Fast shape retrieval using term frequency vectors

Similarity based retrieval of visual shapes is performed using a vector space technique originally developed for document retrieval. In that technique, a document is represented by a vector in which each component is the frequency of occurrence of a specific term (word or phrase) in that document. Similarity between query and database documents is measured by the normalized inner product of the vectors. In the shape retrieval system, contours are partitioned into perceptually significant segments that have a role analogous to words in a document. Short sequences of segments are analogous to phrases. Each shape is represented by a vector of the frequency of occurrences of each term (segments or segment sequences). Fast retrieval is achieved by using a B+ tree and inverted lists. Additionally, a new similarity measure is introduced and shown to result in improved performance over the normalized inner product.