A Multiscale Dual Attention Sparse Dual Network for Time-Frequency Image Denoising of Radar Signal

In realistic scenarios, intercepted radar signals are usually accompanied by severe noise. Denoising the time-frequency image of radar signal (RS-TFI) is crucial to improve signal robustness. To address this challenge, a multiscale dual-attention sparse dual network (MDASDNet) is proposed in this article. The model is designed with a dual network to extract global and local features simultaneously. Three-way parallel multiscale blocks containing dilated convolutions with different dilation rates are designed to enrich the receptive field of the network and compensate for the loss of detailed information during the convolution process. These fuses feature information from different layers. Additionally, a sparse mechanism formed by three-way parallel multiscale blocks and ordinary convolutions with random order is devised to enhance the generalization ability of the network. Meanwhile, to better extract edge and texture features from the image, the dual attention mechanism (AM) is embedded to guide network denoising. The experimental results show that MDASDNet effectively improves the denoising performance of RS-TFI.