Colour Normalization for Colour Object Recognition and Image Retrieval

Finlayson, G. D. and Tian, G. Y. (2000) Colour Normalization for Colour Object Recognition and Image Retrieval. In: Invariants for Pattern Recognition and Classification. World Scientific Publishing, pp. 213-228. ISBN 978-9810242787

Full text not available from this repository. (Request a copy)

Abstract

Colour images depend on the colour of the capture illuminant and the objects in a scene. As such raw image colours are not stable features for object recognition. It follows then that before the colours in images can be compared they must first be pre-processed to remove the effect of illumination. Two types of pre-processing have been proposed: first, run a colour constancy algorithm or second apply an invariant normalization. In colour constancy pre- processing the illuminant colour is estimated and then, at a second stage, the image colours are corrected to remove colour bias due to illumination. In colour invariant normalization image RGBs are re-described, in an illuminant independent way, relative to the context in which they are seen (e.g. RGBs might be divided by a local RGB average). In theory the colour constancy approach is superior since it works in a scene independent way: colour invariant normalization can be calculated post colour constancy but the converse is not true. However, in practice colour invariant normalization usually supports better indexing. In this paper we ask whether colour constancy algorithms will ever deliver better indexing than colour normalization. The main result of this paper is to demonstrate equivalence between colour constancy and colour invariant computation. To date colour constancy algorithms set out to determine the illuminant irrespective of scene content. Here we show that scene content can impinge on the effective illuminant. For example in the presence of interreflection, the light arriving at a surface is a function of the illuminant and the other surfaces in a scene. However, since, by definition, the effective illuminant is a function of the context in which a surface is seen colour constancy computation should also be scene dependent and so is really a type of colour normalization. Experiments demonstrate that colour normalization and scene dependent colour constancy algorithms deliver significantly better indexing than the scene independent colour constancy computation.

Item Type:	Book Section
Faculty \ School:	Faculty of Science > School of Computing Sciences
UEA Research Groups:	Faculty of Science > Research Groups > Visual Computing and Signal Processing Faculty of Science > Research Groups > Colour and Imaging Lab
Depositing User:	EPrints Services
Date Deposited:	01 Oct 2010 13:42
Last Modified:	28 Mar 2025 01:39
URI:	https://ueaeprints.uea.ac.uk/id/eprint/3731
DOI:	10.1142/9789812791894_0011

Actions (login required)

View Item