The research presented in this thesis investigates the problem of detecting faces from constituent features, with minimal assumptions and constraints on posedness and environmental conditions such as scale and lighting. Constituent features include mouths, nostrils, eyes, and eyebrows. Face detection proceeds in four stages: candidate feature detection, candidate feature pairing, partial face context construction, and complete face context construction. At each stage increasing levels of contextual information are employed. The last two stages involve the application of an energy minimizing spring-based facial template, consisting of spatial, anthropometric, rotational, and constituent energy terms. Initial experiments indicate the system outperforms related work in the field, with detection rates in excess of 83%. Less stringent evaluation criteria yields detection rates in excess of 90%. Furthermore, fewer assumptions and constraints are imposed in this research than in any other work reviewed.
展开▼
