Simultaneous Blind De-Convolution of Repeated Astronomical Exposures

Johns Hopkins University, Baltimore MD

Investigators

Abstract

Images are ubiquitous throughout the scientific endeavor. Although many pictures are perfectly adequate in their original form, there are many scientific fields where single frames are not sufficient. Traditional methods of improving such results to enable research with them include combining multiple frames, which often reduces the final quality to around that of the lowest common denominator, and taking long sequences from which only the best are selected and combined, thus throwing away a lot of information. This is a project to find new ways through statistics and image processing to combine repeated exposures to produce images of superior quality, with less blur and higher resolution, and to carry this out with automatic pipeline processing. The value of high quality images that retain the maximum amount of information from all the painstakingly assembled data is inestimable, and will impact pretty well every field of science, including amateur activities and citizen science projects, and will be generally useful in any situation where blurred or faint pictures are a limiting factor. Imaging detectors of increasing size and complexity are nowadays the primary source of data in many scientific experiments. The images they produce, however, are often blurred by distortions that can change rapidly, such as the atmosphere between astronomical objects and telescopes. Eliminating such effects with hardware is either extremely complex (e.g,. adaptive optics) or extremely expensive (e.g., space-based observatories). To detect fainter signals requires multiple exposures, which are traditionally combined by convolving to the lowest acceptable quality, but doing that throws away a lot of the information in the images. There has to be a better way. This project will develop new methodologies in computational statistics and image processing for the optimal combination of repeated exposures to produce images of superior quality, having minimal blur and higher resolution than the originals. This requires developing novel Bayesian algorithms and industrial-strength scalable software. The study combines statistics, computer science and data-intensive science into a focused, powerful research program that should lead to a significant leap in image quality. The new idea is potentially game changing, is particularly well suited to low signal-to-noise images, and should help in the design of new experiments and observing strategies. The algorithms and tools developed will be directly applicable to other research fields as diverse as meteorology and genomics, and will be open-source and publicly available. These next-generation data challenges will be especially valuable training for the graduate student who will be doing much of the work.