A multi-scale enhanced feature fusion model for aircraft detection from SAR images

Aircraft detection has been a challenging task although many efforts have been made due to the diversity of aircraft scale and interference of complicated background in synthetic aperture radar (SAR) images. So, this paper proposes a new method, named ‘multi-scale enhanced feature fusion network, briefly, MSEFF-Net’. Firstly, a nonlinear activation free attention module (NAFAM) is proposed to enhance the feature information of aircraft. Secondly, a feature fusion module (FFM) is designed and a multi-scale feature fusion pyramid network (MFFPN) is proposed to integrate the semantic information of different layers. Finally, a global-to-local context aggregation (GLCA) module is built to aggregate global and local information. The proposed model is validated using two groups of public datasets, SAR-AIRcraft-1.0 and SADD, and is compared with various advanced detection methods (e.g. Faster R-CNN, Cascade-RCNN, YOLO series and RT-DETR series). The experimental results demonstrate that the precision, the recall, and the mAP50 reach 97.4%, 97.6%, 98.9% for SADD dataset; the mAP50 and the mAP50:95 reach 70.1% and 49.0% for SAR-AIRcraft-1.0 dataset, respectively. The results indicate that the proposed method achieves higher accuracy than the other detection methods do.