FACE IMAGE PROCESSING

A new scheme for covariance matrix regularization [*] gives excellent results when applied to appearance-based face processing. Hierarchical Covariance Estimation regularizes the plug-in estimates from statistics by mixing with assumed priors, global statistics and noise. The result is that conventional maximum likelihood classification yields results significantly superior than with unregularized models or with earlier covariance estimators. For example, for face/non-face classification, this graph shows how HCR error rates vary with training set size compared with Regularized Discriminant Analysis and Mixed LOOC1 regularization.

This means that ML face/non-face classification can be applied to face detection by running a window over a picture and finding matches. Some results are shown here:

On the bottom row the left picture includes a false positive and the right picture a false negative. The results are noteworthy not for being significantly better than the sophisticated face detectors now available, but for achieving comparable performance using the most conventional of classical models with well-regularized covariance estimates.

Using conditional density estimation, 3D face structure can be recovered from mugshots without explicit modelling of shape [*]. With regularized normal models, any missing measurements are estimated as the expected value of a conditional density given the known measurements. Having trained a model with greyscales plus depths, the greyscales of a new face image form a "probe" into the model to recover depths.

We are investigating methods for speeding up the training phase of cascaded classifiers that use Haar basis functions for face detection. One such classifier was applied in [*].

Earlier activity in face image processing is summarized here and related pages (follow the links under "IMAGING HUMANS").