Reproduce “Single Image Super-Resolution via a Holistic Attention Network” on OpenResearch

About this paper

Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao and 1 more

The Holistic Attention Network (HAN) reconstructs high-resolution images from low-resolution inputs by introducing novel Layer Attention and Channel-Spatial Attention modules, dynamically emphasizing informative features across hierarchical levels and within feature maps. This architecture outperforms previous state-of-the-art methods in PSNR/SSIM scores and visual quality across multiple scales and degradation types, notably achieving a 0.2 dB PSNR gain on the Manga109 dataset for 4x SR over an RCAN baseline.

Key insights

• Simultaneously modeling interdependencies across hierarchical layers, channels, and spatial positions leads to a more robust and discriminative feature representation for super-resolution.
• The Layer Attention Module (LAM) effectively leverages correlations between features from different network depths, preventing dilution of important shallow information.
• The Channel-Spatial Attention Module (CSAM) enhances feature modulation by jointly considering channel and spatial importance, demonstrating superior recalibration over separate attention mechanisms.

Summary by alphaXiv.