531 Recognition and retrieval via histogram trees Stuart Gibson and Richard Harvey School of Information Systems, University of East Anglia, Norwich, NR4 7TJ, UK. Email:{s.e.gibson, r.w.harvey}@uea.ac.uk Abstract This paper explores a new method for analysing and comparing image his- tograms. The technique amounts to a novel way of backprojecting an image into one with fewer, statistically significant colours. When the method is tested with the COIL-100 and MPEG-7 data sets it is shown to have a per- formance that is as good as the best methods using fewer enteries than the original histogram. Therefore it offers the potential for extending the use of histograms into high dimensional feature spaces. 1 Introduction Histograms are popular features for recognition and retrieval [20]. They form part of the forthcoming MPEG-7 standard for image and video metadata and, despite their known shortcomings as density estimators [26], they are popular because of their computational simplicity. When using histogram-based techniques, important issues are the selection of an ap- propriate feature space, the quantization of the selected space, and the algorithm for com- paring histograms. The choice of feature space is often problem specific (see [4] for an example where colour spaces are compared). The choice of histogram quantisation (bin size) is discussed in, for example, Scott [19] 1 . This paper examines the last issue, the problem of comparing multidimensional histograms. We review the existing options and suggest a new method for representing and comparing multidimensional histograms. There are several reviews of histogram comparison methods ( [13,16] are examples). The methods may be summarized as being based on inter-bin distances, intra-bin dis- tances or feature distances. The inter-bin distances take the form d (h, k)= g( ∑ i f (h i , k i ))/W (1) where h =[h 1 ,..., h n ] and k =[k 1 ,..., k n ] denote the n-bin histograms, g and f are func- tions that vary from method to method, and W is a scaling factor. Table 1 summarizes these for a variety of inter-bin distances. For colour histograms formed in a non-invariant colour space the effect of lighting variation is to shift the modes of the underlying distri- bution. In this case inter-bin distances can be ineffective. The usual solution is to use an intra-bin distance, a perceptual colour space or both. The best known intra-bin distance 1 From which one deduces that many Computer Vision systems often operate with undersmoothed his- tograms.