CARL-RTDETR: A Lightweight Real-Time Detection Transformer for SAR Object Detection in Complex Scenarios

To meet the real-time requirements of synthetic aperture radar (SAR) target detection on resource-constrained embedded platforms, a Compact Adaptive Representation Learning Real-Time Detection Transformer (CARL-RTDETR) model is proposed. The model is tailored to SAR image characteristics, incorporating a cross-stage spectral convolution mixer module that suppresses background interference and noise while enhancing features of small targets. An anchor-free downsampling module is employed to avoid the loss of small target information. Furthermore, a reparameterized context anchor aggregation module is designed to substantially reduce model parameters and computational costs, while improving the robustness of multi-scale features. The model further integrates a layer-adaptive magnitude-based pruning algorithm to enable structured compression. Evaluated on the SARDet-100K dataset, CARL-RTDETR reduces parameters by 47.2% relative to the baseline RT-DETR50, and achieves an average precision (AP) of 56.1% and AP75of 61.0%, representing improvements of 3.0 and 4.6 percentage points, respectively. These results demonstrate the model’s suitability for real-time embedded deployment of SAR target detection in complex environments while maintaining high accuracy.