STC-Net: Combining a Swin Transformer With a CNN for Micro-Doppler Curve Separation

Accurately extracting micro-Doppler (m-D) curves is crucial for parameter estimation and spatial target recognition. However, complex ballistic scenarios often lead to overlapping m-D components and significant noise interference in the time–frequency domain, thus posing significant challenges to the effective separation of individual m-D components. To address this issue, we propose STC-Net, a hybrid deep learning framework that reformulates m-D curve separation as an instance segmentation task. STC-Net synergistically combines the global context modeling ability of the Swin transformer with the fine-grained feature extraction capabilities of convolutional neural networks. Specifically, long-range dependencies are first captured through the hierarchical structure of the Swin transformer. Meanwhile, we propose a feature aggregation module (FAM) to mitigate the loss of information and effectively preserve local details during the interlayer downsampling process. Subsequently, a feature pyramid network is incorporated in conjunction with a path aggregation network to facilitate effective multiscale feature fusion. Moreover, we propose a channel recomposition module (CRM) to deeply fuse low-level and high-level semantic information. Experimental results demonstrate that the proposed STC-Net achieves an F1 score of 76.77% and a mean average precision (mAP) of 78.12% on the validation set, with the proposed FAM and CRM contributing improvements of 2.28% and 3.24% in the F1 score and mAP, respectively. In addition, STC-Net exhibits superior performance and robustness in m-D curve separation tasks implemented under various challenging conditions, including scenarios with increased numbers of m-D components and low signal-to-noise ratios.