Modulation recognition is an essential technology for a receiver to demodulate baseband signal in communication system.Conventional way of doing this involves sending a message of modulation type along with the original signal,which will reduce channel bandwidth utilization and data throughput.The receiver can automatically recognize the modulation type of signal using automatic modulation classification(AMC)which will improve the transmission efficiency,with no need for the message of modulation type.Therefore,AMC has important application value in the fields of civil radio monitoring,military electronic countermeasure and electronic detection.According to the distribution of the received symbol samples in the signal constellation,the signal feature image is generated by computational imaging in the thesis.Through this method,the recognition of modulation type is transformed into image classification.In this thesis,some deep convolution neural networks are constructed and trained by the signal feature image,so that deep convolution neural networks can recognize the signal modulation type.The innovation of this thesis is as follow: Firstly,an efficient feature image generation algorithm is proposed,which can reduce the time of computational imaging while preserving the original image information,secondly,a multi-scale feature classification network is proposed in this thesis,which combines the feature information of different scales by concatenating different levels of feature images,so as to improve the network's recognition accuracy of signal modulation types.The results show that the feature image generation algorithm proposed in this thesis has higher recognition accuracy and lower computational complexity than the other three algorithms.And the multi-scale feature classification network proposed in this thesis has higher recognition accuracy than the other three network structures under the same data set. |