Enhanced cattle identification using Siamese network and MobileViT with EMA attention.

Abstract

Accurate identification of individual cattle is paramount in livestock insurance to combat fraud. However, the performance of existing muzzle recognition methods degrades in complex scenarios involving occlusion or multi-angle views. This study addresses this limitation by first constructing a comprehensive cattle muzzle image dataset encompassing frontal, multi-angle, and occluded conditions. We then propose CattleMuzzleNet, a lightweight recognition model that integrates a siamese network, an enhanced MobileViT backbone, and an Efficient Multi-scale Attention (EMA) mechanism for robust feature extraction and matching. Its efficacy is systematically validated through comparative experiments on feature extraction networks, ablation studies on the attention mechanism, and confidence threshold analysis. Evaluated on a dataset of 31,312 images from 658 cattle, CattleMuzzleNet achieved an accuracy of 97.87% and an F1-score of 98.89%, with a compact model size of 6.9 MB. The results demonstrate high accuracy and robustness in complex scenarios, providing an effective technical solution for identity verification in cattle insurance.