莊永裕Chuang, Yung-Yu臺灣大學:資訊網路與多媒體研究所彭盛韶Sheng-Shau, PengPengSheng-Shau2010-05-052018-07-052010-05-052018-07-052008U0001-2407200820310100http://ntur.lib.ntu.edu.tw//handle/246246/180620主動形狀模型近幾年在電腦視覺方面被充分的注意，在實際應用層面也涵蓋了許多不同方向諸如醫學影像、臉部特徵定位等。模型的特色為它同時結合了形狀及表面顏色這兩種資訊，透過一些訓練後即可用來處理一般問題且達到不錯的成果。但因為模型中一些為了逼近而作的假設，使得有些時候搜尋特徵點時容易進入區域的最小值而導致不好的結果。一方面，支持向量機是近幾年在機器學習及圖型辨識中人們常用的工具。它的優點為訓練過程及判斷過程皆很迅速，但最後的成果往往訓練時使用的樣本有很大的關係。這篇論文中，我們在臉部特徵定位上加上了與前人研究截然不同的觀點。我們結合了主動形狀模型及支持向量機並透過巧妙的運用，持向量機可在臉部特徵定位過程中提供額外寶貴的資訊，而這些資訊將使主動型狀模型遠離掉入區域極小值的機會並使最後臉部特徵定位的果更完善。Active Shape Model(ASM) causes great attention in recent years, it can be used in many different application including medical imaging,ace alignment etc. One of the characteristic of ASM is combining shape and intensity, with carefully training it can locate featuresn given image . But due to some assumption and approxiamation, sometimes the search falls into local minima.n the other hand, Support Vector Machine is a tool that many researcher use in machine learning or pattern recognition area.ts advantage is the speed of training and classification although the performance is highly relative to training samples.n this paper, we propose a novel view in face alignment. With combining ASM and SVM, we obtain additional information instead ofocal intensity. These information help the search away from local minima and improve the performance of alignment.1 Introduction 1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Related work 5.1 Shape and shape model . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.2 Aligning shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Active Shape Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.1 Point Distribution Model . . . . . . . . . . . . . . . . . . . . . . 6.2.2 Local Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.3 Search Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.4 Two dimensional profile used in Stacking Two Active Shape Model 9.3 Basic Thoery of Support Vector Machines . . . . . . . . . . . . . . . . . 10.4 The Eigenface Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Algorithm 15.1 Background technique . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.1.1 Warped Images . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.1.2 Eigenappearance . . . . . . . . . . . . . . . . . . . . . . . . . . 17.1.3 Classificaiton between Correct and Incorrect Alignment . . . . . 18.2 Algorithm comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.2.1 ASM algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 19.2.2 ASM algorithm with SVM local region adjustment . . . . . . . . 21 Experiments 25.1 Face Dataset and Experimental Procedure . . . . . . . . . . . . . . . . . 25.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26.2.1 Model Comparison . . . . . . . . . . . . . . . . . . . . . . . . . 26.2.2 EigenAppearance . . . . . . . . . . . . . . . . . . . . . . . . . . 27.2.3 Result of searched images . . . . . . . . . . . . . . . . . . . . . 28 Conclusion and future work 35 Appendix 37.1 Face landmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37ibliography 39application/pdf1446783 bytesapplication/pdfen-US主動型狀模型臉部特徵定位支持向量機特徵臉人臉追蹤人臉辨識Active Shape Model(ASM)face alignmentSupport Vector Machine(SVM)eigenfaceface trackingface recognitionthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/180620/1/ntu-97-R95944021-1.pdf