Search

Due to the over-parameterization of neural networks, many model compression methods based on pruning and quantization have emerged. They are remarkable in reducing the size, parameter number, and computational complexity of the model. However, most of the models compressed by such methods need the support of special hardware and software, which increases the deployment cost. Moreover, these methods are mainly used in classification tasks, and rarely directly used in detection tasks. To address these issues, for the object detection network we introduce a three-stage model compression method: dynamic sparse training, group channel pruning, and spatial attention distilling. Firstly, to select out the unimportant channels in the network and maintain a good balance between sparsity and accuracy, we put forward a dynamic sparse training method, which introduces a variable sparse rate, and the sparse rate will change with the training process of the network. Secondly, to reduce the effect of pruning on network accuracy, we propose a novel pruning method called group channel pruning. In particular, we divide the network into multiple groups according to the scales of the feature layer and the similarity of module structure in the network, and then we use different pruning thresholds to prune the channels in each group. Finally, to recover the accuracy of the pruned network, we use an improved knowledge distillation method for the pruned network. Especially, we extract spatial attention information from the feature maps of specific scales in each group as knowledge for distillation. In the experiments, we use YOLOv4 as the object detection network and PASCAL VOC as the training dataset. Our method reduces the parameters of the model by 64.7 % and the calculation by 34.9%., Appl Intell

The morphological features of fish, such as the body length, the body width, the caudal peduncle length, the caudal peduncle width, the pupil diameter, and the eye diameter are very important indicators in smart mariculture. Therefore, the accurate measurement of the morphological features is of great significance. However, the existing measurement methods mainly rely on manual measurement, which is operationally complex, low efficiency, and high subjectivity. To address these issues, this paper proposes a scheme for segmenting fish image and measuring fish morphological features indicators based on Mask R-CNN. Firstly, the fish body images are acquired by a home-made image acquisition device. Then, the fish images are preprocessed and labeled, and fed into the Mask R-CNN for training. Finally, the trained model is used to segment fish image, thus the morphological features indicators of the fish can be obtained. The experimental results demonstrate that the proposed scheme can segment the fish body in pure and complex backgrounds with remarkable performance. In pure background, the average relative errors (AREs) of all indicators measured all are less than 2.8%, and the AREs of body length and body width are less than 0.8%. In complex background, the AREs of all indicators are less than 3%, and the AREs of body length and body width is less than 1.8%.

The statistics of disease spores is significant for early strategy design of disease control in precision agriculture. To obtain the statistics information of spores in microscopic images, it is crucial to segment spores from images. In this paper, we research a deep learning based method to segment spores, taking anthrax spores as the research objects. We first built an anthrax spore dataset consisting of more than 40,000 spores with accurate labeled spore boundaries to advance the state of the art technology of spore statistics. Then on consideration of the complex class imbalances in actual anthrax spore images, we investigate how class imbalances and hard examples simultaneously influence the loss during training and we discover that hard examples are more likely to appear at the pixels of rare pixels, such as small class pixels and contour pixels. Based on this discovery, we propose Constrained Focal Loss (CFL), which focuses on small class objects, and has a constrained term related to hard examples. In addition, we further propose CFL*, where high importance is put on the pixels surrounding spore contours to improve classification accuracy. The results show that the mean IoU of the DeepLabv3+ trained with CFL* (called as CFL*Net) achieves 91.0%, higher than original DeepLabv3+ with cross-entropy by 8.6 points, and the DeepLabv3+ with Focal Loss by 10.4 points. Moreover, CFLNet* can achieve better performance than original DeepLabv3+, using less than one-third of the training samples and half of the training steps.

The anthracnose is one of the most serious diseases in the growth period of mango. In order to take preventive measures timely, it is indispensable to calculate accurate statistics on the distribution density of anthrax spores on the farm, which has challenges in accurate instance segmentation of adherent spores. Based on the traditional watershed algorithm, which treats the image as a morphological topography and segments the image by finding the lowest and highest points on the topography, we proposed the marked watershed algorithm combined with morphological preprocessing to realize the segmentation of adherent spores. Firstly, the spore images are preprocessed with morphology technique. Then the gradient values of the spore images are calculated. The segmentation of spores is performed in the gradient image by the watershed algorithm with foreground mark and background mark. The experimental result shows that our proposal has a better segmentation performance for adherent spores than the morphological method and the level set evolution.

The automatic detection and segmentation of pest based on the technology of image processing and computer vision can not only reduce the human effort and improve the detection precision for a better guideline in the prevention and control of agricultural pest, but also provide a method to capture and label the training samples for deep learning automatically. In this paper, we use a mobile robot car to automatically capture the scene image in field, and we propose a method to detect and segment the pests/diseases in the acquired image. Firstly, a Gaussian Mixture Model (GMM) is constructed for the pest/disease from only one template pest image, then we take use of the logarithm similarity to the GMM and Aggregation Dispersion Variance (ADV) based approach to detect the specified pest/disease in plant. It is likely to make a wrong judgment when the pest is close to the lens. In order to avoid such mistake, we also combine the mean and the area as the classifier. Further, we employ the distance regularization level set evolution (DRLSE) driven by the similarity to evolve the contour toward the actual pest/disease contour. Taking the pests belonging to Pyralidae as a case study, the result shows that our method could automatically identify the positive and negative samples of the specific pest from a large number of scene images, and the recognition accuracy was up to 95%. For the positive samples, our algorithm could also segment the pests accurately, which shows that our method can realize the real-time detection of the specific pest, and also provide a feasible scheme for the establishment of pests' data set.

In smart mariculture, an automatic and accurate prediction of key water quality parameters is a significant and challenge issue. This paper focuses on the prediction of pH and water temperature parameters in key water quality parameters. Firstly, the water quality parameters are preprocessed by improved method. Then, the Pearson correlation coefficient method is used to find the correlation between the water quality parameters. Finally, the SRU (Simple Recurrent Unit) deep learning model is used to establish a prediction model for the key water quality parameters, so as to achieve accurate prediction. Meanwhile, we also evaluate the prediction effect of prediction model built by RNN (Recurrent Neural Network) deep learning network. The experimental results show that the proposed prediction method has higher prediction accuracy than the method based on RNN when the time complexity is similar. The proposed method takes 11.3ms to predict every data in average, and the prediction accuracy can reach 98.91%.

Many species of Pyralidae insects are the important pests in agriculture production. However, the manual detection and identification of Pyralidae insects are labor intensive, inefficient, and subjective factors can influence recognition accuracy. To address these shortcomings, an unmanned monitoring robot car is designed. Firstly, the robot gets images by performing a fixed action and detect whether there are Pyralidae insects in the images. Secondly, the detection algorithms obtain the total probability image by using reverse mapping of histogram and multi-template images. Finally, according to the Hu moment characters, perimeter and area characters, the contours can be filtrated, and the recognition results are marked by triangles. The theoretical analysis and experimental results show that the proposed scheme has high timeliness and high recognition accuracy in the natural planting scene.

The statistical information of disease fungal spores is of great significance for precision agriculture. This work proposes a novel image segmentation method to automatically calculate the density and number of anthrax spores from microscopic images. Firstly, we prove that the widely used Otsu threshold has bias towards one class with larger variance, and the bias degree varies with the size difference of classes. The existing Otsu methods only consider the stable unbalanced distributions and are not effective when the data distributions are unbalanced and diverse. To address this problem, we propose a new evaluation index, named as Gaussian Separate Degree (GSD), to evaluate the image clustering results. Based on the index, a simple and effective thresholding method, called as G-Otsu, is proposed to segment anthrax spore images. Secondly, we take advantage of the connectivity, shape and area features of anthrax spores to filter out interferences. Next, we propose a signed logarithmic Gaussian Mixture Model similarity (SLGS) based level set model to accurately evolve the spore contours, which can automatically adjust the evolution direction of contours, ensuring the correct evolution of contours. Lastly, the density and the number of spores can be calculated accurately. The experimental results demonstrate that our method achieves 93.0% mean IoU (intersection-over-union of the calculated spore pixels and the ground truth pixels), and 1.5% overall error rate, which are higher than the existing methods.

Aiming at the problem that contours always are fractured and not precise in the segmentation of multiple visual objects, Distance Regularized Level Set Evolution(DRLSE) is used. Considering the characteristics of background difference image, an adaptive Otsu is proposed to segment difference image. In order to take full advantage of temporal information in videos, frame difference and background difference are introduced into energy function of DRLSE. Firstly, an adaptive Otsu and asymmetric morphological filtering based method are used to obtain better initial contours. Secondly, initial contours are evolved with improved DRLSE that frame difference and background difference are integrated in as priori knowledge, which can avoid that objects contours are evolved to background edge, and reduce over segmentation. The experimental results show that the contours of multiple video targets can be obtained more precisely and rapidly with the method proposed in this paper than with the existing methods.

Aiming at the problem that precision of frame difference for moving object detection is weak, a multi-info fusion model and a novel moving object detection algorithm based on the model are presented in this paper. Firstly, the temporal difference image of two frames in a motion sequence is reconstructed with morphologic operator to obtain target area. Then the spatial-temporal information in the target area is integrated into a fusion image by using the multi-info fusion model. Finally the accurate moving object is detected with automatic threshold segmentation method. The methods of fusion in fusion model are discussed, and a static linear fusion and dynamic self-adaptive fusion based on temporal entropy are presented. The experimental results show that the edge of the obtained moving target with the multi-info fusion method proposed is more accurate than the existing method, and the time complexity is low, which meets the requirement of real-time detection.

Traffic flow detection is an important part of intelligent transportation system and it has a wide range of applications. We analyse the existing methods of traffic flow detection and propose a traffic flow detection method which based on vertical virtual road induction line (VVRIL). Firstly, according to the direction of the vehicle travelling, we set a VVRIL in the middle of the driveway. Secondly, the background image is gained from the video image with Gauss mixture model. We then make differential operation between the background image and video image to get a binary image, which we set the values of the foreground pixels as 1 and that of background pixels as 0. Thirdly, we extract the values of the pixels in the VVRIL of the binary image. Besides, we regard the vehicle maximum length obtained by self-learning as the length of the detection zone and get the information of vehicles in the VVRIL. Finally, we get the number of vehicles through the analysis of vehicle centre coordinates in the VVRIL of each video image. Experimental and theoretical analyses show that the method is accurate enough to meet the requirement of real-time performance.