Study On Impulse Noise Filtering Algorithm Based On Sparse Representation

During acquisition and transmission digital images could be corrupted by impulse noise.As a result, some pixel values are inevitably contaminated whereas others remain noise free. Therefore, switching techniques have been applied for the removal of impulse noise, detect whether the current pixel is corrupted or uncorrupted before image filtering.For noise detection, advanced boundary discriminative noise detection algorithm used global histogram to obtain noise boundary and got good detection results. However, the rate of false detection increases a lot for ABDND when the range of noise boundary is broadened. For noise filtering, the traditional spatial filtering methods can not preserved details of the image well. Sparse representation got good performances in the Gaussian noise filtering, some improved algorithms are used for impulse noise filtering. However, these improved algorithms are based on the dictionary training or noise detection, the objective function is still sparsity of image, resulting in bad performance when the impulse noise level is high.This thesis addresses the issues of image impulse noise filtering. Surrounding the two main problems, noise detection and image filtering, the random value noise boundary detection, regular function filtering and sparse representation used in impulse noise filtering are deeply studied. The main contributions are as follows:(1) Aim at random-valued impulse noise detection in two boundaries, Modification of ABDND (MABDND) is proposed in this paper. In the second stage, it uses the statistic of part histogram to find out pixels of false detection in the first stage, and marks them as uncorrupt pixels. The merit of MABDND is to use the confirmation technique in the second stage to rectify many pixels of false detection in the first stage to keep a low rate both for miss detection and false detection.(2) Combining ABDND global statistical and BDND local boundaries, boundary statistics noise detection method was proposed, which is a combination of local and global statistics, and can accurately determine the final boundary of the noise, so as to achieve the noise detection.(3) A filtering method was proposed to restore image corrupted by random impulse noise, including a noise detector and an image filter. The functional filter is constructed by the regular function of the image and the constraints of noise pixels to restore noise pixels. Due to the selected regular function satisfies the general characteristics of the image, it can be well preserved image details. This method has a high real-time, and also has a good filtering effect.(4) Aim at the sparse representation used in impulse noise filtering. Based on the characteristics of impulse noise, an improved K-SVD based on uncorrupted pixel reconstruction, PK-SVD, is proposed to filter impulse noises. It has the optimal function with uncorrupted pixels to obtain the filtered image to improve the filtering performance, which is computed by the method integrated the hierarchical property into the OMP algorithm. Moreover, PK-SVD uses the iterant K-singular value decomposition to update atoms and coefficients for the dictionary during the training of the dictionary. The PK-SVD applied to impulse noise and mixed impulse-Gaussian noise filtering, and compared with other filtering methods.