Possible improvements

There are several ways that one could potentially improve the performance of this classifier system. One potential modification would be to take positive training example clusters into account. It may be that a significant subset of the training images share a discriminating epitome patch, but the rest do not. In this case, that epitome patch would not score highly overall for its ability to discriminate between positive and negative training example images, even though it is an excellent classifier patch for a subset of the positive training examples. A practical example would be human faces under low illumination versus human faces under normal illumination. Patches that identify faces under low illumination could be missed because they do not have a high probability of mapping into positive images overall. Checking for patches that are excellent discriminators for a subset of the positive training examples could result in a more robust classifier.

Also, the ``winner take all'' nature of the procedure for determining whether an epitome patch gets mapped into an image or not could be altered. In the current implementation, only the epitome patch with the highest posterior probability of being mapped into a given patch in the original image is considered. This is obviously discarding information about other patches that had high, but not highest, probabilities of mapping into this image patch. Allowing ``partial'' mappings weighted by relative posterior probabilities may help retain some of this information.

A scheme for capturing information about the relative spatial distributions of image patches mapped to by discriminative epitome patches may also improve performance. An example of such a scheme can be found in [2].

Finally, the image epitome representation can also be modified in several ways. The results shown in [3] were created using a modified epitome learner that allowed varying block sizes, which allows tesselations of coarser and finer blocks. The results in [3] show a clearly superior preservation of shapes and structures from the original image. This may enhance the performance of our classification application. Also, [3] mentions the fact that block mappings from the epitome to the image need not be limited to simple copying operations. It would be possible to allow blocks to be scaled, rotated, or otherwise transformed in the mapping process. This modification to the image epitome model seems likely to result in a more robust patch-based classifier.