Algorithms for Image Content Indexing and Information Retrieval from Unstructured or Semi-structured Complex Database

Humans take for granted their extraordinary visual abilities, however efforts to give such abilities to computers have meet with limited success thus far. Objects may be difficult for a computer to recognize/index because they are partly occluded, viewed from an unusual angle, because they blend into the background due to low contrast (or camouflage), or more generally because the best features (color, shape or texture) to recognize the objects may vary from image to image. The problems are further compounded by the typical need for fast response time. While computer CPU speeds allow for very fast processing, most image-databases reside on hard-drives. Query of disk-resident image data using linear search results in far too slow retrievals to be of practical use in most military settings. If successful, this STTR will provide both accuracy and speed in a simple, intuitive and maintainable/upgradeable system designed for efficient and timely searches of massive image archives. We will achieve speedup by leveraging of recent advances in indexing enormous datasets, and will investigate principled ways to solve the "multitude of features" problem (i.e shape, color, texture) by using relevance feedback to autonomously and interactively learn optimal weightings of the features on a query by query bias.