SAGA-YOLO: A High-Accuracy Detector for SAR Aircraft in Complex Environments

Synthetic Aperture Radar (SAR), utilizing its ability to actively transmit microwave signals, is widely used for military aircraft target reconnaissance and aviation safety monitoring. However, the complex background of ground-based aircraft targets, combined with the coherent imaging characteristics of SAR, results in speckle noise, which increases the difficulty of detection. To address this, we propose a high-accuracy SAR aircraft detection model, SAR-Adaptive Gated-Attention You Only Look Once (SAGA-YOLO). It includes the following three improvements: Firstly, the model constructs a robust gated convolutional backbone network that adaptively filters out noise and irrelevant clutter by dynamically adjusting feature channels, thereby extracting more robust semantic features of aircraft. Secondly, an attention module, C2-CAS, is introduced at the end of the backbone network to refocus features on the most distinctive scattering regions of the aircraft, enhancing its key structural features. Finally, a feature fusion network, Slim-neck, is introduced at the neck to improve the model’s ability to extract and fuse features of targets at different scales. Experimental results on the SAR-AIRcraft-1.0 and SADD aircraft datasets demonstrate that SAGA-YOLO achieves superior performance; compared to the baseline YOLO11, SAGA-YOLO improves mAP50 and mAP50-95 by 3.0% and 5.8% respectively on SAR-AIRcraft-1.0, and by 0.6% and 18.3% on SADD. Keywords: SAR image, YOLO11, aircraft detection, gated backbone