TFEN: two-stage feature enhancement network for single-image super-resolution

In recent years, deep convolution neural networks have made significant progress in single-image super-resolution (SISR). However, high-resolution (HR) images obtained by most SISR reconstruction methods still suffer from edge blur-ring and texture distortion. To address this issue, we propose a two-stage feature enhancement network (TFEN) for the SISR reconstruction to realize nonlinear mapping from low-resolution (LR) images to HR images. In the first stage, an initial feature reconstruction module (IFRM) is constructed by combining a feature attention enhancement block and multiple convolution layers that simulate degradation and reconstruction operations to reconstruct a coarse HR image. In the second stage, based on the extracted features and the coarse HR image in the first stage, multiple residual attention modules (RAMs) consisting of the proposed spatial feature enhancement blocks (SFEBs) and an attention interaction block (AIB) are cascaded to generate the final HR image. In RAM, the SFEB is designed to learn more refined features for the reconstruction by adopting dilated convolutions and constructing spatial feature enhancement block, and the AIB is built to enhance the important features learned by RAMs through constructing multi-directional attention maps. Extensive experiments show that the proposed method has better performance than some current state-of-the-art SISR networks.