Abstract: To overcome the limitations of traditional face recognition algorithm under the single training sample, a fusion of local direction gradient amplitude and phase difference is proposed. The method firstly extracts the gradient amplitude and phase of the image. 8 edge gradient values of each sub-neighborhood are gained by convolving the gradient amplitude with 8 Kirsch masks respectively. And secondly the local difference on the phase is extracted. Then, the LDGMPD just utilizes the direction of the largest value of the edge gradient and local difference, these two directions are encoded into a double digit octal number to produce the LDGMPD code values, it selects the structure contrast information to carry out the weighted processing on face image sub-block of each LDGMPD and extracts the LDGMPD histogram features of face image. Finally, nearest neighbor classifier is used to classify the faces. Experimental results on AR and CAS-PEAL-R1 face databases validate that LDGMPD algorithm is an outstanding method for single sample face recognition.

Key words: face recognition, single training sample, difference on phase, edge gradient, nearest neighbor classifier

摘要： 针对传统人脸识别算法在单训练样本下效果不佳，提出一种局部方向梯度幅值和相位差分相结合的方法（LDGMPD），首先提取图像的梯度幅值与相位，梯度幅值图像与8个Kirsch模板卷积得到每个子邻域的8个边缘梯度值；然后对相位进行局部差分。局部方向梯度幅值与相位差分仅使用边缘梯度值与相位局部差分值中最大值的方向编码成一个二位八进制数，产生LDGMPD值。再选取结构对比信息对各LDGMPD人脸分块进行加权处理，提取人脸的LDGMPD直方图特征，最后利用最近邻分类器分类识别。在AR和CAS-PEAL-R1共享库上进行实验表明LDGMPD在单样本人脸识别具有较好的效果。

关键词: 人脸识别, 单样本, 相位差分, 边缘梯度, 最近邻分类器