Abstract:

In order to solve the problem that face recognition based on Local Directional Number pattern（LDN） usually only uses gradient information and does not extract enough information, a method called Double Variation and Double Space Local Directional Pattern（DVDSLDP） is proposed. Firstly, this method expands the associated neighborhood information by pixel sampling, and then the relative deviation and absolute deviation are obtained by edge response operator and local forward and backward difference respectively to form double deviation information, which can fully extract the information of the local gradient space. Then the gradient spatial features are cascaded with the grayscale spatial features of the extracted pixels to obtain double spatial features, which are used for pattern coding to get the feature image. Finally, the face feature vector is obtained by weighted cascading the sub-block histograms according to the information entropy, and the nearest neighbor classifier is used to complete the classification. The proposed method is compared with the relevant typical methods, and the results on the ORL, Yale and AR databases show that the feature images with clearer outline and richer texture are obtained by fusing the features of double space. The recognition rate of the DVDSLDP method on the ORL and Yale databases are 99.50% and 94.44%, respectively, especially when there are few training samples, the performance of the proposed method is significantly improved. Meanwhile, in particular, it is worth mentioning that the recognition rate of the proposed method on the AR expression, illumination, occlusion A and occlusion B databases are 99.67%, 100%, 99.33% and 97.33%, respectively, which is significantly higher than other methods, the proposed method shows good robustness.

Key words: double variation, local direction pattern, histogram feature, information entropy weighting, Double Variation and Double Space Local Direction Pattern（DVDSLDP）, face recognition

摘要：

针对局部方向数（Local Directional Number pattern，LDN）类方法的人脸识别通常仅利用梯度信息且信息提取不充分的问题，提出双偏差双空间局部方向模式（Double Variation and Double Space Local Directional Pattern，DVDSLDP）。该方法首先通过像素采样扩大关联邻域信息，再利用边缘响应算子和局部前后向差分获得的相对偏差和绝对偏差以构成双偏差信息，充分挖掘局部梯度空间信息；然后与所提取像素的灰度空间特征级联融合，以获得双空间特征，再进行模式编码得到特征图；最后依据信息熵加权级联各子块直方图获得人脸特征向量，使用最近邻分类器完成分类。针对ORL、Yale、AR人脸库和相关典型方法的对比结果表明：利用双空间特征的融合，获得了轮廓更清晰、纹理更丰富的编码特征图，在ORL和Yale库上分别达到了99.50%、94.44%的识别率，尤其是在训练样本较少时性能提升明显；该方法针对AR库的表情、光照、遮挡A和遮挡B子集分别达到了99.67%、100%、99.33%和97.33%的识别率，明显高于其他方法，表现出良好的鲁棒性。

关键词: 双偏差, 局部方向模式, 直方图特征, 信息熵加权, 双偏差双空间局部方向模式（DVDSLDP）, 人脸识别