Replacing K-SVD with SGK: Dictionary training for sparse representation of images

Sparse representation using trained dictionary is advantageous over the standard parametric bases. Recently, a dictionary training algorithm called SGK has been proposed as an alternative to the well known K-SVD algorithm. Analytically it has been shown that SGK has a superior execution speed in comparison to K-SVD, and it is advantageous to use SGK for constrained sparse coding. Through synthetic data experiments it has also been shown that the SGK dictionary training performances are comparable with K-SVD. However, these claims have not been verified for real data in practice. In this article, by training dictionaries on a face image database, we made a comparison between SGK and K-SVD using the problem of image recovery (inpainting and denoising). We have also introduced a simple technique to avoid incursion of noise into the dictionary.