Fast and accurate AM-FM demodulation of digital images with applications.

摘要

Multidimensional Amplitude-Modulation Frequency-Modulation (AM-FM) models allow us to describe continuous-scale modulations in digital images. In this dissertation; we develop new, one-dimensional and two-dimensional AM-FM demodulation algorithms that provide significant improvements in both accuracy and speed over previously reported non-parametric approaches. An extension to any finite number of dimensions is straight-forward and is briefly discussed. For the case of unknown constant amplitude and constant frequency sinusoids, we show that the proposed, one-dimensional, non-parametric algorithm computes amplitude and frequency estimates with variances that approach the Cramer-Rao Lower Bounds. In a number of cases, the accuracy improvements are in the range of one to two orders of magnitude. Results are shown for both real and synthetic images. We also present two applications to medical imaging: M-mode echocardiography image segmentation and AM-FM based feature extraction of ultrasound carotid artery images for classification. For the M-mode segmentation application, the proposed AM-FM method led to automated video segmentation of M-mode videos coming from different probes. For the classification of ultrasound carotid artery images, the results are at least as good as those from state of the art, multi-feature, multi-classifier systems.; For fast AM-FM demodulation, we present separable implementations that take advantage of Single Instruction Multiple Data (SIMD) components that are present in modern, general-purpose architectures. Moreover, we present a novel approach for the development of SIMD algorithms and show that algorithms taking advantage of the basic model of the underlying SIMD architecture can lead to a significant time performance improvements over current, state of the art, methods for large datasets. Furthermore, we show that the novel SIMD-FFT algorithm, the flagship algorithm developed under this new framework, exhibits significant improvements over widely accepted state of the art standards.