End-to-End Single Image Super-Resolution Based on Convolutional Neural Networks

Single image super-resolution algorithms aim to increase the resolution of an input image without deteriorating its visual perception. With a strong ability to understand the structure of an image, convolutional neural networks (CNNs) are successfully applied to this problem. Previous studies have shown that human perception is mainly influenced by variations in luminance. In this regard, this paper introduces two CNNs that operate only on the luminance channel, at low computational costs. Each model provides an end-to-end mapping between a low-resolution (LR) and a high-resolution (HR) map. Because upsampling is integrated into CNN, the design allows the control of HR image quality. In addition, the neural architectures can be configured with many layers operating with small LR feature maps, to provide fast run-time image processing. The approach is exemplified in two cases: generate the HR map or a residual map for the luminance channel; the residual map should be added to the map upsampled by interpolation. Besides having improved time performance, the two models can produce HR images with high NIQE scores, as shown experimentally.