A New Dictionary Learning Approach Using SVD with Kmeans and its appliance as regularization method for Image Deblurring using Non-Locally Centralized Sparse Representation

摘要

Image deblurring is the process to remove blur from image and get back the original image. Basically, it is the inverse problem, but the blurring operator is generally not directly invertible so no direct solution is available. Many approximate solutions are proposed by using different regularization methods. This paper proposed a new dictionary learning method using Kmeans and SVD and used it as regularization method to solve image deblurring inverse problem using nonlocally centralized sparse representation. In this method SVD is used to improve quality of dictionary. It is observed that this new method gives better average PSNR for colour as well as gray images while SSIM is same as that of NCSR-PCA technique for gray and colour images. Results are tested for 7 colour and 10 gray images affected by Gaussian blur. This can be used as alternate better method to Non-Locally Centralized Sparse Representation (NCSR).