Kelly, Kevin F.2021-11-232021-122021-11-17December 2Giljum, Anthony T.. "Versatile Compressive Sensing: Extensions of Random Projections for Imaging and Machine Vision." (2021) Diss., Rice University. <a href="https://hdl.handle.net/1911/111676">https://hdl.handle.net/1911/111676</a>.https://hdl.handle.net/1911/111676Compressive imaging is a versatile approach to sensing that offers the potential to dramatically reduce data acquisition time by multiplexing the scene with a spatial light modulator below than the Nyquist rate. The trade-off, however, is a high reconstruction time that limits real-time applications. The first half of this thesis will focus on multiplexed single-pixel imaging techniques. To this end, I will give a method for speeding up image reconstruction by foveating the image anywhere in the scene with regions of interest determined entirely after measurement acquisition. The technique is then extended to hyperspectral video and neural network reconstruction. Then, I will describe the use of shallow neural networks to map from random, task-insensitive measurements to highly targeted, task-specific measurements without sacrificing reconstruction capability. This work improves single-pixel object classification accuracies even when the object is against a complex, unknown background. I will then demonstrate an extension of these ideas to hyperspectral video in sum-frequency generation microscopy by measuring the localized blue shift in a sample of CO on a Pt surface as the electrochemical potential is swept. The second half of this thesis focuses on focal-plane array imaging. First, a diffraction limited, wide field-of-view hyperspectral microscope design will be shown. The proposed system is capable of acquiring images at 3 megapixels with over 100 spectral bands in under 30 seconds through spectrally multiplexed illumination. As an example application, the microscope is used for reflection contrast spectroscopy to measure the thickness of very thin graphene and MoS2. Finally, a novel snapshot hyperspectral camera design is shown, which uses a random Bayer filter and invertible blurring to compressively capture the entire hyperspectral datacube at the camera's pixel resolution in a single image. The proposed snapshot camera is compact and requires no moving parts or spatial-light modulatorsapplication/pdfengCopyright is held by the author, unless otherwise indicated. Permission to reuse, publish, or reproduce the work beyond the bounds of fair use or other exemptions to copyright law must be obtained from the copyright holder.Compressive sensingHyperspectral imagingMicroscopyComputational PhotographyComputer VisionSingle-Pixel ImagingThesis2021-11-23