Research on Image Super-Resolution Reconstruction Based on Transformer

Abstract

The existing super-resolution methods have the problems of blurring contour and losing details, which cannot make full use of the internal and external information of the image. On the basis of convolutional neural network, a super-resolution model embedded in Transformer module is proposed. Through the fusion of CNN and Transformer, image feature extraction and corrupted information recovery are performed. The training process is divided into an unsupervised pre-training process on the ImageNet subset and a supervised fine-tuning training process on the DIV2K data set. Contrastive loss is introduced into the loss function to achieve the consistency of super-resolution image category and texture. It is compared with the current mainstream image super-resolution methods on the standard 5 benchmark data sets. Experiments show that compared with other algorithms, this algorithm can better recover the high-frequency information of the image, and generate super-resolution results with clear details and fine structure.