Document Type
Dissertation
Date of Award
Spring 5-31-2006
Degree Name
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Department
Computer Science
First Advisor
Chengjun Liu
Second Advisor
Joseph Y-T. Leung
Third Advisor
James Geller
Fourth Advisor
Andrew Sohn
Fifth Advisor
Roy R. You
Abstract
Biometric authentication systems automatically identify or verify individuals using physiological (e.g., face, fingerprint, hand geometry, retina scan) or behavioral (e.g., speaking pattern, signature, keystroke dynamics) characteristics. Among these biometrics, facial patterns have the major advantage of being the least intrusive. Automatic face recognition systems thus have great potential in a wide spectrum of application areas. Focusing on facial analysis, this dissertation presents a face detection method and numerous feature extraction methods for face recognition.
Concerning face detection, a video-based frontal face detection method has been developed using motion analysis and color information to derive field of interests, and distribution-based distance (DBD) and support vector machine (SVM) for classification. When applied to 92 still images (containing 282 faces), this method achieves 98.2% face detection rate with two false detections, a performance comparable to the state-of-the-art face detection methods; when applied to videQ streams, this method detects faces reliably and efficiently.
Regarding face recognition, extensive assessments of face recognition performance in twelve color spaces have been performed, and a color feature extraction method defined by color component images across different color spaces is shown to help improve the baseline performance of the Face Recognition Grand Challenge (FRGC) problems. The experimental results show that some color configurations, such as YV in the YUV color space and YJ in the YIQ color space, help improve face recognition performance. Based on these improved results, a novel feature extraction method implementing genetic algorithms (GAs) and the Fisher linear discriminant (FLD) is designed to derive the optimal discriminating features that lead to an effective image representation for face recognition.
This method noticeably improves FRGC ver1.0 Experiment 4 baseline recognition rate from 37% to 73%, and significantly elevates FRGC xxxx Experiment 4 baseline verification rate from 12% to 69%. Finally, four two-dimensional (2D) convolution filters are derived for feature extraction, and a 2D+3D face recognition system implementing both 2D and 3D imaging modalities is designed to address the FRGC problems. This method improves FRGC ver2.0 Experiment 3 baseline performance from 54% to 72%.
Recommended Citation
Shih, Peichung, "Facial analysis in video : detection and recognition" (2006). Dissertations. 778.
https://digitalcommons.njit.edu/dissertations/778
