基于拉普拉斯金字塔的Gabor特征人脸识别算法  

Face recognition with Gabor feature based on Laplacian pyramid

作　　者：吴定雄 景小平[1,2] 张力戈 王文彬 

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学,北京101400

出　　处：《计算机应用》2017年第A02期163-166,178,共5页journal of Computer Applications

基　　金：四川省科技支撑计划项目(2014FZ0046)

摘　　要：针对Gabor特征缺乏对人脸图像的全局描述的问题,提出了一种基于拉普拉斯(Laplacian)金字塔的Gabor特征人脸识别算法。该算法同时捕获了人脸图像的局部特征和全局特征信息,提高了在复杂多变环境下的人脸图像识别性能。首先,通过多分辨率分析建立人脸图像高斯金字塔;其次,通过对高斯金字塔相邻层图像相减构建具有多分辨率分析能力的人脸图像Laplacian金字塔;用一组Gabor滤波器对Laplacian金字塔的每层图像进行卷积,提取金字塔中每层图像的Gabor特征谱;最后,对Laplacian金字塔的每层特征谱进行分块处理,并将经过处理后的特征连接起来作为整个人脸图像的特征向量来实现人脸的分类识别。在ORL人脸库上进行仿真实验,在不同的训练样例数下,所提算法的识别率均高于几种流行的识别算法,如当每个人的样例个数为2和3时,该算法的识别率为90.3%和97.4%,比传统的Gabor特征Fisher判别分类算法(GFC)提高了2.5%和1.9%。实验结果表明,所提算法具有较强的人脸特征表达和识别能力,并且对人脸的姿态变化、表情变化以及光照强度变化具有很好的鲁棒性。Focusing on the issue that the Gabor feature lacks global description of tface images, a face recognition algorithm with Gabor feature based on Laplacian pyramid was proposed. This algorithm captures both local and global feature information of face images, and improves the performance of face recognition in complex and changeable environments.Firstly, the face image Gaussian pyramid was constructed by multi-resolution analysis. Then, the Laplacian pyramid with multi-resolution analysis capability was constructed by subtracting the adjacent level image of the Gaussian pyramid. The Gabor feature maps were extracted by convolution of image in each level of the Laplacian pyramid with a set of Gabor filters.Finally, the feature maps in each level of Laplacian pyramid were respectively separated into several blocks from which the Gabor descriptors are built and flatten into an advanced feature vector to be used as the whole face feature representation later.In the face recognition experiments carried out on publicly available face image databases, such as ORL, the recognition rate of the proposed algorithm is higher than that of several popular recognition algorithms under different training samples. When the number of samples per person is 2 and 3, the recognition rate of the proposed algorithm is 90. 3% and 97. 4%, which is improved by 2. 5% and 1. 9%, compared with the traditional Gabor-Fisher Classifier algorithm（ GFC）, the results show that the proposed algorithm is of highly discriminable ability in face feature representation and is robust to face posture, face expression and illumination variations.

关 键 词：人脸识别 Laplacian金字塔 GABOR特征 卷积 多分辨率分析 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]