A parallel computation algorithm for super-resolution methods using convolutional neural networks

Abstract

An acceleration method for interpolation-based super-resolution (SR) methods using convolutional neural networks (CNNs), represented by SRCNN and VDSR, is proposed. In this paper, estimated pixels are classified into a number of types according to upscaling factors, and then SR images are generated by using CNNs optimized for each type. It allows us to adapt smaller filter sizes to CNNs than conventional ones, so that the computational complexity can be reduced for both running phase and training one. In addition, it is shown that the optimized CNNs for some type are closely related to those of other types, and the relation provides a method to reduce the computational complexity for training phase. A number of experiments are carried out to demonstrate that the effectiveness of the proposed method. The proposed method outperforms conventional ones in terms of the processing speed, while keeping the quality of SR images.