以元件為基礎之人臉辨識

Abstract

本論文的目的是利用人臉對稱與鼻尖凸出於雙眼與嘴的假設，算出偏轉人像之概略的頭部旋轉資訊，並將這一資訊使用在修正不同元件component) 在角度改變時所應修正的混合比例上，來提升辨識正確率。以往利用元件間相對位置的資訊時，會想到算出每個人特有的臉部關係做為辨識內容，而忽略了算出臉部角度的資訊也可以提供辨識時混合元件的資訊。本辨識系統基於不同元件在不同人臉角度下應該使用不同重要性的混合比例，利用以元件為基礎之人臉偵測，獲得雙眼、鼻、嘴等四個元件的位置與四個由灰階值構成的特徵向量。由四個元件位置可算出概略的頭部旋轉資訊。再利用一套兩層式處理的辨識方法，第一層是各自元件利用Support Vector Machine (SVM)訓練出辨識某人相似程度的判斷器，第二層則是利用SVM訓練混合頭部旋轉資訊與某人相似程度函數中的系數。根據實驗結果，在精確偵測元件位置與訓練資料較少的情況下，兩層架構比單層架構得到較好的辨識率。

This paper use hypothesis of that face symmetrical and nose which stick out the eyes and mouth to calculate the information of the biased of the head turn, and use this information to modify the weight of different component. When Using the information of the relation between the components, we would ignore the information that the face turn angle can be used to provide the information. When we use the component based face detection, we can obtain the position of the eyes , nose,and mouth and the four feature vectors which are compose of four gray value. According to the position of four component, we can calculate the information of how angle which head turns. After this, we use the two layers recognition method. The layer one, we use each of components and the support vector machine (SVM) to train the recognized machine which can recognize some level of similarity. The layer two, we use the SVM to train the coefficient of the similar degree function and the information of the mixed turn of the head. According the result of experiment in the precisely detect the position of component and less training information state, two layer architecture has better recognition rate than single layer architecture.