Object Recognition Using Colour, Shape and Affine Invariant Ratios

This paper describes a spectral-spatial model (for colour object recognition) which exploits the shape, colour and position of regions on the surface of a rigid object in describing it. Given a model and test image (with colour constancy pre-processing) suitably segmented into colour regions, model and test regions with similar shape and colour are identified. If at least three model regions have matching test regions then the consistency of these matches are verified using distance/area affine invariant ratios. Subsequently, model regions are affine transformed into image space for matching, from which a match probability is determined. Experimental results demonstrate that this model is significantly invariant to illumination changes, affine deformity and partial occlusion.