MULTI-ENVIRONMENT ATR BASED ON BIDIRECTIONAL PERCEPTION ARCHITECTURES
Abstract
Abstract The U.S. Navy has spent considerable resources researching and developing automatic target recognition (ATR) systems to detect and classify underwater sea mines in sonar imagery. In this proposal we plan to couple a self-organizing bidirectional architecture recently developed for object recognition in video with an active learning system, currently being developed in this program. The active learning module will be hierarchically connected to a content addressable memory and grammatical inference for sonar scene understanding to provide top down context from previous environments and configure appropriately the frontend system parameters to optimize ATR performance for the current uncertain environment. The proposed work will concentrate on the active learning component and its interface with the video perception architecture, to create a cognitive architecture for ATR. We will still include the human in the loop to allow training of global parameters with human supervision. We will test the overall system in the realistic ATR sonar data sets provided by NSWC PC.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512103
Entities
People
- José PrÃncipe
Organizations
- Office of Naval Research
- United States Navy
- University of Florida