A Method for Dairy Cow Identification Based on Multi-view Feature Fusion of the Cow's Back

To address the identification difficulties in single-view cow back identification during milking scenarios caused by partial body occlusion, posture variations, and changes in markings and stains, this paper proposes a dairy cow identification method based on multi-view feature fusion of the cow's back. First, a multi-view data set of the cow's back was constructed, containing 31,356 images from 1954 cows. Then, YOLO11n and ByteTrack were combined to achieve adaptive image acquisition of the central, left, and right views of the cow's back in the milking scenario. Next, by comparing the recognition performance of 7 feature extraction backbones (HRNet, EfficientNet, DenseNet, ResNet, ConvNeXt, Swin Transformer, and Swin Transformer V2) on mixed-view data, the 3 best-performing networks were selected, and then trained separately to find the optimal single-view feature extraction model for the central, left, and right views. Finally, the extracted central, left-side, and right-side features were concatenated, and the cow's identity was matched by calculating the cosine distance. The experiments proved that the proposed method achieved an overall accuracy of 99.69%, a mAP of 99.83%, and an accuracy of 94% on samples that could not be identified by single-view models. It achieves accurate identification of cows in milking scenarios and provides an effective method for dairy cow identification.