Research On Several Technologies Of Image Analysis For The Objectification Of Facial Diagnosis In TCM

TCM (Traditional Chinese Medicine) considers that human body is an organicwhole. The pathological changes of internal organs or psychological will be reflectedin the relevant region of the face eventually. Facial diagnosis is the best way toobserve disease. Therefore, facial diagnosis is very important in diagnosis, and it's animportant diagnostic evidence for characterization of disease and prognosis. With thestudy of modernization and objectification in TCM received more and more concernat home and abroad, it's an urgent affair to speed up objectifying and modernizingdiagnostic criteria in TCM. The objective study of facial diagnosis is an importantpart of diagnosis in TCM, and also is of great significance to accelerate the objectivestudy of the four diagnostic methods.Several key technologies of image analysis for the objectification of facialdiagnosis are studied in this thesis. Based on previous studies, we have followingworks:In the first chapter, facial diagnosis in TCM was introduced. Researchbackground, the current situation and problems in objective research, and the contentof this thesis were elaborated.In the second chapter, the acquisition environment of face image and severalcolor spaces were presented. We also elaborated some related algorithms, including:Level Set theory, LDA, sLDA topic model and SVM algorithm.In the third chapter, the face segmentation was studied, and an improved LevelSet algorithm was put forward. My algorithm has the following major improvements:(1) According to the distribution of human face in the HSV and YCbCr-like colorspace, we realized the initialization of active contour automatically.(2) According tothe characteristics of the junction of the face and neck, we proposed a boundaryenhancement method.(3) To improve convergence of the weak edge, we constructed a signed pressure force function based on regional statistics to replace the edgestopping function of Level Set model.(4) We used a Gaussian filtering process tofurther regularize the level set function, and that greatly improved the evolutionefficiency of curve.In the fourth chapter, facial feature points were located.(1) Based on previousworks and priori knowledge of Geometry, the face image was pre-processed and theinterference of external factors was eliminated. An improved integral projection wasproposed, and then it was used to locate the human eye and the horizontal line ofnasal alar base.(2) According to the characteristics of the H component of the eyeregion in the HSV color space, we achieved to locate the pupil by using integralprojection, and combining with the initial results of the human eye location.(3)According to the difference color between the region of eye and skin, we found thecontour of eye and located the inner and outer canthus.In the fifth chapter, identification and classification of facial color were studied.Quantification color histogram was employed to extract features from skin blocks,and sLDA was first introduced to construct a model, which was used to identify andclassify the color of skin.In the sixth chapter, we summarized our works in the thesis. Shortcomings ofworks were pointed out and the objectification research of facial diagnosis wasprospected.