Iterative Network for Image Super-Resolution

Single image super-resolution (SISR), as a traditional ill-conditioned inverse problem, has been greatly revitalized by the recent development of convolutional neural networks (CNN). These CNN-based methods generally map low-resolution to its corresponding high-resolution version with sophisticated network structures and loss functions, showing impressive performances. This paper provides new insight on conventional SISR algorithm, proposes substantially different approach relying iterative optimization. A novel (ISRN) is proposed top We first analyze observation model SR inspiring feasible solution mimicking fusing each iteration in more general efficient manner. Considering drawbacks batch normalization, we propose feature normalization (F-Norm, FN) method regulate features network. Furthermore, block FN developed improve representation, termed FNB. Residual-in-residual structure form very deep network, which groups FNBs long skip connection for better information delivery stabling training phase. Extensive experimental results testing benchmarks bicubic (BI) degradation show our ISRN can not only recover structural information, but also achieve competitive or PSNR/SSIM much fewer parameters compared other works. Besides BI, simulate real-world blur-downscale (BD) downscale-noise (DN). extension ISRN+ both performance than others BD DN models.