Image Super Resolution Using C N N

Abstract

In super resolution of an image, the aim is to extract the missing high frequency details which is not present in the low resolution image. Single Image Super Resolution is used in computer applications, where more image details are required. Many Single Image Super Resolution methods have studied in computer vision community like interpolation-based, Reconstruction-based and Example learningbased. Currently, learning based methods are used in modeling for mapping of Low Resolution and High Resolution patches. These methods have the power for determining the hierarchical description for visual data which can be directly extracted and learned from the data. Recent advancement have seen in deep learning based image Super Resolution methods. A network, Convolutional Neural Network can be trained to identify the relation between Low Resolution and High Resolution patches. Training a very deep network is challenging and difficult because of slow convergence rate, so by using high learning rate and residual-learning, results in fast training for very deep neural network. The proposed algorithm includes both spatial and wavelet domain strategy, which gives better results(1.44 percent higher) as compared to other existing dictionary learning methods in terms of Peak Signal to Noise Ratio(PSNR) in dB on existing datasets on magnification factors 2, 3 and 4.