Radon-Enhanced Spectrograms for Robust Deep Learning-Based Spectrum Sensing in Cognitive Radio Systems

Cognitive Radio (CR) technology offers an adaptive framework for dynamic spectrum utilization, enabling opportunistic access to available frequency bands and mitigating the persistent issue of spectrum scarcity in modern wireless communication systems. A critical challenge in CR is spectrum sensing (SS), which aims to reliably detect Primary Users (PUs), particularly under low and varying Signal-to-Noise Ratio (SNR) conditions. In this paper, we propose a novel deep learning-based SS framework, referred to as WDR-CNN (Wavelet Denoising and Radon-enhanced Convolutional Neural Network), designed to enhance detection performance in low-SNR environments. The proposed model integrates wavelet denoising, spectrogram transformation, and the Radon Transform to construct robust signal representations for classification. Specifically, time-domain signals are first denoised using the Discrete Wavelet Transform (DWT), converted into spectrograms, and subsequently processed via the Radon transform to highlight structural and directional features. These Radon-enhanced images are then used to train CNN for binary classification of PU activity. Experimental results show that WDR-CNN achieves robust detection in low-SNR scenarios with high computational efficiency. Furthermore, its performance is benchmarked against transfer learning-based approaches such as AlexNet, conventional spectrum sensing techniques, and other deep learning-based SS methods. Results indicate that WDR-CNN improves detection accuracy by 11.19% compared to previous SS methods and by 42.19% compared to AlexNet at − 20 dB. Moreover, WDR-CNN outperforms conventional deep-learning SS models with 16.31% and 22.19% gains at − 15 dB and − 20 dB, respectively, while maintaining lower computational complexity. Overall, the proposed model combines higher detection accuracy with the shortest sensing time, making it highly suitable for real-time applications.