A VHDL CNN Accelerator on Zynq-7020 for Embedded Maritime Ship Recognition

Abstract

Unmanned surface vehicles (USVs) increasingly rely on vision-based perception to support safe navigation and maritime surveillance, yet onboard computing must operate under strict size, weight, and power (SWaP) constraints. Deep convolutional neural networks (CNNs) can provide strong recognition accuracy, but their compute intensity and memory footprint are difficult to fit within low-cost embedded platforms. This work investigates an energy-aware deployment strategy for embedded ship recognition using a heterogeneous ARM–FPGA system based on the Xilinx Zynq-7020 SoC. A hardware–software co-design approach is adopted to enable near-sensor processing: CNN inference is executed in the FPGA fabric through a streaming-oriented accelerator, while the ARM processing system provides lightweight configuration, scheduling, and DMA-driven data movement under a PYNQ runtime. The methodology includes developing a custom VHDL accelerator and benchmarking it with representative classification backbones trained for maritime ship recognition, including VGG16 as a widely used reference model. Experimental evaluation on the Zynq-7020 platform shows that the proposed co-design delivers practical inference performance with modest FPGA resources and maintains a balanced trade-off between throughput, resource usage, and power consumption under tight embedded constraints. To contextualize deployment choices for edge maritime perception, the study also discusses the trade-offs between dedicated accelerators and general-purpose embedded computing in terms of latency, throughput, and energy-related efficiency. Overall, the results support the viability of compact Zynq-class devices as continuously operating perception modules for USVs and provide a solid basis for extending the onboard pipeline toward more advanced perception functions (e.g., tracking) in realistic maritime operating conditions.