Exploring lightweight network design for ship detection under resource constraints based on YOLOv8

Synthetic Aperture Radar (SAR) provide all-weather, all-day imaging capabilities and broad-area coverage, making them well-suited for large-scale ship detection. However, their onboard hardware remains subject to constraints in both computational power and energy consumption. In previous ship detection tasks, improvements in model accuracy have often depended on increasingly complex architectures and higher computational costs, which limit the direct application of conventional neural networks on such resource-constrained platforms. To address this limitation, this study presents a lightweight redesign of the YOLOv8 framework. Specifically, depthwise separable convolutions are introduced to reduce parameter count and computational load, the Efficient Multi-Scale Attention (EMA) module is integrated to enhance object perception, and the Efficient Channel Attention (ECA) module is incorporated to strengthen multi-scale feature fusion without adding extra computational burden. Experimental results demonstrate that the proposed model, with 1.27M parameters and 4.35 GFLOPs, maintains reliable detection performance across various scenarios, including nearshore and offshore environments, achieving a balance between model complexity and detection accuracy.