Date of Award

Fall 2010

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer Science

First Advisor

Chengjun Liu

Second Advisor

James Geller

Third Advisor

Andrew Sohn

Fourth Advisor

Usman W. Roshan

Fifth Advisor

Xiaoguang Lu

Abstract

Face recognition as a particular problem of pattern recognition has been attracting substantial attention from researchers in computer vision, pattern recognition, and machine learning. The recent Face Recognition Grand Challenge (FRGC) program reveals that uncontrolled illumination conditions pose grand challenges to face recognition performance. Most of the existing face recognition methods use gray-scale face images, which have been shown insufficient to tackle these challenges. To overcome this challenging problem in face recognition, this dissertation applies multiple features derived from the color images instead of the intensity images only.

First, this dissertation presents two face recognition methods, which operate in different color spaces, using frequency features by means of Discrete Fourier Transform (DFT) and spatial features by means of Local Binary Patterns (LBP), respectively. The DFT frequency domain consists of the real part, the imaginary part, the magnitude, and the phase components, which provide the different interpretations of the input face images. The advantage of LBP in face recognition is attributed to its robustness in terms of intensity-level monotonic transformation, as well as its operation in the various scale image spaces. By fusing the frequency components or the multi-resolution LBP histograms, the complementary feature sets can be generated to enhance the capability of facial texture description. This dissertation thus uses the fused DFT and LBP features in two hybrid color spaces, the RIQ and the VIQ color spaces, respectively, for improving face recognition performance.

Second, a method that extracts multiple features in the CID color space is presented for face recognition. As different color component images in the CID color space display different characteristics, three different image encoding methods, namely, the patch-based Gabor image representation, the multi-resolution LBP feature fusion, and the DCT-based multiple face encodings, are presented to effectively extract features from the component images for enhancing pattern recognition performance. To further improve classification performance, the similarity scores due to the three color component images are fused for the final decision making.

Finally, a novel image representation is also discussed in this dissertation. Unlike a traditional intensity image that is directly derived from a linear combination of the R, G, and B color components, the novel image representation adapted to class separability is generated through a PCA plus FLD learning framework from the hybrid color space instead of the RGB color space. Based upon the novel image representation, a multiple feature fusion method is proposed to address the problem of face recognition under the severe illumination conditions.

The aforementioned methods have been evaluated using two large-scale databases, namely, the Face Recognition Grand Challenge (FRGC) version 2 database and the FERET face database. Experimental results have shown that the proposed methods improve face recognition performance upon the traditional methods using the intensity images by large margins and outperform some state-of-the-art methods.

Share

COinS