AngusRecNet: Multi-module cooperation for facial anti-occlusion recognition in single-stage Angus cattle

In the context of the booming development of modern precision livestock farming, traditional cattle recognition methods exhibit clear limitations when faced with interference from feed residues, dirt, and other obstructions on the face. To address this, this study proposes an innovative deep learning framework, AngusRecNet, aimed at solving the facial recognition problem of Angus cattle under occlusion scenarios. The backbone network of AngusRecNet includes the innovatively designed Occlusion-Robust Feature Extraction Module (ORFEM) and the Vision AeroStack Module (VASM). By combining Asymmetric convolutions and fine spatial sampling, it effectively captures facial features. The neck structure is integrated with the Mamba architecture and the core ideas of DySample, leading to the design of the State Space Dynamic Sampling Feature Pyramid Network (SS-DSFPN), which enhances multi-scale feature extraction and fusion capabilities under occlusion scenarios. Additionally, the proposed Mish-Driven Channel-Spatial Transformer Head (MCST-Head), combining Channel Spatial Fusion Transformer (CSFT) and Smooth Depth Convolution (SDConv), optimizes feature representation and spatial perception in deep learning networks, significantly improving robustness and bounding box regression performance under complex backgrounds and occlusion conditions. Testing on the newly constructed AngusFace dataset demonstrates that AngusRecNet achieves a mAP50 of 94.2% in facial recognition tasks, showcasing its immense potential for application in precision livestock farming. The code can be obtained on GitHub: https://github.com/HLJ11235/AngusRecNet.