Research on herd sheep facial recognition based on multidimensional feature information fusion technology in complex environment

Fu Zhang ^* Xiaopeng Zhao Shunqing Wang Yubo Qiu Sanling Fu ^* Yakun Zhang

• Henan University of Science and Technology, Luoyang, China

To enable intelligent management of large-scale farms, a deep learning-based approach for monitoring individual livestock is proposed, together with a three-phase intelligent monitoring system designed for sheep flocks. The system incorporates multipart recognition, facial classification, and facial expression analysis, thereby facilitating inventory counting, facial identity recognition, and health monitoring for sheep within the flock environment. A multi-site detection network was designed for flock inventory counting, achieving detection of both lamb and adult sheep faces, with an emphasis on the rear of the sheep to facilitate counting. The YOLOv5s path aggregation network topology was modified through the incorporation of a multi-link convolution fusion block (MCFB), thereby improving fine-grained feature extraction for objects of different sizes. To further enhance the detection of dense small targets, a Re-Parameterizable Convolution (Repconv) structure was integrated into the YOLOv5s head. The network's ability to extract key facial features was further improved by replacing the sixth-stage structure in GhostNet with a four-layer spatially separable selfattention mechanism (SSSA) to enhance facial identity recognition. Additionally, model compression experiments were performed on the facial expression analysis network to improve detection speed and efficiency. A transfer learning strategy was employed for weight pre-training, with FPS, model weight, mean average precision (mAP), and test set accuracy as the primary evaluation metrics. The results showed that the enhanced multi-part identification network effectively extracts features from different regions of the sheep flock, achieving an average detection accuracy of 95.84%.In comparison to YOLOv5s, mAP improved by 2.55%. The improved facial classification network achieved a test set accuracy of 98.9%, a 3.1% increase compared to GhostNet. Furthermore, the facial expression analysis network reached a test set accuracy of 99.2%, representing a 3.6% increase compared to EfficientNet, thus enabling real-time facial expression detection.