“Computable Performance Analysis of Sparsity Recovery”
Crime investigation TV shows, such as CSI, commonly feature a digital forensics laboratory capable of recognizing faces and vehicle license plates from extremely blurry shots. Photo and video evidence is displayed on a large projection screen while a recognition system attempts to identify the perpetrator’s identity.
This technology exists in research labs today thanks to advanced signal processing. Various developments in signal processing, particularly in sparsity-based image reconstruction, have recently emerged with the potential to dramatically improve system performance.
Prof. Arye Nehorai is a leader in this research area, and recently delivered a lecture on the “Computable Performance Analysis of Sparsity Recovery,” as part of the Department of Electrical and Computer Engineering Distinguished Lecture Series. Nehorai is the Eugene and Martha Lohman Professor and Chair of the Preston M. Green Department of Electrical and Systems Engineering at Washington University in St. Louis.
As part of his lecture, Prof. Nehorai discussed a movement within the signal processing community to update the classical framework based on the Nyquist-Shannon sampling theorem using a new approach known as compressive sensing. Compressive sensing makes it possible to acquire and represent signals using fewer samples than classical sampling methods, under the key assumption that the signal itself is sparse with respect to some basis. For instance, although a facial image is comprised of many, many pixels, it can still be accurately represented using just a few key features. Indeed, identification of a criminal based on a low-resolution, blurry image, while unthinkable a decade ago, is becoming increasingly viable in part due to modern image processing techniques based on compressed sensing. Other important applications include hyperspectral imaging and anomaly detection.
Prof. Nehorai’s recent work has focused on a challenging and important compressive sensing problem. In particular, while it is known that dramatic savings are possible via compressive sensing, it is often difficult to say exactly how many samples are required for a specific sampling scheme (or sensing matrix). Prof. Nehorai and his collaborators have developed a suite of efficient algorithms, based on convex programming that can rapidly ascertain the number of samples needed under a particular scheme. These algorithms can in turn be used to guide the development of better sensing schemes.
Nehorai’s talk was the first in the three-part Fall 2014 Distinguished Lecture Series. The next talk features Philippe Fauchet, Professor of Electrical Engineering, College of Engineering of Vanderbilt University. He will speak on the topic, “Nanoscale Silicon as an Optical Material.” His lecture will be held October 28, 2014 at 4 pm in PHO 210.
By Gabriella McNevin