Evaluation of content based image retrieval (CBIR) system using precision measure

Abstract

Content based image retrieval (CBIR) is an application of the computer system for image retrieval, where it will aid searching for digital images in a large database. CBIR operates on a totally different principle from keyword indexing. Content based means that the search will analyze the actual content of an image, which will be done automatically by the system, where the cost and time will be reduced greatly. The study sets out to assess the retrieval performance of a CBIR system with regards to the users’ evaluation and precision measure and was modelled after Cranfield Test by Cleverdon. To collect the queries and relevance criteria used during the searching activities, a semi structured interview was carried out where two users from different professional backgrounds; school teacher and graphic designer were interviewed. Users’ relevance judgment of the retrieved images were collected and later used to calculate the precision values. The precision values measured were varied among the two users because their professional backgrounds reflected their final judgments of the relevance criteria and the retrieved images. The precision values differed greatly between the two users and the low precision values were indicated as a poor retrieval performance by Retrievr. In conclusion, Retrievr did an image clustering based on the focus in the image queries whereby queries with fewer objects are more accurately matched and retrieved.