Publications

Abstract : Recent advancements in computer vision have given image understanding and retrieval a new face. But the current image retrieval systems do not meet users’ demand in retrieving alike and meaningful images with respect to the query. This paper proposes a novel method for image retrieval using scene graphs. We generate scene graphs for images and each one of them is a collection of objects (apple, table), attributes ((apple, red), (table, wooden)) and relationships (apple, on, table). When an image is given as a query, images whose scene graphs are similar to that of the query image’s scene graph, are retrieved. An algorithm called SPLIT–SIM is proposed to find similarity between two scene graphs, which grant rewards for similar objects, attributes and relationships. Based on the rewards awarded, a final ranking list of images is generated for retrieval. The proposed algorithm was evaluated at three levels: object level, attribute level as well as relation level and the results demonstrate the efficiency of the proposed semantic similarity measure, significantly improving existing similarity measures.