Perceptually-driven moment thresholds for shape description in image databases

Over the past five years, a huge influx of visual data has arisen within the realm of computing. Consequently, it is evident that a means of automatically indexing this visual information on the basis of content is required. This paper discusses a component of an image retrieval system with the ability to automatically locate and classify an image’s constituent shapes. The images employed are pre-processed, and pre-segmented and analysis is performed on component regions which are separated using color. Shape analysis is performed via invariant moments and the subjectivity of shape classification is taken into account using data obtained from human testing. The subjective data is used to generate a predetermined set of classification thresholds for various simple shapes of which 84% were successfully classified. Key-Words: Shape classificiation, content-based image retrieval, subjective testing, invariant moments.