Blind image restoration method by PCA-based subspace generation

Abstract

Principal Component Analysis (PCA) has been effectively applied for image restoration. Original idea underlying PCA approach has two different roots. One is from the fact that PCA is relevant to variance of pixel intensity by which the missing high frequency components in blurred image should be recovered. The other comes from the idea of source separation based on PCA. In the light of PCA approach we have proposed an image restoration algorithm which contains the following three novel aspects: iterative application of PCA, Gaussian smoothing filtering for image ensemble creation, and no-reference image quality index for iteration number management. This paper aims to investigate and propose a non-iterative PCA-based image restoration with some generalizations. First, through conducted experiments the variance of Gaussian filters as well as the number of created images by them are appropriately determined. Second, weights are introduced to the principal component images. Finally, optimal weights are determined by maximizing the image quality index with no reference. Experimental results by the proposed method provide higher PSNR than the previous iterative PCA approach.