项目名称: 基于图像配准与表示联合优化的自动人脸识别研究
项目编号: No.61471048
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 无线电电子学、电信技术
项目作者: 胡佳妮
作者单位: 北京邮电大学
项目金额: 82万元
中文摘要: 实用的自动人脸识别系统必须有准确的人脸检测与配准作为前端,便于后续进行稳定可靠的人脸特征表示,实现高精度的人脸匹配。以往的研究工作大多在图像配准或图像表示领域独立进行,使得现有方法难以组成稳定可靠的自动识别系统。本项目以申请人最新提出的变换不变性主成分分析(TIPCA)模型为基础,采用图像配准与表示联合优化的技术路线,利用PCA技术把前端的人脸配准与后端的表示学习结合在一起。以典型的人脸检测结果为研究对象,本项目主要研究:(1)面向遮挡或多姿态人脸图像的配准与表示联合优化方法,提高TIPCA模型对遮挡和多姿态人脸的配准鲁棒性;(2)扩展线性表示模型及其迁移学习算法,提高模型在小样本条件下的表示与识别的泛化能力;(3)在检测与配准过程中间引入预配准算法,提高误检测图像自动配准的稳定性。本项目的研究成果将有助于解决小样本条件下的遮挡和多姿态的全自动人脸识别难题,具有较大的应用价值。
中文关键词: 人脸识别;图像配准;图像表示;生物特征识别;计算机视觉
英文摘要: Practical recognition system must have accurate face detection and feature localization at the front end, in which the main purpose of feature location is face image registration. Therefore, different images of the same face are more compact, easy to extract reliable features. Most previous works mostly conduct independent study on feature localization, image registration or image representation. This proposal is based on the newly proposed transform-invariant principal component analysis (TIPCA) model by the applicant and collaborators, assumes that registered image can be efficiently linear representation, integrates the image registration and image representation stages together, and finally conducts research on fully automated face recognition algorithms. The research project introduces occlusion detection and posture correction technology to improve the robustness of TIPCA model against face occlusion and pose changes, and introduce an extended linear representation model and its transfer learning algorithms to improve the generalization ability on representation and recognition under small training sample size conditions. The research outputs of this project will help to solve the problem occlusion and pose-robust automatic face recognition problem with small-size training samples, and have a high application value.
英文关键词: Face Recognition;Image Registration;Image Representation;Biometrics;Computer Vision