Second-Order Channel Attention Multi-scale Grouped Convolution LSTM Networks for Automatic Modulation Recognition

Automatic modulation recognition (AMR) is one of the essential techniques and a di cult challenge to crack in non-cooperative communication systems. Attention mechanisms have been widely applied to deep learning- based (DL) AMR, and its effectiveness has been proven. However, these methods still have problems of high complexity and low accuracy. This letter proposes a high-performance, lightweight framework combined with multi-scale grouped convolution (MGC) and second-order channel attention (SCA), named SCA-MGCLSTM. The MGC structure ensures that channel independent multi-scale depth features are extracted while considerably reducing the number of model parameters. Meanwhile, unlike general attention mechanisms that use first-order information, SCA leverages second-order information from feature maps to obtain more effective attention scores. Experiments on benchmark datasets show that our model outperforms existing deep learning methods regarding training speed and recognition accuracy.