Automatic modulation recognition with vision transformer in radio signals

In this thesis, the automatic modulation classification of radio signals is studied. In the literature, there are various feature-based or deep learning-based studies on this subject. Differently from the literature, this study was conducted using the Vision Transformer (ViT) model, and better results were obtained than the methods in the literature. Vision Transformer models are a new method that is successful in tasks such as image classification. It divides images into small parts and applies an attention mechanism to each part. In this study, signal data is treated as an image and trained without preprocessing. The ViT model demonstrated its success with 98.48% accuracy. Another ViT model, Swin Transformer, was also used, but due to the complexity of the model, it did not perform as well as the classical ViT model. In addition, the DenseNet201 and the generated CNN model were used and showed much better results than these methods. This study has shown important results in terms of showing how successful ViT models are in applications in signals.