Compact colour descriptors for colour-based image retrieval

In many colour-based image retrieval systems the colour properties of an image are described by its colour histogram. Histogram-based search is, however, often inefficient for large histogram sizes. Therefore we introduce several new, Karhunen-Loeve transform (KLT)-based methods that provide efficient representations of colour histograms and differences between two colour histograms. The methods are based on the following two observations; Ordinary KLT considers colour histograms as signals and uses the Euclidian distance for optimization; KLT with generalized colour distance measures that take into account both the statistical properties of the image database and the properties of the underlying colour space should improve the retrieval performance. Image retrieval applications compare similarities between different images. Relevant for the decision is only the local structure of the image space around the current query image since the task is to find those images in the database that are most similar to this given query image. Therefore only the local topology of the feature space is of interest and compression methods should preserve this local topology as much as possible. It is therefore more important to have a good representation of the differences between features of similar images than good representations of the features of the images themselves. The optimization should therefore be based on minimizing the approximation error in the space of local histogram differences instead of the space of colour histograms. In this paper we report the results of our experiments that are done on three image databases containing more than 130,000 images. Both objective and Subjective ground truth queries are used in order to evaluate the proposed methods and to compare them with other existing methods. The results from our experiments show that compression methods based on a combination of the two observations described above provide new, powerful and efficient retrieval algorithms for colour-based image retrieval. (C) 2004 Elsevier B.V. All rights reserved.