In-Situ Adaptation For Underwater Target Detection and Classification Using An Information Theoretic Approach

A critical need of the U.S. Navy is the development of a reliable, efficient and robust underwater target detection and classification system that can operate in real-time with various sonar systems and in different environmental and operating conditions. To maintain performance in such conditions, new solutions are needed to update the detection and classification systems in-situ in response to environmental and operational changes. The main goal of this Phase I research is to develop innovative solutions that offer in-situ learning ability for classification and possible identification of underwater targets using (a) a model-reference mechanism that incorporates input/output relations within a set of a new samples with class/within-class labels and confidence scores, (b) a relevance-feedback mechanism that attempts to capture expert operators high-level decision-making concepts via operators feedback, (c) an information-theoretic selective sampling method to extract the most informative training samples from the new environment, and (d) demonstration of the effectiveness of the algorithms on sonar data sets. The unique advantage of our proposed solutions is the ability to offer system flexibility while preserving the stability of the previously learnt information. Additionally, the system is simple and amenable for real-time implementation on a wide variety of sensor platforms.