Occluded Face Recognition Method Based on Multi-Scale Attention Mechanism

In previous occluded face recognition models, the performance of partially occluded face recognition is significantly degraded due to the problems of limited recognizable features and high model complexity. In this article, We presented an occlusion method for faces based on a multi-scale attention mechanism by replacing the 3×3 convolution in the bottleneck block with a PSA attention module. The method cleans up the features damaged by masking by generating masks to conceal them, and then utilizes the multi-scale attention mechanism to effectively extract spatial information. Experiments show that the method is better than previous methods and effectively improves the accuracy.