Research on Cattle Face Recognition Algorithms Based on Spherical Metric Learning

Accurate cattle identification is crucial for modern cattle farm management. However, cattle face recognition faces c hallenges due to factors such as animal movement and varying ca mera angles, with existing models exhibiting low recognition accu racy for multi-angle and blurred images. This study proposes a n ovel metric learning approach to address multi-angle and blurrin g issues in cattle face recognition. MobileNet V3 is employed as th e feature extraction network, transforming the embedding space i nto a spherical space via spherical projection to enhance represen tational capacity for irregularly distributed samples. An intra-cla ss multi-subcenter proxy method is introduced, and triplet loss co mbined with softmax loss are jointly employed to minimize intra-class distances while maximizing inter-class distances. Experimen tal results demonstrate recognition accuracies of 99.60% and 99.4 4% on multi-angle and blurred multi-angle datasets, respectively.