A New Guided Image Denoising by Principal Component Analysis with Local Pixel Grouping

Abstract

Recently developed local pixel grouping (LPG) using principal component analysis (PCA), called LPG-PCA, has achieved satisfying performance on image denoising, especially at low noise levels. However, such method is not robust to various noise levels and produces ringing artifacts especially at high noise levels. To address this issue, we propose a new framework to further improve the denoising performance of LPG-PCA by introducing a guide image for the PCA transform estimation. In our new framework, we do the block matching in the noisy image, and select the corresponding blocks in the guide image as the training samples for the PCA transform estimation. Compared to the LPG-PCA, the proposed method not only reduces the computational complexity, but also achieves better performance especially at high noise levels. Experimental results show that the proposed method can achieve competitive results compared with some state-of-the-art denoising methods in terms of both quantitative metrics and subjective visual quality.