Selection of Best Super-Resolution Subspace in Face and Automatic Target Recognitions: A Class of Directional Principal Correlation Analysis Approach
หัวหน้าโครงการ
:
วิทยากร อัศดรวิเศษ
ทีมวิจัย
:
วิทยากร อัศดรวิเศษ
หัวหน้าโครงการ
สมชาย จิตะพันธ์กุล
นักวิจัยที่ปรึกษา
วันที่เริ่มโครงการ
:
2 ก.ค. 2550
วัตถุประสงค์
:
The aim of this project is to theoretically and numerically study the performance improvement of Principal Correlation Analysis (PCorrA) by employing variable selection and super-resolution techniques to selection best directional PCorrA and at the same time improve the resolution at its eigenspace-domain. This study will mainly focus on face and automatic target recognition applications.
In this study, various selection criteria will be used as for choosing the best 2D cross-covariance matrix. This selected 2D cross-covariance matrix will be used for singular value decomposition (SVD) to find a set of uncorrelated coordinates. The principal component vectors (PCV), uncorrelated space, are then used for feature extraction for face and automatic target recognition.
There are two techniques used in super-resolution framework for face recognition. In the first technique, each high-resolution image is constructed from a set of low-resolution images and then all high-resolution image ensembles are used as training input in finding PCVs. Secondly, a set of super-resolution PCVs (SR-PCVs) will be directly constructed at the eigenspace-domain space for each 2D cross-covariance matrix. As a result, we will have N set of SR-PCVs for N cross-covariance matrices, where N is the number of possible cross-covariance patterns. Next, only one set of SR-PCVs will be selected from the N sets of PCVs, and used for feature extraction. At this point, two main issues will be studied and is considered as a half of this research proposal.
Another issue that will be included in this study is that better directional selectivity spaces of image will be proposed to build new 2D cross-covariance matrices. Not only that, but we will also investigate on applying directional selectivity spaces to the transformed 1D cross-covariance in order to create multiresolution directional PCA.. Thus, we address here two approaches that can be used to construct directional PCA. These investigations will also be performed for super-resolution improvement at eigenspace-domain.
