Structured illumination can be used to form images without using a lens or a detector array. A series of spatially-structured laser pulses are cast on the scene of interest, and a single-detector power measurement is made on the light each pulse returns from the scene. There has been significant interest in the "ghost imaging" configuration, in which the spatial patterns are randomly generated---e.g., by driving the pixels of a spatial light modulator with independent, identically-distributed pseudorandom inputs---and the sequence of measurements is correlated with reference versions of those patterns to image the scene. This naive reconstruction, however, is far from optimal for standoff imaging, for which rough-surfaced objects create laser speckle in the measurements. We have developed a graphical model that encompasses the probabilistic relationships in structured-illumination standoff imaging along with an approximate message-passing algorithm for belief propagation to perform optimal image reconstruction. This approach lets us accurately model the statistics of speckled images, photon detection, and atmospheric turbulence, as well as incorporate intelligent priors for the scene that capture the inherent structure of real-world objects. The result is state-of-the-art image reconstructions.
Thesis Supervisor: Professor Jeffrey Shapiro