Bayesian Fusion of Contour Descriptions: Application to 3-D Object and Face Recognition

摘要

In this paper a 3-D object's/face's caricature recognition system is proposed. An object's-caricature is recognized through a probabilistic fusion procedure. The innovation introduced is that an object's/face's 2-D caricature's views that are taken in 3-D are fused in terms of their contour features. In addition, these features are directly connected to all off the objects and faces stored in our database. A face/object is thus regarded as the output of a detailed probabilistic Bayesian analysis of the views' contours being inter-independent parameters. The features are the object's/face's border pixels that are extracted from low-level edge information. The faces were separated from background clutter using the C.M. for two clusters (background and foreground) for the fist view and then the Nearest Neighbour classifier. Both kinds of patterns are modelled as distributions, as they are vague due to the non-perfect lighting conditions and different face postures. For all contour features partial hypotheses, expressed as Gaussian probabilistic conditionals were examined in real time, in terms of their plausibility with regard to which object they are most likely connected to. For faces, shoulder noise does not deteriorate the recognition performance as a result of the robust Bayesian reasoning followed. We arrive at a final distribution allocating a certain degree of confidence to a set of the available objects/faces. The objective is three-fold: 1. to recognize a known object/face from a significantly reduced set of all candidate views/faces not presented to the system before, 2. to recognize a strongly altered unknown view that belongs to known object/face, 3. Find the best resembling known object/face for a totally unknown object/face that is presented to the system, 4. Do all the above with minor training, and with comparable success to systems using complex model parameters distributions. The object features were taken from a camera assuming different longitude positions around the unknown object. The faces were taken from the Manchester face database. The applications to forensic science object and person identification are obvious as the system uses very simple characteristics amenable for use with CCTV cameras, fast algorithms and reaches sufficient reliability. An easier expert knowledge integration using probabilistic priors is also provided.