WSF-YOLO: Wavelet-Guided Spatial–Frequency Joint Modeling for Aircraft Detection in SAR Images

Abstract

Synthetic aperture radar (SAR) provides all-weather and day-and-night imaging capabilities, but aircraft detection in SAR images remains challenging because of discontinuous target scattering responses, complex background clutter, shallow-detail attenuation, and unstable localization of small targets. In particular, high-frequency SAR components may contain both target-related scattering discontinuities and clutter fluctuations, making direct frequency enhancement susceptible to background interference. To address these issues, this paper proposes WSF-YOLO, a wavelet-guided spatial–frequency joint detection method based on YOLO11n. First, a wavelet-guided spatial–frequency feature extraction module (WGStem) selectively incorporates frequency information through gated residual fusion, reducing structural-detail loss during downsampling. Second, a spatial–frequency collaborative feature enhancement module (SFC3k2) continuously strengthens aircraft contours, edge responses, and local scattering differences in the intermediate stages of the backbone. Finally, a dynamic-ratio joint regression loss (DRCN-Loss) combines complete intersection over union (CIoU) and normalized Wasserstein distance (NWD) with training-dependent weights to improve localization stability for small targets and low-overlap predictions. Experimental results show that WSF-YOLO achieves 95.6% mean average precision at an intersection-over-union threshold of 0.5 (mAP50) and 71.2% mean average precision averaged over thresholds from 0.5 to 0.95 in increments of 0.05 (mAP50–95) on SAR-AIRcraft-1.0, outperforming YOLO11n by 1.4 and 6.2 percentage points, respectively. Consistent improvements on another public SAR aircraft dataset further demonstrate the effectiveness of the proposed method.