|Ph.D Student||Levis Aviad|
|Subject||Volumetric Imaging of the Natural Environment|
|Department||Department of Electrical Engineering||Supervisor||Professor Yoav Schechner|
|Full Thesis text|
Volumetric imaging is an inverse problem that seeks to retrieve a three-dimensional (3D) density distribution from projections. This is done by fitting an image-formation model, or a synthetically rendered image, to acquired images. X-ray Computed Tomography (CT), for example, is a well-established volumetric imaging approach that is widely used for medical imaging and other applications. In this type of imaging scenario, the medium is fixed and the radiation source is controlled, resulting in a linear image formation model. We develop several generalizations to the linear CT that enable taking volumetric imaging out of the lab to image the natural environment. Our approach is computational photography with complex image formation models, which due to increased computing power and algorithmic advances, can be inverted. We consider imaging scenarios where the lights source is uncontrolled, the medium is not fixed and the projections are non-linear.
Non-linear projections and an uncontrolled light source arise from the task of recovering the 3D volumetric scatterer distribution of the atmosphere. Atmospheric scatterers, such as clouds and aerosols, play a crucial role in the earth's radiation balance impacting all life on earth. We develop novel algorithmic approaches to sense them as they are, in 3D. We exploit light that undergoes multiple-scattering interactions by fitting multi-view multi-spectral intensity and polarization measurements to a radiative transfer image formation model. We hypothesize that our inverse-scattering approach can be applicable outside of atmospheric science to a wide variety of fields including computer vision, graphics, and medical imaging.
We further look into the ocean, to in-situ imaging of plankton. Here, we develop an approach for self-calibrating volumetric imaging of transparent objects. This enables the extraction of statistics that are important for the scientific study of specimen populations, specifically size-distribution. We generalize tomographic recovery to account for all degrees of freedom of a similarity transform and make the computational load manageable to reach good quality reconstructions in a reasonable time.