A novel convolutional neural network architecture for image super-resolution based on channels combination

Several models based on deep neural networks have applied to single image super-resolution and obtained great improvements in terms of both reconstruction accuracy and computational performance. All these methods focus either on performing the super-resolution (SR) reconstruction operation in the high resolution (HR) space after upscaling with a single filter, usually bicubic interpolation, or optimizing parts of the reconstruction pipeline. Then the studies of network-based model advance to attempting to shrink the feature dimension of the nonlinear mapping considering the tradeoff between accuracy and time cost. In this paper, we present an improved convolutional neural network (CNN) architecture based on channels combination, which benefits from both quick training and accuracy gain. In addition, we propose that the feature maps can be extracted in the LR space and an efficient multi-channel convolution layer which learns an array of upscaling filters, specifically trained for each feature map, to upscale the final HR feature maps into the HR output. We explore different settings and evaluate the proposed approach using images from publicly available datasets and show that it performs significantly better (about + 0.3 dB margin on term of PSNR and + 0.03 on term of SSIM than previous works) with better visual appearance.