Deep Learning versus Classical Denoising for Mobile Receivers: A Field-Parameterised Simulation Study

Noise and interference severely degrade signal clarity in mobile communication systems. This poses a critical challenge for reliable data transmission in dynamic propagation environments. This study presents a field-informed evaluation framework for denoising algorithms. The framework is parameterised using statistical properties derived from urban, suburban, and indoor field measurements. This approach replicates realistic mobile channel conditions. The study assessed four denoising algorithms across an input signal-to-noise ratio (SNR) range of 5–20 dB. These included two classical filters (Wiener and Wavelet) and two deep learning models (Deep Neural Network and Convolutional Neural Network). Results demonstrated the superiority of neural networks. The CNN achieved an output SNR of 31 dB from a 20 dB input. It also achieved a bit error rate (BER) of 0.015 and a mean squared error (MSE) of 0.01. Meanwhile, the DNN achieved a 30 dB output SNR with a BER of 0.025. Both deep learning models significantly outperformed the classical filters. The Wiener filter yielded a 28 dB output SNR, a 0.04 BER, and a 0.03 MSE. The Wavelet filter recorded a 32 dB output SNR and a 0.03 BER. The DNN and CNN models required longer processing times. Specifically, they took 0.80 s and 0.95 s, respectively, compared to 0.45 s for the Wiener filter. However, pairwise t-tests confirmed statistically significant performance gains (p < 0.001, Cohen's d > 1.2). These gains render deep learning models a compelling choice for modern software-defined receivers. The study concludes that deep learning offers a robust solution for noise mitigation under field-informed conditions. Furthermore, it establishes a reproducible evaluation methodology applicable to next-generation mobile communication systems.