Research Of Segmentation Algorithms For Noisy Image Based On Level Set

Image segmentation is an important preprocessing step in the field of digital image processing and computer vision.It reduces the amount of data to be processed in advanced processing stages such as image analysis,image recognition and image understanding,while retains the original structural information of the image.Because the result of image segmentation affects each link of image processing,it attracts the attention of researchers so that large numbers of segmentation models and algorithms have been proposed.Level set methods are widely used in image segmentation because they provide more smooth and accurate segmentation results,and embed other prior knowledge easier.However,the diversity and the complexity of visual information,as well as the noise interference in digital images introduced by imaging devices and external environments,bring great challenges to image segmentation methods based on level set.This paper researches on the robustness against noise and the accuracy of segmentation for the segmentation problem of noisy images based on variational level set models.The main research contents and achievements are given as follows:(1)To solve the problem that edge-based level set method is sensitive to noise,an improved edge-based level set method combining local fitting information is proposed to segment noisy images.It analyzes different regional attributes between noise points and edge points.That is,local neighborhoods of noise points are homogeneous regions and edge points' are heterogeneous regions,which lay the foundation for using the local regional attributes of pixels to determine their ownership.Based on this analysis,it uses local fitting means to construct a variable regional coefficient,which solves problems that constant regional coefficient in noisy image segmentation,more specifically,a large coefficient leads to active contour leaking object edges while a small coefficient results in contour curve falling into local minima.Meanwhile,it introduces local fitting variances to improve edge stop function,which enhances its robustness against noise.Noisy image segmentation experiments on synthetic and natural images demonstrate that it is robust against noise and solves the segmentation problem of noisy image to a certain extent.(2)To solve the problem that edge-based level set method is difficult to segment noisy images effectively,a weighted edge-based level set method based on multi-local statistical information is proposed.Based on analyzing the function of both energy terms and their coefficients of traditional variational model in the level set evolution,and the previous work(1),local intensity and frequency information of image pixels are used to determine the attribution of given pixels.It uses the maximum value of local entropy to present a normalized local entropy.Then based on normalized local entropy,local fitting means and local standard deviations,a weighted length coefficient,a weighted regional coefficient and a modified edge stop function are proposed.The weighted length coefficient can effectively suppress the appearance of noise points and retain more edge details;the weighted regional coefficient effectively can deal with noisy images with different types of noise and solve the problem to select nonlinear coefficient selection;the modified edge stop function is more robust to different types of noise.Noisy image segmentation experiments performed on synthetic,natural and medical images show that segmentation results provided by this method are much better in both accuracy and visual effect,which verifies its effectiveness in noisy image segmentation.(3)To solve problems that region-based level set method is difficult to deal with high noisy images and its non-convex energy functional makes active contour falling into local minima,a variational level set method based on adaptive local fitted image is proposed for the segmentation of noisy images.The method uses local fitting means and normalized local entropy to propose an adaptive local fitted image and then introduces it into the data energy term,which enhances the performance of the proposed model against noise.Based on the special convergence of the data term,a data penalty term is presented to optimize the adaptive local fitted image,which reduces the influence on segmentation accuracy introduced by the error of the adaptive local fitted image.To smooth level set function,total variational regularization term is used to regularize it.Besides,the influence of noise on the active contour curve is further reduced.Since the energy functionals of data energy term,data penalty term and total variational term are convex functions,the whole energy functional of the model is also a convex function according to the properties of convex function,and it has a global optimum in the image domain.Large numbers of noisy image segmentation experiments implemented on synthetic,natural,synthetic aperture radar and oil spill images indicate that the proposed method has strong robustness against noise and effectively deals with the segmentation of noisy images and high noise images.