Enhanced Second-order Attention Network for Single Image Super Resolution

Abstract

Deep learning empowered the latest progress in single image super-resolution (SISR) to achieve a better reconstruction performance. The SISR is a technique that produces a high-resolution (HR) image from the low-resolution (LR) edition of same image. The recent methods to solve the problem of SISR are based on Second-order Attention Network (SAN) and achieved a comparable performance. However, the SAN methods produces residual blur in the HR image due to their error function. Therefore, the error function is corrected to get better  performance of the SAN. The presented model has been tested on signature dataset, Yale dataset and some other miscellaneous images. The testing results evidenced that the projected model carry out the enhanced execution than classical SAN[1] for text as well as for images in terms of visual quality.