Research On Bleeding Detection Of Wireless Capsule Endoscopy Images Based On Support Vector Machine

As the changing of eating habits and developing of life-style, diseases of digestive track has become a real big problem for human health. How to alleviate or solve this problem is a major concern in society. Wireless capsule endoscopy is a revolutionary technique with no pain, easy to use and the capability of inspecting the whole small intestine which traditional ways do not provide. However, WCE produces too many pictures of a patient. These images require an experienced doctor to browse one by one, which is a laboring and boring work. This paper aims to help medical staff using computer based techniques. Of all digestive track diseases, intestine bleeding is a common symptom which is also the hint of other diseases. Hence, automatic bleeding detection is an import issue in preserving and diagnosis of digestive track.In this paper, we use support vector machine as classifier to identify images, and use kinds of kernel function to check the feature vector’s performance. The first step is image preprocessing, which extracts the region of interest from the image, execute color balance based on gray world assumption and improve the image’s local contrast. For different capsule endoscopy, we provide different feature vector based on the characteristic of image content. With Pill Cam endoscopy, we use hsv color model, which is similar to human visual perception, to construct a two dimension hue-saturation feature. To eliminate the effect of area of bleeding region, we binary the feature. Considering the property of endoscopy images, a clipped illumination invariant color space is introduced. The binary feature is faster and more precise than histogram based algorithm which is shown in the experiment results. As OMOM endoscopy, we construct the feature in Lab color space, which is more suitable for OMOM endoscopy images, since its wide color gamut and serious color shift. By combining the characteristic of Lab color space and endoscopy images, we propose a maximum in each chromium channel to represent the image’s feature. Cooperating with B-spline kernel function, this method give a good result on OMOM endoscopy images. At the final chapter, we list a series of approach to complete the current algorithm.