A deep learning lightweight model for real-time captive macaque facial recognition based on an improved YOLOX model

Area of Science:

• Primate behavior analysis
• Computer vision
• Biomedical research tools

Background:

• Automated behavior monitoring in macaques is crucial for research and welfare.
• Individual macaque identification in group settings presents a significant technical challenge.
• Existing methods lack the accuracy and efficiency required for real-time applications.

Purpose of the Study:

• To develop a lightweight and accurate facial recognition model for individual macaque identification.
• To enhance the YOLOX framework with attention and feature fusion mechanisms for improved performance.
• To create a deployable solution for on-device, real-time macaque recognition.

Main Methods:

• Integration of Efficient Channel Attention (ECA), Complete Intersection over Union loss (CIoU), and Adaptive Spatial Feature Fusion (ASFF) into the YOLOX object detection framework.
• Training and validation using a large dataset of 179,400 labeled macaque facial images from 1,196 individuals.
• Development of an Android application for smartphone-based, on-device deployment of the ACE-YOLOX model.

Main Results:

• ACE-YOLOX demonstrated superior prediction accuracy and reduced computational complexity compared to classical object detection models.
• The model achieved effective multiscale feature extraction, crucial for facial recognition in varied conditions.
• Real-time processing capabilities were validated, enabling immediate identification.

Conclusions:

• ACE-YOLOX offers a highly accurate and efficient non-invasive tool for individual macaque identification.
• The developed model and application provide a foundational technology for advancing macaque behavioral studies.
• Potential applications include facial expression recognition, cognitive psychology, and social behavior research.