A Synthesis-by-Analysis Network with Applications in Image Super-Resolution

Recent studies have demonstrated the successful application of convolutional neural networks in single image super-resolution. In this paper, we present a general synthesis-by-analysis network for super-resolving a low-resolution image. Unlike Laplacian Pyramid Super-Resolution Network (LapSRN) that progressively reconstructs the sub-band residuals of high-resolution images, our proposed network breaks through the sequential dependency to expand the input and output into multiple disjoint bandpass signals. At each band, we perform the nonlinear mapping in truncated frequency interval by applying a carefully designed sub-network. Specifically, we propose a validated network sub-structure that considers both efficiency and accuracy. We also perform exhaustive experiments in existing commonly used dataset. The recovered high-resolution image is competitive or even superior in quality compared to those images produced by other methods.