自動化形變與邊緣偵測於人臉辨識

Abstract

人臉辨識是近年來很受矚目的議題。主向量分析是眾多成功的人臉辨識方法之一，但是當有明顯的光線與姿態變化時，主向量分析的辨識率不夠準確。許多相關研究提出利用多張訓練樣本來解決光線與姿態變化的問題。本論文提出一套以主向量分析方法為基礎的人臉辨識系統，稱作自動化形變與邊緣偵測辨識器，讓我們可以在有光線以及姿勢變化的情況下，只用一張訓練樣本就可以更正確地辨識人臉的身份。我們的構想是重新描繪一個不受人臉姿勢影響的標準參考模型，並使用不易受光線變化影響的人臉邊緣影像，來作人臉的辨識。我們使用ORL資料庫來驗證系統的有效性。實驗結果證實，只使用一張訓練樣本，利用主向量分析方法，此自動化形變與邊緣偵測辨識器的確可以在有光線與姿勢變化的情況下，得到更佳的辨識效果。

Face recognition has received much attention during the past several years. Principal component analysis (PCA) is one of the most successful methods for face recognition but it is not highly accurate when the illumination and pose of the facial images vary considerably. Many researches have discussed some solutions to solve the illumination and pose problems, but most of them need multiple training images. This paper presents a novel face recognition system based on PCA, named Automatic Pose normalization and Edge map face Recognizer (APER). The idea is to automatically re-render a pose invariant reference model to accommodate varying pose of the images. Face edge images, which are insensitive to illumination changes, are incorporated. The APER requires only a single face image for training per person. The APER and the PCA method are evaluated using ORL database. The experimental results demonstrate that the APER can improve the performance of conventional PCA approach under varying pose and illumination with single training image.