結合費雪特徵臉之非負數稀疏表示法應用於行動裝置之人臉辨識系統

Abstract

本論文基於一般稀疏表示法在人臉辨識所可能產生之過度擬合問題，提出非負數稀疏表示法予以改善，並結合費雪特徵臉（Fisherface）以提升人臉之辨識率。本論文所開發之辨識系統包括四個部分，包括了影像預處理、人臉偵測、特徵擷取與人臉辨識，使用了Viola-Jones人臉偵測器，費雪特徵臉以及非負數稀疏表示法。從實驗結果可知，將此辨識器應用於五種人臉資料庫中，其辨識率可達到95%以上。近年來，應用程式在行動裝置上的發展已經成為趨勢，因此本論文也將此人臉辨識系統植入Samsung Galaxy SII智慧型手機上，用於人臉辨識或身分認證，其速率在一秒內可完成三次人臉影像之辨識，已具即時辨識系統之雛形。

Due to the over-fitting problem caused by sparse representation in face recognition, this thesis proposes the non-negative sparse representation for improvement. In addition, the Fisherface is combined to increase the recognition rate. There are four parts in the proposed face recognition system, image preprocessing, the face detection using Viola-Jones face detector, the Fisherface feature extraction and the face classification based on non-negative sparse representation. From the experimental results applied to five face databases, the proposed classifier achieves an accuracy more than 95%. In recent years, the development of applications on mobile device has become a trend. This thesis also implements the face recognition system into Samsung Galaxy SII smart phone for both face identification and face verification. Most significantly, it completes the face recognition three times a second, which indicates it is indeed a real-time recognition system.