Abstract: In view of the existing intelligent signal detection methods based on deep learning theory, most of them can only detect single signal or signals which don’t overlap in time-frequency domains. This paper proposes a new intelligent detection method based on Mask R-CNN and Criminisi algorithm for time-frequency overlapping multi-signals. First, the signal in the time domain is transformed into a time-frequency image. Then, to solve the problem of missing pixels’ position information in the overlapping part of multiple signals in the time-frequency domain, the Criminisi algorithm to repair and fill the information is applied. Finally, Mask R-CNN is used for training the restored image, and used for detecting the unknown signals. Experimental results show that when the SNR is 0 dB, the average detection rate of overlapping signals in the time-frequency domain reaches 99%. Compared with the method based on convolutional neural network, the average detection rate is increased by more than 20%.