Model-based deconvolution of the human face

Many practical deconvolution problems arise in which explicit knowledge of both the system PSF and the spectral characteristics of the noise are unknown. We describe and present an approach to deconvolution in this situation which is specifically matched to the forensically important problem of face identification. Our approach is to model both human faces and image aberrations in a statistical appearance framework using a representative sample of faces. Deconvolution is then achieved experimentally by moving along known transition curves in a parametric face space. Our preliminary studies demonstrate that the accuracy of the method is superior to maximum-likelihood blind deconvolution at low signal-noise ratios. A hybrid method in which the noisy face image is first projected into the model space and blind deconvolution then applied yields the best overall performance.