Underwater UXO classification using Matched Subspace Classifier with synthetic sparse dictionaries

Abstract

UNDERWATER UXO CLASSIFICATION USING MATCHED SUBSPACE CLASSIFIER WITH SYNTHETIC SPARSE DICTIONARIESClassification of underwater objects such as unexploded ordnances (UXO) and mines from sonar datasets poses a difficult problem. Among factors that complicate classification of these objects are: variations in the operating and environmental conditions, presence of spatially varying clutter, variations in target shape, composition, orientation and burial conditions. Furthermore, collection of large quantities of real and representative data for training and testing in various background conditions is very difficult and impractical in many cases. In order to remedy the lack of data availability, physical models of varying computational complexity are often used to supplement training databases with synthetically created samples which predict the response of known target models.In this thesis, we try to address several key questions for designing robust classifiers for UXO and munitions classification from low frequency sonar. These include: (1) "How can we form discriminative and highly separable features for describing UXO and non-UXO objects in a given dataset?", (2) "When do we reach a point of diminishing returns when utilizing synthetic models in a classifier's training?", and more importantly (3) "Which types of object variations cannot be modeled well by synthetic data?". Although, it may be somewhat ambitious to expect model data to capture all the essential features of proud or buried underwater objects for target characterization, these models can nevertheless provide us with clues on how to augment the training datasets to improve the robustness in different environmental conditions.ii Using empirically validated scattering models developed by University of Washington's Applied Physics Laboratory (APL-UW), fast ray models were acquired to generate the required synthetic training dataset for various UXO and non-UXO objects. A comprehensive analysis is then carried out on the classification performance of two subspace matching classifiers, trained on the synthetic data generated from this physical model, and tested on three real underwater sonar datasets. Both single and multi-aspect classification were considered using a combination of linear subspace models. Our classification hypothesis is that the spectral content of sonar backscatter display unique signatures providing good discrimination between different classes of objects. To develop a robust target classification method that can be applied to discriminate munitions from non-hostile man-made objects and competing natural clutter, the Matched Subspace Classifier (MSC) framework was adopted in conjunction with multidimensional Acoustic Color (AC) feature data extracted from raw sonar returns.Classification results of the MSC system constructed using two different signal subspace learning methods, namely K-SVD and locality preserving (LP) K-SVD are presented and benchmarked against each other. Additionally, a non-linear version of MSC using the...