基於模型方法之表情辨識與模擬

A Model based Method for Facial Expression Recognition and Simulation

羅聖諺(SHENG-YEN LO)

指導教授 ： 劉益宏 ； 陳祝嵩

中原大學/工學院/機械工程學系/碩士(2008年)

https://doi.org/10.6840/cycu200900557

摘要

最近幾年來，人臉表情辨識(Facial Expression Recognition)已被運用在一些科技產品
上，例如數位相機具有判斷受拍者是否為最開心的表情進而決定拍攝(Smile-shot)。因此
本篇論文致力於開發自動化人臉表情辨識與模擬之人機系統，此系統具備有人臉偵測
(Face detection)、眼部偵測(Eye detection)、嘴巴偵測(Mouth detection)、表情模擬(Facial
expression simulation)及表情辨識等功能。本文使用主動形狀模型(Active Shape Model,
ASM)偵測在動態影像上因表情變化而不斷改變的嘴巴形狀，並以形狀模型中收斂後的
形狀參數(Shape parameters)為特徵，建構兩個類別的人臉表情分類器，其可辨識目前使
用者的表情為正向表情或是負向表情，其中正向表情包含高興與驚訝，負向表情包含生
氣與悲傷，平均分類正確率達到98.15%。本篇論文也利用收斂後特徵點位置來達到表
情模擬的目的。另外ASM 向來有耗時的缺點，因此本篇論文另一重點在改善此一耗時
問題，利用有別於傳統ASM 收斂的方式，以不同的預設形狀所建立的灰階分佈模型
(GLDM)來判斷ASM 的初始形狀為何，進而減少收斂時疊代的次數並保持收斂後形狀的
準確性。本論文所提出之改良式ASM 在單一張影像上，從估計初始形狀至收斂的時間
平均達到0.08 秒，平均疊代的次數為3 次，而傳統ASM 從估計初始形狀至收斂的時間
平均為0.2 秒，平均疊代的次數為18 次。因此，本論文所提出之改良式ASM 能有效地
加速整個人臉表情辨識與模擬系統。

關鍵字

灰階分佈模型 ； 人臉表情辨識 ； 人臉表情模擬 ； 主動形狀模型

並列摘要

In recent years, facial expression recognition (FER) has applied in several high-tech
products, for example, some digital cameras can determine to take a picture by the apex of
human expression (Smile-Shot technology). Thus, this thesis aims to develop an automatic
facial expression recognition and simulation system. This system includes some functions
such as face detection, eye detection, mouth detection, expression recognition and simulation.
First, Active Shape Model (ASM) is used in this thesis. It can detect mouth in one still image.
This thesis constructs a two-class Support Vector Machine (SVM)-based expression classifier
where inputs are shape parameters. For SVM, positive class is defined as the positive
expressions including happiness and surprise. On the other hand, negative class is defined as
negative expression including anger and sadness. Average classification rate is 98.15%. To
solve time-consuming problem of classical ASM, Gray Level Distribution Model (GLDM) is
used. It will decrease iterations of convergence and maintain accuracy of shape. By the
improved ASM, the average execution time for one image is reduced to 0.08 seconds, and
average iterations are 3. By the classical ASM, the average execution time for one image is
about 0.2 seconds, and average iterations are 18. Therefore, the proposed improved ASM can
speed up the whole facial expression recognition and simulation system.

並列關鍵字

facial expression simulation ； gray level distribution model (GLDM) ； active shape model ； Facial expression recognition

參考文獻

[1] P. Ekman, W. V. Friesen, The Facial Action Coding System: A Technique For The

Psychologists Press, 1978.

Emotion-Based Agent Architectures, pp.18–26, 1999.

[4] Y. Li, W. Ito, “Shape Parameter Optimization for AdaBoosted Active Shape Model,”

IEEE International Conf. Computer Vision, vol. 1, pp. 251-258, 2005.