基于 AG-CNN 的轻量级调制识别方法

针对传统卷积神经网络在调制方式盲识别过程中，存在模型体积大、运算量高、无法部署至移动端等问题，提出了一种基于双注意力机制与Ghost模块的轻量级CNN模型AG-CNN(attention and Ghost convolution neural network)调制识别方法，该方法首先将调制信号映射至复空间，并根据归一化点密度对映射点进行颜色处理，得到高阶特征密度星座图；将该特征作为AG-CNN模型的输入进行学习训练，最后使用训练好的模型对接收端接收到的未知信号进行识别。实验表明，AG-CNN模型对散点为10 000的密度星座图识别率在99.95%以上，与相同层数的CNN模型相比，卷积层参数量压缩6.01倍，计算量压缩6.76倍，且相较于VGG-16、InceptionV3、ResNet-50、Shufflenet、Efficientnet等卷积网络模型，参数量与浮点数运算数下降明显，且在大幅节省学习参数量、降低模型复杂度的情况下，表现出优秀的分类性能。

Lightweight modulation recognition method based on AG-CNN

In view of the problems of large model volume, high computation and unable to deploy to mobile terminal in the blindrecognition of modulation mode in traditional convolutional neural network, a modulation recognition method of attention and Ghostconvolution neural network (AG- CNN) based on dual attention mechanism and ghost module is proposed. The modulation signal ismapped to complex space, the map points are processed by the normalized point density, and the higher order feature densityconstellation is obtained. The feature is used as input of AG-CNN model for learning training, the trained model is finally used to identifythe unknown signal received by the receiver. The experimental results show that the recognition rate of density constellation map withsampling point of 10 000 is over 99. 95% by AG-CNN model. Compared with CNN model with the same number of layers, theconvolution layer parameter is compressed by 6. 01 times and the calculation amount is 6. 76 times. Compared with VGG-16, InceptionV3, ResNet-50, Shufflenet, Eficientnet and other convolutional network models, the number of parameters and floating-point operationsdecreases significantly, and in the case of saving learning parameters and reducing the complexity of the model, it shows excellentclassification performance.