Blurred Facial Recognition Based on AdaFace
Quan-Bin Zhang, Po-Yang Chi, Shang-Ze Lin, Chia-Xsuan Wu, Yi‐Zeng Hsieh · 2024
Many campuses in Taiwan feature open access, complicating the control and identification of individuals entering. This issue, along with emerging campus safety concerns, underscores the critical need for efficient security measures. The low resolution of surveillance cameras and the inefficiency of manual identification methods exacerbate this challenge. This study suggests the use of artificial intelligence, specifically the Adaface model with its unique Marginal-based loss functions, to enhance the identification process even with low-quality, blurred images. By automating the classification and processing of facial data–distinguishing between registered and unregistered faces–the system can perform real-time comparisons and logging. Our experiments with Adaface have demonstrated improved multi-face recognition capabilities and the ability to track unrecognized faces over time. Nonetheless, variations in scene context can still affect feature visibility and model stability.