Multi-Modality Sheep Face Recognition Based on Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Recognizing individual sheep faces is challenging due to high intra-class similarity.
• RGB image performance varies significantly with lighting conditions and viewing angles.
• Existing methods struggle with the nuances of sheep face identification.

Purpose of the Study:

• To develop a robust sheep face recognition system resilient to environmental variations.
• To enhance recognition accuracy by integrating multi-modal data.
• To leverage both geometric and texture features for improved identification.

Main Methods:

• A dual-branch ResNet18-based model was proposed, processing RGB and depth data separately.
• InceptionV2 layers extracted features from each modality.
• Multi-modal fusion was achieved using the Convolutional Block Attention Module (CBAM) and residual networks.

Main Results:

• The model effectively learned geometric features from depth data and texture features from RGB data.
• Multi-modal fusion significantly enhanced recognition accuracy.
• High accuracy was achieved even under complex lighting and diverse angles.

Conclusions:

• The proposed dual-branch multi-modal model offers a superior solution for sheep face recognition.
• Effective fusion of geometric and texture features is key to overcoming recognition challenges.
• This approach demonstrates potential for applications in livestock management and monitoring.