Individual Identification Of Group-Housed Pig Based On Deep Learning

In recent years,China's livestock breeding industry is developing in the direction of intensification,scale and specialization.The high-density closed breeding mode puts forward new requirements for the real-time monitoring of the abnormal behavior of pigs,and the premise for the monitoring of the abnormal behavior of individual pigs is the identification of pigs.Manual detection is not only time-consuming and labor-consuming,but also inefficient.Therefore,it is of great significance to study the prospect extraction,identity recognition and behavior recognition of individual pigs in group-housed pig images for improving the automation level and efficiency of livestock breeding.This paper focuses on the image segmentation and identity recognition of pig individuals based on deep learning,focusing on the top-down group-housed pig images.The main research results are as follows:(1)In the complex background,it is a big difficulty to segment each individual pig completely from the group pig images with the situation of adhesion.To solve this problem,this paper proposes two segmentation methods,and carries out a comparative experiment.The first method is to use the full convolution network to segment the whole image,and then combine the distance change and watershed algorithm to separate the individual pigs.The second method is to extract the local region of a single pig through object detection algorithm,and then use the full convolution network to segment the local region to obtain the complete image of the individual pig.Compared the accuracy of the two methods and the processing ability of the two methods,the experimental results show that the second method is better.In addition,because the contour is not completely accurate,in some cases,it is quite different from the manual segmentation,so it needs to be subdivided.After image enhancement and region growing segmentation,the segmentation accuracy is improved from 92.31% to 95.15%.(2)In face recognition,human beings will actively cooperate with the face facing camera,while pigs will not.In the top view,pigs can rotate 360 degrees,which brings certain difficulties to identity recognition.Therefore we proposed a pig head and tail recognition method based on Generalized Hough Transform,which adjusts the pig image in the top view to the same direction according to the head and tail.In this method,the contour points are sampled,and every three non collinear sampling points generate a circle.The center point of the circle is mapped to the parameter space by using the generalized Hough transform,and the head and tail of the two contours are distinguished according to the degree of aggregation and dispersion of the center points after mapping.Because of the low efficiency of sampling contour points one by one,interval sampling is considered.In this paper,the head and tail recognition experiments are carried out according to different sampling intervals.The experimentsshow that when the sampling interval is 4,the recognition rate reaches 94.6%.(3)We proposed a pig identification method based on convolutional neural network Firstly,the LBP feature of pig image is extracted.Considering that there is a lot of uncertain detail information in pig head,according to the recognition results of head and tail,the image of head area is cut off and only the trunk part is retained.Then,the GoogLeNet model of pig identification is established.The experiment compares the recognition accuracy of three different loss functions.The result shows that l-softmax has the highest accuracy,reaching 90.7%.