Research On Animal Target Detection Method Based On One-stage Deep Neural Network

With the development of intellectualization and scale of animal husbandry,it is inefficient and increases labor cost to monitor animal health status,abnormal behavior and other aspects by artificial monitoring.Therefore,the use of deep learning monitoring technology instead of manual monitoring gradually emerged.Animal target detection is an important branch of deep learning monitoring technology,which aims to obtain the characteristic information of animal targets by building a deep neural network model,so as to confirm the category and location range of animal targets in images or videos.However,when the illumination is difficult,the environmental background is complex,the animal target is very similar to the background,and the animal population is blocked,the detection of animal target is easy to appear difficult and missed.In view of the above challenges,an animal target detection method based on one-stage deep neural network is put forward.The main research contents are as follows:Through theoretical analysis and comparison of several one-stage and two-stage target detection algorithms,in order to ensure the accuracy and speed of animal target detection,this paper adopts YOLOv3 one-stage algorithm as the basic network framework of animal target detection.Two unsaturated activation functions,Leaky Re LU and Mish function,are combined to improve the expression ability of the model.According to the problem that animal target detection is subject to complex and changeable environment and occlusion,an optimization method of feature extraction network integrated with visual attention mechanism was proposed to focus feature extraction more on irrelevant areas such as useful areas and blurred suppression background,and improve the accuracy of animal target detection by the model.At the same time,aiming at the particularity of the animal data,use the K-means++ clustering algorithm to find the pre-selection box that better fits the size of the animal,and the loss function that can better reflect the situation of overlapping animals is selected.Finally,model training and experiments are carried out on animal dataset.The experimental results show that the one-stage animal target detection algorithm combined with visual attention mechanism can reduce the environmental impact and the difficulty of multi-number animal target detection,raising animals target detection accuracy,reduce the missing rate.Aimed at the problem of occlusion of animal targets,an animal target detection method in occlusion scene is put forward.Mainly from two different way of thinking,the first,from the Angle of the loss function to solve,the repulsion loss function are introduced to better distinguish the overlapping occluding animal targets,make the network in the process of automatic learning improved positioning performance,alleviate the impact of the animal obscured.In the second way,the individual features of the animals were enhanced by integrating the head features of the animals,so that the prediction score of the obscured animals could be raised above the test threshold and not suppressed by the animal target detection model.The experimental results show that the two kinds of animal target detection methods in occluded scenes have higher accuracy and lower missed detection rate than the YOLOv3 method.