MWIR FPAs often contain a non-negligible number of dead pixels, and even clusters of pixels. In addition to the undesirable cosmetic effect that these flaws have on the output image, it can also be detrimental to downstream image exploitation efforts, including target detection and tracking algorithms. It is therefore necessary to mask such defects as early as possible, before it is exacerbated by the imaging pipeline processes, especially sharpening filters and contrast stretching, which are commonly used. This paper presents the results of an investigation into several dead-pixel replacement schemes of varying complexity, starting with simple replacement by the previous neighbor to interpolation using kernels of varying shapes and domain sizes. These are evaluated for accuracy, but also for their cluster-handling performance, latency impact, and suitability for real-time implementation on resource-constrained FPGA matrices. It is shown that asymmetric predictive kernels, optimized using neural nets or genetic algorithms, can offer significant improvements over naïve last-good-neighbor replacement, while affording improved cluster-handling capability, low FPGA resource requirements, and incurring minimal and even zero latency.
|