Abstract
In the recent works, deep learning has been applied to the single-image super-resolution (SISR) with significant results. To enhance performance, most approaches advocate for the construction of deeper and wider networks. Nonetheless, the augmentation of network depth and width could result in considerable memory storage demands and computational expenses. Besides, there is still a lot of room for improvement in edge information processing. In order to tackle these problems, we present an edge-enhanced information distillation efficient transformer network (EDTN) for lightweight super-resolution, in which the edge enhancement block and the efficient transformer module are combined to a united framework. Specifically, this work adopts an edge-enhanced reparameterization convolution block based on the existing reparameterization methods which can pay more attention to the edge information in the image. Meanwhile, this work proposes a novel gradient-variance loss which can recover edge details and retain sharp edge visuals. In order to use the information of different situations more flexibly from the feature, this work employ an efficient transformer in order to preserve high-frequency information similar features across long distances. Experimental results show that the proposed method achieves more remarkable performance under limited resource utilization.
DOI: 10.61416/ceai.v27i2.9108