A Self-Supervised Adversarial Deblurring Face Recognition Network for Edge Devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is vital for intelligent surveillance and health monitoring.
• Facial recognition is crucial for vision-based HAR but struggles with blurry and dynamic images.
• Existing models lack robustness in real-world HAR scenarios due to image quality limitations.

Purpose of the Study:

• To develop a fast and accurate facial recognition model for enhanced HAR.
• To improve facial recognition performance in challenging, dynamic, and blurry conditions.
• To address limitations of current models in real-world HAR applications.

Main Methods:

• Utilized a generative adversarial network (GAN) as the core algorithm.
• Optimized GAN modules by decomposing the global loss function and incorporating a feature pyramid.
• Integrated deblurring techniques to enhance recognition in blurry and dynamic image scenarios.

Main Results:

• Achieved high accuracy and recall rates across multiple facial recognition datasets.
• Reported an average recall rate of 87.40% and accuracy rates of 81.06% (YTF) and 79.77% (WiderFace).
• Demonstrated effective handling of dynamic and blurry facial images in HAR.

Conclusions:

• The proposed model significantly enhances facial recognition in challenging HAR environments.
• The novel approach shows strong potential for real-world applications in intelligent surveillance and human-computer interaction.
• The study successfully addresses key limitations in current facial recognition technology for HAR.