Accelerated fixed-point iterative reconstruction for fiber borescope imaging.

Area of Science:

• Computational imaging
• Embedded systems
• Image reconstruction algorithms

Background:

• Embedded vision systems face limitations in computing capacity and parallelism.
• Iterative image reconstruction algorithms typically require high-precision floating-point processors, demanding substantial computational resources.
• Existing embedded processors struggle to meet the demands of complex image reconstruction tasks.

Purpose of the Study:

• To develop an efficient fixed-point iterative solver for image reconstruction suitable for embedded systems.
• To implement and evaluate the solver on a field-programmable gate array (FPGA) for enhanced performance.
• To demonstrate the solver's effectiveness in correcting artifacts in fiber borescope imaging.

Main Methods:

• Developed a regularized Landweber fixed-point iterative solver.
• Implemented the solver on a field-programmable gate array (FPGA).
• Evaluated reconstruction accuracy and computational speed against floating-point uniprocessors.

Main Results:

• Achieved 1 to 2 orders of magnitude acceleration compared to floating-point embedded uniprocessors.
• Maintained comparable reconstruction accuracy with a similar number of effective iterations.
• Successfully corrected artifacts in fiber borescope images.

Conclusions:

• Fixed-point iterative solvers on FPGAs offer a significant acceleration for embedded computational imaging.
• This approach provides a power-efficient and cost-effective solution for high-performance image reconstruction.
• The developed method is effective for real-world applications like fiber borescope imaging.