控制環境下的人臉特徵辨識

Abstract

在本論文當中，提出了一套全自動的人臉特徵辨識法，可針對在控制環境下拍攝的人臉影像，自動標示出人臉特徵點與特徵輪廓線的位置。在我們的演算法中，將人臉特徵辨識的問題分為兩個階段處理，在初始猜測的階段，藉由模糊理論來大略辨識出嘴巴區塊、眼睛區塊、眉毛區塊、下巴區塊和鼻孔區塊，並依據這個大略辨識的結果給予人臉特徵點一組較佳的初始位置；在人臉特徵辨識的階段，我們針對不同的人臉特徵設計了幾組適合用來辨識人臉特徵的主動輪廓模型(active contour model，又稱snake)，我們所設計的主動輪廓模型不只使主動輪廓模型本身那條主動曲線能描繪出人臉當中的部份特徵輪廓線，其中一部份用來定義主動輪廓模型的元素(snaxel)更能直接對應到MPEG-4中定義的人臉定義參數(Facial Definition Parameter, FDP)，標示出人臉當中的重要特徵點。根據實驗結果顯示，我們的演算法能夠相當準確地標示出在控制環境下拍攝的人臉影像中部份特徵點和特徵輪廓線位置(包含嘴巴、眼睛、和人臉輪廓)。

In this thesis, we propose a new approach of facial feature recognition which can automatically locate specific facial feature points and contours in a 2D color image, which is captured in controlled environments. In our approach, we do facial feature recognition in two stages: first recognize facial features (mouth region, eye regions, eyebrow regions, chin region and nostrils regions) approximately by fuzzy logic. According to this approximate recognition of facial features, we give a good initial guess of facial feature points. Then, for different facial features, we use different active contours to extract more accurate facial feature points and contours. Our modified active contour is especially designed for facial feature recognition. Not only the active contours themselves may well represent facial feature contours, but some of the snaxels also correspond to FDP parameters in MPEG-4 to locate facial feature points directly. According to our simulation results, our approach can locate specific facial feature points and contours (mouth, eyes and facial contour) in a 2D color image with reasonable accuracy.