No abstract

A new class of linear recursive image estimation methods, based on multiresolution image representations, is introduced. Although recursive, the estimators are causal not in the image plane but in a third dimension -that of the scale index. The estimators are efficient computationally and are in general sub-optimal for a class of image models based explicitly on the scale-space underlying the multiresolution description. Methods are also presented for adapting the estimates to local image structure across a wide range of scales. The estimates are robust and outperform many of the techniques reported in the literature in terms of computational efficiency, signal-to-noise gain and subjective appearance. A brief presentation of the theoretical basis of the methods is followed by experimental results and conclusions on the potential of the new approach. INTRODUCTIONThe widely-recognised inadequacies of linear shift-invariant estimators when applied to natural images have led to a multitude of proposed solutions to the problem. These range from simple nonlinear methods (e.g. [l]) to more complex 'multiple-model' approaches which, although not strictly linear, retain many of the properties of linear estimators, at

The authors describe a face tracking and recognition system for video indexing that handles variable face poses (left-right and up-down) and deformations due to speech and facial expressions. The system is based on deformable template matching, and employs person-specific templates at near-frontal poses for recognition, and novel person-independent templates at multiple poses on the view-sphere for tracking. Relative to an earlier version that used multiple person-specific templates at multiple (left-right) poses, the new system speeds up processing by (i) restricting attention to skin-color regions; (ii) performing recognition using the person-specific templates at near-frontal poses only; and (iii) tracking at non-frontal poses using the novel person-independent templates. Registration is also simplified since multiple views of each target individual are no longer required, but at the cost of a loss of recognition functionality at poses far from frontal (the system instead “remembers” the identity of each individual from near-frontal matches and tracks between them).