Iterative Dense Network using Laplacian Pyramid Model for Blind Super Resolution

Abstract

The Deep learning-based approach for solving the ill-posed problem of single image super-resolution reconstruction (SRR) has achieved tremendous success in recent times. However, not much work is carried out in the direction of blind image super-resolution, where the degradation kernel is said to be unknown. This paper addresses the said problem of blind single image super-resolution reconstruction using an alternative learning approach by training two convolutional neural networks. Most of the available model for blind super-resolution considers a fixed degradation kernel for reconstruction, which leads to drop in performance. Therefore a learnable kernel estimation approach is adopted by using a kernel-estimator network. Further, this estimated kernel is used to generate a super resolution image using a Generator network. To successfully model the reconstruction of vital features like edges and texture and to learn the inter-pixel dependencies between multi-level feature maps, we employ a densely residual Laplacian attention block (DLA-Block). The proposed method is extensively tested on real image and synthetic image data-sets. The experimental results have shown out-performance compared to the state-of the- art in terms of high reconstruction accuracy as well as PSNR and SSIM.