Electromagnetic situational analysis is a crucial task in information warfare, and modulation signal recognition by using Deep Learning(DL) is one of the key technologies. In this paper, the modulation signals are firstly transformed into the form of constellation diagrams with color information, and two Convolutional Neural Networks(CNNs), VGG16 and AlexNet, are selected to complete the modulation signal recognition task by using DL. The results show that a recognition accuracy higher than 99% can be achieved when the Signal-to-Noise Ratio(SNR) of noise is greater than or equal to 0 dB. Since the computational performance and storage performance of military devices are more stringent in controlling, the Average Percentage of Zeroes(APoZ) method is adopted to compress the DL model. The results show that with 0 dB SNR, AlexNet can be compressed by 3 466 times and VGG16 can be compressed by 20 156 times for model parametric quantities, and by 2 314 times and 13 475 times for floating-point operations, respectively, without losing recognition accuracy. In summary, the proposed method is both feasible and efficient in modulation signal recognition.