High Performance Image Super-Resolution Using Convolutional Neural Networks and Nonsubsampled Contourlet Transform

Abstract

Disadvantages of deep convolutional neural networks (CNNs) in image super-resolution (SR) are slow convergence rate, long training time, slow run time, complexity and high computational cost. On the other hand, shallow CNNs, while have lack of above-mentioned disadvantages for deep CNNs, do not have ability of learning the high frequency components appropriately. The purpose of this paper is promoting shallow networks so as to increase learning ability for fine details of the image while keeping the previous advantages. We present a learning-based approach which is composition of a feature extraction transformation unit with a training and approximation unit (shallow CNNs). In the transformation unit, to extract fine details of the image, nonsubsampled contourlet transform (NSCT) is used. Transforming the feature space to frequency domain causes that the output coefficients to have enough sparse so that by increasing the sparse coefficients, training of neural network will be easier and the training time will decrease considerably. Advantages of the proposed approach with respect to other approaches are high reconstruction accuracy, high training speed, low runtime, lower network complexity and faster convergence.