SIGNAL PROCESSING AND LEARNING FOR ROBUST, ADAPTIVE SONAR AUTOMATIC TARGET RECOGNITION (ATR)

Abstract Sonar-ATR is one of the key enablers in promoting autonomous operation of Naval systems in a variety of maritime domain awareness (MDA) missions such as mine counter measures (MCM). Yet there are many fundamental scientific problems that need to be solved and technological barriers that need to be cleared before the full impact of Sonar-ATR to the fleet can be realized. Based on our recent contributions in Signal Processing and Machine Learning [1–11], we have identified two basic theoretical strands which we propose can be significantly developed and extended to address the significant outstanding technical challenges in Sonar-ATR: 1) Adaptive sparsity based discriminative signal representations [4, 5, 7, 9, 11], and 2) Graphical models for classification and regression [1–3, 8, 10]. The limitations of working in a fixed feature space are well known in the Sonar-ATR community. Previous notable attempts to address this issue have focused on classifier fusion and learning in-situ states in order to select the optimum feature-selection cum classifier combinations [12–14]. Recently we have developed powerful extension of these ideas based on probabilistic graphical models [1, 3,8]. On the classification side we have developed a new methodology called IGT (Iterative Graph Thickening) wherein feature fusion is performed in a principled manner. IGT not only has the capablility of accomodating heterogeneous feature sets but also has demonstrably robust performance in low training regimes which therefore can find powerful applications in Sonar-ATR. We have also devised novel methods of performing regression via graphical models [2, 10]. Another promising direction that we have investigated builds upon the work of sparsity based classification recently introduced in the machine learning community wherein the use of a dictionary (or basis) matrix comprising of class-specific training sub-dictionaries is advocated [15]. This simple strategy has shown remarkable robustness under heavy amounts of clutter and distortion. Furthermore this strategy furnishes a natural out-of-class measure based on the distribution of sparse coefficients. Extending recent ideas from model-based compressive sensing, we advocate using class-specific priors to capture structure on sparse coefficients in the feature domain that helps explicitly distinguish between signal classes thereby enabling the extraction of sparse coefficients that are truly discriminative in nature [4, 5, 9]. Further, complementing our work on class-specific priors we have recently developed novel approaches to discriminative dictionary learning [7, 11] thereby adding new dimensions to our ability to adapt the feature space to the characteristics of the application domain and the environment.