| Summary: | 碩士 === 國立成功大學 === 資訊工程研究所 === 89 === In an automatic face recognition system , feature extraction , discriminant analysis and decision rule are three crucial research issues which considerably affect the recognition performance. This thesis deals with three issues simultaneously and proposes a new hybrid approach to achieve an effective and efficient recognition system. For the issue of feature extraction , we apply the two dimensional wavelet transform for dimensionality reduction of face image. This method is able to overcome the drawback in traditional face features. Second , using the linear discriminant analysis (LDA) , we could transform the features into a new space with better separability. Finally , we employ the nearest feature line (NFL) to determine the most likely person. With the combination of three schemes , we can construct a robust and high-accuracy face recognition system.
In the experiments , we conduct a series of evaluation on a face database consisting of 128 persons. Compared to conventional eigenface method , we demonstrate the superiority of using wavelet face , LDA and NFL. Using eigenface , the recognition rate of 91.2% is obtained. However, the proposed approach can achieve the recognition rate of 98.5%
1.1研究動機
1.2人臉辨識系統流程
1.3本論文系統架構
1.4論文架構
第二章 人臉特徵參數擷取
2.1傳統特徵參數擷取
2.1.1 特徵臉
2.1.2 類神經網路(Neural Network)
2.2 離散小波轉換
2.2.1 多層解析空間
2.2.2 二維離散小波轉換
第三章 模型訓練及辨識
3.1 線性鑑別式分析(Linear Discriminant Analysis，LDA)
3.2決策法則
3.2.1歐氏距離(Euclidean distance)
3.2.2最接近特徵線(Nearest Feature Line，NFL)
3.2.3最接近特徵平面(Nearest Feature Plane)
3.2.4 特徵空間(Feature space)
第四章 實驗結果
4.1人臉資料庫
4.2 特徵參數之比較
4.3線性鑑別式分析之影響
4.4決策法則之比較
4.5線上調整技術
第五章 即時展示系統
第六章 結論與未來研究方向
6.1結論
6.2未來發展
參考文獻
附錄一、中研院人臉資料庫人臉影像
附錄二、ORL人臉資料庫人臉影像
附錄三、資料庫影像及其行向量分解、列向量分解圖
|
|---|
