An integration framework of feature selection and extraction for appearance-based recognition.

详细信息   

• 作者：Li ; Qi.
• 学历：Doctor
• 年：2006
• 导师：Kambhamettu, Chandra
• 毕业院校：University of Delaware
• 专业：Artificial Intelligence.;Computer Science.
• ISBN：9780542719714
• CBH：3220745
• Country：USA
• 语种：English
• FileSize：8785217
• Pages：142

文摘

Appearances of objects reside in high-dimensional spaces. Many dimensions (pixels or attributes) are not important or even irrelevant to a given recognition task. Reducing the dimension of appearances thus helps improve not only the recognition accuracy but also efficiency. Feature selection and feature extraction are two basic strategies to reduce the dimension of appearances.;Feature selection/detection aims to select pixels most relevant to a given classification task from original appearances. A feature selection method usually contains two components: (i) assigning interest strength to each pixel; and (ii) reducing the redundancy of pixels. A popular method for redundancy reduction is non-maximum suppression that forces interest pixels to distribute uniformly in the entire image plane. This might not be desirable for weakly-textured images such as face images. We propose an imbalance oriented scheme that chooses image pixels whose zero-/first-order intensities can be clustered into two imbalanced classes (in size), as candidates. The strength assignment used in previous interest pixel detectors is usually based on spatial information only. We propose a strength assignment scheme integrating spatial and discriminant information, with the motivation that strong spatial information can be helpful in improving the robustness of the discriminant strength estimation, e.g., in undersampled training scenarios. We use wavelet regularity to represent a pixel.;In traditional appearance based face recognition, a popular scheme is to represent a face instance by its global appearance that is in a high-dimensional space. By feature extraction, it can be effectively encoded by a low-dimensional vector, which can reduce the recognition cost significantly. In a local recognition scheme, an image is represented by a set of repeatable local patches, obtained by a feature selection method. In contrast to the global scheme, the local scheme is much stronger in tolerating localization errors and outlier regions, while much less efficient. We propose a framework for adaptive appearance based face recognition that aims to integrate the robust query of the local scheme (based on feature selection) and the efficient query of the global scheme (based on feature extraction).;We presented comprehensive experimental studies, including (i) the evaluation of repeatability of interest pxiel detectors across rotations and illuminations that shows the superiority of imbalance redundancy reduction over non-maximum suppression; (ii) embryo/face image classification that shows the value of integrated strength based feature selection; (iii) face recognition under constrained localizations and occlusions that shows the advantage of adaptive appearance based recognition framework; and (iv) facial representability evaluation that shows the value of feature distribution.