Local Radon Transform and Earth Mover's Distances for Content-Based Image Retrieval

Content-based image retrieval based on feature extraction is still a highly challenging task. Traditional features are either purely statistical, thus losing spatial information, or purely spatial without statistical information. The Radon transform (RT) is a geometrical transform widely used in computer tomography. The projections transformed embed spatial relationships while integrating information in certain directions. The RT has been used to design invariant features for retrieval. Spatial resolutions in RT are inhomogeneous resulting in non-uniform feature representation across the image. We employ the local RT by aligning the centre of the RT with the centroids of the region of interest and use a sufficient number of projections. Finally the earth mover's distance method is utilized to combine local matching results. Using the proposed approach, image retrieval accuracy is maintained, while reducing computational cost.