Emerging Versatile Context Representation on Modulation Recognition with Vision Transformer

Accurate recognition of intra-pulse modulation patterns is essential for enhancing radar system performance. Tranditional recognition algorithms are typically designed under ideal conditions and handcrafted features, the presence of substantial noise interference in complex electromagnetic conditions can significantly degrade the performance of these algorithms. Although Transformers have achieved great success with computer vision tasks, existing vision transformers are lack of capturing long-range dependencies. To remedy these flaws, this paper proposes an accurate automatic modulation recognition algorithm based on vision transformer. First, we utilize traditional image processing techniques to provide position of signal in time-frequency image and introduce a radius basis function to generate the positional embeddings. This procedure enhances the spatial awareness of our model. Second, dynamic scales of 2D patches are embedded as the input to the encoders. Third, inspired by dilated convolutions, vanilla attention is reformed in the cascade encoders. Experiments and ablation studies exhibit the proposed algorithm performs well under conditions of various signal-to-noise ratio conditions. It is worth mentioning that the classification accuracy can be up to 93.4% when SNR is -10 dB, and reach to 94.4% when SNR is 0 dB, which indicates the proposed algorithm demonstrates superior robustness against noise perturbations compared to other methods. The present work provides a sound experimental basis for further studying automatic modulation classification for the sake of future field application in electronic warfare systems.