Estrus Detection of Sows Based on Improved YOLOv5s Algorithm

Bao Quan; Zhou Xin; Wu Yue; Wang Kun; Xu Xing; Jin Mengyang; Wang Xingbo; Huang Juan; Zhou Weidong

doi:10.14083/j.issn.1001-4942.2026.01.020

Abstract：In order to solve the problems of low efficiency and precision of traditional manual detection,this study proposed an estrus detection method of sows based on improved YOLOv5s,aiming at the characteris-tics of vulva becoming red and swollen during estrus.Specific improvements included adding the Coordinate Attention(CA)mechanism to the backbone network of YOLOv5s,replacing the original interpolation-based upsampling with CARAFE upsampling,and substituting the original CIoU loss function with the EIoU loss function.Furthermore,the ablation experiments and comparative tests with other models were conducted to verify the effectiveness of the proposed method.The results showed that the improved model achieved the pre-cision of 94.5%,recall of 93.3%and mean average precision(mAP@0.5)of 93.8%for sow vulvar morpholo-gy recognition,respectively.Compared with some YOLO series models(including the original YOLOv5s)and the Faster R-CNN model,the improved model maintained the highest precision,recall and mAP@0.5 for sow vulvar morphology recognition.Meanwhile,its model weight,parameter count and computational complexity were 16.3 MB,7.2 M and 16.3 G,respectively,which was slightly higher than those of the original YOLOv5s(15.8 MB,7.0 M,14.5 G)but significantly lower than those of other comparative models.In terms of detec-tion speed,the frame rate(FPS)of the improved model reached 64.3 frames per second(fps),which was slightly lower than those of the original YOLOv5s(73.2 fps)and YOLOv8s(67.1 fps)but notably higher than those of the other models.These results indicated that the improved YOLOv5s model proposed in this study could significantly enhance detection accuracy and promote real-time detection performance,thereby providing technical support for sow estrus detection.