Conformation-based hidden Markov models: application to human face identification

Area of Science:

• Computer Science
• Pattern Recognition
• Machine Learning

Background:

• Hidden Markov Models (HMMs) are effective for classifying structured objects.
• A key limitation of HMMs is their inability to intrinsically handle shape or conformation of visible observation (VO) sequences.
• Predicting the n-dimensional shape formed by VO sequences remains a challenge for standard HMMs.

Purpose of the Study:

• To introduce a novel paradigm, Conformation-based Hidden Markov Models (COHMMs), to address the shape-related limitations of HMMs.
• To develop a formalism capable of classifying VO sequences by considering their intrinsic shape.
• To extend the COHMM framework to both one-level and multilevel structures.

Main Methods:

• COHMMs embed nodes of an HMM state transition graph within a Euclidean vector space.
• The method models noise within the shape formed by the VO sequence.
• A multilevel COHMM framework is presented, addressing sequence probability, statistical and structural decoding, shape decoding, and learning.

Main Results:

• The proposed COHMM formalism was applied to human face identification tasks.
• Experiments were conducted on various benchmarked face databases.
• Multilevel COHMMs demonstrated superior performance compared to embedded HMMs and other standard HMM-based models.

Conclusions:

• COHMMs offer a significant advancement in sequence classification by integrating conformational information.
• The multilevel COHMM approach is particularly effective for complex tasks such as human face identification.
• This formalism provides a robust solution for analyzing and classifying structured data where shape is a critical feature.