Integrated Convolution and Attention Enhancement-You Only Look Once: A Lightweight Model for False Estrus and Estrus Detection in Sows Using Small-Target Vulva Detection

Abstract

Accurate estrus detection and optimal insemination timing are crucial for improving sow productivity and enhancing farm profitability in intensive pig farming. However, sows’ estrus typically lasts only 48.4 ± 1.0 h, and interference from false estrus further complicates detection. This study proposes an enhanced YOLOv8 model, Integrated Convolution and Attention Enhancement (ICAE), for vulvar detection to identify the estrus stages. This model innovatively divides estrus into three phases (pre-estrus, estrus, and post-estrus) and distinguishes five different estrus states, including pseudo-estrus. ICAE-YOLO integrates the Convolution and Attention Fusion Module (CAFM) and Dual Dynamic Token Mixing (DDTM) for improved feature extraction, Dilation-wise Residual (DWR) for expanding the receptive field, and Focaler-Intersection over Union (Focaler-IoU) for boosting the performance across various detection tasks. To validate the model, it was trained and tested on a dataset of 6402 sow estrus images and compared with YOLOv8n, YOLOv5n, YOLOv7tiny, YOLOv9t, YOLOv10n, YOLOv11n, and the Faster R-CNN. The results show that ICAE-YOLO achieves an mAP of 93.4%, an F1-Score of 92.0%, GFLOPs of 8.0, and a model size of 4.97 M, reaching the highest recognition accuracy among the compared models, while maintaining a good balance between model size and performance. This model enables accurate, real-time estrus monitoring in complex, all-weather farming environments, providing a foundation for automated estrus detection in intensive pig farming.

Keywords

• sow pseudo-estrus
• sow estrus
• YOLO
• sow vulva
• machine vision