An Improved Automatic Modulation Classification Method Based on Complex-Valued Capsule Network

The modulation classification in communication systems is an essential technology in military and civilian domains requiring continuous updates and improvements. Furthermore, deep learning (DL) is widely acknowledged as an effective approach for Automatic Modulation Classification (AMC) and has been extensively applied in the field of pattern recognition. In image classification, capsule networks have shown excellent performance by encoding features and establishing relationships between parts and the whole in an image. However, the original capsule network is limited in its ability to extract useful feature information as it tends to contain all the information in the image, failing to leverage the advantages of capsule networks when faced with complex background datasets. In this paper, we propose a signal recognition architecture based on capsule networks and introduce a feature extraction block based on multi-channel and complex-valued networks to improve classification performance. Subsequently, we conduct comparative studies on different neural network models using the RadioML2016.04C dataset. Our results prove the outstanding results achieved by the proposed network in AMC. The proposed model outperforms other models by a 10% improvement in accuracy under low SNR conditions.