Notice

Topic: Enhancement of ICA Representation for Face Recognition

Doctor Student:  Jiajin Lei

 

Advisors:  Dr. Chao Lu

                  (Department of Computer and Information Sciences, Towson University)

                  Dr. Victor Chen
                 (Naval Research Lab)

 

Time:  Dec. 1, Tuesday 3-4 pm

Location: COSC dept Conference room YR459

 

ABSTRACT

Face recognition found a large number of applications. Independent Component Analysis (ICA) has successfully been used in this area. Compared to other “traditional” methods in face recognition, ICA is relatively new, and its many aspects need to be further investigated. In this dissertation, we explored the methods to enhance ICA face representation for improvement of face recognition performance, mainly by using spatiotemporal ICA (stICA). stICA joints face information from both space domain and time domain. This strategy leads to enhancement of ICA representation for face recognition with respect to spatial ICA (sICA) where only still face information is used.

 

The first part of the dissertation briefly reviews and outlines the status of face recognition and its main algorithms. Although face recognition includes broad spectrum of algorithms, here we mainly focused on statistic methods, such as PCA (Principal Component Analysis), ICA, and LDA (Linear Discriminant Analysis). Especially the origin, purpose, definition and principle of ICA were detailed described.

 

The second part of the dissertation deals with spatial ICA and spatiotemporal ICA for face recognition. By using one well known face database and ourselves-built face database we investigated the intrinsic difference and performances of ICA architecture I and ICA architecture II, showed the advantages of ICA over PCA in face recognition. Result indicates while ICA architecture Ι focuses on face local feature, ICA architecture II is more sensitive to face global feature. In the mean time we also explored the methods for ICA feature enhancement by importing algorithms of ICA Subspace Selection and Sequential Forward Floating Selection. In this part model of spatiotemporal ICA was set up and explained. Then the advantage of stICA over sICA was experimentally justified.

 

In the following part, an algorithm for locating special facial components was reported. With the availability of this algorithm, we investigated localized spatiotemporal ICA representation. By localized, we mean faces are patched for special facial portions. Localized spatiotemporal ICA representation can extract unique face features. Although ICA derives many face representations, no one alone fully satisfies our final goal. To further enhance the performance, we proposed Optimizing-Selection-Fusion method. This method selects most discriminant features from different ICA representations, meanwhile the numbers of features from different feature sets are optimized. Our method greatly improves the recognition performance with recognition rate being as high as 94.62%.