Model-based feature extraction and classification for automatic face recognition

Recognising faces through model-based feature extraction and description currently appears to be less popular than statistical or face-based recognition approaches. Certainly there is concern that model-based approaches might not prove reliable in practice. Accordingly, this thesis describes a programme of research for improving model-based recognition through robust feature extraction, selection and combination. First we present a new two stage process for finding eyes. A reformulated evidence gathering process is used to determine the rough location of the eyes by exploiting their natural concentricity. Their location was refined by an improved deformable eye template which does not require internal energy terms and uses few parameters. These parameters were best optimised using a genetic algorithm. The technique produced 91% and 93% successful location rates on face databases of 1000, and 88 faces, respectively. A feature vector composed of 29 geometric, 6 colour and 55 forehead contour measures, was extracted from 44 faces from XM2VTS database. To achieve this, the skin boundary was extracted by region growing using a sample of skin below the eyes. Other features such as the nose, mouth and eyebrows were then located by noting that these features are enclosed by skin but exhibit different statistical properties.A new method, based on intrinsic feature variance, is presented for combining and selecting features which are of potentially disparate magnitude and/or independent sources. Our method provided an increase variance in the classification matrix and facilitated identification of the most discriminating features. Surprisingly, although the eyes were a good initialiser in the search for other face features, their template parameters offered low discriminatory power. Much higher discriminatory power was available through the normalised Fourier descriptors of the forehead contour. We simulated the effect of measurement noise on classification performance and found that errors of 6 pixels on the geometric features resulted in up to 43% classification error. Recognition rates of 77% and 72% were experienced using manual and automatic geometric measures. However, when we combined the geometric measures with perfectly extracted contour measures from its first eight Fourier descriptors we achieved 100% classification.