Image Processing Algorithms Based On Total Variation Model For Wireless Capsule Endoscopy Images

Traditional clinical gastrointestinal disease diagnosis means cannot detect the small intestinal part. Moreover, it could bring great physical discomfort and psychological burden to the patient during the inspection process, and easily lead to cross-infection. In addition to the above mentioned defects, these classical diagnose methods possess low rate of positive poor positioning accuracy and heavy invasiveness. Wireless capsule endoscopy (WCE) is a state-of-the-art technology to diagnose the gastrointestinal (GI) tract diseases with almost no invasiveness, and this device can readily reach the small intestine region, it is a new breakthrough in the field of biomedical. It has become a hot topic and technology in the clinic research and gradually gained clinical recognition for its clinical detection effectiveness. One time capsule endoscopic detection will produce tens of thousands of pictures, selecting abnormal or abnormal figures from these mas-sive data become an onerous task for clinicians. Thus, developing an image processing technology for WCE images to shorten the reading time and improve the diagnosis accu-racy gradually become a hot research in medical image processing and urgent problem to be solved. For the purpose of improving the image quality, we in this paper developed several mathematical models after analyzing a number of issues for the wireless capsule endoscopy color image processing.Due to the gastrointestinal dark complex internal environment and the inevitable outside influence, and interference during the image acquisition and wireless transmis-sion process, the color wireless capsule endoscopy images are often contaminated by noise or blurred by different convolution operator, or degraded by these two effects. The degraded wireless capsule endoscopy images will produce some side effects for the clin-ical aided diagnosis. Therefore, digital image processing technique is urgent needed to improve the wireless capsule endoscopy color images quality by using, for instance, im-age denoising, image deblurring, blind deconvolution and level set image segmentation. In this thesis, image processing algorithms based on total variation model for wire-less capsule endoscopy color images are developed for improving the quality of these images, which include image denoising, image deblurring, image blind deconvolution and variational image segmentation. Specially, we will:(1) Analyze some basic issues and related work about image pre-processing based on total variational framework, which including image denoising, image deblurring, im-age blind deconvolution and image segmentation. Advantages and disadvantages are demonstrated at the beginning of each chapter after some research and analysis about the existing algorithms, and then described the proposed solutions, and last demonstrated the mathematical model. From the theoretical derivation to the analytical details for this al-gorithm, at the end of every chapter, we present the experimental results and comparisons with the classical approaches.(2) Discuss fundamental issues and related work for the current image denoising algorithms and pointed out their drawbacks. In the traditional image denoising methods based on the total variational framework, the regularization parameter is kept invariant in the entire iteration process. An optimal value of regularization parameter is crucial as it affects the denoising performance directly and significantly. Consequently, develop a bisection technique to automatically determine a near optimal value with an exponential convergence rate is an urgent task, and then extended it to the color images consider-ing of the high correlation between channel components. The proposed algorithm is enhanced by incorporating a bisection technique into the algorithm that helps identify a near optimal value for a key regularization parameter. Simulation results are presented to demonstrate the effectiveness of the proposed algorithm for restoring color wireless capsule endoscopy images.(3) Analyze the fast gradient projection algorithm proposed recently by Beck and Teboulle. After that, an improved objective functional is presented to overcome the defects of heavy computation and long elapse time for the traditional methods, reduce algorithmic complexity and improve the convergence rate. Facing to the same situation, the regularization parameter is kept invariant in the entire iteration process and this will affect the experimental results especially for the deblurring procedure. The proposed algorithm is enhanced by incorporating a bisection technique to effectively identify a near optimal value for the regularization parameter of the TV-Frobenius objective func-tion. Simulation results are presented to demonstrate the effectiveness of the proposed algorithm for restoring color wireless capsule endoscopy images.(4) Investigate the image blind deconvolution problems. Classical image restora-tion is devoted to estimating the true image assuming the blur is known. By contrast, blind deconvolution as an image restoration task tackles a much more difficult problem where, in addition to the image being unknown, the blurring mechanism is also unknown. So this topic is actually an ill-posed problem and it is a complicated and challengeable question. Unlike the alternating optimization based algorithms, the new method adopts a joint estimation strategy that estimates the unknown blurring kernel and unknown im-age in an iterative manner where each iteration performs two separate image denoising subproblems that admit fast implementation. Experimental results on gray-scale test im-ages as well as wireless capsule endoscopy color images are presented to demonstrate that the new method is capable of blindly deblurring images with considerably improved performance, for instance, the robust property, more steady and fast convergence rate.(5) Analyze the classical C-V and GAC models and illustrate their disadvantages, for instance, the initial level set curve must keep the signed distance function property, the initial curve must surround the target objects, which is especially sensitive to noise and blur effect and neglect the weak boundary of the deep concave objects. By virtue of eliminating the time-consuming re-initialization procedure neglecting for the property of the level set function during the evolution process, we present two strategies to solving these problems. Two scenarios are considered, namely, first, the distance regularization term which is defined by double-well potential function with two minimum points is in-troduced to our mathematical model for avoiding the re-initialization process. Second, by combining a Tikhonov like regularization term which can guarantee the smooth for the evolution curve over the previous method. This is the same purpose for the mul-tiphase level set image segmentation model. From the comparison resuts between the proposed method and the traditional C-V and the multiphase C-V model, our revised two models for the aforementioned versions possess the following advantages, one is the fast convergence rate and the other is the smooth contour in the experimental results. Several experimental studies in comparison with algorithm are presented to evaluate the considerable performance gain of the proposed algorithm.