As the variation law of frequency with time is the most important difference between different modulated signals, a radio modulation classification and recognition method combining Choi-Williams distribution and improved convolutional neural network model is proposed. In the signal preprocessing stage, in order to better retain the time-frequency characteristics of the signal, the Choi-Williams transform is introduced to transform the original time series signal into time-frequency image, and then the modulation signal classification problem is transformed into an image recognition problem. In the signal recognition stage, the convolutional neural network model is introduced with residual dense blocks and global average pooling layer to overcome the shortcomings of poor generalization ability and long training time of convolutional neural network model. Experimental results show that the proposed method can effectively solve the problem of gradient disappearance, and has the advantages of high recognition rate and strong generalization ability. Especially in the case of low SNR, the performance is even better. When the SNR is -4 dB, the classification accuracy of 8 kinds of signals can reach 100%.