Research On Sheep Detection And Identification Methods In Farm Environmen

In recent years,with the rapid development of artificial intelligence and the improvement of people’s living standards,the application of object detection and classification technology on the farm of image processing to animal husbandry has attracted much attention.In the process of sheep breeding,sheep detection and sheep identification are the key technologies to achieve intelligent and precise breeding management.In this thesis,image processing technology based on deep learning was applied to the farm environment,which has investigated sheep detection and identification methods.The main work of the thesis is as follows:The sheep detection dataset Farm Sheep and the sheep individual identification dataset Sheep Face70 were established.The key frame extraction,annotation,data enhancement and other operations of the farm surveillance video and the video data taken in the field were obtained,and finally 8,000 images were obtained to constitute the sheep detection data set,and 28,000 images constituted the sheep individual identification data set.A sheep detection model based on improved YOLOv4 was constructed.In order to achieve high-speed and accurate detection of sheep on the condition of the farm environment,an improved YOLOv4 model was proposed.Shuffle Net v2,a lightweight network based on depthwise convolution and Channel Shuffle,was selected as the feature extraction network.The convolutional block attention module was introduced in the feature extraction network to strengthen the feature extraction capability of the network from both channel and space.Use depthwise separable convolution to reduce the computational amount of standard convolution;The DIo U non-maximum suppression method was used to reduce the missed detection caused by the mutual occlusion of sheep.Transfer learning methods were introduced to improve training efficiency.The experimental results show that the improved YOLOv4 model has an AP of 94.19% in the Farm Sheep dataset,which is 2.77%higher than the original YOLOv4 network,and the model size is reduced by 202.77 MB.A sheep identification model based on CNN-Vi T was constructed.Combining Convolutional Neural Networks with Vision Transformer,a CNN-Vi T hybrid network was proposed.The residual attention convolution module was designed in the CNN-Vi T network,and depthwise convolution,efficient channel attention mechanism and residual structure were introduced into the module to improve its feature extraction ability.The shallow features extracted by the residual attention convolution module were fed into the Transformer module to extract deep features.The experimental results show that the classification accuracy of the CNN-Vi T network in the Sheep Face70 dataset reaches 98.07%.The sheep detection and identification applet was designed and implemented.Design the system structure and flow diagram according to user requirements.Develop a We Chat applet,and use the network model proposed in this thesis to visualize sheep detection and sheep identification tasks on the mobile terminal.