Hardware-Optimized Minimum-Search for SAR Backprojection Autofocus on FPGAs

Optimized Minimum-Search for SAR Backprojection Autofocus on GPUs Using CUDA

“…The image quality is linearly correlated with the iteration number, matching the simulations of Fahnemann et al [3]. As expected, configurations leading to the same number of effective samples, like Ŝ = 64, R = 1 and Ŝ = 8, R = 2 result in about the same image quality, but differ in runtime, since the latter only evaluates 2 × Ŝ = 16 samples.…”

“…Starting with Φ 0 = 0, φ 0 is varied to minimize the cost function, while keeping the other elements constant. Φ 0 is then updated with the found value, and the process is B. Optimization for parallel architectures [3] As presented in [3], several properties of the presented algorithm can be exploited to reach an efficient implementation on massively parallel architectures like FPGAs or GPUs.…”

“…2) Quadratic interpolation-based minimum search: Different optimization algorithms can be used to find the optimal value for φ m in each step of the autofocus algorithm. In [3] a quadratic interpolation-based minimum search is suggested, which allows parallel evaluation of different candidate values. This is well-suited to the inherently parallel architecture of FPGAs and can also be exploited on GPUs.…”

“…The cost function kernel evaluates all Ŝ sampling points iteratively. Similar to the proposed FPGA architecture [3], this saves memory bandwidth, as the values for g u,v and ĝu,v have to be loaded only once. The iterative processing makes the kernel runtime dependent on Ŝ.…”

“…Fahnemann, et al [3] showed, that the autofocus algorithm by Duersch and Long is very well suited for the application of FPGAs and proposed hardware-related optimizations for an efficient parallel execution, termed Parallel Autofocus Optimization (PAFO).…”