基于曲波系数加权融合的人脸识别

Face recognition based on weighted fusion of curvelet coefficients

为了提高人脸识别率和缩短识别时间,研究了基于曲波变换的人脸识别技术。考虑到传统曲波变换无法将多尺度多方向的曲波特征进行最优表示且其特征维数过大的缺点,提出了一种基于自适应加权融合的曲波变换和独立分量分析(ICA)的人脸识别算法。该算法通过曲波变换提取原始人脸图像的最优尺度和方向上的曲波系数,并对这些特征系数进行多方向上的融合,根据类别可分离性的判据原则对融合后的系数进行加权,以减少特征数量,提高处理速度;通过独立分量分析降维,将这些特征投影到更具表达力的空间,以获取有效特征,减少冗余信息,便于最近邻分类器进行人脸识别。基于在奥利维帝研究实验室(ORL)人脸库、Yale B人脸库和AR人脸库对该算法进行了测试,结果表明,其识别率分别达到98%、97%和98.57%,单幅图片的识别时间分别为65.43,158.94和20.37ms,从而验证了其实用性。

The curvelet transform based face recognition was studied to raise the recognition rate and shorten the recognition time. Considering that the traditional curvelet transform is unable to optimally represent multi-scale, multi-direction curvelet features and its feature dimension is too high, the study put forward a novel face recognition algorithm using the curvelet transform based on adaptive weighted fusion and the independent component analysis （ICA）. The algorithm extracts original face images＇ optimal Curvelet coefficients in multi-scale and multi-direction by Curvelet transform and fuses them along multi-direction, and weights the fused coefficients according to the class separability principle to reduce the feature quantity and raise the processing speed. It also uses ICA to realize di- mension reducing to project the features to more effective space to decrease ineffective and redundant information, so these preproeessed face features can be classified and recognized by tile nearest neighbor classifier, thus the face recognition can be achieved. The proposed algorithm was tested by using the face databases of ORL, Yale B and AR, and the recognition rate of 98%, 97% and 98.57%, as well as the corresponding recognition time of 65.43, 158.94 and 20.37ms on single image, were obtained respectively, showing its higher performance and practicality.