EPLL image denoising with multi-feature dictionaries

Abstract

The expected patch log likelihood (EPLL) is a patch-prior based denoising model that has become a popular tool in image restoration for its outstanding edge preservation and noise reduction abilities. However, without considering the geometrical feature of patches, the EPLL removes numerous important details and frequently suffers from the staircase effect, residual noise and artifacts. To preserve more detailed information and suppress the generation of artifacts, we propose a novel EPLL denoising model with multi-feature dictionaries (MFDs) and adaptive regularization parameters in this paper. In the proposed method, to exploit the statistical and geometrical characteristics of patches, two feature layers are designed before training the Gaussian mixture model (GMM). In the first layer, the within-class variance minimum (WCVM) method is used to learn the statistical features of patches. In the second layer, two geometrical characteristic dictionaries (GCDs) are trained. In the denoising phase, we select the corresponding dictionary for each patch. Then, each patch is denoised independently. Moreover, the mean localized spectral response (MLSR) based regularization parameters are employed in the proposed method, which is spatial self-adaptive and robust to noise. Extensive experiments demonstrate that our approach achieves desired performance in both detail preservation and artifact suppression compared to many existing denoising methods.