Increasing the Accuracy of Spectrogram-based Spectrum Sensing Trained by a Deep Learning Network Using a Resnet-18 Model

This paper presents the evaluation of an image restoration approach based on a bilateral Gaussian filter (BGF) to address signal identification for spectrum sensing. The proposed methodology uses synthetic Long-Term Evolution (LTE) and 5G New Radio (NR) signals generated in MATLAB to build two spectrogram datasets. A Resnet-18 model has been trained with and without enhanced spectrograms to show the benefits of spectrogram preprocessing. Several case studies are considered using noisy and restored spectrograms. A noise recognition of $84.81 \%$ is obtained, showing the benefits of the proposed approach for spectrum sensing<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>.