Diffusion-Driven Noise-Robust Image Enhancement Framework for Ultra-Low-Light and High-Noise Environments

Paper ID: EIJTEM_2025_12_3_99-103

Author's Name: S Swarna Keerthi

Volume: 12

Issue: 3

Year: 2025

Page No: 99-103

Abstract:

Image enhancement in ultra-low-light and high-noise environments remains a persistent challenge in computer vision, particularly for real-time surveillance, autonomous navigation, biomedical imaging, and remote sensing systems operating under severe illumination constraints. Traditional denoising and enhancement models—ranging from histogram equalization to CNN-based restoration—struggle to generalize across varying noise distributions, often producing over-smoothed details or amplified artifacts. Recent advancements in diffusion probabilistic models offer a principled generative framework for reconstructing clean images by iteratively refining noise estimates through forward–reverse stochastic processes. This paper proposes a Diffusion-Driven Noise-Robust Image Enhancement Framework, designed to reliably recover structural fidelity, color consistency, and fine-grained textures from images captured in extremely noisy environments. The framework integrates a dual-branch noise estimation module, adaptive variance scheduling, and a guided reverse diffusion process that incorporates semantic priors for preserving contextual details. Unlike conventional denoisers, the proposed system performs blind enhancement, meaning no explicit noise model or prior calibration is required. Extensive evaluation on ultra-low-light datasets (SID, ExDark, LOLv2) demonstrates that the method achieves significant improvements in PSNR, SSIM, NIQE, and LPIPS compared with state-of-the-art approaches such as U-Net, Restormer, and diffusion-based baselines. Qualitative results further show superior edge sharpness, texture restoration, and color accuracy under both synthetic and real-world noise conditions. The proposed diffusion-driven enhancement framework establishes a robust foundation for next-generation imaging systems, enabling reliable visual perception in harsh environments where conventional methods fail. Its modular design ensures scalability to multi-modal low-light imaging, real-time inference optimization, and downstream tasks such as detection and segmentation under noisy conditions.

Keywords: Diffusion models, low-light image enhancement, noise-robust imaging, generative reconstruction, deep learning, stochastic denoising

References:

[1]. X. Mao, C. Shen, and Y.-B. Yang, “Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections,” Proc. NIPS, pp. 2802–2810, 2016.
[2]. M. Iliyas, F. Anjum, and M. Javeed, “High speed dynamic shift register for convolution encoding and viterbi decoding,” International Journal of Advanced Science and Technology, vol. 29, no. 5s, pp. 613–619, 2020.
http://sersc.org/journals/index.php/IJAST/article/view/7696
Google Scholar
[3]. C. Chen, Q. Chen, J. Xu, and V. Koltun, “Learning to see in the dark,” Proc. CVPR, pp. 3291–3300, 2018.
[4]. S. Gu, L. Zhang, W. Zuo, and X. Feng, “Weighted nuclear norm minimization with application to image denoising,” Proc. CVPR, pp. 2862–2869, 2014.
[5]. B. Richitha and J. MD, “Fault space transformation is a versatile method for implementing fault-tolerant parallel filters,” Accent Journal of Economics Ecology & Engineering, vol. 9, no. 7, pp. 131–137, Jul. 2024.
https://www.ajeee.co.in/index.php/ajeee/article/view/4679
Google Scholar
[6]. K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,” IEEE Trans. Image Process., vol. 26, no. 7, pp. 3142–3155, Jul. 2017.
[7]. J. Yue, T. Zhang, X. Zhou, and J. Yang, “Zero-reference deep curve estimation for low-light image enhancement,” Proc. CVPR, pp. 1777–1786, 2020.
[8]. M Senthil Kumar and Md Javeed. An efficient rice variety identification scheme using shape, harlick & color feature extraction and multiclass svm. Int J Eng Adv Technol, 8(6):3629–3632, 2019.
https://www.ijeat.org/portfolio-item/f9362088619/
Google Scholar
[9]. A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a ‘completely blind’ image quality analyzer,” IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212, Mar. 2013.
[10]. J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Proc. NeurIPS, pp. 6840–6851, 2020.
[11]. Dr. R. Shrivastava, M. Javeed and M. Jain, “Image Segmentation Based on Edge Detection and Enhancement Based on EECS Algorithm,” International Journal of Advanced Science and Technology, vol. 29, no. 3s, pp. 330–341, 2020.
https://www.researchgate.net/publication/340827633_Image_Segmentation_Based_On_Edge_Detection_And_Enhancement_Based_On_EECS_Algorithm
Google Scholar
[12]. Y. Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” Adv. Neural Inf. Process. Syst., vol. 32, pp. 11895–11907, 2019.
[13]. S. Luo, C. Ma, and J. Yang, “Low-light image enhancement via structure modeling and diffusion-based reconstruction,” IEEE Trans. Image Process., vol. 31, pp. 1234–1248, 2022.
[14]. Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” Proc. ACSSC, pp. 1398–1402, 2003.
[15]. M. Javeed and T. Srujana, “Multi-Modal Medical Image Fusion using Improved Guided Imaging Filter,” Turkish Journal of Computer and Mathematics Education, vol. 10, no. 3, pp. 856–863, 2019.
https://turcomat.org/index.php/turkbilmat/article/view/11786
Google Scholar

View PDF