Research And Implementation Of Animal Target Detection Algorithm Based On YOLO

Thanks to the rapid development of deep learning technology,the research of target detection algorithm based on deep learning has become a hot research direction in recent years.In this paper,the target detection algorithm is combined with the animal target detection scene,and the algorithm model is optimized based on the actual business scene of the company to meet the actual production needs.The focus of this paper is mainly on two aspects,one is how to improve the performance of the algorithm to improve the accuracy of animal object detection and classification,and the other is how to deploy the target detection algorithm with huge weight parameters in edge devices with poor hardware performance.First of all,this paper improves the detection effect of small targets and overlapping targets by integrating mixed hole convolution,and then uses a channel pruning algorithm based on K-means clustering algorithm to compress the network model,and then compress the network model.The data set is preprocessed and the method of non-maximum suppression is used to reduce the positioning error.In order to reduce the size of the equipment,reduce the hardware cost and the difficulty of deployment,and for the effect of small target animal detection in wild scenes,this paper proposes a method based on The target detection algorithm of YOLOv4 is the M-HDC-YOLO algorithm.Finally,the detection network framework is transplanted to the embedded platform based on the K210 chip design to achieve the purpose of real-time detection.The main work of this paper is as follows:1.In order to optimize the performance of YOLOv4 algorithm in small target detection,the HDC-YOLOv4 algorithm is proposed,which uses hybrid hole convolution to replace the original convolution layer,and achieves the algorithm by acquiring more feature maps with less parameters.Accuracy improvements.2.In view of the fact that the weight file generated by the existing model is too large and the calculation amount far exceeds the limitation of the hardware platform,this paper uses a compression algorithm to slim down the model,and the compression rate reaches80%.The size of the model is only about 4M,which can be well supported by the hardware platform.The experimental results show that the animal target detection algorithm applied in this paper can correctly judge 10 types of animals collected in wild scenes,and maintain a high network compression rate and recognition speed with only a small loss of accuracy.On the RTX3090 hardware platform,after the model is compressed,the m AP can reach89.26% and the FPS can reach 112;when transplanted to the K210 hardware platform,the m AP is 68.63% and the FPS is 14,which can meet the basic requirements of real-time detection and complete the animal target.Embedded implementation of detection.