Improving Automatic Modulation Classification Accuracy Using Combination of Short-Time Fourier Transform and Radon Transform

Automatic modulation classification (AMC) is an important task in modern cognitive radio communication and signal intelligence systems. While notable performance in terms of accuracy and execution time can be achieved using raw signals in the time domain, there is still room for improving AMC performance through preprocessing techniques. Accordingly, this paper proposes a novel combination of Short-Time Fourier Transform (STFT) and Radon Transform to enhance feature extraction, thereby improving AMC accuracy. By leveraging both time-frequency and spatial properties of signals, several deep network models, such as ResNet, MobileNet, and GoogLeNet, can accurately classify modulation types. Experiments are conducted with these three models (ResNet, MobileNet, and GoogLeNet) using different types of input data. The experimental results demonstrate that the STFT-Radon Transform input data significantly improves AMC accuracy compared to conventional input data. Additionally, the ResNet model achieves higher accuracy than MobileNet and GoogLeNet in the same experimental scenarios.