Advances in Detection and Classification of Underwater Targets using Synthetic Aperture Sonar Imagery

Abstract

In this PhD thesis, the problem of underwater mine detection and classification using synthetic aperture sonar (SAS) imagery is considered. The automatic detection and automatic classification (ADAC) system is applied to images obtained by SAS systems. The ADAC system contains four steps, namely mine-like object (MLO) detection, image segmentation, feature extraction, and mine type classification. This thesis focuses on the last three steps. In the mine-like object detection step, a template-matching technique based on the a priori knowledge of mine shapes is applied to scan the sonar imagery for the detection of MLOs. Regions containing MLOs are called regions of interest (ROI). They are extracted and forwarded to the subsequent steps, i.e. image segmentation and feature extraction. In the image segmentation step, a modified expectation-maximization (EM) approach is proposed. For the sake of acquiring the shape information of the MLO in the ROI, the SAS images are segmented into highlights, shadows, and backgrounds. A generalized mixture model is adopted to approximate the statistics of the image data. In addition, a Dempster-Shafer theory-based clustering technique is used to consider the spatial correlation between pixels so that the clutters in background regions can be removed. Optimal parameter settings for the proposed EM approach are found with the help of quantitative numerical studies. In the feature extraction step, features are extracted and will be used as the inputs for the mine type classification step. Both the geometrical features and the texture features are applied. However, there are numerous features proposed to describe the object shape and the texture in the literature. Due to the curse of dimensionality, it is indispensable to do the feature selection during the design of an ADAC system. A sophisticated filter method is developed to choose optimal features for the classification purpose. This filter method utilizes a novel feature relevance measure that is a combination of the mutual information, the modified Relief weight, and the Shannon entropy. The selected features demonstrate a higher generalizability. Compared with other filter methods, the features selected by our method can lead to superior classification accuracy, and their performance variation over different classifiers is decreased. In the mine type classification step, the prediction of the types of MLO is considered. In order to take advantage of the complementary information among different classifiers, a classifier combination scheme is developed in the framework of the Dempster-Shafer theory. The outputs of individual classifiers are combined according to this classifier combination scheme. The resulting classification accuracy is better than those of individual classifiers. All of the proposed methods are evaluated using SAS data. Finally, conclusions are drawn, and some suggestions about future works are proposed as well.