Adaptive multi-scale YOLO framework with context-aware attention for robust ship detection in SAR imagery

Synthetic Aperture Radar (SAR) is one of the primary sensing models and is broadly used in numerous remote sensing applications including environmental monitoring, maritime monitoring, and climate change research. Unfortunately, ship detection remains difficult due to different sizes of ships, complex background, and noise effects in the coastal areas. In order to overcome such issues in the domain, this work proposes a new DL solution, i.e., YOLO with Shuffle Reparametrized Blocks and Dynamic Head (YOLO-SRBD), built upon the architecture of YOLOv8. In particular, the proposed YOLO-SRBD algorithm uses channel shuffle reparametrized convolution blocks for efficient feature extraction. Additionally, a dynamic detection head is incorporated into the design in order to detect multi-scale targets. Experimental results obtained through experiments conducted using the SAR high-resolution ship dataset indicate that the presented methodology outperforms the existing YOLOv8 system. In particular, while the increase in detection accuracy is marginal (from 89.9% to 91.3%), the average precision significantly improved (from 66.7% to 74.3%).