Hybrid Learning and Model Based Methods for SONAR ATR

Based on our recent contributions in Signal Processing and Machine Learning, we have identified two basic theoretical strands which' we propose can be significantly developed and extended to address outstanding technical challenges in Sonar-ATR: 1) Adaptive sparsi'ty based discriminative signal representations and 2) Deep learning based methods for classification and regression. In recent work,' we developed a robust strategy called pose corrected sparsity (PCS) as a means to classify Sonar targets that can capably handle no'ise, occlusion, and other issues like limited training. PCS incorporates a novel interpretation of a sparsity inducing spike and sla'b probability distribution towards use as a Bayesian prior for class-specific discrimination in combination with a dictionary learni'ng scheme for localized patch extractions.While PCS shows much promise for sparsity based SONAR ATR, it cannot handle Multiview and' multi-channel ATR scenarios. This is because the observation or test vector is a single SONAR image and sparsity is enforced on the corresponding coefficient vector. A key extension in our future work therefore will be the development of customized multi-task sparsity models that capture matrix and tensor sparsity for SONAR ATR. Such multi-task sparsity models can be invaluable for multi-view' and multi-sensor/channel ATR. Finally, we also intend to explore the integration of using prior structure on images with deep learn''ing techniques. While deep learning techniques have been promising in classification and recognition, their potential for SONAR and' other remote sensed ATR problems is severely limited by the fact that most deep learning architectures such as CNNs and auto-encoders make inordinate demands on the amount of training imagery available. Such generous training may be possible to obtain in the form of optical imagery for applications such as face recognition but is completely unrealistic in remote sensing applications where training imagery is severely limited. Our research will hence aim to retain the merits of deep learning architectures while lessening t'heir training burden via the use of informative probabilistic priors.