In situ learning for underwater object recognition Navy SBIR FY2009.1 Sol No.: Navy SBIR FY2009.1 Topic No.: N091-066 Topic Title: In situ learning for underwater object recognition Proposal No.: N091-066-0661 Firm: 3 Phoenix, Inc. 13135 Lee Jackson Hwy Suite 220 Fairfax, Virginia 22033-1907 Contact: Russ Jeffers Phone: (703) 956-6480 Web Site: www.3phoenix.com Abstract: Sea mines are a cost-effective method for hostile forces to attempt to neutralize assets of the U.S. Navy by limiting mobility and creating delay. Mine detection, classification and localization (DCL) is very challenging in littoral environments due to the high clutter, increased background, and dense multipath. 3 Phoenix, Inc. has developed an innovative approach for automatic target detection and classification of sea mines and other underwater targets of interest. The proposed algorithm robustly adapts to changes in environment and has the potential for dramatically reducing false alarm rate, while still maintaining a high probability of detection and classification. A novel, efficient method of training the classifier is formulated and retraining for adaptation is performed intrinsically with weight optimization. The algorithm is generalized to work over several sensor types and sensor modalities. The proposed algorithm has the potential to reduce operator load while reducing false positives in classification. Benefits: The proposed Phase I investigation is expected to yield innovative algorithms to efficiently detect and classify underwater objects in a statistically non-stationary environment. In particular the algorithms will be designed for automatic target recognition of mines in littoral environments with initial training data from similar but not exact environmental settings. The approach presented in this proposal represents the potential to reduce operator load and allow for constant algorithm update to reduce false positives in classification. 3 Phoenix, Inc. anticipates that the results of this effort will demonstrate the feasibility of the algorithms and the path to implementation. The Phase I effort, as a demonstration of feasibility, will provide a preliminary basis on a small, but representative dataset. This preliminary feasibility analysis will establish the potential performance bounds for an automated classification system and provide a basis for estimating the computer/processing requirements. This will serve to identify candidate methods, from the proposed set of algorithms, which may be used as a system to enhance target detection and classification. Many of the critical algorithms will be implemented as tools that can be incorporated in a Matlab-based test bed. This signal-processing test bed will be used in the Phase II research to investigate the performance of the classification system. Return