Walter Scott, Jr. College of Engineering

Graduate Exam Abstract

Abstract: Detection and classification of buried underwater objects using broadband sonar is a challenging signal processing problem. Natural and/or manmade clutter on the seafloor corrupts the returned sonar signals, causing unavoidable problems in the detection and classification processes. Additionally, the environmental and operational changes encountered when moving from one site to another could be substantial, hence creating large changes in the feature space of both mine-like and non-mine-like objects. These issues, coupled with the necessity for capturing highly discriminatory signatures of mine-like versus non-mine-like objects in the sonar returns, are challenging tasks to overcome.

Feature extraction using multiple sonar returns can be extremely valuable for object classification as it allows for the exploitation of information regarding the characteristics of the object. This information helps in determining essential properties regarding the object's composition, size, shape, and density: all of which can be exploited to discriminate between mine-like and non-mine-like objects.

Among the existing methodologies for performing multi-ping classification are: (a) decision-level fusion, which uses a fusion scheme to combine multiple intermediate decisions obtained from a single-ping classifier to produce a final decision, (b) feature-level fusion, which exploits the information in consecutive sonar pings to make a combined decision, and (c) coherence-based feature extraction, which generates a set of salient features from two sonar pings, which are then classified by a single-ping classifier. The latter method alleviates the computational burden and overall complexity of the other fusion processes while simultaneously obtaining good classification results. This method uses Canonical Correlation Analysis (CCA) as a two-ping feature extraction procedure. Dominant canonical correlations between two sonar returns are used as features for discrimination of mine-like objects from non-mine-like objects. These canonical correlations directly measure the linear dependence, or coherence, between the two sonar returns. The coherence patterns of mine-like objects are found to be different than those of non-mine-like objects, hence providing excellent discrimination using only a simple classifier.

In this work, we investigate how to extend this useful two-channel CCA-based feature extraction to the multi-channel case. This extension is referred to as Multi-channel Coherence Analysis (MCA) throughout this thesis. The fundamental idea is to exploit more than two sonar returns in a coherence-based feature extraction with the hope to provide features that are more discriminatory than in the two-channel case. This robust and discriminatory feature set is subsequently used in the classification process to further improve the overall performance.

A new data-driven iterative learning algorithm for performing MCA is also developed that uses a deflation process on the data to extract multi-channel coordinates and corresponding correlations, one-by-one. This algorithm is particularly useful for applications where only a subset of the principle multi-channel coordinates/correlations is desired. The performance of the algorithm is experimentally demonstrated on a simulated data set. The results indicate that after enough iterations, the estimates of the multi-channel coordinate mapping vectors and the corresponding correlations approach their theoretical values within some acceptable numerical error bounds.

A thorough evaluation is then provided which compares the classification performance of various classifiers that use features obtained via CCA and MCA. The data used in this study was collected using the buried object scanning sonar (BOSS) system and contains sonar returns from many buried and proud mine-like and non-mine-like objects. Three different classification systems are benchmarked: a CCA-based single-ping classifier, an MCA-based single-ping classifier, and a CCA-based nonlinear decision-level fusion classifier. These classification systems are developed using back propagation neural networks (BPNN) which are trained and validated using data from a single run through the target field. The optimally trained systems are then tested on data sets from other runs throughout the target field. The performance of each classifier is evaluated in terms of classification rates, receiver operating characteristics (ROC) curves, and their abilities to provide simultaneous detection and classification on entire runs through the target field. It was observed that by using features extracted via MCA, as opposed to CCA, classification performance similar or better than that of the nonlinear decision fusion system can be achieved while simultaneously decreasing the overall classifier complexity.
Adviser: Dr. Mahmood R. Azimi-Sadjadi
Co-Adviser: Dr. Edwin K. P. Chong
Non-ECE Member: Dr. Robert Liebler, Mathematics
Member 3: N/A
Addional Members: N/A
Program of Study: