YOLO-SR: An optimized convolutional architecture for robust ship detection in SAR Imagery

Accurate and efficient ship detection in synthetic aperture radar (SAR) imagery remains a challenging task due to speckle noise, scale variations, and the low contrast of small vessels. In this work, we present YOLO-SR, an enhanced version of YOLOv10 tailored for SAR ship detection, introducing four key innovations: Balanced Detail Fusion (BDF), C2f‐MSDR, DySample, and the Focaler-SIoU loss. Our BDF module adaptively merges shallow, fine‐grained features with deeper semantic features, preventing subtle ship signatures from being overshadowed by irrelevant clutter. Concurrently, C2f‐MSDR replaces standard bottleneck layers with multi-scale dilation residual blocks, expanding the receptive field to handle wide variations in ship size. To improve spatial resolution and retain boundary details, we incorporate DySample, a data-driven upsampling strategy that counteracts the artifacts of naive interpolation. Finally, Focaler-SIoU refines bounding-box regression by integrating distance, orientation, shape, and a focal-like reweighting, thereby emphasizing difficult, small, or partially occluded ships. Experimental results on SAR ship detection datasets confirm that YOLO-SR outperforms state-of-the-art methods in both precision and recall, while retaining competitive inference speeds. These advances offer a robust framework for real-time maritime surveillance, enhancing the detection of both small and large ships under challenging SAR conditions.