While object detection performs well in natural images,it faces challenges in Synthetic Aperture Radar(SAR)images for detecting airports and aircraft due to discrete scattering points,complex backgrounds,and multi-sca...While object detection performs well in natural images,it faces challenges in Synthetic Aperture Radar(SAR)images for detecting airports and aircraft due to discrete scattering points,complex backgrounds,and multi-scale targets.Existing methods struggle with computational inefficiency,omission of small targets,and low accuracy.We propose a SAR airport and aircraft target detection model based on YOLO,named YOLO-SAATD(You Only Look Once-SAR Airport and Aircraft Target Detector),which tackles the aforementioned challenges from three perspectives:1.Efficiency:A lightweight hierarchical multi-scale backbone reduces parameters and enhances detection speed.2.Fine granularity:A"ScaleNimble Neck"integrates feature reshaping and scale-aware aggregation to enhance detail detection and feature capture in multi-scale SAR images.3.Precision:Wise-IoU loss function is used to optimize bounding box localization and enhance detection accuracy.Experiments on the SAR-Airport-1.0 and SAR-AirCraft-1.0 datasets show that YOLO-SAATD improves precision and mAP50 by 1%-2%,increases detection frame rate by 15%,and reduces model parameters by 25%compared to YOLOv8n,thus validating its effectiveness for SAR airport and aircraft target detection.展开更多
基金supported by Key R&D projects in Zhejiang Province[No.2023C01032,2024C01108]Key R&D projects in Hangzhou[No.2024SZD1A09,2024SZD1A03]Key R&D projects in Ningbo[No.2024Z114].
文摘While object detection performs well in natural images,it faces challenges in Synthetic Aperture Radar(SAR)images for detecting airports and aircraft due to discrete scattering points,complex backgrounds,and multi-scale targets.Existing methods struggle with computational inefficiency,omission of small targets,and low accuracy.We propose a SAR airport and aircraft target detection model based on YOLO,named YOLO-SAATD(You Only Look Once-SAR Airport and Aircraft Target Detector),which tackles the aforementioned challenges from three perspectives:1.Efficiency:A lightweight hierarchical multi-scale backbone reduces parameters and enhances detection speed.2.Fine granularity:A"ScaleNimble Neck"integrates feature reshaping and scale-aware aggregation to enhance detail detection and feature capture in multi-scale SAR images.3.Precision:Wise-IoU loss function is used to optimize bounding box localization and enhance detection accuracy.Experiments on the SAR-Airport-1.0 and SAR-AirCraft-1.0 datasets show that YOLO-SAATD improves precision and mAP50 by 1%-2%,increases detection frame rate by 15%,and reduces model parameters by 25%compared to YOLOv8n,thus validating its effectiveness for SAR airport and aircraft target detection.
