A statistical Image of colour space

The distribution of colours in an image has proven to be very useful for object recognition. Building on Swain's colour indexing (1991), colour distributions are now an integral part of many recognition schemes. This is not to say that colour alone suffices but rather that colour is one important cue that aids recognition. In this paper we look at colour distribution based recognition from a rather novel image processing perspective. Specifically we view the distribution of colours in an image as an image and so recast colour distribution matching as a problem of image comparison. Two results are reported here. First that, by compressing images we can improve matching efficiency (recognize objects more quickly). Second, that the degree of compression (and so speedup) that is possible depends on the colour space on top of which the distribution images are built. The more uniform opponent colour encoding can be compressed more effectively compared with conventional rg-chromaticity encoding. We explain this in the following way: image processing is based on the assumption that all image locations are equal so, to treat colour space as an image, each colour location should also be equally likely. This is in fact approximately the case for the opponent chromaticity space. To validate our approach we repeated Swain's object recognition experiments. We show that the distribution of colours in an image (which Swain encoded with 4096 numbers) represented by 8 numbers (the projection coefficients onto an 8-dimensional principal component basis) suffices to achieve the same (almost perfect) recognition rate. Our method delivers a 500-fold speed up in indexing without loss of accuracy. This result scales to a second larger database of 140 images