融合Monogenic幅值和相位的人脸识别方法

Face recognition method fusing Monogenic magnitude and phase

针对仅利用图像滤波幅值信息进行识别而忽视相位信息的问题,提出一种融合Monogenic局部相位和幅值的识别方法。该方法先对相位进行量化和异或,并结合方向和尺度信息得到相位编码(MLXP);其次,分别对相位编码和基于幅值的二值编码进行分块,计算直方图特征;然后,采用基于分块的线性判别进行降维,提高特征的判别能力;最后在评分层实行融合。在ORL和CAS-PEAL人脸数据库上,相位方法 MLXP的平均识别率分别为0.97和0.94,融合Monogenic相位和幅值的方法平均识别率分别为0.99和0.979,超越实验中其他所有方法。实验结果表明,相位利用方法 MLXP是有效的,融合Monogenic相位和幅值的方法不但能够避免传统线性判别中的小样本(3S)问题,而且能以较低的时间和空间复杂度,有效地提高身份的正确识别率。

In order to use the magnitude and phase information of filtered image for face recognition, a new method fusing Monogenic local phase and local magnitude was proposed. Firstly, the authors encoded the phase using the exclusive or （XOR） operator, and combined the orientation and scale information. Then the authors divided the phase pattern maps and binary pattern maps based on magnitude into blocks. After that, they extracted the histograms from blocks. Secondly, they used the block-based Fisher principle to reduce the feature dimension and improve the discrimination ability. At last, the authors fused the cosine similarity of magnitude and phase at score level. The phase method Monogenic Local XOR Pattern （MLXP） reached the recognition rate of 0.97 and 0.94, and the fusing method recognition rate was 0.99 and 0.979 on the ORL and CAS-PEAL face databases respectively and the fusing method outperformed all the other methods used in the experiment. The results verify that the MLXP method is effective. And the method fusing the Monogenic magnitude and phase not only avoids the Small Sample Size （3S） problem in conventional Fisher discrimination methods, but also improves the recognition performance significantly with smaller time and space complexity.