Abstract: Detection and classification of underwater objects in sonar imagery is a complicated problem due to various factors such as variations in operating and environmental conditions, presence of spatially varying clutter, variations in target shapes, compositions, and orientation. This work focuses on coherence-based detection methods for the detection of objects from underwater sonar imagery. A new system-level solution is developed using Gauss-Gauss detection. The standard one channel Gauss-Gauss detector is extended to the two channel case allowing for the the extension to using multiple disparate platforms for detection. The use of multiple sensors allows for significantly better capture of the target characteristics due to the fact that the targets are viewed from different aspects, grazing angles, ranges, frequencies, and sensor modalities and therefore providing an improvement in the detection performance. The two-channel Gauss-Gauss detector is then cast in the canonical correlation analysis framework (CCA). The CCA framework provides an ideal framework for simultaneous coherent detection and feature extraction of target attributes that present in both sensors. These extracted canonical correlations for each pair of regions of interest (ROI's) within a sonar image provide a coherence (or incoherence) measure that can be used to determine if a target is present (or absent) in the processed ROI's. Our detection hypothesis in this multi-sensor coherence-based detector is that presence of objects in the multi-platform sonar data leads to high level of mutual information or coherence measure comparing to that of the background clutter only.
A comprehensive study is carried on the coherence-based detector and a comparison of the detection and false alarm rate performance for two different data sets with different background difficulties is done. Secondly a study is carried out on how the Gauss-Gauss detector performs when the sample support becomes poor and a kernel (non-linear) version of the detector is developed. The data used in this study was collected by Naval Surface Warfare Center in Panama City, FL and consists of two data sets. The first one contains high frequency side scan sonar images obtained from one sensor with varying degree of difficulty and bottom clutter. The second one contains multiple sensors with images registered over the same target field with varying degree of difficulty and bottom clutter. The data-set contains one high resolution sonar and three broadband sonars that are co-registered over the target area giving multiple looks at the target region. Results illustrating the effectiveness of the proposed detection tool will be presented in terms of probability of detection, false alarm, and correct detection rates for various bottom difficulty types.
Adviser: Dr. Mahmood Azimi-Sadjadi Co-Adviser: N/A Non-ECE Member: Dr. F Jay Breidt (Statistics) Member 3: Dr. J. Rockey Luo Addional Members: N/A