最小化总投影误差优化一元回归分类的人脸识别

FACE RECOGNITION WITH URC OPTIMISED BY MINIMISING TOTAL PROJECTION ERROR

针对现有回归分类算法不能很好地考虑总类内投影误差而影响人脸识别鲁棒性的问题,提出最小化总投影误差(TPE)的一元回归分类算法。首先通过各个类投影矩阵计算所有训练数据的类内投影误差矩阵,并且借助特征分解找到一元旋转矩阵;然后利用一元旋转矩阵将每个训练图像向量转换为新的向量空间,并计算出每个类的特定投影矩阵;最后,根据一元旋转子空间中各个类的最小投影误差来完成人脸的识别。在人脸数据库ORL、FERET、扩展YaleB及一个户外人脸数据库上的实验验证了该算法的有效性及鲁棒性。实验结果表明,相比于其他几种先进的回归分类算法,该算法取得了更好的识别性能。

For the problem that existing regression classification algorithms seriously impact the robustness of face recognition due to not well considering totM projection error within classes, we propose a unitary regression classification algorithm which is based on minimising the total projection error （TPE）. First, it calculates the projection error matrix within class for all the training data by projection matrix of each class, and finds unitary rotation matrix with the help of characteristics decomposition. Then, it uses unitary rotation matrix to convert each training image vector to new vector space, and works out the specific projection matrix of each class. Finally, the face recognition is implemented by using the minimum projection error of each class in unitary rotating subspace. The effectiveness and robustness of the proposed algorithm are verified by the experiments on common face databases ORL, FERET, extended YaleB and a wild face database. Experimental results show that the proposed method algorithm achieves better recognition performance than several other advanced regression classification algorithms.