Deep and Shallow Features Fusion Based Deep CNN for Spectrum Sensing in Cognitive Radio

Abstract

The powerful classification capability of deep neural network (DNN) makes the DNN-based spectrum sensing algorithms very attractive in practical applications. However, it is worth noting that most existing DNN-based spectrum sensing algorithms only utilize deep features of the received signals, which may limit the further improvement of sensing performance of those algorithms. On the one hand, DNNs often lose most of the global information in the process of extracting deep features, resulting in that the deep features sometimes will not be the optimal choice for classification; on the other hand, shallow features will retain most of the global information, but they are hard to highlight the detailed information. In view of this, a deep and shallow features fusion based CNN (DSFF-CNN) framework is proposed for spectrum sensing. The DSFF-CNN based algorithm uses the sample covariance matrix (SCM) of the received signal as input and fuses the features of different convolutional layers, allowing to make full use of both deep and shallow features. The experimental results show that the proposed algorithm obtains higher detection probability than the classical spectrum sensing algorithm based on CNN, which verifies the effectiveness of the algorithm.