Search

Because of the limitations of current PCB detection methods, such as low detection accuracy and high rate of false and missed detection, a PCB defect detection algorithm based on improved RetinaNet is proposed, that is, RetinaNet-line. Firstly, the algorithm uses PVT v2B2-Hi as the backbone network of the whole model, which can not only solve the deep feature redundancy caused by the residual network in RetinaNet but also improve the feature extraction ability of the network. Then, the CARAFE up-sampling module is used to enrich the semantic features of defective targets and promote the discrimination ability of the algorithm to defective targets. Finally, HSV color space is used to process the input feature image, and the color distribution is expanded by random perturbation in H, S, and V channels, with the data enhancement completed. In this paper, experiments are conducted on the publicly shared PCB defect detection dataset from the Intelligent Open Laboratory of Peking University, and the results show that the improved algorithm achieves 94.2% in the mAP index, which is better than other networks. [ABSTRACT FROM AUTHOR]

Scoring targets in shooting sports is a crucial and time-consuming task that relies on manually counting bullet holes. This paper introduces an automatic score detection model using object detection techniques. The study contributes to the field of computer vision by comparing the performance of seven models (belonging to two different architectural setups) and by making the dataset publicly available. Another value-added aspect is the inclusion of three variants of the object detection model, YOLOv8, recently released in 2023 (at the time of writing). Five of the used models are single-shot detectors, while two belong to the two-shot detectors category. The dataset was manually captured from the shooting range and expanded by generating more versatile data using Python code. Before the dataset was trained to develop models, it was resized (640 × 640) and augmented using Roboflow API. The trained models were then assessed on the test dataset, and their performance was compared using matrices like mAP50, mAP50-90, precision, and recall. The results showed that YOLOv8 models can detect multiple objects with good confidence scores. Among these models, YOLOv8m performed the best, with the highest mAP50 value of 96.7%, followed by the performance of YOLOv8s with the mAP50 value of 96.5%. It is suggested that if the system is to be implemented in a real-time environment, YOLOv8s is a better choice since it took significantly less inference time (2.3 ms) than YOLOv8m (5.7 ms) and yet generated a competitive mAP50 of 96.5%. [ABSTRACT FROM AUTHOR]

Aiming at the problems of serious image degradation and low target recognition rate in current underwater vehicle target detection methods, an underwater target detection method combining improved RetinaNet and attention mechanism is proposed. Firstly, RetinaNet backbone network is replaced with DenseNet network, which retains more target features and reduces the number of parameters. Secondly, in order to increase the operation speed of the network model, the initial convolution is replaced by the depth separated deformable convolution, thus greatly reducing the parameters of the model. Finally, CBAM attention module is introduced to enhance features in space and channel dimensions, reducing the interference of underwater complex environment to target detection. The experimental results of underwater robot grasping show that compared with the initial RetinaNet methods, The mAP value of the improved method can reach 81.9%, the parameters are 56.8 MB, and the detection speed is 16.8 frames. The improved method has excellent performance in underwater target detection. [ABSTRACT FROM AUTHOR]

Early detection and efficient management practices to control Tuta absoluta (Meyrick) infestation is crucial for safeguarding tomato production yield and minimizing economic losses. This study investigates the detection of T. absoluta infestation on tomato plants using object detection models combined with ensemble techniques. Additionally, this study highlights the importance of utilizing a dataset captured in real settings in open-field and greenhouse environments to address the complexity of real-life challenges in object detection of plant health scenarios. The effectiveness of deep-learning-based models, including Faster R-CNN and RetinaNet, was evaluated in terms of detecting T. absoluta damage. The initial model evaluations revealed diminishing performance levels across various model configurations, including different backbones and heads. To enhance detection predictions and improve mean Average Precision (mAP) scores, ensemble techniques were applied such as Non-Maximum Suppression (NMS), Soft Non-Maximum Suppression (Soft NMS), Non-Maximum Weighted (NMW), and Weighted Boxes Fusion (WBF). The outcomes shown that the WBF technique significantly improved the mAP scores, resulting in a 20% improvement from 0.58 (max mAP from individual models) to 0.70. The results of this study contribute to the field of agricultural pest detection by emphasizing the potential of deep learning and ensemble techniques in improving the accuracy and reliability of object detection models. [ABSTRACT FROM AUTHOR]

Remote sensing image scene classification has become a well-liked research subject because of its use in a number of domains, including object recognition, land use classification, picture extraction, and monitoring. Semantic data, which is crucial in direct measures like oceanography, vegetation, mineral extraction, temperature, and forest management, will be used by the Remote sensing image classification process to provide a class label to each scenario class. By the use of Swarm Intelligence and Support Vector Machines, this work develops an RSI categorization technique in order to obtain the optical images. The recommended SVM-SI strategy inherits the advantages of both methods. The RSI has undergone pre-processing to transform it into a suitable form. The features were also developed using the RetinaNet framework. Moreover, SVM-based classifiers are employed to give the RSIs the proper class labelling, and the design of the SI can improve categorization effectiveness. Empirical investigations have shown that the SVM-SI technique is superior to other cutting-edge methods, and common databases that are easily available to the public are used to validate the success of the strategy. Accuracy of 99.49% was attained using the SVM-SI model and approach that was presented. [ABSTRACT FROM AUTHOR]

Mesoscale convective systems (MCSs) and associated hazardous meteorological phenomena cause considerable economic damage and even loss of lives in the mid-latitudes. The mechanisms behind the formation and intensification of MCSs are still not well understood due to limited observational data and inaccurate climate models. Improving the prediction and understanding of MCSs is a high-priority area in hydrometeorology. One may study MCSs either employing high-resolution atmospheric modeling or through the analysis of remote sensing images which are known to reflect some of the characteristics of MCSs, including high temperature gradients of cloud-top, specific spatial shapes of temperature patterns, etc. However, research on MCSs using remote sensing data is limited by inadequate (in size) databases of satellite-identified MCSs and poorly equipped automated tools for MCS identification and tracking. In this study, we present (a) the GeoAnnotateAssisted tool for fast and convenient visual identification of MCSs in satellite imagery, which is capable of providing AI-generated suggestions of MCS labels; (b) the Dataset of Mesoscale Convective Systems over the European Territory of Russia (DaMesCoS-ETR), which we created using this tool, and (c) the Deep Convolutional Neural Network for the Identification of Mesoscale Convective Systems (MesCoSNet), constructed following the RetinaNet architecture, which is capable of identifying MCSs in Meteosat MSG/SEVIRI data. We demonstrate that our neural network, optimized in terms of its hyperparameters, provides high MCS identification quality ( mAP = 0.75 , true positive rate TPR = 0.61 ) and a well-specified detection uncertainty (false alarm ratio FAR = 0.36 ). Additionally, we demonstrate potential applications of the GeoAnnotateAssisted labelling tool, the DaMesCoS-ETR dataset, and the MesCoSNet neural network in addressing MCS research challenges. Specifically, we present the climatology of axisymmetric MCSs over the European territory of Russia from 2014 to 2020 during summer seasons (May to September), obtained using MesCoSNet with Meteosat MSG/SEVIRI data. The automated identification of MCSs by the MesCoSNet artificial neural network opens up new avenues for previously unattainable MCS research topics. [ABSTRACT FROM AUTHOR]

The field of GeoAI or Geospatial Artificial Intelligence has undergone rapid development since 2017. It has been widely applied to address environmental and social science problems, from understanding climate change to tracking the spread of infectious disease. A foundational task in advancing GeoAI research is the creation of open, benchmark datasets to train and evaluate the performance of GeoAI models. While a number of datasets have been published, very few have centered on the natural terrain and its landforms. To bridge this gulf, this paper introduces a first-of-its-kind benchmark dataset, GeoImageNet, which supports natural feature detection in a supervised machine-learning paradigm. A distinctive feature of this dataset is the fusion of multi-source data, including both remote sensing imagery and DEM in depicting spatial objects of interest. This multi-source dataset allows a GeoAI model to extract rich spatio-contextual information to gain stronger confidence in high-precision object detection and recognition. The image dataset is tested with a multi-source GeoAI extension against two well-known object detection models, Faster-RCNN and RetinaNet. The results demonstrate the robustness of the dataset in aiding GeoAI models to achieve convergence and the superiority of multi-source data in yielding much higher prediction accuracy than the commonly used single data source. [ABSTRACT FROM AUTHOR]

[Objective] Pineapple is a common tropical fruit, and its ripeness has an important impact on the storage and marketing. It is particularly important to analyze the maturity of pineapple fruit before picking. Deep learning technology can be an effective method to achieve automatic recognition of pineapple maturity. To improve the accuracy and rate of automatic recognition of pineapple maturity, a new network model named MobileNet V3-YOLOv4 was proposed in this study. [Methods] Firstly, pineapple maturity analysis data set was constructed. A total of 1580 images were obtained, with 1264 images selected as the training set, 158 images as the validation set, and 158 images as the test set. Pineapple photos were taken in natural environment. In order to ensure the diversity of the data set and improve the robustness and generalization of the network, pineapple photos were taken under the influence of different factors such as branches and leaves occlusion, uneven lighting, overlapping shadows, etc. and the location, weather and growing environment of the collection were different. Then, according to the maturity index of pineapple, the photos of pineapple with different maturity were marked, and the labels were divided into yellow ripeness and green ripeness. The annotated images were taken as data sets and input into the network for training. Aiming at the problems of the traditional YOLOv4 network, such as large number of parameters, complex network structure and slow reasoning speed, a more optimized lightweight MobileNet V3-YOLOv4 network model was proposed. The model utilizes the benck structure to replace the Resblock in the CSPDarknet backbone network of YOLOv4. Meanwhile, in order to verify the effectiveness of the MobileNet V3-YOLOv4 network, MobileNet V1-YOLOv4 model and MobileNet V2-YOLOv4 model were also trained. Five different single-stage and two-stage network models, including R-CNN, YOLOv3, SSD300, Retinanet and Centernet were compared with each evaluation index to analyze the performance superiority of MobileNet V3-YOLOv4 model. [Results and Discussions] MobileNet V3-YOLOv4 was validated for its effectiveness in pineapple maturity detection through experiments comparing model performance, model classification prediction, and accuracy tests in complex pineapple detection environments. The experimental results show that, in terms of model performance comparison, the training time of MobileNet V3-YOLOv4 was 11,924 s, with an average training time of 39.75 s per round, the number of parameters was 53.7 MB, resulting in a 25.59% reduction in the saturation time compared to YOLOv4, and the parameter count accounted for only 22%. The mean average precision (mAP) of the trained MobileNet V3-YOLOv4 in the verification set was 53.7 MB. In order to validate the classification prediction performance of the MobileNet V3-YOLOv4 model, four metrics, including Recall score, F1 Score, Precision, and average precision (AP), were utilized to classify and recognize pineapples of different maturities. The experimental results demonstrate that MobileNet V3- YOLOv4 exhibited significantly higher Precision, AP, and F1 Score the other. For the semi-ripe stage, there was a 4.49% increase in AP, 0.07 improvement in F1 Score, 1% increase in Recall, and 3.34% increase in Precision than YOLOv4. As for the ripe stage, there was a 6.06% increase in AP, 0.13 improvement in F1 Score, 16.55% increase in Recall, and 6.25% increase in Precision. Due to the distinct color features of ripe pineapples and their easy differentiation from the background, the improved network achieved a precision rate of 100.00%. Additionally, the mAP and reasoning speed (Frames Per Second, FPS) of nine algorithms were examined. The results showed that MobileNet V3-YOLOv4 achieved an mAP of 90.92%, which was 5.28% higher than YOLOv4 and 3.67% higher than YOLOv3. The FPS was measured at 80.85 img/s, which was 40.28 img/s higher than YOLOv4 and 8.91 img/s higher than SSD300. The detection results of MobileNet V3-YOLOv4 for pineapples of different maturities in complex environments indicated a 100% success rate for both the semi-ripe and ripe stages, while YOLOv4, MobileNet V1-YOLOv4, and MobileNet V2-YOLOv4 exhibited varying degrees of missed detections. [Conclusions] Based on the above experimental results, it can be concluded that MobileNet V3-YOLOv4 proposed in this study could not only reduce the training speed and parameter number number, but also improve the accuracy and reasoning speed of pineapple maturity recognition, so it has important application prospects in the field of smart orchard. At the same time, the pineapple photo data set collected in this research can also provide valuable data resources for the research and application of related fields. [ABSTRACT FROM AUTHOR]

With advancements in autonomous driving, demand for stringent and computationally efficient traffic sign detection systems has increased. However, bringing such a system to a deployable level requires handling critical accuracy and processing speed issues. A focal loss-based single-stage object detector, i.e RetinaNet, is used as a trade-off between accuracy and processing speed as it handles the class imbalance problem of the single-stage detector and is thus suitable for traffic sign detection (TSD). We assessed the detector's performance by combining various feature extractors such as ResNet-50, ResNet-101, and ResNet-152 on three publicly available TSD benchmark datasets. Performance comparison of the detector using different backbone includes evaluation parameters like mean average precision (mAP), memory allocation, running time, and floating-point operations. From the evaluation results, we found that the RetinaNet object detector using the ResNet-152 backbone obtains the best mAP, while that using ResNet-101 strikes the best trade-off between accuracy and execution time. The motivation behind benchmarking the detector on different datasets is to analyse the detector's performance on different TSD benchmark datasets. Among the three feature extractors, the RetinaNet model trained using the ResNet-50 backbone is an excellent model in memory consumption, making it an optimal choice for low-cost embedded devices deployment. [ABSTRACT FROM AUTHOR]

The detection of defects on small cotter pins that are installed in electric power fittings is an essential part of the inspection task of overhead lines using Unmanned Aerial Vehicles (UAV). It is challenging to detect small defects from a large number of UAV images. In this paper, an efficient and high-performance defect detection model called DDNet is proposed to recognize defects from images of unmanned aerial vehicles. The attention mechanism was adopted in the improved detection model in order to enhance the representation learning of the image. Inspired by the human visual system, the RFB module is added to the FPN module, increasing the receptive field of the entire detection network, which is conducive to the detection of small objects. Then a dataset of cotter pins for model training and testing was introduced. The study demonstrates that the proposed DDNet increases the average precision from 82.0 to 90.1% and reduces the miss rate of defect detection from 14.5 to 7.4% in our dataset compared to the baseline RetinaNet model. We also compared existing frameworks for object detection and discussed other common ways to improve precision. The results showed that our optimized model in this paper improved detection performance, which subsequently proved the practicability and effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Due to the massive surge in the world population, the agriculture cycle expansion is necessary to accommodate the anticipated demand. However, this expansion is challenged by weed invasion, a detrimental factor for agricultural production and quality. Therefore, an accurate, automatic, low-cost, environment-friendly, and real-time weed detection technique is required to control weeds on fields. Furthermore, automating the weed classification process according to growth stages is crucial for using appropriate weed controlling techniques, which represents a gap of research. The main focus of the undertaken research described in this paper is on providing a feasibility study for the agriculture community using recent deep-learning models to address this gap of research on classification of weed growth stages. For this paper we used a drone to collect a dataset of four weed (Consolida regalis) growth stages. In addition, we developed and trained one-stage and two-stage models YOLOv5, RetinaNet (with Resnet-101-FPN, Resnet-50-FPN backbones) and Faster R-CNN (with Resnet-101-DC5, Resnet-101-FPN, Resnet-50-FPN backbones), respectively. The results show that the generated Yolov5-small model succeeds in detecting weeds and classifying weed growth stages in real time with the highest recall of 0.794. RetinaNet with ResNet-101-FPN backbone shows accurate results in the testing phase (average precision of 87.457). Although Yolov5-large showed the highest precision in classifying almost all weed growth stages, Yolov5-large could not detect all objects in tested images. Overall, RetinaNet with ResNet-101-FPN backbones shows accurate and high precision, whereas Yolov5-small shows the shortest inference time in real time for detecting a weed and classifying its growth stages. [ABSTRACT FROM AUTHOR]

Unmanned aerial vehicle (UAV) equipped with efficient vision system is widely used with the continuous maturity of UAV technology. In order to solve the problem of low detection precision due to the large number of small targets and low resolution, an improved RetinaNet algorithm for UAV aerial photography target detection was proposed. To solve the problem of insufficient extraction of small target information in feature maps, a multi-stage feature fusion method was proposed, and the feature mining module was designed by connecting the method with attention mechanism, which can integrate deep semantic information into shallow feature map and enrich small target features. The Anchor-free method based on central point detection was designed in the algorithm. In this way, the network can locate the target by the regression of the center point, rather than by the fixed size of the anchor frame to match, which can make the network more flexible in the regression of small targets. The depth separable convolution method was used to design its lightweight network model with characteristics of both reduced size and improved detection speed. The experimental results show that the average accuracy of the improved algorithm is increased by 8.5% compared with the original RetinaNet algorithm, and the detection speed is increased by 6 frame/s. Compared with other advanced algorithms, the improved algorithm also has performance advantages, and achieves the balance of detection accuracy and detection speed. [ABSTRACT FROM AUTHOR]

Developing an automatic system for detection, segmentation, and classification of skin lesions is very useful to aid well-timed diagnosis of skin diseases. Lesion segmentation is a crucial task for automated diagnosis of skin cancers, as it affects significantly the accuracy of the subsequent steps. Varieties in sizes and locations of lesions, and the lesions with low-contrast boundaries make this task very challenging. In this paper, a three-stage CNN-based method is presented for accurate segmentation of lesions from dermoscopic images. At the first step, normalization, approximate locations and sizes of lesions are estimated. Due to the importance of the normalization stage, three CNN-based networks (Mask R-CNN, RetinaNet, and YOLOv3) are used for the lesion detection. A convolutional network is presented and used to combine the results of the object detection networks with a novel approach. The output of the first stage is a normalized cropped image containing the detected lesion in the center. At the second stage, segmentation, a CNN in a DeepLab3+ structure, is used to extract the lesion from the normalized image. Finally, an active contour method is used as the postprocessing to enhance the boundary of the segmented lesion. The proposed method is evaluated on well-known datasets. Experiments show that the proposed method outperforms all the previous state-of-the-art methods. [ABSTRACT FROM AUTHOR]

RetinaNet has been widely used in the field of ship detection in synthetic aperture radar (SAR) images due to its simple structure and excellent performance. Although various modified RetinaNet models for SAR ship detection have been proposed, there are still some problems, such as how to capture more complete ship features and the unbalanced distribution of positive samples in different tasks, which have not been mentioned in previous studies. Aiming at the above problems, this paper proposes a modified RetinaNet detector for SAR ship detection, namely D 2 ADet. Specifically, a modified feature pyramid network is proposed to capture adaptive features that are more suitable for the ship. Then, a dual-activation domain training method is designed to decouple the sample assignment processes of different tasks to meet their requirements of positive samples for each task. Experimental results show that compared with RetinaNet, D 2 ADet obtains 7.67 % , 11.57 % and 7.28 % gains in AP50, AP75 and mAP on the public SAR ship detection dataset (SSDD), respectively. And D 2 ADet achieves the state-of-the-art performance compared with other methods. [ABSTRACT FROM AUTHOR]

There is a critical need for detection of administrative regions through network map pictures in map censorship tasks, which can be implemented by target detection technology. However, on map images there tend to be numerous administrative regions overlaying map annotations and symbols, thus making it difficult to accurately detect each region. Using a RetinaNet-based target detection model integrating ResNet50 and a feature pyramid network (FPN), this study built a multi-target model and a single-target cascading model from three single-target models by taking Taiwan, Tibet, and the Chinese mainland as target examples. Two models were evaluated both in classification and localization accuracy to investigate their administrative region detection performance. The results show that the single-target cascading model was able to detect more administrative regions, with a higher f1_score of 0.86 and mAP of 0.85 compared to the multi-target model (0.56 and 0.52, respectively). Furthermore, location box size distribution from the single-target cascading model looks more similar to that of manually annotated box sizes, which signifies that the proposed cascading model is superior to the multi-target model. This study is promising in providing support for computer map reading and intelligent map censorship. [ABSTRACT FROM AUTHOR]

A fast and accurate detection is one of the most important prerequisites for the apple harvest robots. However, there are many factors that can make apple detection difficult in a real orchard scene, such as complex backgrounds, fruit overlap, and leaf/branch occlusion. In this study, a fast and stable network was proposed for apple detection using an improved RetinaNet. A picking strategy was also developed for the harvest robot. Specifically, once the apples occluded by branches/wires were regarded as the picking targets, the robot arm would be injured at the same time. Therefore, the apples were labeled with multiple classes, according to different types of occlusions. The Res2Net module was also embedded in the ResNet50, in order to improve the ability of the backbone network to extract the multi-scale features. Furthermore, the BiFPN instead of FPN was used as a feature fusion network in the neck of the network. A weight fusion of feature maps was also made at different scales for the apples with different sizes, thus improving the detection accuracy of the network. After that, a loss function was combined with the Focal loss and Efficient Intersection over Union (EIoU) loss. Among them, Focal loss was used for the classification loss function, further reducing the errors from the imbalance of positive and negative sample ratios. By contrast, the EIoU loss was used for the regression loss function of the bounding box, in order to maintain a fast and accurate regression. Particularly, there were some different relative positions in the prediction and the ground truth box, such as overlap, disjoint and inclusion. Finally, the classification and regression were carried out on the feature map of five scales to realize a better detection of apple. In addition, the original dataset consisted of 800 apple images with complex backgrounds of dense orchards. A data enhancement was conducted to promote the generalization ability of the model. The dataset was then expanded to 4 800 images after operations, such as rotating, adjusting brightness, and adding noise. There was also a balance between the detection accuracy and speed. A series of experimental statistics were obtained on the number of BiFPN stacks in the network. Specifically, the BiFPN was stacked five times in the improved RetinaNet. The ablation experiments showed that each improvement of the model enhanced the accuracy of the network for the apple detection, compared with the original. The average precision of the improved RetinaNet reached 94.02%, 86.74%, 89.42%, and 94.84% for the leaf occlusion, branch/wire occlusion, fruit occlusion, and no occlusion apples, respectively. The mean Average Precision (mAP) reached 91.26%, which was 5.02 percentage points higher than that of the traditional RetinaNet. The improved RetinaNet took only 42.72 ms to process an apple image on average. Correspondingly, each fruit picking cycle was 2 780 ms, indicating that the detection speed fully met the harsh requirement of the picking robot. Only when the apples were large or rarely occluded, both improved and traditional RetinaNet were used to accurately detect them. By contrast, the improved RetinaNet performed the best to detect all apple fruits, when the apples were under a complex environment in an orchard, such as the leaf-, fruit-, or branch/wire-occluded background. The reason was that the traditional RetinaNet often appeared to miss the detection in this case. Consequently, the best comprehensive performance was achieved to verify the effectiveness of the improvements, compared with the state-of-the-art detection network, such as the Faster RCNN and YOLOv4. Overall, all the apples in the different classes can be effectively detected for the apple harvest. The finding can greatly contribute to the picking strategy of the robot, further avoiding the potential damage by the branches and wires during harvesting. [ABSTRACT FROM AUTHOR]

Pneumonia is an acute respiratory infectious disease caused by bacteria, fungi, or viruses. Fluid-filled lungs due to the disease result in painful breathing difficulties and reduced oxygen intake. Effective diagnosis is critical for appropriate and timely treatment and improving survival. Chest X-rays (CXRs) are routinely used to screen for the infection. Computer-aided detection methods using conventional deep learning (DL) models for identifying pneumonia-consistent manifestations in CXRs have demonstrated superiority over traditional machine learning approaches. However, their performance is still inadequate to aid in clinical decision-making. This study improves upon the state of the art as follows. Specifically, we train a DL classifier on large collections of CXR images to develop a CXR modality-specific model. Next, we use this model as the classifier backbone in the RetinaNet object detection network. We also initialize this backbone using random weights and ImageNet-pretrained weights. Finally, we construct an ensemble of the best-performing models resulting in improved detection of pneumonia-consistent findings. Experimental results demonstrate that an ensemble of the top-3 performing RetinaNet models outperformed individual models in terms of the mean average precision (mAP) metric (0.3272, 95% CI: (0.3006,0.3538)) toward this task, which is markedly higher than the state of the art (mAP: 0.2547). This performance improvement is attributed to the key modifications in initializing the weights of classifier backbones and constructing model ensembles to reduce prediction variance compared to individual constituent models. [ABSTRACT FROM AUTHOR]

Target detection at the hive gate of a beehive can be used to effectively monitor invasive beehive species. However, in the natural environment, there is often a multi-target and multi-scale problem at the hive gate, making it difficult for beekeepers to accurately detect the internal state of the hive. (1) To solve the above problems, this paper proposes an improved RetinaNet target detection network, DY-RetinaNet, for the identification of common species at the hive doors of beehives in natural environments, i.e., Chinese bees, wasps, and cockroaches. (2) First, to solve the multi-target multi-scale problem presented in this paper, we propose replacing the FPN layer in the initial model RetinaNet with a symmetric structure BiFPN layer consisting of a feature pyramid, which allows the model to better balance the feature information of different scales. Then, for the loss function, using CIOU loss instead of smooth L1 loss makes the network more accurate for small target localization at multiple scales. Finally, the dynamic head framework is added after the model backbone network, due to the benefits of its multi-attention mechanism, which makes the model more concerned with multi-scale recognition in a multi-target scenario. (3) The experimental results of the homemade dataset show that DY-RetinaNet has the best network performance, compared to the initial model RetinaNet, when the backbone network is ResNet-101-BiFPN, and the mAP value of DY-RetinaNet is 97.38%. Compared with the initial model, the accuracy is improved by 6.77%. The experimental results from the public dataset MSCOCO 2017 show that DY-RetinaNet is better than the existing commonly used target-detection algorithms, such as SSD, YOLOV3, Faster R-CNN, Mask R-CNN, FCOS, and ExtremeNet. These results verify that the model has strong recognition accuracy and generalization ability for multi-target multi-scale detection. [ABSTRACT FROM AUTHOR]

To improve the detection precision in actual application scenes, a high resolution feature extraction network High-Resolution Representation Network(HRNet)and a Guided Anchoring mechanism were proposed to improve the RetinaNet algorithm, the high resolution information of the feature maps were maintained during feature extraction and the anchors in network were generated adaptively, which improved the detection precision of the algorithm. The results show that the improved algorithm achieves a mean average precision of 0. 905 on Caltech pedestrian dataset, which is 6. 0% higher than the standard RetinaNet algorithm. The detection speed reaches 19 FPS on video with the image size of 1 280 × 720 pixels per frame, the algorithm achieves the balance between the detection precision and the detection speed. [ABSTRACT FROM AUTHOR]

Industrial X-ray analysis is common in aerospace, automotive or nuclear industries where structural integrity of some parts needs to be guaranteed. However, the interpretation of radiographic images is sometimes difficult and may lead to two experts disagree on defect classification. The automated defect recognition (ADR) system presented herein will reduce the analysis time and will also help reducing the subjective interpretation of the defects while increasing the reliability of the human inspector. Our convolutional neural network (CNN) model achieves 0.942 mAP@IoU = 0.50, which is considered as similar to expected human performance, when applied to an automotive aluminium castings dataset (GDXray). On an industrial environment, its inference time is less than 400 ms per 16 GB DICOM image (16 bits), so it can be installed on production facilities with no impact on delivery time. In addition, an ablation study of the main hyper-parameters to optimise model accuracy from the initial baseline result of 0.75 mAP up to 0.942 mAP, was also conducted. [ABSTRACT FROM AUTHOR]

Changes in the ungulate population density in the wild has impacts on both the wildlife and human society. In order to control the ungulate population movement, monitoring systems such as camera trap networks have been implemented in a non-invasive setup. However, such systems produce a large number of images as the output, hence making it very resource consuming to manually detect the animals. In this paper, we present a new dataset of wild ungulates which was collected in Latvia. Moreover, we demonstrate two methods, which use RetinaNet and Faster R-CNN as backbones, respectively, to detect the animals in the images. We discuss the optimization of training and impact of data augmentation on the performance. Finally, we show the result of aforementioned tune networks over the real world data collected in Latvia. [ABSTRACT FROM AUTHOR]

The Internet of Things (IoT) represents a potential paradigm increase in the number of linked devices, referred to as things. Administering these things remains a difficulty. The advancement of Internet of Things technology has generated innovative techniques to resolve these issues. Within the context of the low speed, low fusion precision, and other issues of present Internet of Things data fusion techniques, an Internet of Things data fusion procedure based on an intelligent optimization method is developed with the goal of improving the Internet of Things heterogeneous data fusion effect. Although IoT systems truly provide the potential to enhance efficiency, ownership, transparency, and efficacy, they also experience a larger number of potential flaws. This study examines the security of IoT systems under the framework of smart-world critical setups. This study specifically performs an in-depth analysis of the risks in IoT-based critical infrastructures from the viewpoint of apps, networking, operating systems, software, rmware, and hardware. Additionally, this study highlights the three essential IoT-based cyber-physical infrastructures, viz. intelligent passage, intelligent build-up, and intelligent gridding. The current state of IoT research focuses mostly on the processing of scalar sensor data events and pays scant attention to the issues faced in multimedia-based events. In addition, this work conducts a comparative analysis of various ML techniques, such as the support vector machine, convolutional neural network, naive Bayes, and k-nearest neighbor techniques. By assessing sensors, networking, services, and implementation services offered by the IoT, this study analyzes the existing methods for the Internet of Multimedia Things. [ABSTRACT FROM AUTHOR]

Detecting defects in industrial-quality inspection is an important task, but defect detection remains challenging due to limited size, various types of defects, and imbalanced samples in images. In this paper, we propose a novel automatic defect detection network that efficiently detects defects. We introduce the Automated Sample Assignment (SCID) algorithm, which automatically divides the distribution of positive and negative samples based on similarity scores using the Gaussian mixture model. The Automated Sample Assignment (SCID) algorithm solves the problem of sample imbalance in defect detection networks caused by manually setting the intersection over the union (IOU) threshold and assigning positive and negative labels to samples. Then, instead of the pooling method used by most backbone networks, we use a loss fusion pool named LE-Pool to retain more detailed information. Finally, we use the soft weighted attention mechanism of the detection head (S-head) to identify noise information. Experimental results on multiple defect detection datasets show that the proposed method is superior to other detection methods in terms of accuracy. [ABSTRACT FROM AUTHOR]

The real-time perception of hand gestures in a deprived environment is a demanding machine vision task. The hand recognition operations are more strenuous with different illumination conditions and varying backgrounds. Robust recognition and classification are the vital steps to support effective human-machine interaction (HMI), virtual reality, etc. In this paper, the real-time hand action recognition is performed by using an optimized Deep Residual Network model. It incorporates a RetinaNet model for hand detection and a Depthwise Separable Convolutional (DSC) layer for precise hand gesture recognition. The proposed model overcomes the class imbalance problems encountered by the conventional single-stage hand action recognition algorithms. The integrated DSC layer reduces the computational parameters and enhances the recognition speed. The model utilizes a ResNet-101 CNN architecture as a Feature extractor. The model is trained and evaluated on the MITI-HD dataset and compared with the benchmark datasets (NUSHP-II, Senz-3D). The network achieved a higher Precision and Recall value for an IoU value of 0.5. It is realized that the RetinaNet-DSC model using ResNet-101 backbone network obtained higher Precision (99.21 %for AP0.5, 96.80%for AP0.75) for MITI-HD Dataset. Higher performance metrics are obtained for a value of γ= 2 and α= 0.25. The SGD with a momentum optimizer outperformed the other optimizers (Adam, RMSprop) for the datasets considered in the studies. The prediction time of the optimized deep residual network is 82 ms. [ABSTRACT FROM AUTHOR]

Background: The correct identification of pills is very important to ensure the safe administration of drugs to patients. Here, we use three current mainstream object detection models, namely RetinaNet, Single Shot Multi-Box Detector (SSD), and You Only Look Once v3(YOLO v3), to identify pills and compare the associated performance.Methods: In this paper, we introduce the basic principles of three object detection models. We trained each algorithm on a pill image dataset and analyzed the performance of the three models to determine the best pill recognition model. The models were then used to detect difficult samples and we compared the results.Results: The mean average precision (MAP) of RetinaNet reached 82.89%, but the frames per second (FPS) is only one third of YOLO v3, which makes it difficult to achieve real-time performance. SSD does not perform as well on the indicators of MAP and FPS. Although the MAP of YOLO v3 is slightly lower than the others (80.69%), it has a significant advantage in terms of detection speed. YOLO v3 also performed better when tasked with hard sample detection, and therefore the model is more suitable for deployment in hospital equipment.Conclusion: Our study reveals that object detection can be applied for real-time pill identification in a hospital pharmacy, and YOLO v3 exhibits an advantage in detection speed while maintaining a satisfactory MAP. [ABSTRACT FROM AUTHOR]

Today's deep learning architectures, if trained with proper dataset, can be used for object detection in marine search and rescue operations. In this paper a dataset for maritime search and rescue purposes is proposed. It contains aerial-drone videos with 40,000 hand-annotated persons and objects floating in the water, many of small size, which makes them difficult to detect. The second contribution is our proposed object detection method. It is an ensemble composed of a number of the deep convolutional neural networks, orchestrated by the fusion module with the nonlinearly optimized voting weights. The method achieves over 82% of average precision on the new aerial-drone floating objects dataset and outperforms each of the state-of-the-art deep neural networks, such as YOLOv3, -v4, Faster R-CNN, RetinaNet, and SSD300. The dataset is publicly available from the Internet. [ABSTRACT FROM AUTHOR]

The existing detection method of wearing safety helmet is difficult to detect in the complex posture of construction personnel,and its accuracy is not high.This paper proposed a detection method of wearing safety helmet based on posture estimation.This method used the residual network in the OpenPose pose estimation model to optimize feature extraction and obtained the skeleton point information of the constructor.This paper proposed a three-point positioning method to determine the head area through the position information of bone points to reduce the detection range.This method used RetinaNet to detect the wearing of the helmet in the head area to solve the problem of extreme imbalance between the helmet and the construction background.The experimental results show that this method has significantly improved detection accuracy than other methods,and is more adaptable to the environment. [ABSTRACT FROM AUTHOR]

With the large-scale construction of urban subways, the detection of tunnel cracks becomes particularly important. Due to the complexity of the tunnel environment, it is difficult for traditional tunnel crack detection algorithms to detect and segment such cracks quickly and accurately. The article presents an optimal adaptive selection model (RetinaNet-AOS) based on deep learning RetinaNet for semantic segmentation on tunnel crack images quickly and accurately. The algorithm uses the ROI merge mask to obtain a minimum detection area of the crack in the field of view. A scorer is designed to measure the effect of ROI region segmentation to achieve optimal results, and further optimized with a multi-dimensional classifier. The algorithm is compared with the standard detection based on RetinaNet algorithm with an optimal adaptive selection based on RetinaNet algorithm for different crack types. The results show that our crack detection algorithm not only addresses interference due to mash cracks, slender cracks, and water stains but also the false detection rate decreases from 25.5–35.5% to about 3.6%. Meanwhile, the experimental results focus on the execution time to be calculated on the algorithm, FCN, PSPNet, UNet. The algorithm gives better performance in terms of time complexity. [ABSTRACT FROM AUTHOR]

Object detection in aerial images remains a challenging problem due to low image resolutions, complex backgrounds, and variations of sizes and orientations of objects in images. In recent years, several multiscale and rotated box-based deep neural networks (DNNs) have been proposed and have achieved promising results. In this article, a new method for designing loss function, called adaptive saliency biased loss (ASBL), is proposed for training DNN object detectors to achieve improved performance. ASBL can be implemented at the image level, which is called image-based ASBL, or at the anchor level, which is called anchor-based ASBL. The method computes saliency information of input images and anchors generated by DNN object detectors, weights of different training examples, and anchors differently based on their corresponding saliency measurements. It gives complex images and difficult targets more weights during training. In our experiments using two of the largest public benchmark data sets of aerial images, DOTA and NWPU VHR-10, the existing RetinaNet was trained using ASBL to generate an one-stage detector, ASBL-RetinaNet. ASBL-RetinaNet significantly outperformed the original RetinaNet by 3.61 and 12.5 mean average precision (mAP), respectively, on the two data sets. In addition, ASBL-RetinaNet outperformed ten other state-of-the-art object detection methods. [ABSTRACT FROM AUTHOR]

Purpose: Urinary particles are particularly important parameters in clinical urinalysis, especially for the diagnosis of nephropathy. Therefore, it is highly important to precisely detect urinary particles in the clinical setting. However, artificial microscopy is subjective and time consuming, and various previous detection algorithms lack the adequate accuracy. In this study, a method is proposed for the analysis of urinary particles based on deep learning. Methods: We used seven cellular components (i.e., erythrocytes, leukocytes, epithelial, low‐transitional epithelium, casts, crystal, and squamous epithelial cells) in the microscopic imaging of urine as the detection targets. After the extraction of features using Resnet50, feature maps of different sizes are obtained in the last few layers of the feature pyramid net (FPN). The feature maps are then input into the classification subnetwork and regression subnetwork for classification and localization respectively, and detection results are obtained. First, we introduce the basic model (RetinaNet) to detect the cellular components in urinary particles, and the features of the objects can then be extracted more effectively by replacing different basic networks. Lastly, the effects of different weight initialization methods and different anchor scales on the performance of the model are investigated. Results: We obtained the optimal network structure based on the adjustment of the loss functional parameters, thereby achieving the best results in the test set of urinary particles. The experimental data yielded an accuracy of 88.65% with a processing time of only 0.2 s for each image on a GeForce GTX 1080 graphics processing unit (GPU). Our results demonstrate that this method cannot only achieve the speed of the first‐stage target detector, but also the accuracy of the two‐stage target algorithm in the analysis of urinary particles. Conclusion: This study developed new automated analysis urinary particles based on deep learning, and this method is expected to be used for the automated analysis and detection of urinary particles. Moreover, our approach will be useful for the detection of other cells in the clinical setting. [ABSTRACT FROM AUTHOR]

Object detection using deep learning methods in the field of intelligent transportation has become a research hotspot. Currently, most of the classic object detection algorithms, whether which are the single-stage object detection models based on regression or the two-stage object detection models based on candidate regions, use a large number of predefined priori boxes called "anchor" to enumerate the possible positions, sizes and aspect ratios so as to search the objects. It will cause serious imbalance between positive and negative samples, and the performance and generalization ability of the models are also limited by the anchor's design. Aiming at the above problems of the anchor-based object detection algorithms, a single-stage object detection network, called RetinaNet, is used to establish the anchor-free based object detection models for vehicles, pedestrians, and cyclists in traffic scenes. Pixel-by-pixel prediction is adopted to handle object detection and add central prediction branches to improve the detection performance. Experiments show that, compared with the original RetinaNet algorithm based on anchor, the improved algorithm can better recognize vehicles, pedestrians, and cyclists in traffic scenes. [ABSTRACT FROM AUTHOR]

Since the motherboard images of different mobile phones have a multi-resolution imaging mode, the shape of defective components become multi-scale. Conventional defect detection methods mainly include image fusion methods or statistical model extraction methods, but the robustness of these methods still needs to be improved. To solve this problem, an automatic learning representation model, called RetinaNet object detector, is proposed. Firstly, feature pyramid network (FPN) is used to extract the multi-scale feature classification and location of defective components, and MobileNetV2 is introduced to compress and accelerate the RetinaNet model. Secondly, focus loss is used to resolve the class imbalance and increase the difficulty of detecting the contribution of the samples to the loss during the training. The experimental results show that RetinaNet can effectively detect defective components of different scales, and has high detection accuracy. Compared with other object detectors, RetinaNet a-chieves an average accuracy (mAP) of over 95%. These results demonstrate the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

• Manual pest idenatification and eradication procedures are time-consuming and labour-intensive, emphasising the need for automation. • An integrated approach using deep learning and computer vision helps to automatically locate and identify the severity levels of pest infestations. • An open-source annotation tool called LabelMe is used to annotate bounding boxes around the infected areas of muskmelon leaves. • The study develops, evaluates, and compares the performance of the RetinaNet-based detector and the Faster R-CNN-based detector under the Detectron2 framework. • The RetinaNet-based detector achieves higher mean average precision, enabling faster and more accurate detection of miner infestations. Corp protection against pests is known to play a crucial role in developing efficient crop management strategies for Precision Agriculture. A recent estimation by Food and Agriculture Organization (FAO) shows that the perennial loss due to crop pests and diseases amounts to nearly 40% of agricultural crop production at a global level. Identifying pests and diseases and eradicating them without automation is laborious and time-consuming. Automation in detecting and identifying miners at the onset and their eradication is possible using deep learning (DL) and computer vision. This study aims to develop a Detectron2 -based framework to detect and localize miner infestations on muskmelon leaves by developing a detection model that integrates DL and a computer vision library to enhance detection capabilities. The approach develops, experiments, and compares a region-based detector (Faster Region-based Convolutional Neural networks (R-CNN)) with a region-free (RetinaNet) by training and validating the bounding box annotated custom dataset of leaf miner infected muskmelon leaves imaged using a smartphone camera. The results show that the RetinaNet-based detector outperforms the Faster R-CNN-based detector in recognizing the infestation severity levels, significantly increasing mean average precision and acquiring faster detection speeds. [Display omitted] [ABSTRACT FROM AUTHOR]

This paper implements a novel approach to automatically detect and classify rockfalls in Lunar Reconnaissance Orbiter narrow angle camera (NAC) images using a single-stage dense object detector (RetinaNet). The convolutional neural network has been trained with a data set of 2932 original rockfall images. In order to avoid overfitting, the initial training data set has been augmented during training using random image rotation, scaling, and flipping. Testing images have been labelled by human operators and have been used for RetinaNet performance evaluation. Testing shows that RetinaNet is capable to reach recall values between 0.98 and 0.39, precision values between 1 and 0.25, and average precisions ranging from 0.89 to 0.69, depending on the used confidence threshold and intersection-over-union values. Mean processing time of a single NAC image in RetinaNet is around 10 s using a GeForce GTX 1080 Ti and GeForce Titan Xp, which is in orders of magnitudes faster than a human operator. The processing speed allows to efficiently exploit the currently available NAC data stack with more than 1 million images in a reasonable timeframe. The combination of speed and detection performance can be used to produce lunar rockfall distribution maps on large spatial scales for utilization by the scientific and engineering community. [ABSTRACT FROM AUTHOR]

Broccoli is an example of a high-value crop that requires delicate handling throughout the growing season and during its post-harvesting treatment. As broccoli heads can be easily damaged, they are still harvested by hand. Moreover, human scouting is required to initially identify the field segments where several broccoli plants have reached the desired maturity level, such that they can be harvested while they are in the optimal condition. The aim of this study was to automate this process using state-of-the-art Object Detection architectures trained on georeferenced orthomosaic-derived RGB images captured from low-altitude UAV flights, and to assess their capacity to effectively detect and classify broccoli heads based on their maturity level. The results revealed that the object detection approach for automated maturity classification achieved comparable results to physical scouting overall, especially for the two best-performing architectures, namely Faster R-CNN and CenterNet. Their respective performances were consistently over 80% mAP@50 and 70% mAP@75 when using three levels of maturity, and even higher when simplifying the use case into a two-class problem, exceeding 91% and 83%, respectively. At the same time, geometrical transformations for data augmentations reported improvements, while colour distortions were counterproductive. The best-performing architecture and the trained model could be tested as a prototype in real-time UAV detections in order to assist in on-field broccoli maturity detection. [ABSTRACT FROM AUTHOR]

Aiming at recognizing small proportion, blurred and complex traffic sign in natural scenes, a traffic sign detection method based on RetinaNet-NeXt is proposed. First, to ensure the quality of dataset, the data were cleaned and enhanced to denoise. Secondly, a novel backbone network ResNeXt was employed to improve the detection accuracy and effection of RetinaNet. Finally, transfer learning and group normalization were adopted to accelerate our network training. Experimental results show that the precision, recall and mAP of our method, compared with the original RetinaNet, are improved by 9.08%, 9.09% and 7.32%, respectively. Our method can be effectively applied to traffic sign detection. [ABSTRACT FROM AUTHOR]

Over the past several centuries, the iron industry played a central role in the economy of Sweden and much of northern Europe. A crucial component of iron manufacturing was the production of charcoal, which was often created in charcoal piles. These features are visible in LiDAR (light detection and ranging) datasets. These charcoal piles vary in their morphology by region, and training data for some feature types are severely lacking. Here, we investigate the potential for machine automation to aid archaeologists in recording charcoal piles with limited training data availability in a forested region of Jönköping County, Sweden. We first use hydrological depression algorithms to conduct a preliminary assessment of the study region and compile suitable training data for charcoal production sites. Then, we use these datasets to train a series of RetinaNet deep learning models, which are less computationally expensive than many popular deep learning architectures (e.g., R-CNNs), allowing for greater usability. Together, our results demonstrate how charcoal piles can be automatically extracted from LiDAR datasets, which has great implications for improving our understanding of the long-term environmental impact of the iron industry across Northern Europe. Furthermore, our workflow for developing and implementing deep learning models for archaeological research can expand the use of such methods to regions that lack suitable training data. [ABSTRACT FROM AUTHOR]

Object detection plays an important role in autonomous driving, disaster rescue, robot navigation, intelligent video surveillance, and many other fields. Nonetheless, visible images are poor under weak illumination conditions, and thermal infrared images are noisy and have low resolution. Consequently, neither of these two data sources yields satisfactory results when used alone. While some scholars have combined visible and thermal images for object detection, most did not consider the illumination conditions and the different contributions of diverse data sources to the results. In addition, few studies have made use of the temperature characteristics of thermal images. Therefore, in the present study, visible and thermal images are utilized as the dataset, and RetinaNet is used as the baseline to fuse features from different data sources for object detection. Moreover, a dynamic weight fusion method, which is based on channel attention according to different illumination conditions, is used in the fusion component, and the channel attention and a priori temperature mask (CAPTM) module is proposed; the CAPTM can be applied to a deep learning network as a priori knowledge and maximizes the advantage of temperature information from thermal images. The main innovations of the present research include the following: (1) the consideration of different illumination conditions and the use of different fusion parameters for different conditions in the feature fusion of visible and thermal images; (2) the dynamic fusion of different data sources in the feature fusion of visible and thermal images; (3) the use of temperature information as a priori knowledge (CAPTM) in feature extraction. To a certain extent, the proposed methods improve the accuracy of object detection at night or under other weak illumination conditions and with a single data source. Compared with the state-of-the-art (SOTA) method, the proposed method is found to achieve superior detection accuracy with an overall mean average precision (mAP) improvement of 0.69%, including an AP improvement of 2.55% for the detection of the Person category. The results demonstrate the effectiveness of the research methods for object detection, especially temperature information-rich object detection. [ABSTRACT FROM AUTHOR]

• Proposing an object detection method for object deviating from center or multi object crowding. • The functions of PA-FPN, FreeAnchor and Soft-NMS are complementary and the combination of them can get the maximum benefits for the original network. RetinaNet is a typical representative of single-stage object detection, which can solve the problem of sample imbalance. However, due to the lack of region proposal extraction process in single-stage object detection, the effect of RetinaNet in dealing with the problem that object deviating from center or multi object crowding is not good. To solve this problem, we use a variety of optimization methods for RetinaNet to improve the accuracy of object detection. Firstly, FreeAnchor is introduced on the basis of RetinaNet, which can autonomously learn to match the target category; secondly, ResNeXt50 is taken as the backbone to improve the accuracy without increasing the parameter complexity; thirdly, Bottom-up Path Augmentation module is used to enhance the transmission of location information and further optimize the recall rate; finally, soft-NMS method is used to effectively reduce the false positive detection results and improve the average accuracy of object detection. We use the MS COCO data set to verify the new model. The mAP value of the new model reached 40.8, which is 4.3 more than baseline. It shows that the optimization methods are complementary to each other, which can effectively improve the object detection accuracy while ensuring the speed. [ABSTRACT FROM AUTHOR]

Small mammals, and particularly rodents, are common inhabitants of farmlands, where they play key roles in the ecosystem, but when overabundant, they can be major pests, able to reduce crop production and farmers' incomes, with tangible effects on the achievement of Sustainable Development Goals no 2 (SDG2, Zero Hunger) of the United Nations. Farmers do not currently have a standardized, accurate method of detecting the presence, abundance, and locations of rodents in their fields, and hence do not have environmentally efficient methods of rodent control able to promote sustainable agriculture oriented to reduce the environmental impacts of cultivation. New developments in unmanned aerial system (UAS) platforms and sensor technology facilitate cost-effective data collection through simultaneous multimodal data collection approaches at very high spatial resolutions in environmental and agricultural contexts. Object detection from remote-sensing images has been an active research topic over the last decade. With recent increases in computational resources and data availability, deep learning-based object detection methods are beginning to play an important role in advancing remote-sensing commercial and scientific applications. However, the performance of current detectors on various UAS-based datasets, including multimodal spatial and physical datasets, remains limited in terms of small object detection. In particular, the ability to quickly detect small objects from a large observed scene (at field scale) is still an open question. In this paper, we compare the efficiencies of applying one- and two-stage detector models to a single UAS-based image and a processed (via Pix4D mapper photogrammetric program) UAS-based orthophoto product to detect rodent burrows, for agriculture/environmental applications as to support farmer activities in the achievements of SDG2. Our results indicate that the use of multimodal data from low-cost UASs within a self-training YOLOv3 model can provide relatively accurate and robust detection for small objects (mAP of 0.86 and an F1-score of 93.39%), and can deliver valuable insights for field management with high spatial precision able to reduce the environmental costs of crop production in the direction of precision agriculture management. [ABSTRACT FROM AUTHOR]

The number of wheat ears is an essential indicator for wheat production and yield estimation, but accurately obtaining wheat ears requires expensive manual cost and labor time. Meanwhile, the characteristics of wheat ears provide less information, and the color is consistent with the background, which can be challenging to obtain the number of wheat ears required. In this paper, the performance of Faster regions with convolutional neural networks (Faster R-CNN) and RetinaNet to predict the number of wheat ears for wheat at different growth stages under different conditions is investigated. The results show that using the Global WHEAT dataset for recognition, the RetinaNet method, and the Faster R-CNN method achieve an average accuracy of 0.82 and 0.72, with the RetinaNet method obtaining the highest recognition accuracy. Secondly, using the collected image data for recognition, the R2 of RetinaNet and Faster R-CNN after transfer learning is 0.9722 and 0.8702, respectively, indicating that the recognition accuracy of the RetinaNet method is higher on different data sets. We also tested wheat ears at both the filling and maturity stages; our proposed method has proven to be very robust (the R2 is above 90). This study provides technical support and a reference for automatic wheat ear recognition and yield estimation. [ABSTRACT FROM AUTHOR]

Because small targets have fewer pixels and carry fewer features, most target detection algorithms cannot effectively use the edge information and semantic information of small targets in the feature map, resulting in low detection accuracy, missed detections, and false detections from time to time. To solve the shortcoming of insufficient information features of small targets in the RetinaNet, this work introduces a parallel-assisted multi-scale feature enhancement module MFEM (Multi-scale Feature Enhancement Model), which uses dilated convolution with different expansion rates to avoid multiple down sampling. MFEM avoids information loss caused by multiple down sampling, and at the same time helps to assist shallow extraction of multi-scale context information. Additionally, this work adopts a backbone network improvement plan specifically designed for target detection tasks, which can effectively save small target information in high-level feature maps. The traditional top-down pyramid structure focuses on transferring high-level semantics from the top to the bottom, and the one-way information flow is not conducive to the detection of small targets. In this work, the auxiliary MFEM branch is combined with RetinaNet to construct a model with a bidirectional feature pyramid network, which can effectively integrate the strong semantic information of the high-level network and high-resolution information regarding the low level. The bidirectional feature pyramid network designed in this work is a symmetrical structure, including a top-down branch and a bottom-up branch, performs the transfer and fusion of strong semantic information and strong resolution information. To prove the effectiveness of the algorithm FE-RetinaNet (Feature Enhancement RetinaNet), this work conducts experiments on the MS COCO. Compared with the original RetinaNet, the improved RetinaNet has achieved a 1.8% improvement in the detection accuracy (mAP) on the MS COCO, and the COCO AP is 36.2%; FE-RetinaNet has a good detection effect on small targets, with APs increased by 3.2%. [ABSTRACT FROM AUTHOR]

Transverse aeolian ridges (TARs) are unusual bedforms on the surface of Mars. TARs are common but sparse on Mars; TAR fields are small, rarely continuous, and scattered, making manual mapping impractical. There have been many efforts to automatically classify the Martian surface, but they have never explicitly located TARs successfully. Here, we present a simple adaptation of the off-the-shelf neural network RetinaNet that is designed to identify the presence of TARs at a 50-m scale. Once trained, the network was able to identify TARs with high precision (92.9%). Our model also shows promising results for applications to other surficial features like ripples and polygonal terrain. In the future, we hope to apply this model more broadly and generate a large database of TAR distributions on Mars. [ABSTRACT FROM AUTHOR]

The development of double haploids (DHs) is a straightforward path for obtaining pure lines but has multiple bottlenecks. Among them is the determination of the optimal stage of pollen induction for androgenesis. In this work, we developed Microscan, a deep learning-based system for the detection and recognition of the stages of pollen development. In a first experiment, the algorithm was developed adapting the RetinaNet predictive model using microspores of different eggplant accessions as samples. A mean average precision of 86.30% was obtained. In a second experiment, the anther range to be cultivated in vitro was determined in three eggplant genotypes by applying the Microscan system. Subsequently, they were cultivated following two different androgenesis protocols (Cb and E6). The response was only observed in the anther size range predicted by Microscan, obtaining the best results with the E6 protocol. The plants obtained were characterized by flow cytometry and with the Single Primer Enrichment Technology high-throughput genotyping platform, obtaining a high rate of confirmed haploid and double haploid plants. Microscan has been revealed as a tool for the high-throughput efficient analysis of microspore samples, as it has been exemplified in eggplant by providing an increase in the yield of DHs production. [ABSTRACT FROM AUTHOR]

Traffic signs are a key element in driver safety. Governments invest a great amount of resources in maintaining the traffic signs in good condition, for which a correct inventory is necessary. This work presents a novel method for mapping traffic signs based on data acquired with MMS (Mobile Mapping System): images and point clouds. On the one hand, images are faster to process and artificial intelligence techniques, specifically Convolutional Neural Networks, are more optimized than in point clouds. On the other hand, point clouds allow a more exact positioning than the exclusive use of images. The false positive rate per image is only 0.004. First, traffic signs are detected in the images obtained by the 360° camera of the MMS through RetinaNet and they are classified by their corresponding InceptionV3 network. The signs are then positioned in the georeferenced point cloud by means of a projection according to the pinhole model from the images. Finally, duplicate geolocalized signs detected in multiple images are filtered. The method has been tested in two real case studies with 214 images, where 89.7% of the signals have been correctly detected, of which 92.5% have been correctly classified and 97.5% have been located with an error of less than 0.5 m. This sequence, which combines images to detection–classification, and point clouds to geo-referencing, in this order, optimizes processing time and allows this method to be included in a company's production process. The method is conducted automatically and takes advantage of the strengths of each data type. [ABSTRACT FROM AUTHOR]