Image clustering using content-based techniques

A growing number of applications now involve the storage and retrieval of digital images, but it is accepted that there is limited value in storing those images if one cannot easily retrieve them. The limitations of methods based on text-labelling are by now well known and have led to a burgeoning of research projects to develop content-based search methods. We have shown that content-based techniques, and texture in particular, can be used to cluster images, giving a reasonable correlation with assignments made by visual inspection. Options for improving the accuracy of the clustering include: synthesis of purer examples; outlining regions within the image; appropriate weighting of the components of the texture feature vector; and experimenting with alternative classifiers, particularly where texture is important.