基于紧致全姿态二值SIFT的人脸识别

Compact complete pose binary SIFT for face recognition with pose variation

姿态变化是人脸识别中的关键问题之一,全姿态二值SIFT(CPBS)提取等间隔采样姿态人脸图像的二值化SIFT特征,并用于任意姿态人脸识别,获得了良好的识别性能。但是,CPBS特征量很大,计算成本很高。提出了紧致全姿态二值SIFT(CCPBS)的人脸识别方法。选取间隔45°的人脸图像作为训练集,首先提取ASIFT特征进行融合。然后用基于稀疏表示的方法进行特征选择,有效地滤除相似或相同的特征,减少数据冗余。进一步对选择的特征进行二值化,即可得到CCPBS。人脸识别通过计算待识别人脸和CCPBS之间的汉明距离来完成。在CMU-PIE和FERET人脸库上实验结果表明,提出的算法无需人脸对齐和标记,即可以取得很高的正确识别率,明显优于其他算法。与CPBS相比,识别率仅降低很少的同时,特征量降低了22.11%和32.63%。

Pose variation is one of challenging problems in face recognition. Complete pose binary SIFT（CPBS） was proposed to extract binary ASIFT from face images of five poses, which is utilized for face recognition and demonstrates good performance. However, CPBS has a large data and requires high computational cost. Here, the compact complete pose binary SIFT （CCPBS） is presented to address the issue. Five face images with poses of frontal view, rotation left/right 45 and 90 degree respectively are selected as gallery images of a subject. Firstly, the ASIFT descriptors of these image are pooled together. Then the algorithm based on sparse representation are pro- posed to fiher out the ASIFT descriptors with similar characteristics. After that, the binary ASIFT descriptors are extracted and the CCPBS can be obtained. Face recognition is finished by hamming distance between the probe face image and the CCPBS. Compared ex- periments are carried out on the CMU - PIE face databases and FERET face databases. Experimental results show that our approach can obtain higher recognition ratio without face alignment or landmark fitting, which is much better than state-of-the-art algorithms. Com- pared with CPBS, the recognition ratio is reduced slightly with the reduced data of 22.11% and 17.0% .