| In the past decade,with the rapid development of animal husbandry,the information technology of pig breeding has achieved rapid development.In order to accelerate the development of breeding information technology in pig breeding,realize high efficiency pig individual identification and counting,and improve the shortcomings of manual statistics in the traditional pig breeding industry,this paper proposes a pig individual identification and counting method based on computer vision.The main research contents of this paper can be summarized as follows:(1)To study the working principle and characteristics of deep learning-based target detection algorithms,compare the advantages and disadvantages of different algorithms,and choose YOLOV3 as the target detection algorithm model for this experimental study.The overall network structure of YOLOV3 algorithm model is studied deeply,and it is applied to the experiment.(2)Parameters and network optimization of YOLOV3 algorithm model were carried out in the experiment.The original 20 types of detection were reduced to the detection of single pigs.K-means algorithm was used to re-cluster and calculate the anchors in the network to make the model more in line with the experimental requirements of pig detection.It not only greatly simplifies the training model,but also improves the detection rate of pigs and the detection speed of the algorithm,so as to achieve the goal of real-time monitoring.(3)Under the same hardware configuration and data set,the model selected in this experiment was compared with the recognition results of Faster RCNN and SSD models.The results showed that the average accuracy(AP),accuracy(P)and recall rate(R)of piglets in the sow farm reached 89.65%,89.65% and 95.99%.The average accuracy(AP),accuracy(P)and recall rate(R)of sows were 95.16%,96.00% and 96.00% respectively.The average accuracy(AP),accuracy(P)and recall rate(R)of finishing pigs were 92.03%,92.23% and 89.57% in the fattening farms.Compared with Faster RCNN,the recognition rate of this model is 6 times higher than that of Faster RCNN.Compared with SSD,the average accuracy of this model is improved by 12.28%.In the comparison of the Yolov3 model before and after the improvement of the fattening pig farm,the FPS,AP and IOU of the improved Yolov3 are increased by 2.5,1.45% and 6.55%,indicating that the recognition rate and accuracy of this model have reached a higher level.(4)In this paper,a counter is added on the basis of the original model,so that it can automatically conduct quantity statistics after the completion of identification.This study provides a theoretical basis and technical support for the realization of monitoring the number of live pigs in large-scale breeding and the promotion of breeding information technology. |
