Anti-inference ability of CNN mainstream models based on face mask detection

Abstract

Face Mask detection is important under epidemic condition, it influences prevention and controlling of epidemic. Picture recognition is mature technology with the development of convolution neural network (CNN). Camera with this function can be easily established for face mask detection, providing unmanned detection for safety concern. Picture noise occurs for long-term using of detection equipment. This research paper aims to explore behaviors of CNN mainstream models under face mask detection, observing their capability of resisting noise interference. AlexNet, visual geometry group (VGG), residual network (ResNet), EfficienNet and Vision Transformer (Vit) are selected as mainstream model. Gaussian (White) noise and Impulse (Salt and Pepper) noise are select to implement noise interference in pictures. Training methods and comparison of testing results are presented in this paper. Result of comparison shows Vit gives best performance for anti-interference ability.