We also introduced the selective symmetry transform to extend the concepts of symmetric enclosure to non-circular contours (such as the semi-elliptical ones found in human faces). This permits the selective detection of particular blob shapes. Most other facial contour estimation techniques are not motivated by low-level perception theory and utilize template matching or deformable models without a measure of symmetric enclosure.
Traditionally, there have been two approaches to face recognition: feature-based techniques and holistic techniques. We propose a way of bridging the gap between the two techniques by synthesizing mug-shot images from arbitrary input images with a high-quality 3D normalization. The 3D normalization involves a deformable 3D model of the average human face which produces superior normalization for out-of-plane rotations than other techniques. Thus, we combine the robust face detection of a feature-based stage with the precise classification capabilities of a linear, holistic, recognition stage.
We proposed the use of mixed-histogram fitting to correct for illumination. The gradual mixture of different histogram correction functions allows a smooth illumination correction for both sides of a face. Typically, illumination correction was performed on the face as whole. We window the histogram analysis and use weighted mixtures of the histogram transfer functions to achieve superior illumination correction.
Finally, we proposed a sensitivity analysis relating the localization accuracy to the recognition certainty. This form of analysis relates the performance of two different components of face recognition: localization and classification.