Optimized Sampling Approaches for Compressive Sensing in Multi-Dimensional Datastreams

点击翻译按钮获取中文摘要

The current state-of-the-art in imaging hardware involves the very precise synthesis and fabrication of semiconducting materials into extended cameras that can now contain up to 64M pixels with a cost that can exceed £1M per device. In most cases, these high sensitivity cameras are implemented to detect signals that are very close to the noise level and as an added complexity are typically looking to characterise dynamic events (i.e. they need to be able to quantify the motion of fast moving objects). The data per image frame in these systems can easily exceed 1TB, meaning that cameras currently have to operate in short bursts, have delayed responses due to the extended transfer of the data, and it can take days/months/years for image analytics to operate and identify key elements in the datastream. Obviously as the global economy pushes towards more automation and the use of remote sensing devices, these limitations have to be overcome.One approach that can alleviate a large number of the problems associated with speed and precision in state-of-the-art imaging systems, is the use of Compressive Sensing (CS) methods. In the CS approach, a small subset of random pixels in the image in acquired and used to reconstruct the full dataset. This immediately reduces the amount of data and increases the imaging speed by the amount of sub-sampling that is used. The goal of this PhD project is to determine the level of sub-sampling that be used to reconstruct images from such diverse sources as satellites, night vision goggles and scanning transmission electron microscopes. By developing and implementing new algorithms for the specifics of the image contrast mechanism and its resolution limits, the goal is to develop a coherent framework that can be used in the design of optimized imaging hardware with embedded algorithms.