THE 2-D HIDDEN MARKOV MODEL FOR IMAGES,
IT'S EXTENSIONS, AND APPLICATIONS
 

ABSTRACT

Digital image is an important medium in modern communication.  As a special type of data, images have posed long-standing challenges for pattern recognition and statistical/machine learning.  Tasks simple for humans, such as annotation and segmentation, are notoriously difficult for the computer to perform automatically, partly because only the human vision can effortlessly treat an image as an entity and make decisions in a global and effective manner.  We developed a two-dimensional hidden Markov model (2-D HMM) so that learning algorithms can utilize long-range (or more accurately large-region) information in images.  In this model, a hidden layer of states is introduced as a mediator to efficiently and flexibly characterize the dependence among observable local feature vectors residing on a 2-D grid.  The states are assumed to follow a Markov mesh (a special MRF).  Feature vectors are conditionally independent given the states; and the conditional distribution varies with the state.  The 2-D HMM is further extended to the 2-D multiresolution HMM (MHMM), which enhances the capability of integrating information in large regions.  A variation of 2-D HMM for the purpose of reducing computational complexity has also been explored.  In this talk, we will introduce the aforementioned models and their estimation methods.  In addition, we will demonstrate the applications of the 2-D HMM methodology in two substantially different contexts: supervised/unsupervised segmentation and image annotation.  For segmentation, the 2-D HMM enables simultaneous optimization on the classes of many pixels.  For annotation, the 2-D MHMM serves as a state-of-the-art tool to profile hundreds of manually annotated categories of images.  Descriptive words are assigned to a new image based on its likelihoods under the profiling models.
 

***********************
MONDAY, MARCH 29, 2004
212 NEW WEST
3:30 P.M.
***********************