Effectiveness Evaluation of Attention Mechanism Strategies in Deep Learning-Based Single Image Super-Resolution

Abstract

Single image super-resolution (SISR) has experienced rapid advancements through the integration of attention mechanisms into deep neural networks. Diverse attention strategies, including channel attention, spatial attention, second-order attention, and self-attention, have been incorporated into SISR architectures with varying degrees of success. A systematic and fair comparison of these attention strategies under unified experimental conditions remains absent from the current literature. This study conducts a controlled benchmark evaluation of five representative SISR algorithms—EDSR, RCAN, SAN, SwinIR, and HAT—each embodying a distinct attention paradigm, across four standard datasets (Set5, Set14, BSD100, Urban100) at ×2 and ×4 scaling factors. The evaluation encompasses distortion-oriented metrics (PSNR, SSIM), perceptual quality metrics (LPIPS), and computational efficiency indicators (parameters, FLOPs, inference time). Experimental results reveal that hybrid attention mechanisms in HAT achieve the highest reconstruction fidelity, with 1.25 dB PSNR gain over the non-attention baseline EDSR on Urban100 at ×2 scale, while window-based self-attention in SwinIR offers the most favorable accuracy-efficiency trade-off. The findings provide practical guidance for selecting attention strategies in SISR applications under different resource constraints.