Enhanced Cattle Identification using Siamese Network and MobileViT with EMAAttention

Mingshuo HanBaoshan LiQi Li^*Yueming WangMei YangChang Gao

• Inner Mongolia University of Science and Technology, Baotou, China

In the field of cattle insurance, the application of muzzle print recognition technology is essential for the precise identification of individual cattle and the reduction of fraudulent actions. The effectiveness of current recognition systems is reduced when faced with muzzle prints that are partially concealed or taken from various angles. This work developed a library of cattle muzzle pattern photos featuring frontal views, various angles, and occlusion conditions. A new cattle muzzle pattern recognition model, named CattleMuzzleNet, was created by combining a dual neural network architecture with an improved MobileViT algorithm and integrating an EMA attention mechanism. The model's methodology entails extracting visual features, comparing these features with existing images for similarity, and subsequently producing recognition results. The CattleMuzzleNet model served as the basis for our experimental framework, which included feature extraction network comparison studies, attention module ablation experiments, and algorithm confidence threshold comparison tests. The CattleMuzzleNet model, with a size of only 6.9 MB, is remarkable for achieving an accuracy rate of 97.87% and an F1 score of 98.89%, based on an assessment of a test dataset of 658 animals and 31,312 photos. The experimental findings demonstrate that a similarity level of 0.5 enables CattleMuzzleNet to attain optimal performance, hence offering technical help for the cattle farming insurance sector.