A Low-Complexity Convolution Network for Upscale Super-Resolution Image

We proposed a modified architecture to reduce the computation of the ESPCN network for super-resolution generation. Based on the deep separable convolution concept, the ESPCN-based network was embedded in the 1x1 convolution layer to reduce the computations. Experiment results showed that the proposed method greatly reduces the number of parameters and computational operators, compared with the original network. The numbers were reduced by about 65% while maintaining the image quality as upscale x16. Using the DIV2K dataset for training and verification, the super-resolution image became better than that by the conventional spatial algorithms with improved PSNR values and reduced blurring.