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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Livestock Management

Background:

• Accurate cattle identification is vital for livestock insurance to prevent fraud.
• Current muzzle recognition systems struggle with occlusions and multi-angle views, limiting their practical application.

Purpose of the Study:

• To develop a robust cattle muzzle recognition system capable of handling complex scenarios.
• To improve the accuracy and reliability of cattle identity verification in livestock insurance.

Main Methods:

• A comprehensive cattle muzzle image dataset was created, including frontal, multi-angle, and occluded images.
• CattleMuzzleNet, a lightweight model integrating a siamese network, enhanced MobileViT backbone, and Efficient Multi-scale Attention (EMA), was proposed.
• Model performance was validated through comparative experiments, ablation studies, and confidence threshold analysis.

Main Results:

• CattleMuzzleNet achieved 97.87% accuracy and a 98.89% F1-score on a dataset of 31,312 images from 658 cattle.
• The model demonstrated high accuracy and robustness in complex scenarios, outperforming existing methods.
• The developed model has a compact size of 6.9 MB, making it efficient for practical deployment.

Conclusions:

• CattleMuzzleNet provides an effective technical solution for cattle identity verification in insurance.
• The proposed model significantly enhances the robustness of muzzle recognition in challenging real-world conditions.
• This advancement contributes to reducing fraud and improving the efficiency of livestock insurance processes.