Parallel CBAM-Edge: decoupling series attention structures to enhance edge fidelity in super-resolution reconstruction

Abstract

Single Image Super Resolution (SISR) focuses on reconstructing high-resolution images from low-resolution observations. Although attention mechanisms such as Convolutional Block Attention Modules (CBAM) have been widely adopted in super-resolution networks, their inherent “channel-to-space” serial architecture exhibits fundamental limitations in pixel-level reconstruction tasks. Spatial attention can only process features filtered through channel attention, leading to suppression of high-frequency information like edges and textures during the channel stage, which hinders effective enhancement in subsequent spatial processing. To address this issue, this study proposes a parallel CBAM-Edge module that decouples traditional serial architectures by establishing independent high-frequency detail extraction branches and CBAM branches optimized for super-resolution tasks, while introducing adaptive gating mechanisms to dynamically fuse dual branch features. Additionally, to enhance edge reconstruction quality, a high-frequency loss function is introduced to directly supervise the restoration process of high-frequency details. Experimental results on lightweight EDSR networks demonstrate that direct embedding with standard CBAM causes a 20.7% plunge in Structural Similarity Index (SSIM). The proposed parallel architecture not only repairs this degradation but also improves perceptual quality (LPIPS decreases from 0.2857 to 0.2742) while maintaining high SSIM comparable to the detail branch (0.9708). The comprehensive performance on four benchmark datasets—Set5, Set14, Urban100, and BSD100—validated the effectiveness of the parallel attention mechanism in enhancing edge fidelity during super-resolution reconstruction.