Residual Wavelet Coefficients Prediction using deep Convolutional Neural Network for Single Image Super-Resolution

This paper is about learning based super resolution (SR) by exploiting wavelet coefficients using convolutional neural network (CNN). The objective of SR is to generate a high quality and visually pleasing image from the low-resolution (LR) image(s) without extensive training. Wavelet coefficients are sparse in nature, this sparsity is further enhanced by learning residuals (difference of SR and LR wavelet sub-bands) of wavelet coefficients. Sparsity helps to speed up the training procedure as it contains significant information within only few coefficients. We train a CNN using residuals of wavelet sub-bands instead of raw image pixels. We learn an end-to-end mapping to predict unknown wavelet coefficients using CNN by reducing the root mean square error of wavelet coefficients from the four sub-bands. SR involves perfect reproduction of high frequency details like textures and sharp edges present at the original scale, which makes it as a severe inverse ill-posed problem. The proposed CNN is profited from the sparsity and is named as Residual Wavelet Super Resolution (RWSR) Net. The RWSR is approximately thirty times faster than the well-known SRCNN method and provides better performance. The performance of the proposed work is assessed through quantitative metrics and it is observed to give competitive results when compared to the existing state-of-the-art methods.