Prototype for Meta-Algorithmic, Content-Aware Image Analysis
Abstract
Report developed under the Defense Advanced Research Projects Agency (DARPA) Visual Media Reasoning (VMR) program. In this effort, several techniques were evaluated, including image segmentation and classification, and feature (algorithm) ranking, within a Content-Based Image Retrieval (CBIR) framework. The effort also examined CBIR performance in object recognition and classification. In this context, automated segmentation algorithms were developed, in particular of active contour-based segmentation techniques, and applied to the extraction of specific objects including weapons, humans, and planes. Self-nomination is the process by which an algorithm (feature-types), "optimal" for a given specific object type, is selected within a pool of available ones. The selection process is carried out by assigning higher weights based on the level of performance of each algorithm. In this effort, two approaches were proposed: the first was based on dictionary learning whereas the second used a multiple kernel learning technique. Both approaches were studied in detail and their results on a sample dataset are presented.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2015
- Accession Number
- ADA621858
Entities
People
- D. Newell
- K. Skadron
- Rituparna Sarkar
- S. Ozer
- Scott T. Acton
Organizations
- University of Virginia