Using Graphics Processors to Accelerate Synthetic Aperture Sonar Imaging via Backpropagation

This paper describes the use of graphics processors to accelerate the backpropagation method of forming images in Synthetic Aperture Sonar (SAS) systems. SAS systems coherently process multiple pulses to provide a higher level of detail in the resolved image than is otherwise possible with a single pulse . Several models are available to resolve an image from the pulse return data; the backpropagation model is the most accurate and flexible, but also the most computationally intensive. Less flexible and accurate algorithms are frequently used because the time to resolve an image via backpropagation is held to be intolerable. A GPU-based implementation of backpropagation was developed at GTRI and inserted into sonar and radar algorithm research testbed systems. The GPU accelerated implementation formed a 4000 x 4400 SAS image from 60 seconds of sonar data in 7 seconds using 8 GPUs. This was 275x faster than a C-based implementation executing on an 8-core i7 platform, and provides an otherwise timeprohibitive technique to sonar and algorithm researchers for use in prototyping environments I. SONAR BACKPROPAGATION The classical approach to both radar and sonar synthetic aperture image reconstruction is known as backprojection, or backpropagation. The scheme is straightforward. Each point in the scene being imaged contributes reflections to throughout the recorded data. Any point has a corresponding locus of echo returns in the observed data. In order to compute the value of a single output pixel in the reconstructed image, all that is required is to integrate the data along this locus while multiplying by the complex conjugate of the expected locus. For each output point this operation has the form of an inner product, and the reconstructed image can the thought of as resulting from a spatially-varying correlation operation. The inner product serves as a measure of how similar the measured data is to the expected locus, and a strong correspondence results in a bright output pixel. To obtain the equation for the backpropagation, consider the ideal locus for a single point scatterer at [ ] x y z  x , denoted by ( , , ) t u x  , which also depends on the time of recording t and the alongtrack position of the sensor u . The integration is performed over the locus given by the curve L :