An illumination and highlights invariant colour model for image matching

Abstract

Colour can provide critical information for a variety of computer vision tasks such as image matching, object recognition and image retrieval. However, for it to be useful in practice, the colour model used to represent the intrinsic properties of the imaged objects must be insensitive of imaging conditions such as lighting geometry, illumination colour and highlights. In this paper, we present a colour model for image matching and object recognition that is invariant for illumination and highlights. The colour model is defined as the ratios of the colour differences between neighbouring pixels for each colour component. Based on the dichromatic reflection colour model, it is shown that the proposed colour model is invariant to lighting geometry, illumination colour, highlights and diffuse lighting. Experimental results show robust image matching using the proposed colour model on objects that are illuminated under different illumination colours and lighting geometry. The proposed colour model can be used as a prepossessing step for applications where limited or no constraints on the imaging process can be imposed.