The Research Of Image Regularization Method Based On Nonlinear Structure Tensor

Image restoration is one of the most important branch in the field of computer vision. It has great significance for improving the image quality and restoring the image information. The essence of image restoration is the solution of the inverse problem. Regularization method is the main choice to solve the question by changing the ill-posed inverse problems to benign problems. Among them, Total variation model can restore the image detail information of the edge and smooth the noise at the same time has been attended widely in recent years. Structure tensor is an effective tool for image analysis in the field of image processing due to its ability to fuse the local structure information of the image. This paper proposes a novel image restoration method by combining Structure tensor with regularization method. The main research work and contributions are as follows.1) In-depth investigation is conducted in terms of the image restoration such as application background, basic principle and the development of domestic and foreign.The main restoration methods are classified and summarized, and the existing problems are also expounded. It has establishes research direction for the later study.2) A structure tensor smoothing method based on atomic decomposition is proposed. At first, informed the main smoothing methods are analyzed and pointed out that the smoothing methods can not preserve local structure information in structure tensor data leading to reduce the image analysis ability of the structure tensor. This paper construct anisotropic nonlinear structure tensor according to atomic decomposition. The benefits of this method as follow: to exploit the Spatial direction information to regular the tensor fields better. To calculate the weight with euclidean distance directly instead of promoting the nonlocal means filtering to the tensor fields. Experimental results indicate that new anisotropic structure tensor can be able to analysis the local structure and extract useful information from image well.3) A nonlocal regularization method based on the nonlocal structure tensor is proposed. The construction of regularizer is the most important subject in the research of the regularization method. The traditional theory is portraying the continuous and discontinuity by the local derivative information of the image and exploits the information of the image inadequate. This paper add more accurate local neighborhood geometry information to the nonlocal regularization method according to introducing the nonlocal structure tensor to the construction of nonlocal regularizers. Experimental results indicate that the new nonlocal total variation model has excellent ability to keep the spatial structure information for removing the noise and retaining the edge information at the same time.