Comprehensive Comparison of Region-Based Image Similarity Models

We perform a multiperspective comparison of the retrieval performance of six variants of the region-based image similarity models. Our objective is to examine the effect on the retrieval performance, of: (1) the region matching approach, (2) the region weighting strategy, and (3) the number of regions. Common for the six variants is that: (1) images are uniformly partitioned into regions (i.e., rectangular blocks of equal size), at five different resolutions; and (2) from each region, color, shape, and texture features are extracted, and used for computing the region similarity. The difference between the variants is either in the region matching approach, or in the region weighting strategy. Regarding the region matching, the correspondence between pairs of regions of the two images is established based on either: (1) their spatial closeness, (2) their visual similarity, or (3) a combination of these. Regarding the region weighting, weights, as a function of distance between corresponding regions, either: (1) decrease linearly, (2) decrease exponentially, or (3) are constant. The evaluation of the six variants is performed on 5 test databases, containing 64,339 images, in 749 semantic categories. In total, 313,020 queries are executed, based on which the average (weighted) precision, recall, rank, and retrieval time are computed. Both the number of queries and the variety of the evaluation criteria make the evaluation more comprehensive than in the case of any of the existing works, dealing with the region-based image similarity models. Results of the evaluation reveal that, contrary to the expectations, the simplest variant results in the best overall retrieval performance.