Cross-Dimensional Attention Fusion Network for Simulated Single Image Super-Resolution

Abstract

Single image super-resolution (SISR) is a task of reconstructing high-resolution (HR) images from low-resolution (LR) images, which are obtained by some degradation process. Deep neural networks (DNNs) have greatly advanced the frontier of image super-resolution research and replaced traditional methods as the de facto standard approach. The attention mechanism enables the SR algorithms to achieve breakthrough performance after another. However, limited research has been conducted on the interaction and integration of attention mechanisms across different dimensions. To tackle this issue, in this paper, we propose a cross-dimensional attention fusion network (CAFN) to effectively achieve cross-dimensional inter-action with long-range dependencies. Specifically, the proposed approach involves the utilization of a cross-dimensional aggrega-tion module (CAM) to effectively capture contextual information by integrating both spatial and channel importance maps. The design of information fusion module (IFM) in CAM serves as a bridge for parallel dual-attention information fusion. In addition, a novel memory-adaptive multi-stage (MAMS) training method is proposed. We perform warm-start retraining with the same setting as the previous stage, without increasing memory consumption. If the memory is sufficient, we finetune the model with a larger patch size after the warm-start. The experimental results definitively demonstrate the superior performance of our cross-dimensional attention fusion network and training strategy compared to state-of-the-art (SOTA) methods, as evidenced by both quantitative and qualitative metrics.