Due to the fact that modulation recognition is applied as a key technology in wireless and non-cooperative communication software-defined radio systems, an automatic modulation recognition method has thus been proposed based on the time-frequency attention module (TFA) Swin Transformer to enhance the robustness of signal feature recognition as well as improve the accuracy of unknown signal modulation recognition at the receiving end. With the time-frequency attention module combined with Swin Transformer, the proposed method effectively improves the accuracy of signal modulation recognition. Considering that the variation of signal frequency over time is an important characteristic for distinguishing different modulation types of radio signals, the one-dimensional radio signal is firstly converted into a two-dimensional time-frequency image, which serves as the input for the Swin Transformer model. On this basis, a time-frequency attention module is introduced for an enhancement of the model’s ability to recognize signal features. The experimental results show that, compared to traditional algorithms, the proposed model is characterized with a significant advantage in recognition performance, with a lower training cost compared to deep neural networks.