CNN Architectures on FPGA for Maritime Object Detection: A Review of Methods and Challenges

Abstract

Optical maritime surveillance requires real-time ship detection under severe power and latency constraints. Convolutional neural networks (CNNs) provide state-of-the-art accuracy, yet their compute and memory footprints remain challenging on embedded platforms. Field-Programmable Gate Arrays (FPGAs) address this regime by enabling low-precision arithmetic, deep pipelining, and on-chip data reuse. This review synthesizes CNN architectures for optical ship detection and their realization on FPGAs. It contrasts layer-by-layer and streaming execution styles, surveys lightweight detector choices suited to edge constraints, and collates acceleration mechanisms such as quantization, pruning, operator fusion, and memory-centric tiling. It then compares reported systems in terms of throughput, latency, and energy. Finally, it points open challenges including dataset shift and transferability, environmental factors, hardware software co-design and articulates persistent open problems relevant to maritime perception on embedded FPGAs.