Automated Image Understanding

From MilcordWiki

Jump to: navigation, search

Overview

Information processing and knowledge representation are core components of knowledge management applications in C4ISR. Given the explosive growth of airborne image data collection systems, our Automated Image Understanding assists an image analyst in managing the information overload in imagery.

Need

  • Vast amount of imagery call for automated solutions to aid the analyst
  • Information beyond atomic units (e.g., house, road)
  • Help analyst identify areas of interest (e.g., residential versus industrial neighborhoods
  • Natural language statements (e.g., an urban parking lot) for integration and use

Approach

Our approach extracts complementary features (road density, building density, and vegetation cover) from diverse sources (multi-spectral satellite imagery, aerial photography, zoning data), stores these features in a geospatial database, discovers the relationships between the extracted features using taxonomy classes of the domain ontology, employs belief networks that uses the extracted features to classify the imagery for automated image understanding.

Automated Zoning Classification of Aerial Imagery
Automated Zoning Classification of Aerial Imagery
  • Use existing extraction algorithms to extract atomic information
  • Apply model to determine dependencies
  • Build relationships among instances
  • Reason about what is seen using dependencies and relationships using Knowledge Base ontology
  • Provide an interface to end-user

Benefits

  • Time saving pre-scanning of imagery
  • Automated cueing based on inherent properties rather than features
  • Improved overall scene understanding—summary information of multiple connected images (rather than looking at one image at a time)

Applications

  • Queuing Applications: Catalog imagery streams (e.g. SHARP (SHAred Reconnaissance Pod) imagery)
  • Search Applications: Find a particular type of target in the pod stream or archive
  • Competitive Advantages:
    • Unlike feature or object extraction at the atomic level, we focus on their relationships and dependencies to provide scene or image understanding
    • Relieves the image analyst from the mundane monitoring task more suitable for machines to focus on situation assessment that requires human judgment

References

  • Windholz, T., Caglayan, A., Guler, S., and Doucette, P. (2007) “Automated Image Understanding", Technical Report, DTIC AD No. ADB334596, Milcord LLC, Waltham, MA, Dec. 2007.
Retrieved from "http://wiki.milcordme.com/pubwiki/index.php/Automated_Image_Understanding"
Personal tools