Multi-Scale Pixel-Attention Feedback Link Network for Single Image Super-Resolution
Yanliang Ge, Shuang Tan, Hongbo Bi, Xiaoxiao Sun · Pattern Recognition and Image Analysis · 2022
Abstract In recent years, deep learning methods have been widely used in single image super-resolution. However, as the network model increase in depth and complexity, the high-frequency information is lost during the feature transmission. The information of shallow features cannot be fully used for image reconstruction. And in many existing super-resolution reconstruction networks, the information flows have merely feedforward that cannot fully exploit their low-level features. In this paper, we propose a new multi-scale pixel-attention feedback link network (MPFSR) for single image super-resolution (SISR) that attempts to achieve better performance from the feedback link and multi-scales. Specifically, to make the network reuse the feature information of different depths comprehensively, we propose a multi-scale downsampling module (MDSM). And in the upsampling part, the pixel attention mechanism and the residual dense block was used, namely the Pixel attention feature extraction module (PAFEM), which can adaptively focus attention on the region with the most abundant pixel information. The feedback link adds to the MPFSR, which aims to feedback the reconstructed high-level features to previous layers to refine the low-level features lacking sufficient contextual information. This feedback can spread the underlying information completely upward in order to better optimize and integrate upper and lower feature information and enhance the utilization of shallow features. In the process of upsampling and downsampling, every layer has the same proportion of the image linked by the skip connection so that the features of each layer can be integrated. In the experiments, we validate that the proposed network achieves significant accuracy compared to the outstanding methods for single image super-resolution.