GPU Image Processing

Introduction

The project aims to develop techniques to use computer graphics cards for high-performance execution of specific types of calculations in image processing, namely image correlation and image resampling. The aim is to provide a performance boost in critical areas for near real-time provision of GIS-ready satellite imagery.

Background

The number of new Earth Observation satellites is growing at an ever-increasing rate while the sensors they carry offer more and more detailed information. As a consequence, the amount of image data is rapidly expanding and it is Spacemetric's goal to be a major supplier of production systems for these new missions, a market in which very rapid and accurate processing of large volumes of imagery is an important differentiator. The goal is to minimise the overall elapse time from the arrival of the satellite image at a ground station to its availability in a map-ready form.
Practically all satellite image applications require map-conformal, or what can be called “GIS-ready” products as their input. These are images equivalent to maps where the relative and absolute positions of objects conform to a known map coordinate system at a known accuracy. In some cases, such as security and disaster response applications, these image-maps additionally need to be made available in as timely a way as possible, in principle within a few minutes of download from the satellite.

The opportunity

Significant performance improvements have been made in recent years in the processing of satellite imagery. Some of these improvements are simply a natural consequence of the continuing emergence of better computer hardware. However, other advantages offered by the latest hardware are only fully accessible through a fundamental re-engineering of critcal software processes. Fully exploiting even these possibilities is an established strategy for Spacemetric. For instance, Spacemetric has implemented true multi-threaded processing for image rectification so as to make full use of mutli-core CPUs. While this has helped give Spacemetric a technical lead, other performance enhancing technologies also need to be exploited in a similar way.
High-performance graphics cards are now universal in mid- to high-range computer hardware with the video gaming market being a major driver behind their development. Their architecture is specifically conceived for massively parallel processing of screen graphics. This was until recently their only application area, but the emergence of effective programming languages, such as CUDA, means that this processing capacity can now be used in other applications. Image correlation and image resampling are computationally intensive activities in the typical image processing chain where performance improvements have a direct impact on data timeliness. Other vendors of systems for image processing have also begun to make use of GPUs, so it is important that Spacemetric investigates and exploits this technology as fully as possible.