BayeUNet-NLM: A Bayesian U-Net with Uncertainty-Guided Non-Local Means Post-Processing for Extreme Low-Light Denoising
Year: 2026
Type: article
Abstract: Extreme low-light image denoising, where photon shot noise and read noise dominate, remains a critical challenge for computer vision and computational photography. While deep learning models, particularly U-Net architectures, have shown remarkable performance, they often produce over-smoothed results and lack robust uncertainty quantification. This limitation is fundamentally rooted in the ill-posed, one-to-many nature of the low-light enhancement problem, which deterministic models fail to capture. This paper introduces BayeUNet-NLM, a novel two-stage framework that marries the representational power of a Bayesian deep network with the interpretable, structure-preserving capability of a classical denoiser. In the first stage, a Bayesian U-Net with Monte Carlo Dropout at test time generates both a denoised image and a per-pixel uncertainty map. This map quantifies the model's confidence, highlighting regions where data is ambiguous (e.g., due to extreme noise or missing textures). In the second stage, this uncertainty map is explicitly used to guide the classical Non-Local Means (NLM) post-processor, adaptively modulating its local bandwidth and its donor-reliability weights: high-uncertainty regions receive stronger non-local regularization from similar, more confident patches elsewhere in the image, while low-uncertainty regions are preserved. This synergy allows the model to recover fine details with high confidence while robustly denoising highly uncertain regions without introducing artifacts. Extensive experiments on benchmark datasets (SID, LOL) demonstrate that BayeUNet-NLM outperforms the deep learning and traditional methods included in our comparison in both quantitative metrics (PSNR, SSIM) and qualitative visual quality, particularly in preserving textures and avoiding over-smoothing. Our work highlights the value of Bayesian uncertainty for guiding hybrid denoising pipelines.
Institution: University of Bechar
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7236725747
