Image denoising in fibre-optic and nano-optoelectronic imaging systems is a critical challenge for highresolution signal recovery in minimally invasive medical diagnostics and industrial nanoscale inspection. Fibre bundles inherently introduce honeycomb-like modular artefacts and dark-spot defects due to discrete core sampling, while nano-optoelectronic sensors suffer from additive noise during weak-signal detection. This paper systematically investigates deep learning architectures to suppress such degradation without increasing hardware complexity. Three representative models are evaluated: a residual-based DnCNN-Plus, a multiscale U-Net, and a proposed hybrid Ultimate U-Transformer integrating global self-attention into a U-shaped encoder-decoder. Experimental results demonstrate that while all models achieve robust denoising, the Ultimate U-Transformer significantly outperforms convolutional counterparts in reconstruction fidelity. By combining long-range dependency modelling with local feature extraction, the hybrid architecture attains a peak PSNR of 24.6 dB and SSIM of 0.88, representing a substantial gain of approximately 6.6 dB over the DnCNNPlus baseline. Parameter sensitivity analysis further confirms the superior adaptability of Transformer-based processing for complex, non-uniform noise patterns common in fibre-optic and nano-optoelectronic image transmission. This work establishes a robust algorithmic and theoretical framework for real-time, high-fidelity image reconstruction, directly supporting the advancement of integrated nano-optoelectronic sensing and fibre-bundle imaging systems.
Yuefeng Cai (Thu,) studied this question.