Measuring image similarity using the geometrical distribution of image contents

To measure the similarity of images using the spatial distribution of primary features such as colour, shape and texture is difficult because the image has to be segmented and features are extracted from localized areas. Little research has been done in this area. However, such information is vital to content-based image retrieval as it contributes to the similarity measurement in the human visual system. Based on our previously proposed signature using the Radon transform, we propose a decimation to reduce the projections using principal component analysis and use correlation in measuring the similarity.