Real-Time Sows Behaviors Detection Algorithm Based On Deep Learning

Recording various behaviors of the sows can closely track their health status,detect their abnormalities and provide suitable assistance for them in time,so that the sows can grow more physically and mentally.The traditional method is mainly to use pure manual observation or manual inspection of video recording forms,which requires a lot of manpower and material resources.As a non-intrusive monitoring solution,machine vision technology plays an important role in the field of behavior tracking and identification.In recent years,it has become a research hotspot to use machine vision technology to conduct research on the behavior detection of sows.However,the current detection methods are based on the accurate identification of pigs as a precondition,and they mainly focus on one behavior detection of pigs.The paper proposes a real-time detection algorithm for sows’ behaviors based on deep learning named SBDA-DL,which detects three behaviors of sows including drinking behavior,urinating behavior and mounting behavior.The algorithm mainly includes the following four parts.(1)Pre-processing the behaviors images of sows,which includes image center interception,image compression processing,image noise processing and data enhancement processing.(2)Using MobileNet classification network model.It contains continuous depthwise separable convolutions,batch normalization and ReLU nonlinear activation to the feature maps.(3)Using SSD detection network to perform Sliding Window operation on multi-scale feature maps,and then,using Softmax classification machine to classifies the default boxes,and finally the detection window is determined by Non-Maximum Suppression.(4)Optimizing the SSD detection network.We accelerate the network detection speed by compressing network models,increase network detection accuracy rate by choosing positive and negative sample proportions,classifying drinking behavior and replacing activation function to linear function.In this experiments,the sows were filmed on Lizhi farms for 4 mounths.And then,we intercepted the behaviors image,labeled their behaviors and marked their specific behaviors location manually.It forms a data set containing 1912 pieces of images,of which 1338 were used for training and 574 for testing.The accuracy precision of drinking behavior,urinating behavior and mounting behavior of sows are 96.5%,91.4% and 92.3%.The mean accuracy precision is 93.4%.The behaviors detection speed of ordinary microcomputers is 7fps.The detection algorithm uses a special deep learning network structure to directly detect the behavior of the sow,which improves the detection accuracy of the behavior while achieves the processing speed of the real-time detection.The detection algorithm basically meets the requirements of farms,and it can assist the staff in daily monitoring.