Deep Learning-based Ship Detection on FPGAs

Detecting ships from remote sensing imagery is a crucial application for maritime surveillance. Object detection algorithms based on Deep Neural Networks (DNNs), such as YOLO, have enabled sophisticated accuracy for ship detection tasks. However, the deployment of DNN models on embedded accelerators raises several issues, including performance degradation, latency delays, and energy consumption due to the computational complexity and model size of these models. Thus, given the limitations of computational resources in embedded devices, the question of how to determine the balanced tradeoff between performance, inference latency, and energy efficiency has been a concern. In this study, we demonstrate the feasibility of the Versal ACAP FPGA (VCK190) for YOLO-based ship detection compared to Jetson TX2 GPU. Our results show that despite providing similar performance, the FPGA board can predict a YOLO model in less time, roughly half the time it takes the GPU to implement the same model in addition to the superior power efficiency.