CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleFeature Correlation Filter and Its Application in Face Recognition
Thesis Advisor李子青
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword特征相关滤波器 相关滤波器 人脸识别 Feature Correlation Filter Correlation Filter Face Recognition
Abstract相关滤波器方法已经被广泛地应用在自动目标检测和识别以及生物特征识别等领域中。已有的研究工作主要集中于如何构建描述能力更强的优化目标函数, 并在此基础上设计更加复杂的相关滤波器, 本文旨在从一个不同的视角提供一种新的设计思路。与已有工作不同, 本文将相关滤波的概念从像素空间引入到特征空间, 在大大增加了相关滤波器设计灵活性的同时获得更好的识别能力。本文的主要的工作和贡献有: ( 1 )将相关滤波器的概念拓展到特征空间, 提出一种基于特征的相关滤波器方法, 称为“ 特征相关滤波器”。不同于传统的相关滤波器直接对图像像素值进行操作的方法, 本文的特征相关滤波器在特征空间中实现并在特征空间中进行运用。这样可以通过选用性质优良的特征来应对物体识别中通常遇到的光照姿态变化等的难题。另外, 不同于近年来一些学者在非线性相关滤波器设计方面的工作,本文提出的特征相关滤波器设计方法, 可以完全保留传统相关滤波器的两大优点: 闭合形式解和平移不变性。传统的相关滤波器方法,只是本方法的一种特例。另外, 由于通常特征的维数远远低于图像的维数(像素数), 本方法可以显著地节省系统的存储空间。 ( 2 )在特征相关滤波器的基础上, 针对人脸识别中表情和姿态变化的问题, 提出了基于部件的特征相关滤波器方法。在该方法中,判决结果通过融合多个部件特征相关滤波器的输出得到。同时, 还充分利用相关滤波器可以输出的最佳匹配位置的信息, 引入人脸几何结构变化作为相似度度量一种方法, 提高了识别精度。最后,为了处理局部剧烈变化和遮挡的问题, 本方法使用一种局部相似匹配的机制, 即在进行判决时, 对各个部件的输出进行加权, 使得相似度高的部件获得更高的权重, 增强了算法的鲁棒性。 总得来说, 本文提出了特征相关滤波器这一新的相关滤波器合成和使用框架。并应用在人脸识别这一特殊问题上, 验证了该方法的有效性。
Other AbstractThe correlation filters have been extensively studied and widely used in the areas of automatic target detection/recognition (ATD/ATR) and biometrics. Unlike most of the existing works on correlation filters which try to construct sophisticated correlation filters by defining more complicated optimization object functions with stronger descriptive abilities, in this thesis, we address this issue from another point of view: extend the concept of correlation filter from image pixels into feature spaces, which can significantly improve the flexibility of the design of correlation filters and would possibly achieve better performance. The main contributions of this thesis include following issues: (1) Whereas the conventional correlation filters perform directly on image pixels, we propose a novel method, called “Feature Correlation Filters (FCF)”, by extending the concept of correlation filter into feature spaces. This FCF method can deal with the common problems of illumination and pose variation in object recognition by using appropriate feature representations in correlation feature synthesis. Unlike some recent work on nonlinear correlation filers, FCF completely preserves the two key benefits of conventional correlation filters: shift-invariance, and closed-form solution. The conventional correlation filter is a special case of our proposed feature correlation filter. Moreover, since the size of feature is often much smaller than the size of image, the FCF method can largely reduce the storage requirement in recognition system. (2)To address the problems of expression and pose variation in face recognition, we propose “Part-based Feature Correlation Filter”. In this method, the final decisions are made by fusing the outputs of several part-based feature correlation filters. And we also consider the geometric structure of facial components as another measurement of similarity, which can improve the recognition accuracy. Finally, to deal with severe local distortion and occlusion, we introduce a “partial similarity” mechanism into the fusion process, which gives higher weights to the correlation filters that produce similar outputs, and lower weights to those produce dissimilar outputs. This mechanism could improve the algorithm’s robustness in the cases of severe local distortion and occlusion. In this thesis, we propose a novel method of correlation filter design and synthesis, and apply it in the problem of face recognition to demonstrate the effectiveness and robustness of this method.
Other Identifier200528014628028
Document Type学位论文
Recommended Citation
GB/T 7714
朱向欣. 特征相关滤波器及其在人脸识别中的应用[D]. 中国科学院自动化研究所. 中国科学院研究生院,2008.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20052801462802(1329KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[朱向欣]'s Articles
Baidu academic
Similar articles in Baidu academic
[朱向欣]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[朱向欣]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.