Research On Lightweight Pig Detection Network Based On Deep Learning

Swine production is the mainstay of China’s animal husbandry and plays a crucial role in it.In recent years,influenced by social and environmental factors,the swine production industry has seen a significant withdrawal of free range farmers,further accelerating the standardization,scale,and modernization process of swine production in China.With the advent of the digital era,artificial intelligence has been widely applied in many fields such as home furnishing,manufacturing,finance,healthcare,etc.Integrating swine production with artificial intelligence is the trend,and building intelligent pig farms requires gradually intelligent and unmanned farming to adapt to industrial development.Among them,the detection task of pigs,as the foundation of many other pig recognition tasks,is very important for the intelligence of pig breeding.However,currently,pig detection algorithms still have problems of insufficient accuracy and speed in practical applications.This project takes the object detection network YOLOv5 as the overall framework,further improving the detection speed and accuracy,and designing a lightweight pig detection network.First of all,this paper makes lightweight improvements to the backbone of pig detection network,and designs three lightweight improvement strategies.The lightweight improvement of strategy 1 is based on group convolution and channel shuffle,which reduces computational complexity and integrates the features of different group inputs.Strategy 1 also proposes four lightweight design ideas to assist in designing network structures better.Strategy 2 adopts depthwise separable convolution.In order to make the activation function obtain more sufficient features,depthwise convolution and pointwise convolution are used to raise the feature extraction part to high dimensions,and some high dimension activation function are improved accordingly to obtain features better.Strategy 3 combines the group convolution and lightweight structure design ideas of Strategy 1,introducing the Ghost Module input feature map into two groups.One group uses intrinsic convolution to generate features,and the other group uses depthwise convolution in strategy 2 to reduce computational complexity.In addition,decoupled fully connected attention is added to the Ghost Module so that the convolution structure can capture the correlation between pixels in remote space,greatly enhancing the expression ability of the model.The experiment conducted accuracy and speed tests on three improvement strategies,and the results showed that the strategy 3 which combined the first two lightweight improvement ideas,had the smallest decrease in accuracy and the largest increase in speed compared to the original detection network.Secondly,this article addresses the issue of reduced detection network accuracy after lightweight,pruning the feature pyramid structure and incorporating coordinate attention.The original detection network uses PAFPN to fuse top-down Semantic information and bottom-up location information.The improved structure,on the basis of the original structure,deletes nodes that have little effect,and adds hop connections to compensate for more pig feature information.The improved structure will delete the nodes that have little effect and increase hop connections on the basis of the original structure.Integrate channel information with coordinate information using coordinate attention mechanism to further enhance feature extraction capabilities.The ablation experiments on feature pyramid and coordinate attention have shown that both have a positive effect on improving the accuracy of the detection network.Finally,the lightweight improvement is combined with precision improvement,and the new detection network is named PDNet.Through the final comprehensive experimental testing,the effectiveness of each module improvement and the new pig detection network in this article has been proven.