Intelligent recognition and counting of pig faces based on improved YOLOv8 and MCGIRNet

In the context of the intelligent development of pig breeding, pig face recognition and counting, as a key link in intelligent management, face problems such as similar facial features of pigs and overlapping and occlusion of targets. In view of the limitations of existing methods, this study makes multi-dimensional improvements to YOLOv8, while building a multi-constraint ghost inverted residual network to optimize feature extraction capabilities, and integrating the two to build a pig face intelligent recognition and counting model. Comparative experiments show that the model is significantly better than the comparison models in the core indicators of pig face recognition and detection. In the PetFace pig subset and the FIRC dataset, the mAP reaches 0.966 and 0.955 when the IoU threshold is 0.5. The latter has only 15 M parameters, which is the best among all comparison models. The model’s recognition consistency finally stabilized at 92.7% in a backlight environment, and reached 94.8% in a dark light environment. The detection and recognition accuracies in the head-turning posture reach 0.938 and 0.919, taking into account detection accuracy and scene adaptability. The research model can effectively deal with the multi-interference problem of pig face recognition in breeding scenarios, accurately capture the core features of pig faces, and significantly improve the accuracy of pig face recognition and the reliability of stock counting. This study provides scientific basis and technical support for intelligent and refined management of pig breeding.