Omnibearing Tongue Image Feature Extractions And Classification Via Multiple Kernel Learning

Tongue Diagnosis is one of the most important diagnostic approaches in Traditional Chinese Medicine (TCM). Recently, the development of computer science makes automatic and objective TCM Tongue Diagnosis possible. This thesis is part of automatic tongue diagnosis research. We focused on omnibearing tongue image feature extractions and classification, including image preprocessing, feature analysis, and classification based on Multiple Kernel Learning (MKL).In the first part, tongue image is preprocessed to improve the algorithm accuracy. We first introduce the instrument to obtain a standard digital tongue image and the color correction method with color cards. Then a semi-automatic method of tongue segmentation based on watershed and region merging algorithm is proposed to remove the non-tongue part of image. After preprocessing, the tongue image can be effectively used for the following analysis.Number of tongue features, including color, texture, geometric feature and tangible material feature are analyzed in the second part. We apply manifold learning method to explore tongue color, and separate the tongue coating and substance. Gabor energy values of sub images are calculated in texture feature. For geometric feature, we introduce the classification method of tongue shape, and put forward an algorithm to calculate the thickness of tongue. Red points, petechia, cracks, and tooth marks are detected in the part of tangible material feature. We also develop an automated tongue analysis system based on these feature analysis algorithm.In the last part, the data of tongue features is used for classification. Matrix reconstruction is applied before classification to fill the missing data and eliminate the large errors. Then, we apply MKL to learn and classify data. Experiments are carried out and the experimental results prove that our methods are effective and practical.