A FCN-based signal extraction for spectrum sensing with considering simultaneously parameters estimation

Abstract

In this paper, An end-to-end, pixels-to-pixels deep learning method of signal extraction and parameters estimation for spectrum sensing was developed. A novel spectrum density map was designed for counting signal number at each time, computing signal-to-noise ratio(SNR) and extracting signal area jointly. The density map would be estimated by a Fully Convolutional Networks (FCN) which can accept arbitrary size input and produce corresponding sized output. The strengths of pixels-to-pixels, Multi-Task learning of FCN were leveraged by designing a special label combining three types of information: the pixels-to-pixels signal area, the signal number and the SNR at each time. We adapt feature extraction network (Vggnet-19) into fully convolutional networks to train a highly effective signal extraction detector which achieves high accuracy in spectrum detection and parameters estimation compared with CNN-based signal extraction approach. Evaluation the presented approach on our datasets, the model demonstrates effectiveness and robustness and the MIOU, pixel accuracy achieve 96.4% and 99% respectively.