A.L. Priorov, V.A. Volokhov, E.V. Sergeev, I.S. Mochalov

It is known that typical digital imaging devices include lenses and semiconductor sensors to capture the projected scene. It should be noted that these elements are introduced many distortions, such as the geometric distortion, blur and noise. Therefore, in order to get high-quality digital images, it is necessary to develop algorithms to solve problems of noise reduction, sharpening and color correction. The main contribution of this work is to design additive white Gaussian noise (AWGN) suppression algorithm. So we select a number of standard approaches that solve the problem of filtering AWGN in digital images: local processing algorithms; non-local processing algorithms; pointwise processing algorithms; multipoint processing algorithms. Each of these approaches to filtering of digital images has certain advantages and disadvantages associated with the quality of digital image recovery algorithms and computational cost. The aim of this work is to develop a parallel filtering procedure based on principal component analysis and non-local processing of digital images. The proposed algorithm consists of several stages. The first stage is based on block-based image representation in principal components domain with subsequent processing of transform coefficients - blocks. Then performed inverse transform to the spatial area and inserting blocks in correct positions of the processed image. First stage resulting image is an estimation image for the second stage Wiener filter in the basis of principal components. A non-local processing technique proposed in 2005 by Buades, Coll and Morel is used at the third stage. The main idea of this approach is the fact that the formation of the initial evaluation of the pixel image by using all the pixel noisy image with the specially designed weights of these pixels. Coefficient calculation is based on a comparison of similarity of a square drawn around the estimated pixel, with the areas described around the analyzed pixels. Thus, similar to the pixel area offers a lot of weight, and very different - small, so in the final evaluation pixels, some pixels are analyzed contribute more, while other smaller contribution. The fourth step is based on the formation of the final "accurate" estimate of the original image using the procedure "mixing pixels", obtained in the second and third stages of processing. Based on these researches it can be concluded that proposed algorithm can achieve good results in image denoising. The advantages of the algorithm are an ability to save local features, high-quality processing of the main boundaries of objects and an adaptability to the analyzed data. The main disadvantage of the algorithm is the high complexity.