A Robust Approach for Image Super-Resolution using Modified Very Deep Convolution Networks
Alok Kumar, Sandeep Kumar Shukla, A. Sharma, Pranay Yadav
Abstract
This research presents a Modified Very Deep Convolutional Network (MVDCN) for Single Image Super Resolution (SISR). The proposed method is based on modified CNN, in which different image features for training also apply up-sampling as well as residual images, which is a fundamental step of SISR, and the depth of the network is 20. For the improvement of the presented method, it applies the fusion of the bi-cubic method with the proposed modified residual image attributes with Very Deep Convolutional Networks (VDCN). The presented method shows better results in terms of the two base parameters of the proposed method. These are PSNR and SSIM. These two parameters are major parameters for the result analysis of image super resolution (ISR). There are different data sets available for the training and testing of the presented method, such as test datasets “Set5’ [15] and ‘Set14’ [26]. Both are primarily used as benchmarks by various researchers; however, in other works, the data set “Urban100” is very interesting because it contains many challenging images that fail by many of the existing methods. Final data set ‘B100’, natural images from Berkeley University. The proposed MVDCN shows better results as compared to other previous methods.