Face Anonymization in Intelligent Experiment Education

Abstract

Face anonymization in intelligent experimental education is crucial for privacy protection. This paper presents a novel, real-time face blurring system for smart experimental settings. Our key contributions include: 1) customized YOLOv8 (Multi-Scale Feature Fusion YOLOv8) algorithm achieving 96% accuracy at 22.67 fps for 1080p video. 2) An annotation dilation preprocessing method, Contour-Adaptive Occlusion Refinement (CAOR), to address instrument occlusion issues for training. 3) A specialized dataset of 51 experimental videos with dense annotations. Our system tackles the unique challenge of preserving experimental details while anonymizing faces. We introduce two metrics, Sensitivity of Blur Accuracy (SOBA) and Over Blurred Rate (OBR), to evaluate performance. Our work demonstrates robustness across physics, biology, and chemistry experiments, maintaining a low mis-blur rate of 0.02 for instruments.