CGAF-CSRNet: A Residual Network Model for Automatic Modulation Classification Based on Complex GADF-Mapped Features

In recent years, deep learning techniques have been widely applied in communication signal processing, with automatic modulation classification (AMC) emerging as a key technology in wireless communication systems. However, conventional AMC methods face technical bottlenecks due to the high similarity of time-domain features and low separability in the feature space among different modulation schemes, thereby limiting further improvements in classification accuracy. To address these challenges, this paper proposes a novel residual network model for AMC based on Complex Gramian Angular Field (CGAF) mapped features, referred to as CGAF-CSRNet. The main contributions of this work are twofold. First, we design the CGAF algorithm to obtain a joint time–frequency feature representation of one-dimensional (1D) signals. Second, we design a channel-split residual network classifier that employs a feature channel decomposition-recombination mechanism to effectively fuse shallow and deep features, significantly improving the discriminability of challenging samples. We conduct experiments on the public datasets RadioML2016.10a and RadioML2016.10b. The results show that the proposed model outperforms other state-of-the-art competing models, and it can achieve average accuracies of 63.12% and 65.2%, with peak accuracies of 93.88% and 94.55%, respectively.