Interactive flag identification using image retrieval techniques

We propose an interactive system for identifying flags in photos taken from natural scenes. The system is interactive in two respects. First, because segmentation can be a difficult problem, users are asked to crop the flag portion from a photo. Second, the user makes the final decision by selecting one of the top choices obtained from the machine classification system. The proposed system utilizes a color-based image retrieval technique. For experimental purposes a large number of flag images are synthetically generated from a small number of original ones in order to increase the reference image database. A nearest neighbor classifier produces a sorted list of candidate choices. Recognition accuracy of these choices varies from 82% to 93% depending on whether the correct flag is among the first 8 or 18 top choices, respectively, from a set of 186 flags.