Single Image Super-Resolution via a Holistic Attention Network

Informative features play a crucial role in the single image super-resolution task. Channel attention has been demonstrated to be effective for preserving information-rich features in each layer. However, channel attention treats each convolution layer as a separate process that misses the correlation among different layers. To address this problem, we propose a new holistic attention network (HAN), which consists of a layer attention module (LAM) and a channel-spatial attention module (CSAM), to model the holistic interdependencies among layers, channels, and positions. Specifically, the proposed LAM adaptively emphasizes hierarchical features by considering correlations among layers. Meanwhile, CSAM learns the confidence at all the positions of each channel to selectively capture more informative features. Extensive experiments demonstrate that the proposed HAN performs favorably against the state-of-the-art single image super-resolution approaches.

PDF Abstract ECCV 2020 PDF ECCV 2020 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Super-Resolution BSD100 - 2x upscaling HAN+ PSNR 32.45 # 3
SSIM 0.8431 # 5
Image Super-Resolution BSD100 - 3x upscaling HAN+ PSNR 29.41 # 1
SSIM 0.8116 # 2
Image Super-Resolution BSD100 - 4x upscaling HAN+ PSNR 27.85 # 5
SSIM 0.7454 # 8
Image Super-Resolution BSD100 - 8x upscaling HAN+ PSNR 25.04 # 2
SSIM 0.6075 # 1
Image Super-Resolution Manga109 - 2x upscaling HAN+ PSNR 39.62 # 2
SSIM 0.9787 # 3
Image Super-Resolution Manga109 - 3x upscaling HAN+ PSNR 34.87 # 2
SSIM 0.9509 # 2
Image Super-Resolution Manga109 - 4x upscaling HAN+ PSNR 31.73 # 5
SSIM 0.9207 # 7
Image Super-Resolution Manga109 - 8x upscaling HAN+ PSNR 25.54 # 3
SSIM 0.8080 # 3
Image Super-Resolution Set14 - 2x upscaling HAN+ PSNR 34.24 # 5
SSIM 0.9224 # 4
Image Super-Resolution Set14 - 3x upscaling HAN+ PSNR 30.79 # 4
SSIM 0.8487 # 2
Image Super-Resolution Set14 - 4x upscaling HAN+ PSNR 28.99 # 6
SSIM 0.7907 # 12
Image Super-Resolution Set14 - 8x upscaling HAN+ PSNR 25.39 # 3
SSIM 0.6552 # 2
Image Super-Resolution Set5 - 2x upscaling HAN+ PSNR 38.33 # 3
SSIM 0.9299 # 11
Image Super-Resolution Set5 - 3x upscaling HAN+ PSNR 34.85 # 3
SSIM 0.9300 # 2
Image Super-Resolution Set5 - 4x upscaling HAN+ PSNR 32.75 # 3
SSIM 0.9016 # 7
Image Super-Resolution Set5 - 8x upscaling HAN+ PSNR 27.47 # 3
SSIM 0.7920 # 3
Image Super-Resolution Urban100 - 2x upscaling HAN+ PSNR 33.53 # 3
SSIM 0.9398 # 3
Image Super-Resolution Urban100 - 3x upscaling HAN+ PSNR 29.21 # 4
SSIM 0.8710 # 2
Image Super-Resolution Urban100 - 4x upscaling HAN+ PSNR 27.02 # 11
SSIM 0.8131 # 11
Image Super-Resolution Urban100 - 8x upscaling HAN+ PSNR 23.20 # 2
SSIM 0.6518 # 3

Methods