Hartung, Steven Frederick (2013) High speed, high volume, optimal image subtraction for large volume astronomical data pipelines. PhD thesis, James Cook University.

Preview

PDF (Thesis) - Submitted Version Download (9MB)

Abstract

This thesis examines the application of massively parallel processing to the computationally intensive method of image comparison known as Optimal Image Subtraction (OIS). Of particular interest is the Dirac delta function basis spatially-varying OIS, a technique particularly applicable to images with asymmetric and non-Gaussian point spread functions (PSFs). This study demonstrates the ability to accelerate the spatially-varying convolution components of OIS by two to three orders of magnitude. This order of speed-up is needed to apply these techniques to real-time analysis of gigapixel survey telescopes that acquire an image every 30-60 seconds.

Many useful image processing techniques in astronomy require a massive number of computations per pixel. Among them is an image differencing technique known as OIS, which is very useful for detecting and characterizing transient phenomena. Like many image processing routines, OIS computations increase proportionally with the number of pixels being processed, and the number of pixels in need of processing is increasing rapidly.

Image differencing in astronomy, also commonly referred to as image subtraction, is a tool for transient object discovery and characterization. It is applicable to a wide variety of astronomical studies involving asteroids, extra-solar planet transits, variable stars, AGNs, supernovae, and almost any other object or phenomena that change in photometric flux intensity or position on human time scales (from sub-second to decadal intervals between observations). Effective image subtraction requires matching PSFs between images of the same field taken at different times using a convolution technique. OIS uses a fitting method to evaluate the changes in PSF between the two images to be subtracted, and then applies a best-fit convolution kernel prior to subtraction. Particularly suitable for large-scale images is a computationally intensive spatially-varying kernel. The underlying algorithm is inherently massively parallel due to unique kernel generation at every pixel location. The spatially-varying kernel cannot be efficiently computed through the Convolution Theorem, and thus does not lend itself to acceleration by Fast Fourier Transform (FFT).

Computing power continues to grow in approximate agreement with Moore's Law (~2x every 1.5-2 years), and astronomy data acquisition rates are growing as fast or faster. Driven by advances in camera design, image sizes have increased to over a gigabyte per exposure in some cases, and the exposure times per image have decreased from hours to seconds, or even down to video frame rates of multiple exposures per second. All of this is resulting in terabytes of science data per night in need of reduction and analysis. New telescopes and cameras already under construction will increase those rates by an order of magnitude. The application of emerging low-cost parallel computing methods to OIS and other image processing techniques provides a present and practical solution to this data crisis. Utilizing many-core graphical processing unit (GPU) technology in a hybrid conjunction with multi-core CPU and computer clustering technologies, this work presents a new astronomy image processing pipeline architecture. The resulting pipeline can process standard image calibration and differencing in a fashion that is scalable with the increasing pixel volume. Acceleration of several processing operations implemented to date has demonstrated an order of magnitude improvement over many existing publicly available codes, and up to three orders of magnitude in the case of the spatially-varying convolution in OIS. With the increased computing power available in mind, new techniques are also presented to improve subtraction quality by limiting noise inducing side-effects, addressing common problems with very bright stars, and dealing with special complications presented to OIS by drift-scan cameras.

This work presents results of accelerated implementation of the spatially-varying kernel image convolution in multi-cores with OpenMP and many-core GPUs. Typical speedups over ANSI-C were a factor of 50 and a factor of 1000 over the initial IDL implementation, demonstrating that the techniques are a practical and high impact path to terabyte-per-night image pipelines and petascale processing.

A new technique of image PSF cross-correlation is introduced. Implemented here as a means of limiting noise amplification by the preferred method of OIS. The new technique requires a similar computing load to the existing OIS for each input image, effectively tripling the computing requirement. However, considering the acceleration achieved through parallel programming, this is no longer a major concern. The PSF cross-correlation method also has significant potential as a far more general method for improving signal-to-noise for point-source targets in images not intended for OIS subtraction.

The techniques developed for multi-core CPU and GPU have been combined with a standard computer clustering technology in order to provide unprecedented scalability in an image subtraction application. The application is called ISAAC Petascale Image Processing (IP2) and is being developed in collaboration with the Infrastructure for Astrophysics Applications Computing (ISAAC) project within the International Center for Computational Science (ICCS) at the University of California at Berkeley (UCB) and the Lawrence Berkeley National Laboratory (LBNL).

Item ID:	30043
Item Type:	Thesis (PhD)
Keywords:	image differencing; image subtraction; transient object discovery; data management; astronomy image processing pipeline architecture; ISAAC Petascale Image Processing
Date Deposited:	04 Nov 2013 01:19
FoR Codes:	02 PHYSICAL SCIENCES > 0201 Astronomical and Space Sciences > 020102 Astronomical and Space Instrumentation @ 100%
SEO Codes:	97 EXPANDING KNOWLEDGE > 970102 Expanding Knowledge in the Physical Sciences @ 50% 97 EXPANDING KNOWLEDGE > 970108 Expanding Knowledge in the Information and Computing Sciences @ 50%
Downloads:	Total: 4573 Last 12 Months: 12
	More Statistics