A Study On Image Denoising With Local And Global Priors

With the rapid development of network technology and the popularization of hard-ware resources,multimedia data has experienced explosive growth in recent years.As the most fundamental and common form of multimedia data,image is the main carrier of message passing on the network.However,images used in real life are not often pure data.Since an image is inevitably polluted by noise in image acquisition,storage and transmission processes,poor quality of image directly affects subsequent image process-ing and image analysis tasks.The main target of image denoising is to filter noise as well as remain structure information and texture information.Image denoising is an important means to improve image quality,and is a research focus in computer vision for several years.Aiming at removing the Additive White Gaussian Noise caused by processes of image acquisition and image transmission,an approach based on local and global priors is proposed by this paper.Main innovations of this paper are as follows:1.A denoising framework which combines local prior and global prior together is proposed to recover noise-free values of pixels.The local prior depends on neighborhood relationships of a search window to help maintain edges and smoothness,and the global prior promotes the consistency of texture structures.Then,normalized contribution leads to final result.The proposed framework is irrelevant to specific noise type or noise level,thus,it can be applied to different noise reduction tasks.2.The proposed method generates the global prior from a hierarchical L₀sparse representation to remain the global consistency of geometric structures in the denoised image.With the global similarity constraint,denoised results can gain better performance in detail textures.Compared with the nonlocal prior generated by a larger search window in other algorithms,global prior is more effective in the preservation of detail information.3.The proposed method adopts Principal Component Analysis?PCA?to make cor-relations between target pixel and auxiliary pixels more meaningful.We constructs more representative feature to describe characters of pixels by projecting the original man-ual feature to PCA subspace,to simplify the similarity measurement and improve the accuracy.Feature constructed by PCA is more robust than manual feature.Moreover,the dimensions of new feature is far smaller than original feature,which can reduce the amount of calculation and reduce computational complexity.This paper performs contrast experiments on the benchmark image set.We com-pares the denoised results both in objective evaluation and subjective analysis,and as-sesses the performance of different denoising methods.Experiments on the benchmark image set show that the proposed approach can achieve superior performance to the state-of-the-art approaches both in accuracy and visual perception in removing the zero-mean additive white Gaussian noise.