In modulation recognition, the feature information of a single image is insufficient, the degree of discrimination is not high enough, and the recognition range is limited. In this paper, a modulation recognition feature enhancement method based on the feature fusion of time-frequency map and constellation map is proposed. The deep learning neural network is used to extract the features of signal image and construct the feature space. Through multi-dimensional feature fusion, the advantages of different features are mined and integrated to enhance the robustness of the model algorithm. In addition, the method of model migration is used, which only needs to train the classifier, which greatly saves the training time and resources, and has strong real-time and practicability. The simulation results show that under the condition of about 0 dB, compared with a single feature image, the average recognition rate of the signal can be improved by about 25% by using the feature fusion enhancement method. Through the model migration, the training of convolutional neural network is omitted, and the training time required is about 10% of that before migration, and the memory consumption is about 7.3% of that before migration. At the same time, the loss of recognition rate of the model is controlled within 5%.