CAJ | 학술논문

提出一种人脸识别方法用于解决姿态变化对识别准确率的影响。首先检测人脸图像的SIFT特征，然后根据SIFT特征计算人脸图像间的多示例距离；基于此多示例距离，用保局投影将人脸图像映射至流形空间，最后在流形空间中采用K近邻方法进行人脸识别。该方法有三个特点：(1)采用SIFT特征减小了未知姿态对识别准确率的影响；(2)通过保局投影将特征变换到流形空间一个点，避免了复杂的SIFT特征匹配策略；(3)借助流形方法滤除高维特征中的噪声。实验结果表明与已有方法相比，在人脸姿态不确定的情况下，该方法能提供较为理想的识别准确率。
제출일충인검식별방법용우해결자태변화대식별준학솔적영향。수선검측인검도상적SIFT특정，연후근거SIFT특정계산인검도상간적다시례거리；기우차다시례거리，용보국투영장인검도상영사지류형공간，최후재류형공간중채용K근린방법진행인검식별。해방법유삼개특점：(1)채용SIFT특정감소료미지자태대식별준학솔적영향；(2)통과보국투영장특정변환도류형공간일개점，피면료복잡적SIFT특정필배책략；(3)차조류형방법려제고유특정중적조성。실험결과표명여이유방법상비，재인검자태불학정적정황하，해방법능제공교위이상적식별준학솔。
This paper present an approach to pose-invariant face recognition. Firstly, SIFT features are detected in each face image and the distance between two faces is determined in a multi-instance manner according to the SIFT features. Based on the distance, we project features into manifold space via Locality Preserving Projection. The recognition is finally done in manifold space through KNN algorithm. Our approach has three advantages: (1) the influence of unknown pose is diminished by using of SIFT features; (2) SIFT features are projected into manifold, which avoid complex strategy for SIFT feature matching; (3) Manifold is helpful for removing unknown noise from high- dimensional features. Experiments show that our approach provides exciting recognition rate under unknown pose compared with existed methods.