Estrus Detection Method of Parturient Sows Based on Improved YOLO v5s

Abstract

Quickly and accurately identify estrous sows to ensure timely breeding is the key to keep sow breeding performance. Aiming at the problems of low sensitivity and accuracy rate in sows estrus identification, according to the interactive characteristics of sows and bionic boars, an estrus detection method of sows based on improved YOLO v5s was proposed. Firstly, the automatic inspection robot was used to collect the video data of sows estrus behavior. The Mosaic data augmentation was used to expand the data set to enrich the data representation and enhance the robustness of the detection model, and the estrus detection model based on YOLO v5s was constructed and then optimized by sparse training, iterative channel pruning and fine-tuning to realize model compression and acceleration. DIOU-NMS was used to replace the GIOU-NMS to improve the recognition accuracy and keep high detection accuracy with lightweight-model. The results showed that the average accuracy of the algorithm was 97.8%, the average detection time of each picture was 1.7ms, and the average detection time of each video frame was 6ms. Analyzing the interactive behavior characteristics of estrous and non-estrous sows at the end of lactation, it was found that the interactive duration and frequency of estrous sows were significantly higher than that of non-estrous sows (P<0.001). On this basis, it was found that when 20s was used as the threshold of estrus detection, the sensitivity rate of estrus detection was 90.0%, the accuracy rate was 89.6%, and the specificity was 89.1%, the method can be used to rapidly and accurately detect estrus sows.