Investigating the combined appearance model for statistical modelling of facial images.

Abstract

The combined appearance model is a linear, parameterized and flexible model which has emerged as a powerful tool for representing, interpreting, and synthesizing the complex, non-rigid structure of the human face. The inherent strength of this model arises from the utilization of a representative training set which provides a-priori knowledge of the allowable appearance variation of the face. The model was introduced by Edwards et al in 1998 as part of the Active Appearance Model framework, a template alignment algorithm which used the model to automatically locate deformable objects within images. Since this debut, the model has been utilized within a plethora of applications relating to facial image processing. In essence, the ap pearance model combines individual statistical models of shape and texture variation in order to produce a single model of correlations between both shape and texture. In the context of facial modelling, this approach produces a model which is flexible in that it can accommodate the range of variation found in the face, specific in that it is restricted to only facial instances, and compact in that a new facial instance may be synthesized using a small set of parameters. It is additionally this compactness which makes it a candidate for model based video coding. Methods used in the past to model faces are reviewed and the capabilities of the statistical model in general are investigated. Various approaches to building the intermediate linear Point Distribution Models (PDMs) and grey-level models are outlined and an approach decided upon for implementation. The respective statistical models for the Informatics and Modelling (IMM) and Extended Multi-Model Verification for Teleservices and Secu- rities (XM2VTS) facial databases are built using MATLAB in an approach incorporating Procrustes Analysis, Affine Transform Warping and Principal Components Analysis. The MATLAB implementation's integrity was validated against a similar approach encoun tered in literature and found to produce results within 0.59%, 0.69% and 0.69% of those published for the shape, texture and combined models respectively. The models are consequently assessed with regard to their flexibility, specificity and compactness. The results demonstrate the model's ability to be successfully constrained to the synthesis of "legal" faces, to successfully parameterize and re-synthesize new unseen images from outside the training sets and to significantly reduce the high dimensionality of input facial images to produce a powerful, compact model.