What are the latest developments with regard to facial recognition algorithms? Toews and Arbel’s paper presents a formulation that detects, localizes, and classifies images from different viewpoints and with partially disguised (occluded) features.

The authors utilize an object class invariant (OCI) model to identify the same object in different scenes or images, but employ a Bayesian probabilistic model to classify objects into categories. In particular, the paper presents the results for classifying images of people, with regard to gender. The error rate realized is approximately 16 percent.

The OCI and Bayesian models are trained by running them against a facial image database. Once the parameters of the models are estimated, a classification experiment is conducted against a second facial image database and the results are tallied. Note that these databases contain multiple images of the same person from different perspectives. The new method presented here is viewpoint invariant.

The paper cites 64 recent references on the subject matter. The authors illustrate their technique in a very clear fashion; even those without expertise in this area should be able to follow the discussion, at least to a certain point.

This paper is a must-read for those who research image analysis. Anyone in the fields of robotics or computer vision should also take a look at this paper. Anyone who hopes to gain a basic understanding of image analysis can glean a good sense of the scope and complexity of this problem by studying the analytic framework presented here.