Recognition of Down Syndrome Based on 3D Models

基於人臉立體模型之唐氏症識別

黃海祐(Hai-Yo Huang)

指導教授 ： 陳永昌

國立清華大學/電機資訊學院/電機工程學系/碩士(2010年)

摘要

摘要

唐氏症識別，除了用染色體做精確檢查外，主要還是由醫師、專家做外觀上的判斷。這篇論文，主要是從單張正面的人臉影像，做出人臉立體模型，並從此模型上擷取特徵，以自動化的方式訓練並識別唐氏症。

新生兒出生時，約有1/800的機會因為第二十一號染色體分裂不完全，而罹患唐氏症 (trisomy 21)。唐氏症外觀上的特徵有: 形狀異常的五官，較為扁平的側臉特徵，傾斜的眼睛，較小的鼻子、耳朵、下顎，突出的舌頭，以及因為自制問題引起的異常肥胖。此系統的主要目的，便是協助工作人員從影像列中快速判定唐氏症的存在。使用者並不需要有醫學上的專業知識，而自動化的判別作業交由電腦來處理，可有效節省人力資源。

實驗中先對正面的人臉影像，做主動外觀模型 (AAM) 特徵點擷取。然後藉由這些特徵點，推算出一個初步的，形狀參數強化(enhanced shape parameters)的立體模型。鑒於唐氏症患者側面通常較正常人更為扁平，我們先從有側面影像資料的範例，建立正面區域影像強度差與高度的模型，再用此模型，調整初步模型的深度。最後，從此模型上擷取特徵點間的距離，角度，比例等，跟原來3DMM 的形狀參數 (shape parameter)，一起做特徵選擇後，由SVM做最後的判別工作。結果顯示，此法在唐氏症患者的辨識上，有九成以上的成功率。

關鍵字

唐氏症 ； 3D ； 人臉

並列摘要

Abstract

Down syndrome, besides from being identified by chromosome tests, is mainly diagnosed by experts according to the subject’s appearance. In this article, 3D models are built from single frontal face images; then, features are extracted and selected automatically to train a classifier used to identify Down syndrome from images.

Down syndrome, or trisomy 21, is a chromosomal disorder caused by the presence of an extra 21st chromosome. The incidence of Down syndrome is estimated at 1 per 800 births. The common physical features of Down syndrome include abnormally shaped facial features, flat facial profile, upward, outward slanting palpebral fissures, small nose, ear, chin, a protruding tongue, and abnormal facial fat distribution. The purpose of this system is to automatically identify Down syndrome from a set of frontal images, without the need of an expert. Time and labor are thus saved.

In this experiment, features are first extracted from frontal face images using active appearance models. Then, an initial 3D model is obtained by fitting these sparse feature points to a 3DMM, with descriptive shape parameter terms being fitted first. Since individuals with Down syndrome usually have a flat facial profile, a PCA
model coupling frontal image intensity difference with depth is trained. This model is used to adjust the initial model. Finally, geometric features are extracted from the 3D
model. Together with shape parameters, they are feature-selected and sent to a SVM classifier for training. The experimental results show that the accuracy is above 90%.

並列關鍵字

Down Syndrome ； 3D ； face

參考文獻

[2] Hung-Yao Chen, “Recognition of Down Syndrome Based on Active Appearance Model”, MS thesis, National Tsing Hua University, Hsinchu, Taiwan, R.O.C., July 2009

[4] M. B. Stegmann, "The AAM-API: An Open Source Active Appearance Model Implementation", Proc. Medical Image Computing and Computer-Assisted Intervention, Page(s) 951-952, 2003

[6] Chih-Chung Chang and Chih-Jen Lin, “LIBSVM : a library for support vector machines”, 2001. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm

[7] Volker Blanz, Albert Mehl, Thomas Vetter, and Hans-Peter Seidel, “A Statistical Method for Robust 3D Surface Reconstruction from Sparse Data”, In Proceedings, IEEE 2nd International Symposium on 3D Data Processing, Visualization and Transmission,, Page(s):293 – 300, 2004

[8] Volker Blanz, “Face Recognition based on a 3D Morphable Model”, In Proceedings, 7th International Conference on Automatic Face and Gesture Recognition, Page(s):617 - 624 , 2006