Relevance tuning in content-based retrieval of structurally-modeled images using Particle Swarm Optimization

Similarity of images in content-based image retrieval (CBIR) is a subjective measure varying by the user, and requires tuning according to the user's preference. Another issue in CBIR is the need of partial image matching. Structural modeling of the images can be promising in finding a small query image within a large database image. In this work, a graph-based image modeling which assigns image regions to labeled nodes and their adjacency to weighted edges is used. Also, the image similarity measure is tuned according to the user's evaluation, by way of parameter selection using Particle Swarm Optimization (PSO)[1][2]. In the experiments, a small-scale CBIR system based on graph modeling of images was developed. Using the system, it was confirmed that images including the query image of different size and rotation angle could be successfully retrieved. Also, the user's preference in weighting the different aspects of similarity in the feedback information was found to be successfully incorporated in the retrieval after parameter optimization using PSO.