采用新型纹理特征的2DLDA人脸识别算法

针对现有基于纹理特征的人脸识别算法中纹理特征维数偏大且对噪声较敏感等不足，提出了用于描述人脸图像大尺度局部特征的中心四点二元模式（Center Quad Binary Pattern, C-QBP）和用于描述图像小尺度局部特征的简化四点二元模式（Simplified Quad Binary Pattern, S-QBP）两种互补的新型纹理特征。在此基础上，实现基于新型纹理特征的2DLDA人脸识别算法。首先对人脸图像进行多级分割，再对所产生的图像块提取C-QBP和S-QBP纹理特征，构建纹理特征矩阵。最后，采用2DLDA子空间学习算法实现基于新型纹理特征的人脸识别。实验结果表明，本文所提出的人脸识别算法的识别率明显高于其他基于纹理特征和子空间学习的人脸识别算法。当每一类训练样本数统一设置为5，特征维数为48×4时，在ORL人脸库上，本文所提出的人脸识别算法的识别率达98.68%；在YALE人脸库上，特征维数为48×36时，识别率达99.42%；在FERET人脸库上，特征维数为48×26时，识别率为91.73%。 

2DLDA face recognition algorithm using novel texture features

Since the existing texture features-based face recognition methods are suffered from large texture feature dimensions and noises, two novel complementary texture features, named center quad binary pattern (C-QBP) and simplified quad binary pattern (S-QBP), are proposed. Based on the proposed C-QBP and S-QBP, the two dimensional linear discriminant analysis (2DLDA) subspace learning algorithm is further employed to realize face recognition. More specifically, a multi-level block division method is firstly performed on the input image to produce multiple image blocks. Then, the C-QBP and S-QBP feature histograms are extracted from each image block for establishing the texture matrix of the input image. Finally, the 2DLDA subspace learning algorithm is applied to find an optimal texture subspace for face recognition. Experimental results have shown that the proposed face recognition method is superior to the state-of-the-art texture feature and subspace learning based face recognition methods. Specifically, when each training class holds 5 images, the face recognition rate of the proposed approach is 98.68% on the ORL database with a 48×4 feature dimension, 99.42% on the YALE database with a 48×36 feature dimension, and 91.73% on the FERET database with a 48×26 feature dimension.