Research On Nonlocal Image Denoising Methods And Their Applications

Image denoising is a basic topic in the field of image processing and computer vision. The aim of image denoising is to remove noise while preserving image details as much as possible. The existing image denoising methods are classified into local methods and nonlocal methods, the nonlocal means (NLM) methods as a brand-new image denoising strategy was proposed in recent years. This dissertation profoundly analyzes some drawbacks of NLM and proposes several improved NLM algorithms. The proposed algorithms obtain better image denoising results than the original NLM method and several existing NLM methods. Moreover, we pay close attention to the problem of impulse noise removal and extend NLM method to remove impulse noise from images. The proposed algorithm is capable of effectively suppressing any type of impulse noise and mixed impulse noise, in contrast to many existing methods which are specialized for impulse noise removal. The main research contents in this dissertation are outlined as follows:1. The preselection based NLM denoising methods are analyzed intensively, and it is pointed out that they all have deficiencies in terms of feature extraction from image patches. We employ singular value decomposition in the image gradient domain and propose an adaptive efficient NLM image denoising method. Mainly, our contributions to the preselection based NLM methods are:1) Robust image structure descriptor is designed;2) Relationship between the size of the similar sets and denoising results is analyzed;3) The similar patches can be selected automatically;4) The similar weight coefficient parameter can be determined adaptively. Experimental results show that the proposed method outperforms original NLM method and other existing NLM methods based on preselection in terms of denoising results and running speed.2. This dissertation profoundly analyzes some drawbacks of NLM method in similarity measure and proposes an improved NLM image denoising method based on adaptive Gaussian kernel. A new similarity measure method is designed by making use of local image structure information obtained from adaptive Gaussian kernel. It includes two parts:1) Rotation matching based image patch similarity comparison;2) Adaptive gaussian kernel based similar distance. The proposed algorithm can robustly measure similarity between image patches even if they appear in the rotated instances. Hence, more candidates can be found for the weighted average and yield improved results. Moreover, the similar weight coefficient parameter is optimized to obtain better denoising results. Experimental results demonstrate the robustness of the proposed similarity measure and its potential with respect to original NLM method and other excellent methods.3. In this dissertation, we also focus on the problem of impulse noise removal and propose a universal impulse noise filter. For detection, a robust local image statistic, called the extremum compression rank-order absolute difference (ECROAD), is designed to detect impulse noise in an image. For filtering, a universal impulse noise filter is proposed by combining ECROAD results with NLM filtering framework. The proposed impulse weight is able to avoid the effect of noisy pixels in computing similarity weight and restoring candidates. The patch-based similarity measure can provide higher correlation between the corrupted pixel and neighborhood pixel. Higher correlation gives rise to better noise suppression and edge preservation. Experimental results demonstrate that the proposed filter is efficiently able to suppress any type of impulse noise and mixed impulse noise, and outperforms other universal impulse noise filters and some existing filters which are specialized for different impulse noise models.4. We present two filters, noise adaptive edge-preserving filter (NAEPF) and noise adaptive switching bilateral filter (NASBF), for removal of salt-and-pepper noise. For detection, two extreme intensity values are used to identify possible noise pixels. For filtering, NAEPF is first developed for noise suppression and details preserving. It adopts three different filtering techniques for noise removal and detail-preserving, switching filtering, edge-preserving filtering, and noise adaptive median filtering. Switching filtering can retain noise-free pixels unchanged. Edge-preserving filtering can preserve more image details. Noise adaptive median filtering can suppress high-level salt and pepper noise. Then, NASBF is proposed by combining noise detection results with NAEPF estimation. Experimental results demonstrate that NAEPF outperforms some of the existing filters and the NASBF can obtain better performances than those state-of-the-art filters.