A Lightweight Deep Residual Attention Network for Single Image Super Resolution

In this paper, sparse coding has been used in learning-based single image super-resolution (SR) for text images to improve the accuracy of optical character recognition (OCR). For single image SR, we create a data-driven model with deep residual attention. The deep residual attention algorithm is built on a new deep architecture that has a high representational capability. In the proposed method the architecture consists of a residual network and a dual attention network. The feature recalibration is achieved by using Channel Attention and Spatial attention technique. The proposed model uses a residual map to recover lost high-frequency features and aids in overcoming the lower spatial resolution issue. Extensive testing on the Set5, Set14, BSD100, and Urban100 datasets shows that DRA is more effective and efficient for single image SR than other reported techniques. The experiments show encouraging results for optical character recognition (OCR) application.