基于非线性特征和线性映射的多视角低分辨人脸识别算法

摘要

A regression-based method is proposed to deal with real world face recognition with difference of image resolution,pose variation and only one gallery image per person.The regression models from the specific non-frontal low resolution images to ker-nel principal component analysis (KPCA)features of the corresponding frontal high resolution images are learnt by procrustes a-nalysis.Then the frontal high resolution KPCA features of the corresponding nonfrontal low resolution test facial image are esti-mated by the learnt regression models.The estimated features are fed to nearest neighbor classification to get the identity.As ex-periments on benchmark database shown,the face recognition rates of the proposed method are 4%-36% higher than those of lin-ear based comparison methods,while the online test time of the proposed method is only 1.087 ms slower than that of the fastest comparison method.%针对分辨率变化、视角变化和认证集单样本等实际条件下的人脸识别问题，提出了一种基于回归的人脸识别算法。该算法采用核主成分分析法（kernel principal component analysis）分别提取侧面低分辨率和正面高分辨率人脸特征，利用 Procrus-tes 分析建立每一种侧面视角低分辨率 KPCA 特征和正面高分辨率 KPCA 特征间的映射关系，从而获得对应的回归模型。根据这些回归模型，即可得到测试侧面低分辨率人脸对应的正面高分辨率 KPCA 特征，并通过最近邻分类器进行识别。在标准图库上的实验表明，与基于线性模型的人脸识别对比算法相比，本文所提算法识别率提高了4％至36％，而在线测试时间仅比最快的对比算法多1．087 ms。