Department of Computer
New Jersey Institute of
Newark, NJ 07102
Office: GITC 4306
(Face/Iris Recognition, Color Image Feature Extraction and
Classification, Classifier Fusion)
(Statistical Learning, Kernel Methods, Innovative Kernel
Functions/Models, Similarity Measures)
(Object/Face/Iris/Eye Detection, Motion Analysis and Video
Image and Video
Analysis (Image Search and Retrieval, Image
Category Classification, Color Image Analysis, New Color
Spaces, Gabor Image Representation)
C. Liu: "Face Detection Method and Apparatus", United
States Patent 7,162,076, January 9, 2007.
C. Liu and H. Wechsler: "Feature Based Classification",
States Patent 6,826,300, November 30, 2004.
Pattern Recognition, Machine Learning, and Image
Processing -- We develop new color models,
advanced pattern recognition and machine learning
methods, and fuse them to address large-scale and
grand-challenge problems, such as the face
recognition grand challenge (FRGC) problem and the
Caltech 256 image categories image search and
By fusing our new kernel methods (kernel
Fisher analysis, kernel PCA with fractional power
polynomial models), new color models, new
similarity measures, we achieve the best face
verification performance for the government
organized FRGC competition.
By fusing new color models with popular
image descriptors, such as the scale-invariant
feature transform (SIFT) and the local binary
patterns (LBP), we are able to develop new image
descriptors with improved image search and image
category classification performance.
Computer Vision -- We develop new
statistical methods for more accurate and
efficient target detection from image and video.
One example is an efficient support vector
machine (eSVM). The eSVM, which introduces a
single value for all the slack variables
corresponding to the training samples on the wrong
side of their margin, defines a much smaller set
of support vectors and hence improves
computational efficiency without sacrificing
Another example is feature local binary
patterns (FLBP). The FLBP method, which
encodes both local and feature information,
improves upon the popular LBP approach for texture
description and pattern recognition.
Yet another example is the Bayesian
discriminating features (BDF) method. The
BDF method, when trained on images from only one
database yet works on test images from diverse
sources, displays robust generalization
performance for face detection.
The eSVM, FLBP, and BDF methods have been
successfully applied to automatic target detection
on large-scale and challenging databases, such as
eye detection and face detection.
Biometrics and Security -- We have
developed advanced face recognition, face
detection, iris detection and recognition, image
search, and image category classification
technologies for homeland security, justice and
law enforcement, and business applications.