ReCoN: A Reconfigurable CNN Acceleration Framework for Hybrid Semantic Segmentation on Hybrid SoCs for Space Applications

Abstract

Recent advancements in deep learning present new opportunities for enhanced scientific methods, autonomous operations, and intelligent applications for space missions. Semantic segmentation is a powerful computer-vision process using convolutional neural networks (CNNs) to classify objects within an image. Semantic segmentation has numerous space-science and defense applications, from semantic labeling of Earth observations for insights about our changing planet, to monitoring natural disasters for damage control, to gathering intelligence for national defense and security. Despite these advantages, CNNs can be computationally expensive and prohibited on traditional radiation-hardened space processors, which are often generations behind their commercial-off-the-shelf counterparts in terms of performance and energy-efficiency. FPGA-based hybrid System-on-Chips (SoCs), which combine fixed-logic CPUs with reconfigurable-logic FPGAs, present numerous architectural advantages well-suited to address the computational capabilities required for high-performance, intelligent spacecraft. To enable semantic segmentation for on-board space processing, we propose a hybrid (hardware/software partitioned) approach using our reconfigurable CNN accelerator (ReCoN) for accelerating CNN inference on hybrid SoCs. When evaluated on the Xilinx Zynq SoC and Xilinx Zynq UltraScale+ MPSoC platforms, our hybrid approach demonstrates an improvement in performance and energy-efficiency up to two orders of magnitude compared to a software-only baseline on the hybrid SoC. Furthermore, fault injection and wide-spectrum neutron beam-testing was performed to characterize the ReCoN architectural response to injected errors and susceptibility to neutron irradiation.