Neural component search for single image super-resolution

Abstract

Deep learning has become the mainstream method in the field of single image super-resolution (SISR), and the neural architecture search has been gradually applied to build SISR networks in a non-hand-crafted way. However, the existing methods can only search the structure of models and the searching speed is slow. To solve this problem, a neural component search (NCS) method is proposed. When searching for SISR networks, the color space and the composition of loss functions during training are also parts of the search space. Under a specific computational constraint, the peak signal noise ratio (PSNR) or structural similarity (SSIM) can be used as the reward to search out an optimal super-resolution network. In addition, a super graph is designed with the idea of parameter sharing to sample adaptive residual dense networks (ARDNs), thus the NCS can complete the search of SISR networks at faster speed compared to existing methods. Experimental results indicate that ARDNs searched by the NCS is competitive with the hand-crafted state-of-the-art networks, and ARDNs achieve favorable performance against state-of-the-art methods with similar computational consumption.