Abstract: Affected by changing illumination, expressions and posture of the face image to the performance of Face Recognition(FR), this paper presents a FR method based on data fusion. The proposed method transforms face image with 2Dimensional-Discrete Wavelet Transform(2D-DWT) into three layers and each image consists of one low frequency sub-image and nine high frequency sub-images. It considers that the low frequency sub-image can directly represent the essence of the face while part of the high frequency sub-images still contain discriminative information, and selects sub-images including abundant of human face information under Fisher projection. It designs three data fusion methods correspond to the pixel level, feature level and decision level. Experimental results on ORL and YALE A face databases show that the proposed method is efficient and its accuracy rate is better than Principal Component Analysis(PCA) method and wavelet transform face recognition method.

Key words: Face Recognition(FR), data fusion, wavelet decomposition, Fisher projection, subspace analysis, feature extraction

摘要： 考虑人脸表情、光照变化和姿态对人脸识别性能的影响，提出一种基于数据融合的人脸识别方法。应用二维离散小波变换对人脸图像进行3次小波分解，使每幅人脸图像得到1幅低频子图和9幅高频子图，低频子图可以直接代表人脸的本质，而部分高频子图仍含有鉴别信息，因此，利用Fisher投影对得到的高频子图进行投影，选取出包含鉴别信息较多的高频子图，并设计 3种数据融合方案。分别在数据级、特征级和决策级实现融合处理，并在ORL和YALE A人脸库上完成实验，结果表明，与主成分分析法和小波变换人脸识别方法相比，该方法能有效提高识别率。

关键词: 人脸识别, 数据融合, 小波分解, Fisher投影, 子空间分析, 特征提取

CLC Number: 

• TP18