Single Image Super-Resolution using Residual Image based on CNN with a Deconvolutional Layer

This paper proposes a method for recovering the intrinsic details of an image that cannot be reconstructed by interpolation, named as residual images, through a convolutional neural network with a deconvolutional layer. The predicted residual image is added to an interpolated LR image to reconstruct the lost details. In both the qualitative and quantitative comparison to SRCNN, the proposed framework performed in a better manner. The proposed framework did not produce the false edges seen in the results of SRCNN. Furthermore, the proposed method resulted 0.18 dB higher PSNR in average, compared to SRCNN.