ELTE: An enhanced Local Texture Estimator based on Convolutional Attention and Transformer for Image Super-Resolution

Single image super-resolution represents a technique to reconstruct a high-resolution image from a low-resolution image. In recent years, the technology of implicit neural representation has become a new hotspot in the field of image super-resolution, the theory of which is approximating the continuous function used to represent the real state of an image by neural networks, and the local texture estimator is one of the representative researches. The principle of the local texture estimator is to predict new pixels for any query coordinate according to the adjacent features. However, this work is based on pixel distance only and cannot fully reflect the context information of the image. To overcome these problems, an enhanced local texture estimator (ELTE) is proposed, which integrates the convolutional attention mechanism and Transformer technology into the local texture estimator to achieve arbitrary-scale super-resolution. In this paper, we firstly propose a Structure Transformer Layer (STL) which can effectively reconstruct the local structure information from feature maps, and then introduce a Spatial Awareness Layer (SAL) to obtain the long-range dependencies ignored by STL. Lastly, the Convolutional Attention Mechanism Module (CBAM) has been added in order to further focus on more useful texture information. Through a wide range of experiments, it has been shown that the proposed approach has aachieved promising performance in both qualitative and quantitative aspects.