Self-Supervised Image Denoising with Regularized U-Net and Sub-Sampled Training.
Saved in:
| Title: | Self-Supervised Image Denoising with Regularized U-Net and Sub-Sampled Training. |
|---|---|
| Authors: | Gor, Ashishkumar1 ashishgor.ce@ddu.ac.in, Bhensdadia, C. K.1 ckbhensdadia@ddu.ac.in |
| Source: | IAENG International Journal of Computer Science. May2026, Vol. 53 Issue 5, p1650-1665. 16p. |
| Subjects: | Image denoising, Mathematical regularization, Artificial neural networks, Signal denoising, Diagnostic imaging, Deep learning, Machine learning |
| Abstract: | Image denoising plays a crucial role across diverse domains--including photography, surveillance, medical imaging, astronomy, and satellite imagery, where noise suppression is essential for reliable downstream tasks like segmentation and classification. Although deep learning-based methods have advanced denoising performance, their dependence on clean reference images, sensitivity to noise variations, and risk of artifact generation remain major limitations. To address these challenges, we propose a self-supervised denoising method inspired by the Neighbor2Neighbor framework but enhanced with three key contributions: (i) a dual-scale training strategy that leverages a noisy image Y and its uniformly down-sampled counterpart Y/2, (ii) spatial and feature domain regularization terms that promote structure preservation and stability, and (iii) a U-Net-based architecture adapted to this training paradigm. Unlike existing approaches, our method learns effectively from single noisy inputs without access to clean images. The model is trained on 50,000 noisy ImageNet validation images with Gaussian (σ = 15, 25, 50) and Poisson (λ = 30, 50) noise. Evaluations on CBSD68, Kodak24, and Urban100 show consistent improvements over recent self-supervised baselines including Noise2Void, Self2Self, Blind2Unblind, and Neighbor2Neighbor, achieving up to 1.2 dB PSNR gain on Kodak24 at σ = 25. Experiments on the real-world SIDD dataset further confirm the adaptability of the framework under practical smartphone noise. Ablation studies confirm the benefit of both dual-scale training and regularization, while robustness analysis demonstrates stable performance under ±25% hyperparameter variation and across different noise intensities. Additional tests with Salt-and-Pepper and Speckle noise further highlight the adaptability of the framework. Beyond standard benchmarks, we demonstrate cross-domain applicability on brain MRI and astronomical images, along with efficient inference enabled by a compact U-Net design, underscoring the framework's practical deployability. Future work will focus on adaptive regularization, integration of physics-informed priors, and deployment in practical imaging domains such as astronomy and medical diagnostics. [ABSTRACT FROM AUTHOR] |
| Copyright of IAENG International Journal of Computer Science is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
| Abstract: | Image denoising plays a crucial role across diverse domains--including photography, surveillance, medical imaging, astronomy, and satellite imagery, where noise suppression is essential for reliable downstream tasks like segmentation and classification. Although deep learning-based methods have advanced denoising performance, their dependence on clean reference images, sensitivity to noise variations, and risk of artifact generation remain major limitations. To address these challenges, we propose a self-supervised denoising method inspired by the Neighbor2Neighbor framework but enhanced with three key contributions: (i) a dual-scale training strategy that leverages a noisy image Y and its uniformly down-sampled counterpart Y/2, (ii) spatial and feature domain regularization terms that promote structure preservation and stability, and (iii) a U-Net-based architecture adapted to this training paradigm. Unlike existing approaches, our method learns effectively from single noisy inputs without access to clean images. The model is trained on 50,000 noisy ImageNet validation images with Gaussian (σ = 15, 25, 50) and Poisson (λ = 30, 50) noise. Evaluations on CBSD68, Kodak24, and Urban100 show consistent improvements over recent self-supervised baselines including Noise2Void, Self2Self, Blind2Unblind, and Neighbor2Neighbor, achieving up to 1.2 dB PSNR gain on Kodak24 at σ = 25. Experiments on the real-world SIDD dataset further confirm the adaptability of the framework under practical smartphone noise. Ablation studies confirm the benefit of both dual-scale training and regularization, while robustness analysis demonstrates stable performance under ±25% hyperparameter variation and across different noise intensities. Additional tests with Salt-and-Pepper and Speckle noise further highlight the adaptability of the framework. Beyond standard benchmarks, we demonstrate cross-domain applicability on brain MRI and astronomical images, along with efficient inference enabled by a compact U-Net design, underscoring the framework's practical deployability. Future work will focus on adaptive regularization, integration of physics-informed priors, and deployment in practical imaging domains such as astronomy and medical diagnostics. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 1819656X |
