SSC-YOLO11: An Improved YOLO11 Model for SAR Small Object Detection in Complex Backgrounds

Abstract

Synthetic aperture radar (SAR), characterized by its all-day and all-weather imaging capabilities, has been widely utilized in both military reconnaissance and civilian remote sensing domains. In recent years, deep learning techniques have achieved remarkable success in SAR image object detection. However, challenges such as missed detections and false alarms remain prevalent, particularly in scenarios involving small and multiscale objects. To address these issues, in this letter, we propose an improved small object detection model for SAR images in complex backgrounds based on YOLO11, named SSC-YOLO11, aiming to enhance detection accuracy and robustness in challenging SAR scenarios. First, an improved C3k2 module optimized with convolutional re-parameterization (CORP) is incorporated into the backbone network. By leveraging a multibranch structure and online convolutional re-parameterization (OREPA), this module significantly enhances feature extraction and representation. Second, an additional ultrashallow pyramid level 2 (P2) is introduced to better preserve details of small objects during downsampling, effectively improving the accuracy of small object detection. Finally, a lightweight dynamic upsampler (DySample) is integrated into the neck network. By adaptively adjusting the upsampling process through a scope factor, DySample effectively mitigates critical information loss and further enhances multiscale feature fusion. Experimental results demonstrate that SSC-YOLO11 delivers substantial performance gains on the HRSID, MSAR-1.0, and OGSOD-1.0 datasets, achieving mAP@0.5 of 91.9%, 91.0%, and 79.6%, and mAP@0.5:0.95 of 67.7%, 59.5%, and 49.8%, respectively, which represent improvements of 3.5%, 3.2%, and 2.1% in mAP@0.5 and 3.7%, 2.5%, and 1.3% in mAP@0.5:0.95 over the original YOLO11.