MIA-YOLO: A Multiscale Interaction Network for SAR Aircraft Detection in Complex Scenes

Abstract

Aircraft detection in synthetic aperture radar (SAR) images is essential for airport surveillance, airspace security, and military reconnaissance. However, the coherent imaging mechanism of SAR, varying observation angles, and complex airport backgrounds make aircraft detection challenging. Aircraft targets may suffer from blurred contours, weakened wing structures, unstable scattering centers, and large variations in scale, morphology, and scattering response. Meanwhile, strong-scattering objects such as runways, buildings, and apron facilities can produce aircraft-like bright responses, leading to false alarms and missed detections. Motivated by these challenges, this study proposes MIA-you only look once (YOLO), a detection model for aircraft targets in complex SAR imagery. Based on YOLOv11s, the proposed method constructs a C2f with multi-scale dilated large selective kernel (C2f-MDLSF) module to fuse local scattering details, target structures, and contextual information, thereby improving multiscale target representation. An improved convolution attention module (ICAM) module is designed to model interactions between target regions and relevant contextual regions, enhancing target-region responses while suppressing irrelevant strong-scattering interference in complex backgrounds. An ADown module is incorporated to preserve fine-grained features during downsampling, alleviating feature degradation in small targets and weak-structure targets. Experiments on SAR-AIRcraft-1.0 and SAR aircraft detection dataset (SADD) show that MIA-YOLO achieves higher detection accuracy than competing methods in complex scenes. The model also maintains consistent performance across both datasets.