Deep USRNet Reconstruction Method Based on Combined Attention Mechanism

Single image super-resolution (SISR) based on deep learning is a key research problem in the field of computer vision. However, existing reconstruction algorithms often improve quality through single network depth, ignoring problems reconstructing texture structure and easy overfitting training. Therefore, this paper proposes unfolding (USRNet) method under integrating channel attention mechanism, which expected to resolution restore high-frequency information image. Thus, appears sharper. First, by assigning different weights features, focusing more important features suppressing unimportant details such as edges textures are better recovered, generalization ability improved cope with complex scenes. Then, CA (Channel Attention) module added USRNet, depth increased express features; multi-channel mapping introduced extract richer enhance effect model. The experimental results show that USRNet has faster convergence rate, not prone overfitting, can be converged after 10,000 iterations; average peak signal-to-noise ratios Set5 Set12 datasets side length enlarged two times are, respectively, 32.23 dB 29.72 dB, dramatically compared SRCNN, SRMD, PAN, RCAN. algorithm generate high-resolution images clear outlines, better.