Progressive representation recalibration for lightweight super-resolution

Abstract

Recently, the lightweight single-image super-resolution (SISR) task has received increasing attention due to the computational complexities and sizes of convolutional neural network (CNN)-based SISR models and the explosive demand in applications on resource-limited edge devices. Current algorithms reduce the number of layers and channels in CNNs to obtain lightweight models for this task. However, these algorithms may reduce the representation ability of the learned features due to information loss, inevitably leading to poor performance. In this work, we propose the progressive representation recalibration network (PRRN), a new lightweight SISR network to learn complete and representative feature representations. Specifically, a progressive representation recalibration block (PRRB) is developed to extract useful features from pixel and channel spaces in a two-stage approach. In the first stage, PRRB utilizes pixel and channel information to explore important feature regions. In the second stage, channel attention is further used to adjust the distribution of important feature channels. In addition, current channel attention mechanisms utilize nonlinear operations that may lead to information loss. In contrast, we design a shallow channel attention (SCA) mechanism that can learn the importance of each channel in a simpler yet more efficient way. Extensive experiments demonstrate the superiority of the proposed PRRN.