통계학적 모델을 기반으로 한 인체 피부 영상의 분할 및 모델링에 관한 연구

다국어 초록 (Multilingual Abstract)

Skin analysis has been applied in many fields of both image-based and modeling-based analysis, involving computer assisted diagnosis for dermatology, topical drug efficacy testing for the pharmaceutical industry, and quantitative product comparison for cosmetics. Computer vision attempts to artificially approximate the capabilities of human vision using computers. Image segmentation is a subset of computer vision that attempts to partition an image into logical parts that correspond to objects or regions of interest. The research in the thesis is primarily in the field of human skin related image segmentation and 3D skin modeling.
The skin is a complex organ, which covers the entire body surface as a flexible shield with the primary function to protect the organism and act as permeability barrier to the environment. In computer graphics, the study on skin is used in human simulation, such as facial expression animation, human body modeling and rendering. In computer vision, skin related research is applied for face recognition, fingerprint identification, or people tracking. On the other respects, skin research helps for medicine development, and cosmetic effect evaluation. In the thesis, the research focus on computer-aided skin image segmentation and 3D modeling for cosmetic effect evaluation and display the skin condition with 3D model clearly.
In the thesis, I propose several algorithms for skin image processing and applied different algorithm for 3D modeling based on stereo images. Main research lists in the following:
First, I proposed a global uneven illumination correction algorithm for skin pores detection as a preprocessing. Skin surface is influenced strongly by the light direction especially with high light reflection on the nose and cheek. Before segmentation algorithm, we will consider for illumination correction in original images. K-means algorithm was applied to skin pores detection.
Second, in order to evaluate medicine effect in promotion of hair growth and nutrition, I proposed a framework for hair counting and measurement. My main contribution is an improvement of iterated threshold algorithm for hair segmentation, and also gives the convergence analysis to my algorithm. Skeleton technique solved the conjunction point problem in crossed hairs to hair counting. The error of hair counting has been controlled fewer than 10%. The work is more effective and effortless than counting and measurement manually.
Third, I study on 3D skin reconstruction based on stereo image. In order to obtain reliable disparity map, I proposed the two stage algorithm for dense matching. And in the condition where images without same focus skin regions, I used semi-dense matching algorithms for wrinkle structure extraction. Consider skin properties, I also use MRF model for 3D reconstruction in the finial part. The Evaluation of those algorithms was given in the end of section.
My research about human skin involves in image segmentation and 3D modeling. The proposed algorithms are effect to evaluate the cosmetic functions and the 3D modeling system based on two images can be used in face animation and wrinkle measurement.

목차 (Table of Contents)

• I. Introduction 1
• 1.1 Research on Human Skin 1
• 1.2 Computer Vision 4
• 1.2.1 Image Segmentation 8
• 1.2.2 Images Acquisition and System Constructed 10
• 1.3 Contributions and Organization of This Thesis 11
• II. Background 15
• 2.1 Segmentation Approaches 15
• 2.1.1 Level Set Methods 17
• 2.1.2 Threshold Methods 19
• 2.1.3 K-means Algorithm for Image Segmentation 21
• 2.2 Basic Concepts in Two-View Geometry for 3D reconstruction 23
• 2.2.1 2D Projective Geometry 23
• 2.2.2 Two-View Geometry 28
• III. Skin Pores Detection for Image-Based Skin Analysis 33
• 3.1 Prior Work and Contribution 33
• 3.2 Global Illumination Correction 35
• 3.3 Skin Pores Detection 36
• 3.4 Experiments and Analysis 39
• 3.5 Concluding Remarks 44
• IV. A Framework for Automatic Hair Counting and Measurement Based on An Improved Iterative Thresholding Method 46
• 4.1 Prior Work and Contribution 46
• 4.2 An Improved Iteration Method for Hair Segmentation 48
• 4.2.1 Iterative Thresholding Method and Convergence Analysis 48
• 4.2.2 An Improved Iterative Thresholding Method 50
• 4.3 Hair Counting and Measurement 52
• 4.4 Experiments and Analysis 55
• 4.4.1 Image Segmentation Experiments 56
• 4.4.2 Hair Counting Evaluation and Measurement 57
• 4.5 Concluding Remarks 62
• V. 3D Skin Modeling Based on Stereo Image Pairs 63
• 5.1 Introduction and Prior Work 64
• 5.2 Semi-Dense Matching for Disparity Map 68
• 5.2.1 Semi-Dense Matching Algorithm 69
• 5.2.2 Experiment and Analysis 72
• 5.3 Dense Matching for Disparity Map 75
• 5.3.1 Two Stages Stereo Dense Matching Algorithm 75
• 5.3.2 Experiment and Analysis 79
• 5.4 Markov Random Field Model for Stereo Images Reconstruction 85
• 5.4.1 MRF Model and Gibbs Probability Distribution 85
• 5.4.2 Proposed Model 87
• 5.4.3 Experiment and Analysis 88
• 5.5 Conclusion and future works 90
• VI. Conclusion 93
• References 95
• Acknowledgements 104