Work

Classifier-Guided Deep Learning for Image-Domain Noise Classification and Restoration After Simulated OFDM Image Transmission

Year: 2026

Type: article

Abstract: This paper presents a classifier-guided image-domain restoration framework for images reconstructed after simulated Orthogonal Frequency Division Multiplexing (OFDM) transmission. The framework uses a ResNet50 classifier to identify the dominant post-demodulation degradation type and select a corresponding specialist denoiser from a fixed bank of pre-trained models. Two denoising strategies are evaluated under identical conditions: a supervised Auto-encoder trained with paired clean–degraded images and the self-supervised Noise2Void (N2V) approach, which does not require clean targets. The simulated OFDM link includes BPSK modulation, Rayleigh fading, additive white Gaussian noise (AWGN), and LMMSE equalization. Controlled Gaussian, salt-and-pepper, impulse, and compound degradations are then applied in the image domain after demodulation. Consequently, the proposed framework is intended as an image-restoration post-processor rather than a symbol-domain OFDM channel-noise mitigation method. Using the public Chest X-Ray Images (Pneumonia) dataset, the ResNet50 classifier achieves mean accuracies of 94.2% and 90.3% in two evaluation scenarios. The classifier-guided Auto-encoder improves PSNR by 0.9 dB over a fixed Auto-encoder baseline, while oracle label selection provides an additional 0.3 dB gain. The supervised Auto-encoder achieves 32.8 dB and 30.5 dB PSNR, outperforming self-supervised N2V under the same paired-data simulation setting. Communication-level BER and EVM are reported separately to characterize the simulated OFDM link. As all degradations are synthetic approximations rather than measured channel data, the results should be interpreted as a simulation-based proof of concept, with validation on additional datasets and measured channel recordings left for future work.

Source: IEEE Access

DOI: 10.1109/access.2026.3724769

Authors: Maie M. Aboghazalah, Abdelnasser Saber Mohamed, Amr A. Hassanain, Nacer Eddine Hammami, Mai Helmy

Institutions: Nahda University, Northern Border University, Sphinx University, Fahd bin Sultan University, Damanhour University

Language: English

Topic: Image and Signal Denoising Methods

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,075

OpenAlex ID: W7203695273

FWCI:

Cites: 24

Cited by:

Related to: 10