was 59.32%, and NB average classification accuracy was 59.48%. The average classification accuracy of the two deep learning models, ResNet101 and EfficientNetV2, was 78.04%, and the average classification accuracy of ResNet101 was 81.61%. Comparing the above six classification models, the EfficientNetV2 classification model performed optimally at all heights. And the improved EfficientNetV2-C had an average accuracy of 90.59%, which was 8.98% higher compared to the average accuracy of EfficientNetV2. The SVM classification accuracies of UAVs at three flight altitudes of 15, 45, and 91 m were 81.33%, 83.57%, and 81.00%, respectively, in which the accuracy was the highest when the altitude was 45 m, and the classification results of the SVM model values were similar to each other, which indicated that the imbal‐ ance of the input data categories would not affect the model's classification effect, and the SVM classification model was able to solve the problem of high dimensionality of the data efficiently and had a good performance for small and medium-sized data sets. The SVM classification model could effectively solve the problem of the high dimensionality of data and had a better classification effect on small and medium-sized datasets. For the deep learning classification model, however, as the flight altitude increases from 15 to 91 m, the classification performance of the deep learning model decreased due to the loss of image feature information. Among them, the classification accuracy of ResNet101 decreased from 81.57% to 78.04%, the classification accuracy of EfficientNetV2 decreased from 84.40% to 81.61%, and the classification accuracy of EfficientNetV2-C decreased from 97.65% to 90.59%. The classification accuracy of EfficientNetV2-C at each of the three altitudes. The difference between the values of precision, recall, and F1-Score results of classification was small, which indicated that the improved model in this study could effectively solve the problems of unbalanced model classification results and poor classification effect caused by data imbalance. [Conclusions] The improved EfficientNetV2-C achieved high accuracy in wheat lodging type detection, which provides a new solu‐ tion for wheat lodging early warning and crop management and is of great significance for improving wheat production efficiency and sustainable agricultural development. [ABSTRACT FROM AUTHOR]

To address the problems of reduced feature map resolution in deep convolutional neural networks, which leads to poor performance in detecting small changes in remote sensing images and difficulty in effectively distinguishing external interference to produce false changes, a Siamese attentiongated fusion encoding-decoding network for remote sensing image change detection is proposed. A triple attention network module is introduced in the encoding part to further solve the problem of false changes in the change detection image. An attention-gated fusion module is proposed to selectively fuse features from multiple levels. A deep supervision strategy is directly introduced in the decoding part to enhance the feature extraction capability of the change detection network. The effectiveness of the proposed network is verified through experiments. [ABSTRACT FROM AUTHOR]

Rapid and accurate segmentation of COVID-19 lesions from CT images is an important step to realize computer-assisted diagnosis and treatment of COVID-19. Therefore, a CT image segmentation method of COVID-19 lesions based on dual-attention mechanism is proposed. Firstly, the attention gate module is introduced to enhance the focus on the focal region in space and reduce the influence of image brightness imbalance and low contrast on the segmentation accuracy. Secondly, the SE-Res module combined with residual element was introduced to enhance the channel of the lesion region, extract the fine structural features, and improve the segmentation performance of the network for the lesion shape change and ground glass boundary region. Experiments on the public datasets of COVID-19 show that the Dice, PPV and IoU achieved by the proposed method are 0.908 8, 0.915 2 and 0.858 9, respectively, which are 0.75%, 0.11% and 0.65%higher than previous studies, respectively. The proposed method can improve the segmentation accuracy of the lesions with large shape changes and ground glass boundaries, and the overall performance is better than current mainstream models. [ABSTRACT FROM AUTHOR]

Accuracy and efficiency of strawberry picking depend mainly on the recognition of strawberry fruit and segmentation of fruit stem in the elevated cultivation environment. In this study, an improved YOLOv5s recognition (ATCSPYOLOv5s) was proposed for the elevated strawberry fruits under elevated cultivation environment, while a Yolov5s-seg segmentation model was applied to detect and segment fruit stems. The detection accuracy of the recognition model was attributed to the improved backbone network of YOLOv5s. The self-attention mechanism was introduced to establish the selfattention cross-stage feature fusion network. This innovative structure was focused on the correlation of inherent features without the dependence on the external cues, thus improving the ability of the network to extract target features. The better performance was achieved to identify the small targets, blocked or overlapping strawberry fruits. The detection and segmentation of strawberry stem included the backbone network, feature fusion network, recognition head, and segmentation head of segmentation module. The mask information of strawberry and stem was captured at the same time. The interest region was determined to contain the target fruit stem. Then, the accurate segmentation of strawberry stem was effectively realized in the overhead cultivation using linear combination of recognition and segmentation. The experimental results showed that the precision (P), recall (R), and mean average precision (mAP) of strawberry fruit recognition by ATCSP-YOLOv5s model were 97.24%, 94.07%, and 95.59%, respectively. The detection speed was 17.3 frames per second. The P, R and mAP values of strawberry frontlight processing on sunny days by ATCSP-YOLOv5s were 97.89%, 95.16% and 95.71%, respectively, which were 3.06, 6.22 and 4.48 percentage points higher than those by YOLOv5s. The P, R and mAP values of strawberry backlight processing on sunny days were 96.53%, 93.97% and 95.34%, respectively, which were 3.26, 7.43 and 4.6 percentage points higher than those of YOLOv5s. The P, R and mAP values of strawberry image processing on cloudy days were 96.93%, 94.88% and 95.56%, respectively, which were 2.30, 7.01 and 4.50 percentage points higher than those of YOLOv5s. The performance of ATCSP-YOLOv5s recognition model was compared with that of Faster RCNN, YOLOv4, YOLOv5s, YOLOv6s, YOLOv7, and YOLOv8. The experimental results showed that the P values of ATCSP-YOLOv5s recognition model was 6.73, 5.92, 4.96, 4.77, 3.35 and 3.43 percentage points higher than Faster RCNN, YOLOv4, YOLOv5s, YOLOv6s, YOLOv7 and YOLOv8, respectively. The R values were higher by 11.68, 4.75, 7.13, 3.82, 3.76 and 3.01 percentage points, respectively. The mAP values were 8.64, 7.03, 4.53, 4.27, 4.31, and 3.55 percentage points higher, indicating the better recognition performance. In addition, the mAP of the ATCSP-YOLOv5s recognition model was above 90%, when detecting strawberry images under different lighting conditions, indicating the better robustness. The P, R and mAP values of all images that processed by the YOLOv5s-seg segmentation model were 82.74%, 82.01%, and 80.67%, respectively. The P values of strawberry images that processed by the YOLOv5s-seg segmentation model on sunny frontlight, sunny backlight, and cloudy days were 85.32%, 81.26%, and 81.89%, respectively. The R values were 83.65%, 82.03% and 83.20%, respectively, while the mAP values were 82.31%, 81.53% and 82.04%, respectively, and the segmentation accuracy was 98.29%, indicating the better segmentation accuracy and universality. The comprehensive experiment showed that the ATCSP-YOLOv5s recognition and YOLOv5s-seg segmentation model can be expected to rapidly and accurately identify the strawberry fruit, and then segment the target fruit stem. This finding can provide the theoretical and technical support for the automatic operation of strawberry picking robot, particularly on the target sensing and efficient picking of strawberry picking robots. [ABSTRACT FROM AUTHOR]

Efficient detection of pepper clusters can be crucial to realize the automated harvesting in the development of pepper industry. However, the existing detection of pepper cluster cannot fully meet the large-scale in recent years, due to the low efficiency, generalization, and real-time algorithm. In this study, an improved YOLOv5s model was proposed to accurately detect the pepper clusters in real time. 400 images were collected from the Xuande Pepper Industrial Park, Qinglong County, Guizhou Province, Qianxinan Buyi and Miao Autonomous Prefecture, Guizhou Province, China. 148 images were also downloaded from the Internet, and then screened to enhance data diversity. The collected 548 images were then labeled using the LabelImg tool. In addition, the data enhancement was performed on each image and labeled bounding box. Two new images were randomly generated using random rotation and brightness adjustment. As such, a total of 1644 images were obtained as the dataset. Firstly, the EfficientViT network was optimized using the lightweight convolutional neural network (CNN) MobileNetV3 Block Convolution (MBConv) module and the ReLU with lightweight self-attention mechanism. CSPDarkNet53 network was used to replace the backbone of YOLOv5s. The global receptive field of model was enhanced to extract the important features. The better performance was suitable for the targets with the different scales, shapes, and background information. Secondly, Optimal Transport Assignment (OTA) was selected as the label assignment strategy. The optimal label assignment plan was obtained from the global perspective to achieve the effective label assignment. The robustness and accuracy of the model were improved for the complex scenes and target distribution. The WIoU loss function was also used as the bounding box localization of the detector, instead of the CIoU one. At the same time, the high-quality anchor frames were achieved to generate the benefit gradients for the better overall performance of the detector. A common experiment of target detection was performed on a homemade dataset, in order to verify the effectiveness of the improved model. The experimental results show that the improved model was achieved in a mean accuracy (mAP) of 97.3%, a parameter count of 5.9 M, and a detection speed of 131.6 frames per second (FPS). Specifically, the mAP and detection speed of the improved YOLOv5s model increased by 1.9 percentage points, and 14.5%, respectively, whereas, the number of parameters decreased by 15.7%, compared with the original. In addition, the mAP of the improved YOLOv5s model was 8, 16.9, 8.6, and 1.5 percentage points higher than that of EffcientDet-D1, SSD512, RetineNet-R50, and YOLOXs, respectively. The better performance of the improved YOLOv5s was achieved in terms of detection accuracy, detection speed, and number of model parameters. Therefore, the improved YOLOv5s model can be expected to accurately and rapidly detect the pepper clusters. [ABSTRACT FROM AUTHOR]

Automatic navigation has improved the efficiency of farmland operation for the high crop yield. Visual navigation can be expected to realize the automatic navigation in the farmland operation machinery, because of its low hardware cost and wide application. However, the traditional algorithm of image processing cannot fully deal with the interference of light and shadow changes, occlusion, and weeds under the complex environment in an orchard. There is a high demand for better robustness and accuracy in the navigation path. In this study, an accurate and rapid detection of navigation lines was proposed for the orchard using the improved YOLOv7 deep learning model. Firstly, the attention mechanism module (Convolutional Block Attention Module, CBAM) containing the channel and spatial attention module was added to the detection head network of the original YOLOv7 structure. This attention mechanism enabled to more accurately capture the essential features of the target for better representation and generalization of the network. The efficient extraction and enhancement of the trunk key features were realized to weaken the environmental background interference. Secondly, the SPD-Conv (space-to-depth, nonstrided convolution, SPD-Conv, i.e., a convolution-free step or pooling) was introduced between the ELAN-H and Repconv modules. As such, the low-resolution images or small-size targets were also detected to reduce the missed and false detection. The data set was collected from Beijing Xingshou Agricultural Professional Cooperative in Changping District, Beijing, China. 28 videos of orchards and a total of 1588 images were obtained under different lighting conditions. The targets were labeled using Labelimg software, with a total of 11043 apple trunks. The detection accuracy of the improved YOLOv7 model was 95.21% after training on the dataset, where the detection speed was 42.07 frames/s. The improved model increased by 2.31 percentage points, and 4.85 frames/s, respectively, compared with the original. Therefore, the improved model can be expected to more accurately identify the trunks of fruit trees, suitable for the apple orchard and jujube garden with the dense tree trunks. The ablation experiments were performed on each improvement point. The model accuracy of 93.97% was achieved after the introduction of the CBAM attention mechanism module, which was 1.07 percentage points over the original. The CBAM attention mechanism and SPD-Conv modules were introduced with a precision of 95.21%, which was a 1.24 percentage point improvement than before, indicating the better effectiveness of each improvement module. The trunk root accurate extraction of midpoint coordinates was crucial for the fitting of the navigation line, particularly with the trunk root midpoint as the navigation positioning base point. The coordinates of the improved YOLOv7 training for the fruit tree trunk were set at the bottom of the rectangular frame, instead of the trunk root midpoint as the locating reference point. Finally, the locating reference points were fitted on both sides of the fruit tree line and navigation line using the least squares method. The average line error of 4.43 pixels was achieved in the error analysis of 769 locating reference points extracted from the randomly selected 100 images from the data set and the manual marking of the trunk midpoints. The internal and external parameter matrixes of the camera were calculated to convert the pixel coordinates into the camera coordinates using Matlab software. The average actual error of 8.85 cm was obtained, indicating the reasonable and effective midpoint at the bottom of the rectangular frame as the locating reference point. The average deviation of the fitting and manual observation navigation line was 2.45 pixels in the 100 images, while the actual deviation was 4.90 cm, which fully met the accuracy requirements of navigation in the orchard. Three videos were randomly selected for the speed detection analysis of navigation lines. The total average time of processing one frame image was 0.044 s, which fully met the speed needs of navigation in the orchard. [ABSTRACT FROM AUTHOR]

Roundness is an important parameter to describe the shape of particle contour at the mesoscopic level, and it quantitatively describes the influence of particle corners on the contour. The existing roundness calculation method can not exactly recognize corners of particle contour, and involves a large amount of calculation. To solve these problems, an optimal selection algorithm was proposed to select the key points of corners of the particle contour. Firstly, the particle contour was transformed into a polar coordinate expansion curve, then the curve was segmented by using the minimum points, and the segmented curve was fitted by using the Fourier fitting algorithm. Secondly, the convex parts in the fitted contours were selected, and those containing the corner circles were selected according to the convexity of the convex parts. Finally, the key points used to fit the corner circles in the convex parts were selected, and the radiuses of the corner circles were figured out in order to calculate the roundness. K-S (Krumbein-Sloss)standard particle chart and other scholars' research results were used to verify the algorithm. The experimental results show that the accuracy of the number of corner circles calculated by the optimization selection method and the accuracy of roundness classification have both reached 100% . Further more, the influence of different filtering parameters on contour segmentation was analyzed when using average filtering algorithm to smooth the particle contour, and the optimized parameter selection range is also given. [ABSTRACT FROM AUTHOR]

Roundness is an important parameter to describe the shape of particle contour at the mesoscopic level, and it quantitatively describes the influence of particle corners on the contour. The existing roundness calculation method can not exactly recognize corners of particle contour, and involves a large amount of calculation. To solve these problems, an optimal selection algorithm was proposed to select the key points of corners of the particle contour. Firstly, the particle contour was transformed into a polar coordinate expansion curve, then the curve was segmented by using the minimum points, and the segmented curve was fitted by using the Fourier fitting algorithm. Secondly, the convex parts in the fitted contours were selected, and those containing the corner circles were selected according to the convexity of the convex parts. Finally, the key points used to fit the corner circles in the convex parts were selected, and the radiuses of the corner circles were figured out in order to calculate the roundness. K-S (Krumbein-Sloss)standard particle chart and other scholars ' research results were used to verify the algorithm. The experimental results show that the accuracy of the number of corner circles calculated by the optimization selection method and the accuracy of roundness classification have both reached 100% . Further more, the influence of different filtering parameters on contour segmentation was analyzed when using average filtering algorithm to smooth the particle contour, and the optimized parameter selection range is also given. [ABSTRACT FROM AUTHOR]

Computer vision systems are affected by foggy weather, resulting in poor quality images captured. To solve this problem, this study proposes a guided filtering dehazing algorithm based on channel difference model and its FPGA design. The channel difference model is obtained by separating the bright channel and dark channel of foggy image, and the model is used as a guide map for guided filtering to smooth the foggy image. Finally, a high boost filtering operation is performed to obtain a dehazed image. The hardware architecture is designed and implemented on FPGA. The experimental results show that the image scene after dehazing has uniform illumination, high degree of texture information recovery and high color fidelity. For an image of 480×270 size, the integrated frequency of the circuit is 108.448 MHz, the throughput is 323.47 MB·s-1, and the time to complete the entire dehazing is 0.001 2 s. These results indicate that the proposed algorithm and its hardware design can effectively improve image visibility and dehazing speed. [ABSTRACT FROM AUTHOR]

Aiming at the problem that target aircraft cannot be accurately and efficiently identified and followed in complex maintenance scenarios due to the large number and volume of aircraft, an aircraft number recognition algorithm was proposed based on YOLOv4 target detection. Firstly, the YOLOv4 detection algorithm was used by the method to obtain aircraft type and prediction block image information and then aircraft skeleton thinning was realized through image information. The location information of the aircraft number was obtained according to the aircraft type, and the OCR( optical character recognition) algorithm was used to identify the target aircraft number information. The experimental results show that the algorithm can recognize the aircraft number with an accuracy of 75% and a frame rate of 30 frames, which can achieve the purpose of real-time. [ABSTRACT FROM AUTHOR]

The trackside signal light is an important guidance for a running train in a rail transportation system. However, the combined trackside signal light included two colors is also used as a specific sign in addition to the usual single color signal light. In order to solve the recognition problem of the combined trackside signal light, a method based on image processing technology is employed to recognize the combined trackside signal light located within 150 m in front of train. As a result, the candidate region characteristics of the combined trackside signal light are extracted by color segmentation, morphological processing, and Hough transform. On the other hand, the combined trackside signal light can be positioned on the right of running rail train. The central spacing between two single lights can determine whether or not existing combined trackside signal light to remove the possible interferences from other lights or environment. The experimental results indicate that this method can accurately locate and recognize the combined trackside signal light, and the color correction ratio is 94.14% for green and yellow light, 96.21% for two green lights, 86.67% for two yellow lights. [ABSTRACT FROM AUTHOR]

Interactive image segmentation refers to bringing in a small number of user guidelines to segment the target object during the segmentation process and is one of the most fundamental tasks in image processing. Existing methods usually require the construction of non-quadratic energy functions and generally suffer from the lack of unique solutions and low segmentation accuracy. To further improve the segmentation quality, this paper proposed an interactive image segmentation algorithm combining local linear embedding and seed information（SILLE). The algorithm toke into account the local and a priori information of the pixel points and incorporated the information of the labeled seed points into the newly constructed energy function to obtain a unique and optimal solution of the energy function with an efficient and fast minimization scheme, thus obtaining more accurate segmentation results. Finally, comparison of different methods on different datasets verifies the effectiveness and feasibility of the algorithm. [ABSTRACT FROM AUTHOR]

Overlapping occlusion has seriously limited the yield estimation of bagged grapes in recent years, due to the ever-increasing grape volume after bagging and the large surface area of grape leaves. The unstable speed of manual video shooting can also lead to the loss of bagging grape targets. In this study, a yield estimation was proposed for the bagged grape using self-correcting Non-Maximum Suppression (NMS)-ByteTrack. Firstly, the bagged grapes were detected in the video using object detection (YOLOv5s). The NMS operation was also post-positioned to the tracking stage to retain the fruit detection boxes that filtered under the occlusion. Specifically, the detection boxes of bagged grapes were detected by object detection (YOLOv5s), whereas, the prediction boxes of bagged grapes were predicted to calculate the intersection over union using the Kalman filter. If the intersection over the union was less than the given threshold, the detection box was filtered to reserve the detection boxes closer to the predicted position. Then, the final detection box was obtained using the NMS operation. Secondly, the camera motion compensation and improved Kalman filter algorithm were added to automatically correct the position of the fruit prediction boxes and track them using ByteTrack. Specifically, the camera motion compensation was used to first extract the background key points in the bagged grape pictures of the previous frame and the current frame except for the tracking target. The sparse optical flow was used to match the extracted background in the key points of the bagged grapes. Then the affine transformation matrix of background motion was calculated by the RANSAC algorithm, and the Kalman filter was utilized to predict the bagged grape prediction boxes of the current frame. Finally, the affine transformation matrix was obtained to convert the prediction boxes in the coordinate system from the previous to the current frame, in order to realize the position information correction for the prediction boxes of the current frame. In addition, the improved Kalman filter algorithm can be used to change the aspect ratio in the state vector, and then to express the position of the bagged grape tracking boxes in the Kalman filter directly as the width, in order to more accurately estimate the position of the bagged grape tracking boxes. As such, a line-counting strategy was proposed to automatically count the bagged grapes. The strategy was also to set the counting line in the middle of the video, and also retain the idle frames without the bagged grapes in the first few seconds of the video. Furthermore, the bagged grapes with the same ID were only counted once to avoid repeated collision line counting. The successful collision between the center point of the tracking box of bagged grapes and the counting line was achieved in the automatic counting of bagged grapes. The dataset was collected from the Paidengte Agricultural Science and Technology Demonstration Park, Bishan District, Chongqing of China. The mobile phone cameras of Redmi K40 and OPPO Reno6pro+ were selected to capture the pictures of the same bagged grape at 8:00, 12:00, and 18:00 from different angles, such as frontal, sideways, and overhead shots. The total shooting time was about 6 h, the shooting height was about 1.5 m from the ground, and the shooting route was line by line. A total of 500 images and six videos were obtained for the bagged grapes. Among them, 500 images were expanded to 2000 images after saturation, brightness enhancement, brightness reduction, and mirror operations. Then, 2000 images were randomly divided into the training and validation sets in the ratio of 8:2, where six valid videos were used as the test set. Relevant experiments were conducted using this dataset. The experimental results showed that the yield estimation achieved a significant improvement in the tracking performance of bagged grapes using self-correcting NMS-ByteTrack. The multi-object tracking accuracy and precision, as well as the identification F1-score, were 64.6%, 82.4% and 80.8%, respectively, which increased by 1.7, 1.0 and 4.1 percentage points, respectively, compared with the ByteTrack. The number of ID switch was reduced by 3 times. Then, the average counting accuracy reached 82.8% in terms of counting performance, compared with manual counting. In addition, a comparison was also made with the five tracking methods. A better tracking performance was achieved, compared with the rest. The applicability of this estimation was also verified in tracking and counting bagged grapes. Therefore, the yield estimation can be expected to effectively promote the tracking and counting of bagged grapes in real scenarios using self-correcting NMS-ByteTrack. More accurate yield estimation of bagged grapes was also achieved in this field. [ABSTRACT FROM AUTHOR]

Termitomyces albuminosus is one species of agaric fungus in the family Agaricaceae in food production. However, it is a high demand to increase the current model detection accuracy, particularly under the complex planting environment, such as the variety of light and dark, the low recognition of soil and termitomyces albuminosus, the dense growth distribution and the serious shelter. In this study, target detection was proposed using an improved YOLOv5s. Firstly, RFBSE (Receptive Field Block Squeeze and Excitation) module was integrated into the backbone network. The human sensory field was then simulated to enhance the consistent contribution of different pixels to neural nodes in the process of network feature extraction. The edge features were highlighted to focus on the key areas. The irrelevant disturbances were suppressed using the module, such as the background. The attention mechanism of the channel was applied to gain the weight of different channels for the adaptive calibration of channel characteristic response. The channel was also enhanced to contain the important characteristic information of termitomyces albuminosus. As such, the high characteristic expression was achieved in the termitomyces albuminosus. Secondly, a multi-branch sampling DCSPP (Double Conv Spatial Pyramid Pooling) Pooling module was designed to perform the multiple sampling, in order to fuse the multiple receptive fields and then strengthen the relation between local and global information. The expression ability of the feature layer was enriched to improve the detection accuracy. Thirdly, the RFP (Recursive Feature Pyramid) structure was adopted in the neck network. The number of samples cannot increase, due to the usual interclass occlusion between termitomyces albuminosus. Previously, the network paid attention to the same image twice, because the context information around the occlusion samples was very important, and the feedback feature layer generated in the FPN structure was re-fed back to the backbone network for the Recursive computation. The neuronal activation was able to learn the correspondence and selectively inhibit, in order to improve the detection ability of the dense occluded sample of termitomyces albuminosus. The semantic information transmission was also promoted to enhance the context information near the occluded sample. At the same time, the cascaded RFP structure and the network fusion structure were lightened to reduce the calculation of parameters and memory usage. The ablation results showed that the RFBSE module, multi-branch pool module, and recursive pyramid structure shared different effects on the model. Specifically, the average precision mAP, precision, and recall rate of the final model reached 90.8%, 86.5%, and 84.8%, respectively. Each index of the improved model was improved, compared with the original one. The higher quality of the bounding box was achieved to detect the occluded target, where the mAP, accuracy, and recall were improved by 2.7, 3.8, and 3.9 percentage points, respectively. At last, the detection test of termitomyces albuminosus was carried out in different environments and shelter conditions by hardware platform model deployment. The visualized results showed that the detection rate of the model was more than 90%, which verified the validity of the model. The experimental results show that the improved model can be expected to accurately and rapidly identify the termitomyces albuminosus in a complex environment. The finding can provide technical support for the development of the termitomyces albuminosus harvesting robots. [ABSTRACT FROM AUTHOR]

Autonomous navigation is an important guarantee for intelligent agricultural machinery to complete harvesting operations. Navigation line recognition is an important step in achieving autonomous navigation of intelligent agricultural machinery. Real-time and accurate detection of harvesting navigation lines can effectively accelerate the progress of operations and reduce crop loss rates. Visual sensors have the advantages of low cost and rich information acquisition, making them widely used in route detection. Frequent changes in lighting under cloudy weather make it difficult for a single feature to cope with wheat field environments. In this study, a wheat harvesting boundary detection method based on multi texture feature fusion was proposed. In the harvested area, the characteristics of low stubble and no top wheat ears result in differences in uniformity, density, and other aspects when harvesting under both light and backlight compared to the non-harvested area. At the same time, due to the exposed wheat stubble in the harvested area, it appears more regular on the image, and its gradient direction is more consistent than that in the non-harvested area. Therefore, a two-dimensional feature vector composed of image entropy and directional gradient was constructed to classify the harvested and unharvested areas of the wheat field, and then the harvesting boundary was extracted. Due to the susceptibility of image entropy features to perspective phenomena, this paper compared the entropy features of different regions in the image when fixing the camera installation angle and height, and ultimately selected the middle region of the image as the region of interest. After analyzing the characteristics of image entropy feature extraction, a histogram statistical method based on sliding windows was proposed to accelerate the speed of image entropy feature extraction. By further dividing the window to be calculated into several sub windows, sliding and combining these sub windows significantly reduced the computation. The entropy feature extraction in this paper took 0.53 seconds, which is 49.52% faster than the traditional method of directly extracting entropy from the entire window. After extracting twodimensional features, based on the distribution characteristics of the feature histogram and combined with the maximum entropy threshold segmentation algorithm, the wheat field image was preliminarily classified. For entropy features, low noise points affected the threshold segmentation effect. By comparing the impact of discarding different proportions of low entropy points on the segmentation threshold, this paper discarded 10% of low entropy points, enhancing the threshold segmentation effect. The method of removing small connected regions was used to eliminate the misclassified regions in the initial classification binary image. Then, the Canny operator was used to extract edge contour points distributed near the harvesting boundary. Due to the consistent direction between the harvesting boundary and the agricultural machinery's forward direction, the Ransac algorithm was used to restrict the fitting line in the region, obtaining the accurate harvesting boundary. In order to verify the feasibility of the method proposed in this article, wheat field images under different lighting conditions were collected, including 2 200 weak light images, 758 local strong light images, and 1 134 strong light images. After selecting the region of interest, the size of the image to be processed was 240 pixels×1 280 pixels. The algorithm in this paper took an average of 0.88 s to process each image on laptops. The detection rates under weak light, local strong light, and strong light were 90.41%, 88.26%, and 89.68%, respectively, with an average of 89.45%. Compared with the traditional method of using Adaboost algorithm for machine learning, the detection speed of this algorithm was improved by 73.89%, and the detection rate was increased by 46.19 percentage points, 46.00 percentage points, and 49.64 percentage points under weak light, local strong light, and strong light, respectively. Compared with traditional algorithms, the method proposed in this paper significantly improved real-time performance and detection rate, which can provide reference for the detection of field agricultural machinery navigation lines. [ABSTRACT FROM AUTHOR]

Aiming at the problem of identifying the picking points of the lotus pods robot, a method of calculating the picking points and picking posture of the robot was proposed. A two-stage network based on YOLO (you only look once) and Deeplab V3 + was proposed, and Mobilenet V2 feature extraction network was applied to Deeplab V3 + . Finally, the segmentation results were processed by image processing, and the picking points and picking posture of lotus seeds were further calculated. The experimental results of 50 original images show that the success rate of the algorithm is 88. 89%, and the average frame rate is 34. 41 FPS. The obtained algorithm can provide effective information to the automatic harvesting machinery of lotus pods, which has the characteristics of lightweight and high efficiency, and promotes the application of computer vision and neural network in modern agriculture. [ABSTRACT FROM AUTHOR]

Wearing medical masks of various materials and types will lead to different prevention and control the effect of virus spread. Segmentation to masks in image data sets will benefit analyzing the differences in prevention and control effects of different types of masks. Currently, there are many facial mask detection algorithms, but there is a lack of further segmentation. Therefore, this study proposes a facial mask segmentation method based on image processing and deep learning. An improved contrast-limited adaptive histogram equalization preprocessing method is proposed to reduce the poor effect of traditional preprocessing caused by dark part of the image through brightness adaptive adjustment. SSD network is used for mask pre-detection, and CA- Net is used for mask segmentation. With U-Net as the backbone network, CA-Net adds spatial attention module, channel attention module and scale attention module to realize the comprehensive attention guidance of space, channel and scale of feature map at the same time. The segmentation result is processed by binary image closing operation to get the final segmentation results. The Dice evaluation of preliminarily segmentation index can reach to 79.18%±3.44%, and the Dice coefficient is increased to 84.03%±2.81% by adding preprocessing and post-processing operations, indicating that the proposed method could significantly improve the visual segmentation results. [ABSTRACT FROM AUTHOR]

In order to improve the communication quality of visible light downlink, an automatic alignment system for visible light communication based on OpenMV is designed in this study. The system realizes automatic recognition of indoor white LED is based on image processing and color recognition. The system uses gray-scale centroid method to determine the centroid coordinates of white LED, and the PID algorithm is used to drive steering gears to achieve the tracking of the indoor white light LED. The experimental results show that the automatic alignment system of visible light communication can realize the automatic alignment of white LED under various motion trajectories at the receiving distance of 2 m, and the maximum alignment angle error is less than 0.534°. In addition, whether the white LED is blocked or there is an interfering light source, system can still maintain the good tracking effect, indicating that system has good robustness. [ABSTRACT FROM AUTHOR]

Robot technology has been gradually applied in modern agriculture in recent years. Among them, a duck egg is one of the most important agricultural products during food processing. However, the current collection of duck eggs can usually require a large amount of manual labor, leading to the time-consuming and labor-intensive task. A smart robot has been developed to automatically collect the duck eggs, in order to improve the production efficiency for the harvesting cost-saving. Specifically, an important technical challenge of harvesting robots can be the rapid and accurate identification and positioning of duck eggs, especially under complex environments, such as occlusion, crowding, and darkness. In this study, duck egg detection was proposed for complex environments using an improved YOLOv7 model. A convolutional attention module (CBAM) was added to the backbone network. The network information transmission was enhanced for better sensitivity to the specific features, while the interference of complex environments was reduced on the duck egg recognition. At the same time, the depth-wise separable convolution (DSC) was utilized to adjust the spatial pyramid pooling (SPP), in order to reduce the number of model parameters and operation costs. A high accuracy was achieved to identify and locate the duck eggs under complex environments, thus providing technical support for the development of harvesting robots. Some materials (such as feathers, straw, and sediment) were also used to simulate the complex environment of the duck house. A duck egg image collection platform was then constructed to evaluate the accuracy and environmental adaptability of the model. Actual duck eggs were photographed in the duck house of Wuhan Yujia Bay Duck Farm using an Honor HLK-AL10 camera for both simulated and actual conditions. The duck egg image data was collected, including the multiple angles, positions, different distances and occlusion forms. A total of 2 600 JPG format images were divided into the training set (1 560 images), validation set (520 images), and test set (520 images), according to a 6:2:2 ratio. At the same time, data augmentation was performed on the training set, in order to improve the robustness and generalization ability of the model. The following procedures were used: 1) To add 12% Gaussian noise. 2) To add 2.5% salt and pepper noise. 3) To set the image gains a =0.3 and a =0.5 to change image brightness. After that, the improved YOLOv7 model was trained with an iteration number of 150. Test results show that the improved YOLOv7 model increased the F1 score by 8.3, 10.1, 8.7, and 7.6 percentage points, respectively, and the F1 score reached 95.5%, compared with the common detection models, such as SSD, YOLOv4, YOLOv5_M, and YOLOv7. The occupied memory space was only 68.7 M, while the average time was 0.022 s for the single-image detection, and the average precision value (mAP, Mean Average Precision) was 85.2%. There were no missed or false detections for the feather occlusion or clustering of duck eggs, with an average confidence level of 93.6% and 85.7%, respectively. The more accurate detection was achieved in the improved YOLOv7 model under the complex environments, indicating superior performance. [ABSTRACT FROM AUTHOR]

Image segmentation has been widely used for the rapid and accurate detection of plants in the various robots of modern agriculture in recent years. However, fully supervised learning cannot obtain the sufficient, effective and low-cost mask labels (manual labeling) as training samples in the segmentation task of plant image instances, due to the diversity of plant species and forms. In this study, an automatic labelling-based instance segmentation network (AutoLNet) was proposed to improve the segmentation accuracy. The weak tags were also used to train the weak supervised deep learning model. Finally, the network model was used for the image segmentation of maize seedling stage. The top view of maize seedling stage was collected by unmanned aerial vehicle (UAV). Data enhancement was then used to improve the sample diversity. A weak label self-generation module was added in front of the backbone network using the weak supervised instance segmentation model. As such, the module was composed of color space conversion, contour tracking, and the minimum peripheral rectangle. The color threshold range of corn plants was firstly set to remove the background area of the image, in order to eliminate the influence of ground shadow and land on the foreground information. The foreground corn plant area was also expanded to remove the small noise points for the binary image with only foreground corn plants. Secondly, the edge detection was carried out on the binary image after threshold segmentation. The contour point set was then set for the foreground corn plants. Finally, the minimum peripheral rectangle of the foreground object was generated automatically in the original image using the coordinates of the contour point set. The final boundary frame was obtained to filter the threshold value. The weak label was generated automatically. The weak tags were used instead of manual tags to participate in network training. The image instance segmentation of maize seedling stage was realized without the manual tags, which was greatly reduced the labor cost that required for data annotation. The test results showed that the distance intersection ratio and cosine similarity between the selfgenerated and manual tags reached 95.23% and 94.10%, respectively. The quality of the tags was fully met the high requirements of weak supervision training. The average accuracy of AutoLNet's output prediction frame and mask reached 68.69% and 35.07%, respectively. By contrast, the average accuracy of Autolnet's output prediction frame and mask increased by 10.83 and 3.42 percentage points, respectively, compared with the manual label models (DiscoBox and Box2Mask). The average accuracy of the forecast frame increased by 11.28 and 8.79 percentage points, respectively, whereas, that of the mask increased by 12.75 and 10.72 percentage points, respectively. The accuracy of weakly supervised learning was improved to reduce the projection and paired loss during training in the AutoLNet, compared with the fully supervised model (CondInst and Mask R-CNN). The average accuracy of prediction frame and mask in AutoLNet reached 94.32% and 83.14% of the CondInst model, 7.54 and 3.28 percentage points higher than those of prediction frame and mask R-CNN mode. Once the intersection ratio threshold was greater than or equal to 0.5, the segmentation effect of AutoLNet was better than that of the fully supervised model Mask R-CNN, similar to the CondInst. Consequently, the improved AutoLNet can be expect to automatically obtain the corn plant labels in the image using the label self-generation module. Manual labeling process was improved using the label self-generation module. Case segmentation of corn seedling images was realized for the cost saving without manual labeling. The finding can provide the solution and technical support to the high precision and low-cost segmentation task of maize seedling image instance in field environment. [ABSTRACT FROM AUTHOR]

Abstract：Weeds pose a great threat to modern agricultural production. The unrestrained growth of weeds has also a significant impact on crop yields and quality. Therefore, it is a high demand to effectively manage and control weeds in recent years, in order to optimize agricultural production activities. The specific area of weed management can be expected to serve as the promising research direction for weed control operations, with the rapid development of unmanned aerial vehicle (UAV) platforms and artificial intelligence (AI). Accurate and efficient identification of weeds can be one of the most key steps to realizing automated weed management in the field. However, the current efficient models of weed identification are often required a significant amount of labeled agricultural data. The major bottleneck can be also limited in the model. In this study, a semi-supervised semantic segmentation network was proposed, called UANP-MT (Uncertainty Aware and Network Perturbed Mean Teacher), in order to reduce the dependence on labeled data. The PSPNet network model was also selected to leverage the teacher-student network concept from MT (Mean Teacher). Several key improvements were then incorporated during this time. Firstly, the improved model reduced the errors that caused by random network factors. The teacher network output was also augmented to take the mean, thereby ensuring the robustness of the network prediction. Secondly, the improved model constructed the uncertainty constraints for the different uncertainties between networks and assigns weightings, in order to balance the differences in the output between different networks. This approach enhanced the confidence and reliability of the prediction, ultimately leading to the higher recognition accuracy of an improved model. A series of ablation experiments were conducted to evaluate the effectiveness of the improved model. These experiments included the setting values of network parameters, the selection of feature extraction network backbones, and the model performance testing on datasets with varying volumes of data. Some optimal parameter settings were identified for the improved model, with the ResNet50 as the preferred choice for the model backbone. The comparison experiments showed that the UANP-MT model outperformed the original supervised network, in terms of the three evaluation indicators of F1 score, pixel accuracy (PA), and intersection over union (IoU), when the labeled data was less than that of the original supervised network, indicating the superiority of the semisupervised network. Furthermore, the effectiveness of the improved model was further validated to compare with the classic semantic segmentation networks, including SegNet, U-Net, and Deeplabv3+ on the CaiXin weed dataset. UANP-MT model achieved better performance with an f_score of 81.83%, a PA of 95.84%, and an IoU of 90.70% when only 1/4 of the labeled data was used for training. These evaluation metrics were superior to those of the Deeplabv3+ model by percentage point of 4.71, 7.94, and 8.27, respectively. The UANP-MT model was achieved in the high-quality weed detection and recognition with the low labeled data, thus significantly reducing the dependence on labeled data with the manpower and time cost saving. The findings can provide valuable implications for the subsequent weed recognition research and the development of automated weeding operations on UAV platforms. [ABSTRACT FROM AUTHOR]

Due to the fast speed or long imaging distance, the road traffic monitoring image is blurred, and there are some problems such as incomplete, wrong extraction and low contrast when extracting the image. There- fore, a road traffic fuzzy image enhancement method based on multi-scale Retinex is proposed. The image is de- composed into low frequency component and high frequency component by Shearlet transform. Linear mapping is used to map the low-frequency coefficients to a specific interval for normalization, and multi-scale Retinex is used to enhance the low-frequency components. Robust median method and hard threshold shrinkage method are used to remove noise and illumination interference in each scale and direction of high frequency component. The fuzzy contrast and linear membership function are used to fuse the two parts of the image, and finally the enhanced clear image is obtained. The experimental results show that the proposed method can optimize the image visual effect, highlight the image detail features, and has good noise suppression performance. Compared with the genera- tive adversarial network method and the feature pyramid method, the peak signal-to-noise ratio, structural simi- larity, information entropy, average gradient and spatial frequency of the image are better. [ABSTRACT FROM AUTHOR]

The existing crack image segmentation methods cannot describe the characteristics of smaller cracks, which leads to incomplete crack extraction results in pavement images. Therefore, a pavement crack image segmentation method based on mathematical morphology is proposed. Use the spatial domain processing method to enhance the original image pixels, obtain a new pavement crack image through forward transformation and inverse transformation, and classify the noise type in the image; use the gray value to represent the noise, and use the 3× in the mean filter method 3 The template represents the crack element and noise element of the discrete image, and denoises the pavement crack image; uses the classical mathematical morphology theory to express the structural elements of the pavement crack image, and judges the form change of the crack in the process of expansion and corrosion processing, to achieve crack segmentation. The experimental results of single crack image and complex crack image show that the proposed method can effectively extract the crack feature value, the feature extraction rate is 99.5%, and it has high segmentation efficiency under different noise conditions, which can meet the needs of pavement crack detection. [ABSTRACT FROM AUTHOR]

The extensive range of olive cultivars available in the market exhibit minor differences in their phenotypic traits. Nonetheless, their quality attributes, particularly the oil content and fatty acid composition, significantly differ among distinct cultivars, resulting in the emergence of use iinferior products as superior products in the market. The accurate and quick identification of olive cultivars holds significant importance in enhancing the production and quality of olives. As such, delving into the study of olive cultivar identification is crucial for the advancement of the olive industry. This study presents a novel approach to address the challenge of identifying olive cultivars in natural conditions. Specifically, a bilinear attentional EfficientNet model is proposed, which incorporates the bilinear network design concept and attention mechanism. The model is trained and evaluated using four commonly planted olive cultivars (i.e., Frantoio, Leccino, Picholine, and Ezhi 8) in Longnan, Gansu. The experimental results demonstrate the effectiveness of the proposed model in accurately and quickly identifying different olive cultivars. A bilinear network has been suggested to comprehensively extract feature information from olive images, for the limited phenotypic differences across different olive cultivars. In light of this, the selection of a feature extraction network has been made with consideration for both speed and accuracy, leading to the selection of the EfficientNetB0 network. To tackle the challenge of identifying olive cultivars under natural conditions which are prone to intricate background interferences, a novel approach combining convolutional block attention module (CBAM) with bilinear network has been proposed. This approach facilitates the model in selectively focusing on the salient features responsible for cultivar identification during the feature extraction process of olive images. Upon conducting experiments, the bilinear attention EfficientNet model presented in this study has exhibited an overall accuracy of 90.28% and an inference time of 9.15 ms in identifying four distinct olive cultivars. These experiment results demonstrate that the proposed model better achieved better rapaid and accurate identification of olive cultivars under natural conditions. The present study proposes a bilinear attention EfficientNet model for identifying olive cultivars and utilizes gradient-weighted class activation mapping (Grad-CAM) to analyse its performance. The results demonstrate that the proposed model exhibits a greater attention towards fruit regions, as well as some leaf regions, within olive images. These findings are in agreement with the expert knowledge and experience of human practitioners. The analysis of heat maps generated from misidentified olive images revealed that inadequate focus on the fruit and leaf regions, which are pivotal for successful identification of cultivars, was the primary contributing factor to misidentification. The outcomes of the ablation experiments indicated that the bilinear network and the CBAM exhibited a positive impact on the precision of olive variety recognition. To ascertain the efficacy of the method elucidated in this manuscript, a set of comparative experiments has been formulated. The primary objective of these experiments is to juxtapose the proposed bilinear attention EfficientNet model against the conventional cultivar identification models, including bilinear ResNet34, EfficientNet-SE attention, bilinear ResNet18, bilinear VGG16, and bilinear GoogLeNet. The experimental results obtained from the comparison analysis provide that the proposed bilinear attention EfficientNet model exhibits superior performance in terms of overall accuracy for the identification of olive cultivars. The bilinear attention EfficientNet model's accuracy for the identification of olive cultivars exceeded that of bilinear ResNet34, EfficientNet-SE attention, bilinear ResNet18, bilinear VGG16, and bilinear GoogLeNet models by 12.78%, 11.53%, 11.11%, 10.70%, and 5.00%, respectively. The present study establishes a foundation for resolving the challenge of accurately identifying olive cultivars in natural conditions. Moreover, it offers a valuable point of reference for the identification of cultivars of diverse crops. [ABSTRACT FROM AUTHOR]

Soil continuous cropping has been an ever-increasingly serious threat to modern agriculture in recent years. The soil disinfection method can be expected to effectively alleviate the soil's continuous cropping. Furthermore, the soil steam disinfection can be sown in a short period of time after disinfection, due to the non-toxic, pollution-free, compared with the chemical methods. But the low disinfection and heating efficiency has been limited in the process of red loam steam disinfection in practice. In this study, the disinfection pipe structural parameters were optimized to clarify the influence on the soil disinfection heating efficiency. Firstly, image processing was used to construct a soil pore structure discrete model. Secondly, the simulation was implemented to analyze the steam flow field in the disinfection pipe and the single-factor soil steam disinfection. Finally, the experiment was performed on the optimized structural parameters of multi-factor disinfection pipes using the soil vapor disinfection test platform. The results show that: 1) The circumferential and axial outlets of the disinfection pipe posed a significant effect on the heating efficiency of steam disinfection (P <0.01). The optimal structural parameters of the disinfection pipe were achieved: 2 mm pore size of the outlet pore, 3 circumferential pores, 2 axial pores. The average disinfection duration of the test when the average soil temperature reaches 80 °C is 394 s, and the average error rate between the test results and the predicted results is 5.3%; 2) The soil heating rates of each treatment reached the peak when the steam disinfection was carried out for 200 s. The efficiency of disinfection and heating gradually decreased after 200 s of disinfection. Future practical operations need to be guided by the variation of soil heating rate. Specifically, the steam valve should be gradually closed, the steam flow should be suspended, and the excess water in the soil should be removed to provide the soil pore permeability when the heating rate decreased gradually. Therefore, an intelligent control system can be expected to be combined, when conducting steam disinfection operations on Yunnan red loam in the later stage. The steam flow rate should be gradually reduced when disinfecting for 200 s. The soil temperature was redistributed and gradually decreased, as time increased. Once the soil temperature was below 60 °C, the steam flow rate should be increased again, and the flow rate should be reduced and increased in a cyclic manner to ensure that the soil temperature was always above 60 °C, in order to achieve the goal of efficient disinfection. The findings can provide a theoretical basis for efficient disinfection operations. 3) The range of single-pipe high-temperature areas in each treatment was mainly concentrated in the horizontal direction 0̚150 mm and the vertical direction 0̚200 mm. This study can provide a theoretical basis for the structural design of the disinfection pipes in the steam sterilizers and the intelligent disinfection operation, and lay a theoretical research foundation for the design of multi-pipe spacing of end actuators. [ABSTRACT FROM AUTHOR]