Parallel Optimisation and Implementation of a Real-Time Back Projection (BP) Algorithm for SAR Based on FPGA

Yue Cao, Shuchen Guo, Shuai Jiang, Xuan Zhou, Xiaobei Wang, Yunhua Luo, Zhongjun Yu, Zhimin Zhang, Yunkai Deng2022 · DOI

TL;DR

This paper presents a highly optimized, real-time Synthetic Aperture Radar (SAR) imaging system implemented on an FPGA. By developing a fully parallelized Back Projection (BP) algorithm architecture and leveraging FPGA's parallel processing capabilities, they achieve significantly faster image generation compared to existing methods, paving the way for more responsive and capable SAR systems.

What Problem Does This Paper Address?

Traditional SAR systems, while powerful, face a bottleneck in the image generation and information processing pipeline. This delay hinders real-time applications, especially for time-sensitive targets, and limits the ability to dynamically adjust platform flight or operating modes based on user feedback. The core challenge is to significantly accelerate the SAR image processing, specifically the computationally intensive Back Projection (BP) algorithm, to enable real-time performance on board airborne or spaceborne platforms. This requires overcoming the algorithm's high computational complexity and memory access demands.

Methods

The authors tackle the BP algorithm's computational burden by designing a fully parallel processing architecture specifically for FPGAs. Key methodological contributions include:

  • Algorithmic Flow Optimization: They redesigned the BP algorithm's flow to be more amenable to FPGA pipelining, reducing memory interactions and eliminating branch jumps that hinder parallel execution. This involved expanding iteration cycles and adopting a data-flow processing approach.
  • Fixed-Point Operation Optimization: To balance accuracy and resource usage, they meticulously analyzed the impact of different bit-widths on SINC interpolation and echo accumulation. They settled on a 15-bit fixed-point precision for critical calculations, finding it sufficient for the required accuracy while managing memory and processing demands.
  • Parallel Processing: The BP algorithm was parallelized across three dimensions:
  • Distance Calculation: 16 parallel threads were used.
  • SINC Interpolation: 8 parallel threads were employed for upsampling.
  • Floating-Point Multiplication and Addition: 8 parallel threads were used. This multi-dimensional parallelism, leveraging both SIMD and MIMD models, resulted in a significant increase in processing units per clock cycle.

  • Distributed Storage: To address the high data access throughput requirements, they implemented a distributed storage strategy combining DDR, SRAM, and registers. This approach strategically utilizes the strengths of each memory type to optimize data fetching and storage for different stages of the BP algorithm.

  • FPGA Implementation: The entire optimized algorithm was implemented on a Xilinx XC7VX690T FPGA, taking advantage of its inherent parallel processing capabilities and high clock speeds.

Datasets

The paper uses actual flight data for validation. The specific dataset details (e.g., source, acquisition parameters beyond range and azimuth size) are not explicitly detailed, but the results are demonstrated on a 900x900 pixel image.

Key Insights and Ideas

  • FPGA as a Game-Changer for Real-Time SAR: The study powerfully demonstrates that FPGAs, with their inherent parallelism and reconfigurability, are exceptionally well-suited for accelerating computationally intensive algorithms like BP for real-time SAR imaging.
  • The Power of Multi-Dimensional Parallelism: Breaking down the BP algorithm into independent computational dimensions and parallelizing each significantly amplifies processing throughput. The combination of 16-way, 8-way, and 8-way parallelism is a core enabler of their speedup.
  • Fixed-Point Precision is Key to Resource Management: The careful analysis of fixed-point precision highlights a crucial trade-off. Achieving high accuracy doesn't always require full floating-point precision, and selecting the right fixed-point representation (15-bit in this case) is vital for efficient FPGA implementation.
  • Distributed Storage is Non-Negotiable for High Throughput: The sheer volume of data movement in BP imaging necessitates a sophisticated storage architecture. The paper shows how combining different memory types (DDR, SRAM, registers) is essential for meeting the demanding data access bandwidth.

Key Results

  • Processing Time: The system achieved a processing time of 1.1 seconds for a 900x900 pixel SAR image at a 200 MHz clock rate.
  • Performance Comparison: This processing time is significantly faster than other platforms, including CPUs, GPUs, and DSPs, as shown in Table 4, where comparable image sizes take orders of magnitude longer.
  • Resource Utilization: The FPGA implementation achieved high resource utilization, with storage resources exceeding 80% and logic resources nearing 60%.
  • Power Consumption: The system consumed 21.073 W.
  • Image Quality: The output image quality was validated using metrics like PSNR (27.26 dB) and SSIM (0.8652), demonstrating distinct features and obvious focusing effects, comparable to MATLAB processing results.

Limitations and Future Work

  • Limitations:
    • The paper focuses on an airborne SAR scenario. While the technology is presented as applicable to spaceborne platforms, specific adaptations might be needed.
  • Detailed information about the specific flight data used for validation is limited.
  • Future Work:
    • The authors suggest that this technology lays the foundation for developing new, high-performance SAR systems for real-time imaging.
  • The work can be applied to solve multi-mode, multi-resolution, and multi-geometry signal processing problems in an integrated manner, implying future extensions to handle more complex SAR scenarios.

Summary updated

BibTeX
@misc{cao2022parallel,
  author = {Yue Cao and Shuchen Guo and Shuai Jiang and Xuan Zhou and Xiaobei Wang and Yunhua Luo and Zhongjun Yu and Zhimin Zhang and Yunkai Deng},
  title = {{Parallel Optimisation and Implementation of a Real-Time Back Projection (BP) Algorithm for SAR Based on FPGA}},
  year = {2022},
  doi = {10.3390/s22062292},
}
Figure