Content-based Image Retrieval Using Shape-Size Pattern Spectra

This paper presents the results of using the shape, size and color features of an image for content-based image retrieval. The use of granulometries has been applied to model the size and shape of the connected components of the image. Granulometry are computed by successively sieving an image using filters of increasing size parameter so that information can be obtained about the components that filter through. A shape-size granulometry using sieves of increasing shape and sizeparameters represents the shape and size distributions of an image. The results of granulometry are stored in a 2-D pattern spectrum that is implemented using attribute thinnings and openings. Additionally, the pattern spectra is extracted from the red, green and blue color bands of the images, concatenated and compared using the L1 Norm and the Jensen-Shannon divergence (JSD). Our results show that incorporating color information improves the performance by 23% and L1 Norm outperforms JSD by 20%.