| Abstract |
Image super-resolution(SR) reconstruction aims at reconstructing corresponding high-resolution images from low-resolution images. Methods based on deep CNN have achieved distinguished performance recently. Among these methods, those based on the single-layer features cannot get top reconstruction results since shallow features are ignored, and those based on multi-layer features need to discriminate good and poor features. To address this issue, we propose a new model, BySEDenseSR (Bypass Convolution Squeeze Excitation Dense). Specifically, We use a squeeze excitation block that learns to assign appropriate weights based on the quality of features. This provides an effective way to suppress the effects of poor features. Furthermore, using bypass convolution, BySEDenseSR reduces the forwarding noise, which improves the stability of middle-layer features. Experiments show that compared to LapSRN, BySEDenseSR improves its peak signal-to-noise ratio (PSNR) by 0.16dB, 0.08dB and 0.03dB respectively on Set5, Set14 and BSD100, at a magnification of 4×. © 2019 IEEE. |
