Anotace:
Addressing the challenge of feature extraction for Low Probability of Intercept (LPI) radar signals under low signal-to-noise ratio conditions, this study introduces a new method for intra-pulse modulation recognition of LPI radar signals based on an enhanced MobileNet architecture. Initially, a Time-Frequency Image (TFI) preprocessing technique suitable for LPI radar signals is proposed, which significantly improves the recognition accuracy of subsequent networks for intra-pulse modulation of LPI radar signals. Subsequently, the MobileNet network is modified by integrating Hybrid Dilation Convolution (HDC) and Efficient Channel Attention (ECA) modules, resulting in the development of an improved MobileNet. This enhanced network expands the receptive field of feature maps and improves the network's ability to capture channel and positional information. Additionally, a label smoothing strategy is utilized to optimize the network training process, reducing overfitting and enhancing sample clustering performance. Simulation experiments indicate that this method not only yields a high recognition accuracy rate but also outperforms existing comparative networks with fewer parameters.