Identification of distant insignia in a visual database

Searching a large visual database for the item(s) most similar to a test object is a tedious task that has many applications. This task is further complicated if the test object (extracted from an image) is degraded by its distance and other factors. A distant target is typically subject to blurring, noise, and other distortions, so that when it is matched to database items it may turn out to be less similar to its own ideal model than another database item. To reduce the likelihood of such misidentifications, we propose an approach where the database is enlarged to include appropriately degraded versions of the ideal items in the original database. The extracted target is likely to match closely one of the distorted versions of its corresponding database item. In particular we discuss the problem of ship identification from the smokestack insignia, i.e., extracting the ship's insignia from its color photo and finding its closest matches in a database containing thousands of registered shipping line insignias. This is a problem of interest to international regulatory agencies, insurance companies and others. We discuss how the database insignias, as well as the extracted insignias, are processed and indexed, and how the distorted versions of ideal insignias are generated.