Research Of Global Sparse Gradient Based Image Processing Methods

As the fact that the human visual system is very sensitive to the changes of a scene and the most information of the scene is contained in its gradient,the gradients information of images have important significance in the fields such as image processing,computer vision and artificial intelligence and play a key role.In this dissertation,we do some research on the applications of image gradient information in some basic but important problems in image processing,including image denoising,image enhancement and image segmentation.Some novel models as well as their corresponding efficient algorithms are proposed.The main contributions are summarized as follows.1.In view of the classical gradient operators are sensitive to noise,a global sparse gradient model is proposed firstly to obtain a more accurate and robust gradient field from a noisy image.The global sparse gradient model makes use of points within the neighborhood(even global image)of an evaluated point to estimate its gradient and imposes a sparse regularization on the gradient field simultaneously.The model is solved by the proximal forward-backward splitting algorithm.Then two global sparse gradient based model are considered for image denoising.Due to using the absolute value of the gradient image as a guiding road map in the diffusion process,the P-M anisotropic diffusion equation is an ill-posed problem and can give spurious oscillations in the denoising process.At first,the global sparse gradient coupled anisotropic diffusion model is considered,furthermore,the existence and uniqueness for the solution with initial-boundary value problem are proved.In order to take full advantage of of direction information of the global sparse gradient,a tensor matrix is constructed by the global sparse gradient.Then,the diffusion equation is guided by tensor matrix for image denoising.Experimental results demonstrate that the proposed methods have better performance both in objective measurement and visual evaluation than existing PDE-based methods.2.Non-local means algorithm has obtained nice denoising results by making full use of the self-similarity and structural information redundancy of images.But the weight function of non-local means algorithm cannot accurately measure the similarity between image patches in the case of noise.This chapter proposes firstly an adaptive global sparse gradient model to improve the accuracy of the gradient field.Then,an adaptive global sparse gradient based non-local means denoising algorithm is proposed,in which the adaptive global sparse gradi-ent is introduced to redefine the weigh function.Experimental results demonstrate that our proposed method has a better performance both in objective measurement and visual evalu-ation than the non-local means algorithm and other improved algorithms using information of gradients.3.We propose a new variational Retinex model for image enhancement which comprise three parts:fidelity term,space regularization term(prior terms)and guidance vector field term.We assume that illumination is spatially smooth,reflectance is piecewise constant and the gradient of illumination layer is expected to approximate a guided gradient field.The new variational Retinex and global sparse gradient model compose a global sparse gradient guided variational Retinex for image enhancement.To solve it,an alternating minimization algorithm are developed.The experimental results are presented to illustrate the effective-ness of the proposed model.4.In view of the problem that the GVF Snake model is difficult to convergence to long and thin indentations,the global gradient flow model is proposed to obtain more effective gradient field to guided image segmentation.Then global gradient flow guided snake model is introduced.Finally,the proposed Snake model is used on synthetic and noisy images.Experimental results show that the model can effectively set the initial outline and accurately obtain the target contour.