Research On Pig Face Recognition Algorithm Based On Deep Learning

With the rapid development of precision farming in intelligent agriculture,the promotion process of automated,large-scale and intensive pig farming has accelerated significantly.The foundation of achieving scientific pig farming lies in identification of individual pig identities,which has important research significance for disease prevention and control of pig farming,tracing quality of pork products,and animal husbandry insurance In terms of traditional identification in live pigs,many managers adopt color labeling,tattoo branding,RFID identification and other ways.These methods not only waste manpower,but also easily cause the stress response of pigs,which does not meet the needs of animal welfare.With the development of deep learning,the application of convolutional neural network to build pig face recognition model has performed well.However,many existing models cannot be widely used in breeding industry due to their complex calculation and large parameter quantity.In addition,the environment of pig house is highly complex and the interference information is large.Therefore,this paper draws lessons from the implementation process of face recognition,and adopts three processes of pig face detection,pig face alignment and pig face recognition to study the individual pigs.The main work includes the following aspects:(1)Research on pig face detection model construction based on YOLOv4.In order to accelerate the detection speed and reduce the number of parameters of original model,the backbone network of YOLOv4 was replaced by lightweight model Mobile Net-v3,and depth separable convolution was introduced.In order to ensure the detection accuracy,CBAM attention mechanism was added into PANet.The introduction of multi-attention mechanism could selectively strengthen key areas of pig face and filter out weak correlation features to enhance the overall model effect.CIo U loss function was introduced for the regression accuracy of pig face boundary frame.Experimental results show that m AP of Mobile Net-YOLOv4-CPNet model reaches 98.15%,the detection FPS reaches 106.3 frames/s,and the model parameter size is only 44.74 MB,which improves the parameter optimization by 82% compared with the YOLOV4 model.(2)Research on pig face key point detection algorithm based on Openpose and pig face alignment method.To reduce the spatial distribution differences of pig facial images,this paper uses positional relationship of feature points for pig facial alignment operations,making the recognition model more powerful in feature extraction and high fault tolerance.In order to speed up the detection speed and avoid the phenomenon of gradient disappearance and degradation in the training,the feature extraction network of original Openpose model was replaced by Res Net-18 with adaptive soft threshold,and the average detection error rate was lower.Compared with common detection model,R-Openpose key point detection algorithm can effectively capture the key points of pig face,and through the output of left and right eye key point position coordinates,implement rotation alignment of pig face at different angles,so as to prepare for later pig face recognition work.(3)Research on pig face recognition model based on Shuffle Net V2.In order to meet the requirements of lightweight model and the high complexity of pig house environment,Shuffle Unit module of lightweight model Shuffle Net V2 was optimized.DW of main branch was expanded to enhance the sensitivity field and the redundant point by point convolution at the end was trimmed to reduce the complexity of model.At the same time,CSP mechanism was introduced into Stage network to improve the learning ability of the network by branch paths and sparse connections.The integration of Sim AM attention-free mechanism after each Stage enables the network model to learn more differentiated neurons.In order to realize intra-class compactness and inter-class separability of pig identification,Center Loss and Softmax Loss were used for joint training.Experimental results show that the accuracy of Shuffle Net＿LS pig face recognition model is 96.84%,3.24 percentage points higher than that of Shuffle Net V2 model.It also verifies validity and necessity of pig face alignment operation for pig face recognition task.