Ship Detection in SAR Images Using Attentive YOLOv8 Architecture

Use of remote sensing technology in recent years has seen a rapid growth especially by government organisations to identify objects in marine ecosystems in remote sensed images. Ship detection is one such task that has seen extensive experimentation inorder to build efficient techniques to produce reliable and accurate detection results. Ship targets in Synthetic Aperture Radar (SAR) images suffer from significant scattering, a complicated background, and ambiguous contour information. The goal of our work presented in this paper is to use the most recent iteration of the You only look once (YOLO) model to detect both inshore and offshore ships in SAR images with the introduction of attention mechanism. We utilised two datasets namely HRSID and our own custom datasets to train different version of YOLOv8, YOLOv9 and Yolov10 models along with our proposed model. The experimental outcomes showcase advancements, with the introduction of the Shuffle Attention mechanism resulting in a remarkable 6.72% increase in precision and a 3.58% boost in ${\mathrm {mAP}} \text{@}$ 0.5: 0.9 evaluation metrics when validated on custom dataset. Our study contributes to the refinement of lightweight baseline YOLO model, demonstrating the efficacy of incorporating shuffle attention mechanisms and a methodology to produce ship detection results on large resolution SAR images in smallest inference time with limited computing resources.