摘要： 根据姿态与表情变化对人脸识别的影响，采用对图像的旋转、尺度变化保持不变性的SIFT算子作为人脸特征，建立人脸各个子区域的相似性测度，并通过混合高斯建立不同变形条件下相同样本与不同样本的相似性概率模型。在此基础上，利用各子区域特有的识别能力获取子区域概率权值，结合基于贝叶斯公式建立的概率框架确定识别结果。实验结果表明，与直接用SIFT算子进行人脸识别的方法相比，该方法在姿态变化较大及表情变化较大的情况下识别率有明显提高。

关键词: 人脸识别, 尺度不变特征变换描述子, 贝叶斯概率模型, 姿态, 表情, 子区域

Abstract: To handle the influences brought by the change of pose and expression, Scale Invariant Feature Transform(SIFT) descriptors, which is rotating and scale invariant, is applied to measure the similarity between corresponding sub-regions of two faces, and the probabilistic similarity models of the same or different faces under various deformations are built with Gaussian Mixture Model(GMM). Then, a probabilistic frame which is based on Bayesian formula is established to get the recognition results, combining with the weight of each sub-region which is decided by their peculiarities. Experimental results indicate that the proposed method outperforms the traditional SIFT-based method when the variation of the pose or expression is large.

Key words: face recognition, Scale Invariant Feature Transform(SIFT) descriptor, Bayesian probabilistic model, pose, expression, sub-region

中图分类号: 

• TP18