Face Super-Resolution via Multi-Scale Feature Fusion

Abstract

Abstract Due to equipment limitations, faces collected by public monitoring platforms often have low resolution. To solve this problem, a face super-resolution model based on deep learning with multi-scale feature fusion is proposed. The low-resolution faces will be processed by this model to obtain faces with high definition, which greatly improves recognition rates. This model preprocesses the input face and then performs super-resolution reconstruction at multiple scales. At the same time, it uses the improved UNET structure to collect high-dimensional information of face images of different scales and returns to the preprocessed image for super-resolution reconstruction. It uses both high and low-frequency information efficiently and adopts a residual attention fusion module to focus on human facial features. The results show that the super-resolution face built by this model has greatly improved clarity, clearer features, and richer detail. The model is also faster than others with similar reconstruction effects.