Research On Image Denoising And Smoothing Methods Based On Nonlocal Sparsity

It is challenging to make the computer understand the environment through im-age or video in computer vision. Image denoising and smoothing are two fundamental problems in image processing field. During acquisition and transmission process, noise is inevitably introduced into images. The image distortion would interrupt the follow-ing tasks in computer vision, such as feature extraction and scene understanding. Image denoising aims to separate noise from noisy images, and it is important to restore the in-formation faithfully, especially in areas including edges and textures. If the clean image can be expressed by some model, the denoising procedure can allow better restoration quality. Image smoothing aims to separate structures and details, in order to extract the meaningful structures which are important for human visual system. If the smoothing result is of high quality, it can serve better for the subsequent tasks in computer vision. As fundamental image processing techniques, image denoising and smoothing can favor a lot of applications, such as space exploration, medical science, archaeology, machine vision, military sensing and satellite image processing. However, as the contents of im-ages vary a lot, there are many challenges in image denoising and smoothing. Firstly, it is still challenging to model the image using mathematical techniques. It is a common skill in image processing that modelling the image under some assumptions. Secondly, it is challenging to study the principle of human visual system, which makes it difficult to define the meaningful information for human beings.This paper focuses on the problems of image denoising and smoothing, and pro-poses nonlocal similarity based image processing methods. The new methods make use of the information, which can promote the denoising and smoothing performance of mathematical models in a data-driven manner. The works are listed below:1. Adaptive sparse coding on PCA dictionary method is proposed for image de-noising. By analyzing the statistical property of sparse code errors on PCA dictionary, a new nonlocal sparse coding model is proposed using maximum a posteriori estimation. The Laplace function is adopted for approximating the distribution of sparse codes. In the new model, the regularization parameter balancing the fidelity term and the nonlo- cal constraint can be adaptively determined, which is critical for obtaining satisfying results. To solve the new model effectively, a filter-based iterative shrinkage algorithm containing the filter-based back-projection and shrinkage stages is proposed. The filter in the back projection stage plays an important role in solving the model. As demon-strated by extensive experiments, the proposed method performs optimally in terms of both quantitative and visual measurements.2. An image smoothing method with low-rank and sparse gradient priors is pro-posed. By analyzing the structures and textures in natural images, a low-rank constraint on groups of similar patches is proposed. The low-rank constraint lies in nonlocal sim-ilarity in natural images. Then a new image smoothing model is proposed combining with global gradient sparsity. The low-rank prior can constrain the high correlation of similar patches in the same group. Therefore, the small-scale details with high-contrast can be removed, and the long and slender structure edges can be kept. An alternative iteration method is given for approximately solve the new model. The image can be smoothed in the consistent manner, and the main structural edges can be kept while details being smoothed.3. A nonlocal gradient concentration method for image smoothing is proposed. Observing the gradient maps of natural images, a nonlocal gradient concentration pri-or is proposed for the gradients of smoothed images based on the nonlocal similarity. This new prior is combined with Lo gradient minimization prior, thus a new optimiza-tion model is obtained. To solve the new model effectively, an alternatively iterative algorithm is adopted. Therefore, details throughout the whole image can be removed automatically in a data-driven manner. As variations of gradients among similar patches can be suppressed effectively, the new model performs excellent for edge preserving, detail removal and visual consistency. Comparisons with state-of-the-art smoothing methods demonstrate the effectiveness of the new method.4. Applications of image smoothing in intelligent image processing are studied. The smoothed image can preserve structural information which is critical for human visual system. Thus the smoothing has high value in promoting content-aware im-age processing tasks. Firstly, the image smoothing is applied in applications including image resizing and editing. Secondly, a new scale-space is constructed using image smoothing, and its application in saliency detection is also studied. The experimental results show that both the smoothing itself and the scale-space work well in the various applications.