Pig Face Open Set Recognition and Registration Using a Decoupled Detection System and Dual-Loss Vision Transformer.

Area of Science:

• Agricultural Technology
• Computer Vision
• Animal Science

Background:

• Accurate animal identification is crucial for effective pig farming, especially in dynamic herd environments.
• Pig face recognition faces challenges like high similarity, lighting variations, and occlusions, hindering monitoring.
• Existing methods struggle with open-set conditions where new individuals are frequently introduced.

Purpose of the Study:

• To develop a robust and adaptable pig face recognition system for dynamic farming environments.
• To address the challenges of open-set recognition in pig identification.
• To improve the accuracy and efficiency of pig monitoring and management.

Main Methods:

• A three-phase Pig Face Open-Set Recognition (PFOSR) system was developed.
• Phase 1: YOLOv8 for pig face detection and a Vision Transformer (ViT) with dual-loss (Sub-center ArcFace, Center Loss) for recognition.
• Phase 2: Registration of known pigs' feature embeddings into a gallery. Phase 3: Real-time recognition and dynamic registration of unknown pigs using cosine similarity.

Main Results:

• The PFOSR system achieved high open-set recognition performance with AUROC of 0.922, OSCR of 0.90, and F1-Open of 0.94.
• Closed-set recognition yielded strong results: precision@1 of 0.97, NMI of 0.92, and mAP@R of 0.96.
• The system demonstrated scalability and efficiency in managing dynamic farm environments.

Conclusions:

• The proposed PFOSR system offers a scalable and efficient solution for pig identification in dynamic farming settings.
• The system effectively handles challenging conditions, including occlusions and lighting variations.
• This approach enhances pig management through accurate and adaptable facial recognition technology.