A two-stage framework for cattle face re-identification with feature-contour and fine-texture enhancement

To address the challenge of accurate identification in low-texture cattle breeds such as Angus, this study presents a structurally optimized two-stage cattle face re-identification framework tailored for complex farming environments. In the detection stage, a lightweight model named Pruned Edge-enhanced Real-Time Detection Transformer (PrunedEdge-DETR) is proposed to enhance contour sensitivity while significantly reducing computational cost via grouped Taylor pruning. For the recognition stage, a dual-branch feature extraction network called Dual-channel fused MobileViT (DCFuseViT) is designed, which fuses local contours, fine-grained textures, and global semantic features to improve feature discriminability. Additionally, a feature library–based matching mechanism named EffiLibCatReID is introduced to enable fast and scalable identity retrieval. Experimental results on the AngusDataset-128 demonstrate that the proposed method achieves a re-identification accuracy of 97.8% with an average inference time of 9.61 ms per image, confirming its real-time performance and deployment potential. The source code is available at: https://github.com/HLJ11235/cattle-re-id.git .