Authors: HauSan Wong Kent K T Cheung ChunIp Chiu Yang Sha Horace H S Ip
Publish Date: 2007/01/27
Volume: 33, Issue: 1, Pages: 91-108
Abstract
Histogram feature representation is important in many classification applications for characterization of the statistical distribution of different pattern attributes such as the color and edge orientation distribution in images While the construction of these feature representations is simple this very simplicity may compromise the classification accuracy in those cases where the original histogram does not provide adequate discriminative information for making a reliable classification In view of this we propose an optimization approach based on evolutionary computation Back Evolutionary algorithms in theory and practice Oxford University Press New York 1996 Fogel Evolutionary computation toward a new philosophy of machine intelligence 2nd edn IEEE Piscataway NJ 1998 to identify a suitable transformation on the histogram feature representation such that the resulting classification performance based on these features is maximally improved while the original simplicity of the representation is retained To facilitate this optimization process we propose a hierarchical classifier structure to demarcate the set of categories in such a way that the pair of category subsets with the highest level of dissimilarities is identified at each stage for partition In this way the evolutionary search process for the required transformation can be considerably simplified due to the reduced level of complexities in classification for two widely separated category subsets The proposed approach is applied to two problems in multimedia data classification namely the categorization of 3D computer graphics models and image classification in the JPEG compressed domain Experimental results indicate that the evolutionary optimization approach facilitated by the hierarchical classification process is capable of significantly improving the classification performance for both applications based on the transformed histogram representations
Keywords: