Higher-Dimensional Signal Processing via Multiscale Geometric Analysis

摘要

This project pursued a general theory for complex-valued multiscale signal and image modeling, processing, and analysis that is matched to singularity-rich data. Higher-dimensional signals that feature geometric manifold structures were of particular interest in developing theory and a practical toolset for analysis and processing. We pursued a three-pronged approach in creating new multiscale transforms, new geometric statistical models, and new manifold-based signal representations. The results of our research include (1) the Dual Tree Quaternion Wavelet, an efficient transform and analysis tool that features near shift invariance and linear computational complexity; (2) a geometric hidden Markov tree wavelet model, which accounts for geometric regularity by capturing the dependencies between complex wavelet coefficients along a contour; and (3) surflet representations of signal discontinuities with near optimal rate-distortion performance. These new tools have led to significant performance gains immediately applicable to a number of important Navy-relevant applications, including target detection and classification, image segmentation and fusion, and computer network traffic modeling.