DCENet: A Dilated Convolution and Attention-Based Network for LPI Radar Signal Modulation Recognition

This letter presents a modulation recognition model to address LPI waveform recognition (LWR) through time-frequency images (TFIs). The model introduces a Dilated-MBConv module to analyze the structural disparities between TFIs and natural images. By introducing dilated convolutions with enlarged receptive fields, the module enables the network to extract both fine-grained and spatially contextualized features for reliable discrimination. To further enhance representational capacity, a Convolutional Block Attention Module (CBAM) is incorporated to selectively emphasize informative regions while suppressing redundant activations. Experimental results demonstrate that model outperforms state-of-the-art methods in LPI waveform recognition and validate the transferability of the Dilated-MBConv module.