利用人臉姿態估測及人臉姿態合成技術之人臉辨識系統及其應用

Abstract

在本篇論文中，我們提出了改善人臉姿態估測的方法，藉以提升人臉辨識系統的辨識率。本論文所提出的方法可以分成三個部份。第一個部份是利用主動輪廓線來偵測人臉的下巴曲線：利用人臉下巴統計出來的實驗數據，本論文提出將此主動輪廓線自動初始化，藉以估測任意姿態的下巴曲線，同時也提高估測的正確性。第二部分是人臉姿態的估測和合成：利用測得的下巴曲線，以及其他的人臉特徵。配合模擬退火的最佳化理論，可估測出任意人臉姿態的資訊，並藉由此資訊來合成一個正面姿態人臉。最後一個部分是人臉辨識：我們利用合成正面姿態的人臉來做辨識，可改善人臉辨識系統中，當人臉非正面姿態而導致辨識率不佳的問題。從實驗結果中可以發現，當有各種不同姿態的人臉影像要辨識時，相較於傳統的演算法的平均辨識率約只有40%，本論文所提出的方法則可大幅改良辨識率到80%左右。將人臉辨識系統應用在監控系統時，傳統的演算法平均約只有25%辨識率，而本論文所提出的方法則仍有約70%的辨識率。

In this dissertation, an improved face pose estimation algorithm to increase face recognition rate is proposed. There are three main stages. The first stage is chin curve estimation by using active contour model. The active contour model is auto-initialized to approach the chin curve under various face poses according to statistical experimental results. The second stage is the face pose estimation and synthesis. Using the chin contour information along with other facial features, simulated annealing algorithm is adopted to estimate various face poses. Using the face pose information, input face image with arbitrary face poses can be synthesized to be frontal. The last stage is face recognition. The synthesized frontal face pose image is utilized to solve the problem that face recognition rate is dramatically reduced when non-frontal face pose images are presented. From experimental results, it can be seen that the face recognition rates of traditional algorithms are only about 40%, while the proposed method greatly improves the recognition rate to about 80%. When face recognition system is applied in surveillance system, the recognition rates are 23% and 70% for traditional algorithm and the proposed system respectively.