Recognition of Punctured Convolutional Codes Based on Multi-scale CNN

Punctured convolutional codes are widely applied in satellite communication systems and mobile communication systems. Blind recognition of channel coding plays a significant role in wireless communication technologies such as cognitive radio and radio spectrum detection. This work proposes a deep multi-scale convolution neural network (CNN) which is composed of a multi-scale feature extractor and dilated convolution layers for punctured convolutional codes recognition. The multi-scale feature extractor can better extract the features of different punctured matrices from codeword sequence with convolution kernels of different sizes. The dilated convolution layers expand the receptive field by using different dilation factors. In addition, mixture of experts is adopted to improve model stability and increase classification accuracy. Experimental results demonstrate that the proposed model performs consistently better than existing models on punctured convolutional codes. The proposed multi-scale CNN also shows better performance on common convolutional codes with code rate R = 1/2 and diverse constraint lengths.