Interpretable and Lightweight Dual-Branch Recognition of Compound Radar Jamming via Layerwise Saliency Generation and Adaptive Fusion

With the rapid evolution of electronic countermea sure (ECM) technologies, compound active jamming poses severe challenges to conventional radar systems operating in complex electromagnetic environments. Existing recognition methods often fail to reliably discriminate jamming signals or sustain stable performance when multiple jamming types coexist, signals overlap, or the jamming-to-noise ratio (JNR) is low. To overcome these limitations, we propose CRJ-DBNet—an interpretable, lightweight, and deployment-efficient network for compound radar jamming recognition. The model builds upon ShuffleNet v2 as its discriminative backbone. A Layerwise Saliency Generator (LSG) with a shared-encoding, dual-output design replaces the original Stage 2 module, producing both gating maps for residual modulation and attention maps for visualization and fusion, all at a consistent spatial scale. These multi-level attention maps are fused along the channel dimension via an Adaptive Attention Map Fusion (AAMF) module and are used only for auxiliary supervision without feedback to the backbone, thereby preserving a stable and interpretable discriminative pathway while minimizing inference cost. On an extended compound jamming dataset with same-type multi-instance samples, CRJ-DBNet achieves 92.48% overall accuracy (OA) and 99.00% overall F1-score (OF1), outperforming lightweight classification baselines and a YOLOv5-based detection baseline with 0.50 GFLOPs and 1.23 M parameters. Cross-JNR, parameter-shift, open-set unknown-type, and cross-domain RF experiments further demonstrate the robustness and transferability of the proposed method.