Learned deformable templates for object recognition

An algorithm is described which learns a geometric template description of an object from a set of training images containing example objects. The template description learned can then be used for recognition (and location) of further examples of the object. Each template consists of an elastic mesh of spatially arranged localised image features such as edgelets or corners. A template description is encoded in the form of a chromosome which details the location and type of all of the feature tokens comprising the template. A genetic algorithm is used to optimise a template description for a particular object or class of objects. A training set of images containing examples of the object forms the environment against which templates are assessed. The fitness, or ability of templates to describe the given object, is measured after an iterative matching process and uses an objective function which rewards feature matches whilst penalising geometric distortions of the template mesh. An application of the learning algorithm is described which derives a deformable template for a forward-looking face from a set of example images of faces. Templates are learned automatically without the need to design them by hand as has previously been necessary.