Contour-texture preservation transformer for face super-resolution

Abstract

In the field of computer vision, face super-resolution (FSR) technology is an important tool for enhancing the performance of basic tasks such as face recognition and video surveillance. However, when faced with complex face images, existing FSR methods often rely on the Transformer model to improve image quality through its powerful global modeling capabilities. Yet, they tend to be slightly insufficient in local feature extraction due to a lack of adequate local detail capture capabilities. To alleviate these problems, we propose a novel contour-texture preservation Transformer (CTP) method for FSR. This method consists of two key components: the multi-scale attention enhancement block (MSAEB), which captures and fuses image features of different scales to improve the detail level of feature representation, and provides high-quality input for the contour-texture Transformer enhancement block (CTTEB). Additionally, CTTEB integrates convolution operations to enhance local feature extraction and improve feature expression. The feed-forward network (FFN) we introduced ensures the full fusion of global and local information. By combining MSAEB and CTTEB into a residual progressive attention group (RPAG), the network gradually extracts and fuses multi-scale features, ultimately achieving dual preservation of contour structure and texture details. Experiments show that our method achieves the best results on LFW, FFHQ, CelebA, and Helen, with a 0.32 dB increase in PSNR, a 0.0126 increase in SSIM, and a 0.0056 increase in FSIM over the second-best model. Experiments on real-world datasets SCface and Chokepoint confirm that the CTP method excels in both FSR reconstruction and face recognition, verifying its effectiveness.