Research On The Analysis Of Endoscopy Images Based On Mean Shift

Endoscopy exam is one of the most important methods to detect gastric cancer. Computer aided analysis of endoscopy images is helpful to improve the accuracy of endoscopy tests. In this thesis, some methods of analysis endoscopy images are discussed and Mean Shift algorithm is applied.First, an endoscopy image database is designed, which help to analyze the color distribution of endoscopy images in RGB, HSV and YCbCr color spaces with histogram. The color distribution is the most discreted in RGB, which is suitable to segmentation. And it is so intergrated in YCbCr that the dimension of the feature vector can be fewer than the others. Then for preprocessing, a mask is selected to get region of interest (ROI) and specular reflection is reduced by the thresholds in HSV. Through an auto segmentation based on Mean Shift algorithm, similar pixels in a endoscopy image merge into same regions. Color feature vectors, three dimension joint histogram in YCbCr, are extracted from these regions. Meanwhile, according to multiscale, rotational invariance and complex color information of endoscopy images, Color Wavelet Covariance (CWC) is selected for texture feature extraction.In this paper, Mean Shift-Gray Level Co-occurrence Matrix algorithm (MS-GLCM), a improved algorithm for computing Gray Level Co-occurrence Matrix (GLCM) based on Mean Shift, is presented to solve the problem that computing GLCM costs too much time in CWC procedure. This new algorithm decreases the computing time cost and the number of function calls as well as partly prevent from data redundancy. It is applied and tested in CWC.Finally, after a comparison between Perceptron and Adaboost algorithms, a stronger classifier to detect abnormal regions in endoscopy images is trained by Adaboost. A series of comparison experiments about region segmentation, feature extraction and classifier confirm that the proposed methods have a satisfactory accuracy, False positive rate (FP) 18.89%, False negtive rate (FN) 36.91% and the mean error 23.50%.