Image denoising via sparse data representation: A comparative study

Noise can be an important factor that may significantly degrades the quality of a digital image. This paper investigates the efficiency of sparse data representation in order to recover as much as possible the noise free content of an image when this image is corrupted by additive Gaussian noise. To decompose data into sparse representation the orthogonal matching pursuit approach is used. The experiments undergo several degree of corrupted pixels, ranging from 25 % to 75 %, and the orthogonal matching pursuit approach is compared with three state-of-the art techniques, namely anisotropic diffusion, Srini-Ebenezer filtering and phase preserving denoising method, respectively. We shown throughout experiments, that the sparse data representation achieved higher peak signal-to-noise-ratio values compared to the other approaches indicating the superiority of orthogonal matching pursuit approach in noise removal application when the degree of corrupted pixels covers half of the image. However, its performance is limited and comparable with the Srini-Ebenezer filtering approach for large number of corrupted pixels.