项目名称: 真实光照环境下人脸识别关键问题的研究
项目编号: No.60873092
项目类型: 面上项目
立项/批准年度: 2009
项目学科: 电工技术
项目作者: 房斌
作者单位: 重庆大学
项目金额: 30万元
中文摘要: 在真实光照环境下获得的人脸图像往往比实验室模拟光照条件下的人脸具有更大的可变性与不可控制性(如未知光源的自然光、物体的阴影等)。真实光照的影响削弱了人脸关键特征的分辨能力,光照问题的解决已成为人脸识别研究和应用面临的一大难题。本项目将重点研究真实光照环境下人脸识别的关键技术,其目的是进一步提高人脸识别系统的精度和鲁棒性。本课题的主要研究思路是在人脸图像全局空间的变换域中提取光照不变量特征,以克服传统方法直接在像素域下提取全局不变量特征的不足;结合流形学习与非负矩阵因子分解的优点,在非负矩阵因子分解算法中以流形学习中的距离机制代替原来的欧氏距离度量,研究非负矩阵因子分解新算法,以抽取具有流形结构的人脸局部子模式;分析符合流形结构的局部特征相似度度量准则,改变欧氏距离度量没有考虑人脸子模式各分量相关性的缺点。本项目的开展不仅能促进模式识别理论和算法的研究,还有望为实际应用提供技术支撑。
中文关键词: 人脸识别;真实光照;光照不变性特征;流形结构;相似度度量
英文摘要: Human face images captured under varying natural lighting conditions render morevariability within face patterns than those obtained in lab with simulated lighting suchas unknown lighting sources and shadow of objects. This problem makes it difficulty toface recognition because key features of faces are covered and unidentified. Thisproject aims at finding novel approaches to tackle the natural lighting problem in orderto improve recognition accuracy and robustness. The main ideas focus on extractingillumination invariant of face images in transformed space other than in pixel domain,and combining manifold learning and non-negative matrix factorization to reveal latentmanifold structure of face patterns. In addition, reasonable distance measure other thanEuclidean distance measure will be investigated for appropriate manifold analysis oflocal feature similarity. The implementation of this project will not only prompt basicmethodology and algorithm study of pattern recognition in the special research field of face recognition, and provide technical support for real application as well.
英文关键词: Face Recognition; natural lighting; illumination invariant; manifold structure; similarity measure