Non-contact yak estrus detection using ocular thermal infrared images and an attention-enhanced DenseNet model

Yak estrus detection in plateau pastoral areas still relies on inefficient manual observation and contact-based devices with limited adaptability, which restricts the development of intelligent breeding systems. This study proposes a non-contact estrus detection method based on ocular thermal infrared images for accurate and efficient recognition. A high-quality thermal infrared dataset of Qinghai yaks was constructed to capture periocular temperature distribution characteristics under real grazing conditions. An improved ECA-DenseNet-SC model is developed to address feature redundancy, limited gradient propagation, and high computational cost in traditional DenseNet. The model integrates Efficient Channel Attention (ECA), depthwise separable convolutions, and a feature branching and cross-layer fusion mechanism to enhance feature representation while maintaining efficiency. Experimental results show that the proposed model achieves a precision of 97.21%, recall of 97.15%, and F1-score of 97.18%. Compared with classical networks including DenseNet, MobileNet, and ResNet, the proposed method demonstrates superior performance in estrus recognition while maintaining a lightweight architecture.