Image analysis for realistic electromagnetic imaging systems

Abstract

This thesis focuses on the choice and implementation of different image and signal processing algorithms adapted to address specific hardware challenges in realistic electromagnetic imaging systems and their applications. A wide range of imaging systems and frequencies with related image processing needs is studied: (1) low-frequency medical brain activity imaging at kHz frequencies for epileptic seizure detection; (2) near-field microwave scanning in the several hundreds of MHz range for non-destructive silicon chip defect detection; (3) Synthetic Aperture Radar (SAR) in the 8-12GHz band for automatic focusing of ground maps; (4) multispectral infrared and visible images for embedded target detection; and (5) passive broadband imaging from 100GHz to several THz for concealed weapon detection.

As with many imaging systems and object recognition applications, there exists a need for pre-processing raw data provided by the imager to improve the quality of the measured scene. In the cases studied in this work, the following limit the image quality and processing requirements: high data dimensionality; low signal-to-noise ratio; scanning position drifts; unknown target characteristics; low number of pixels; broadband integration; and detector non-uniformities. In the widely published fields of image processing and imaging systems, the design of image processing algorithms that match hardware limitations has been lacking, and this is precisely what is addressed in this thesis. The techniques used in this work, as they apply to the five imaging systems listed above, include: (1) dimensionality reduction based on principal component analysis and Laplacian Eigenmaps, applied to epilepsy detection and multispectral IR imaging; (2) two dimensional interpolation using windowing and filtering, applied to near-filed scanning, SAR and THz imaging; (3) manifold learning and classification, applied to epilepsy detection, multispectral IR and THz imaging; (4) dynamic background correction and contrast enhancement, applied to THz imaging; and (5) thresholding, applied to data obtained by all five imaging systems. These image and signal analysis techniques are presented in this thesis as an approach to finding a set of solutions addressing hardware and sensor platform limitations. The trade-offs between performance and optimality of an algorithm solution were considered, with the need for pseudo-real time analysis in most cases.