Research On Object Detection Algorithm Of Cattle With Lightweight Based On Deep Learning

With the development of my country’s animal husbandry,more and more breeding companies have begun to rise,and some traditional breeding companies are facing the risk of being eliminated if they do not change the status quo.In this years,cattle husbandry as a traditional industry has faced a lot of challenges.In traditional animal husbandry,cattle breeding basically depends on artificial feeding.In order to ensure the health of cattle and prevent the large number of infectious diseases such as mad cow disease and bovine pneumonia,cattle farms need A large number of workers are on standby 24 hours a day to check the health of the cows and count the number of cows,and follow up in time if there is an abnormality.This traditional breeding method requires a lot of manpower,material resources and financial resources.For improving the information management ability of cattle farms and decreasing breeding spending,improve the economic benefits of pastures,and promote healthy breeding.The research content of this article is to conduct object detection and quantitative statistics for cattle herds under the framework of deep learning.Due to the limited space of cattle farms,the cattle herd is relatively dense and the cattle body is large,and the cattle bodies will be missed due to mutual occlusion;the similar background of the cattle herd and the cowshed will cause certain difficulties in the feature extraction of the cattle body.In the past ten years,deep learning methods have gradually emerged in the field of machine learning,and have been applied by more and more researchers in all walks of life.There are also researchers who conduct research on cattle object detection based on deep learning.In view of the above-mentioned research difficulties in the detection of the cattle herd,the main research work of this paper is as follows:(1)Proposing a cattle object detection algorithm based on E-DCN-Cascade RCNN.First,use the Expand-Deformable Convolutional Networks(E-DCN)to extract the texture features of the cow’s body,and make the convolution area always cover the cow’s body,which improves the distinction between the cow’s body and the background of the cowshed.;Then,combined with E-DCN and Feature Pyramid Network(FPN),the feature information of each layer is deeply fused,which enhances the pattern feature that blocks the cow’s body,and obtains the feature map of each layer of the cow’s body,which solves the problem of Missed detection caused by occlusion.Since there is no public cattle data set,on the cattle data set produced in the laboratory,when the IOU threshold is set to 0.75,AP value of this algorithm is 4.2%higher than that of Cascade RCNN.Experimental results show that this algorithm has better detection results.(2)Proposing a lightweight method of cattle detection model based on DC-SMKD.There are a large number of convolutions in the feature extraction backbone network,which is very suitable for using separable convolutions for feature sparseness and reducing the size of the model.Since most of the cow back pattern information is a relatively large block pattern,after separable convolution,the characteristic parameters will be reduced to a certain extent,but the cow back pattern information will not disappear.After separable convolution,the reduction of bovine body feature parameters will lead to a decrease in m AP.The teacher model is subsequently used to perform smooth multi-layer perceptual knowledge distillation on the model after separable convolution to improve the model’s performance after separable convolution.Detection accuracy.Comparing this algorithm with traditional SSD object detection algorithms,the network model of this algorithm is 80 M smaller than that of SSD,but AP is 1.34% higher than SSD.It can be seen that the network model obtained by this algorithm is smaller than that of SSD,and the detection effect is better.