But lately, O blindness, I worshipped images just brought from the furnaces,
gods made on anvils and forged with hammers:
I've done some experiments on image classification, i.e. having the computer classify images into a set of categories (defined in advance) after having seen and processed a few examples of images from each category. The basic idea is to use techniques from image retrieval to provide some sort of descriptions of images (e.g., histograms, texture segmentation, etc.), then feed these into some machine learning algorithm (e.g. SVM if the images are described with vectors; or use nearest-neighbour learning, which can be combined with any similarity measure from image retrieval).
The results of these experiments aren't quite exhilarating, but they aren't totally useless either. Using histograms and similar descriptions, such as autocorrelograms, in combination with Support Vector Machines (SVM) as the learning algorithm, worked best. Surprisingly, various similarity measures based on segmentation (WALRUS; and IRM from the SIMPLIcity system), with nearest neighbours as the learning method, didn't do any better than the simple vector-based descriptions.
On a collection of 1172 images, divided into 14 classes, the best classifiers achieved an accuracy of about 75 %. There isn't a lot of literature about image classification, but this result seems reasonably comparable with other published results (e.g. Huang et al., 1998).
Unfortunately, most of the materials here are in Slovenian (see the other index page), with the exception of:
