Signal modulation recognition is a key technology in signal receiving, which is used in adaptive modulation, spectrum monitoring and management, electronic countermeasures, military reconnaissance and other fields. At present, most modulation recognition technologies use a single feature domain or classifier. However, such methods have the problem that the extracted features describe the signal incompletely. To tackle this problem, we propose a novel modulation recognition framework based on multi-feature fusion parallel network. It consists of Swin Transformer (SwinT) feature extraction network, multi-layer perceptron (MLP) feature extraction network and fusion decision modulation classifier. This comprehensive approach can realize modulation recognition in multi-dimensional domains under various signal-to-noise ratios (SNRs). First, the short-time Fourier transform (STFT) method converts modulated signals into high-order time-frequency spectrum images. Meanwhile, the statistical features such as high-order cumulants (HOC), wavelet entropy and spectral line feature are extracted from the time-domain signal. Then, the high-order time-frequency spectrum images are applied to the SwinT, which extracts the time-frequency features, and the valid information on statistical features is extracted by MLP. Finally, the features of multi-dimensional domains are fused to recognize modulation types. Experimental results show that the recognition rates of RML1000 and RML2018.01A datasets are 92.8% and 63.66%, respectively, which demonstrate that the proposed algorithm is robust and efficient. Compared with other methods, the proposed method can achieve better performance, especially at low SNR.