Convolutional neural networks for wavelet domain super resolution

Proposed a super resolution method with higher reconstruction accuracy than before.Cast super resolution as a problem of estimating sparse wavelet detail coefficients.Estimated sparse wavelet coefficients using a convolutional neural network (CNN).Trained our CNN with fewer samples than the one used for estimating pixels directly.Concluded that it is better to estimate wavelet coefficients rather than pixels. We present a single image super resolution technique in which we estimate wavelet detail coefficients of a desired high resolution (HR) image using a convolutional neural network (CNN) on the given low resolution (LR) image. Detail coefficients are necessarily sparse for natural images, unlike pixel intensities, and are thus better suited to be CNN output. This allows us to train a CNN with far fewer samples and lesser training time and yet achieve better reconstruction quality with lesser run time compared to a recent state-of-the-art technique that directly estimates the HR pixels using a CNN.