International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 12 | Dec 2025
p-ISSN: 2395-0072
www.irjet.net
Enhanced Low-Light Image Restoration Using a Lightweight and Optimized Deep Neural Network Architecture for Improved Visibility and Performance Asst. Prof. Vijayalaxmi1, Bhagyashree 2 1Assistant Professor, Master of Computer Application, VTU CPGS, Kalaburagi, Karnataka, India 2 Students, Master of Computer Application, VTU CPGS, Kalaburagi, Karnataka, India
------------------------------------------------------------------------***-------------------------------------------------------------------------
Abstract- Low-light conditions often result in images with
enhancement due to their ability to model rich hierarchical features. Despite this progress, many existing models remain computationally demanding, making them impractical for real-time use or deployment on hardware-limited devices such as mobile and edge systems. To address these limitations, this work presents a lightweight and optimized deep neural network specifically designed for efficient lowlight image enhancement. The model extends the U-Net architecture with Residual Blocks to improve gradient propagation and incorporates a Convolutional Block Attention Module to emphasize both channel-level and spatially important features. Using residual learning, the network refines local details while maintaining global brightness and structural coherence. A composite loss function combining L1, perceptual, and SSIM losses ensures balanced improvements in pixel-level accuracy, perceptual realism, and structural fidelity. Trained on paired low- and normal-light images, the proposed approach achieves notable gains in PSNR and SSIM. A web-based interface further enables real-time processing and visualization, demonstrating the practicality of the method for real-world low-light imaging scenarios.
poor contrast, high noise, and reduced visibility, posing challenges for applications such as surveillance, photography, autonomous driving, and medical imaging. This study proposes a lightweight and optimized deep neural network for low-light image restoration, leveraging an attention-enhanced U-Net architecture with Residual Blocks and a Convolutional Block Attention Module (CBAM) to capture channel-wise and spatial dependencies. Residual learning preserves brightness and structural integrity, while attention mechanisms emphasize critical regions for effective feature extraction. A hybrid loss function combining L1 loss, perceptual loss, and SSIM loss ensures pixel accuracy, content fidelity, and structural consistency. Experimental results demonstrate significant improvements in PSNR and SSIM. A web-based interface enables real-time processing and tracking, highlighting the method’s practicality for real-world low-light imaging.
Keywords: Low-light image enhancement, U-Net, Residual Blocks, CBAM, attention mechanism, perceptual loss, SSIM, lightweight deep learning, realtime image processing.
2. RELATED WORKS
1. INTRODUCTION
Article [1] "A survey on image enhancement for Low-light images" by J. Guo et al. in 2023: This paper comprehensively reviews traditional algorithms combined with machine learning techniques for low-light image enhancement. It discusses the strengths and limitations of various approaches in improving visibility and noise reduction, emphasizing the integration of classical and data-driven methods. The survey highlights the challenges and recent advancements in handling real-world low-light scenarios using deep learning.
Low-light imaging remains a persistent challenge in computer vision and digital photography. When illumination is insufficient, images typically suffer from low contrast, elevated noise levels, and significant distortion of structural and colour details. These degradations directly undermine the performance of downstream tasks such as object detection, surveillance analysis, autonomous navigation, and medical interpretation, all of which depend on clear and reliable visual information. Conventional enhancement methods, including histogram equalization, gamma correction, and basic demonising filters, can offer partial improvements. However, they struggle with complex illumination patterns and often fail to preserve fine details, leading to over smoothing, artefacts, or unnatural visual appearance. Deep learning has reshaped image restoration by enabling models to learn robust, data-driven mappings from degraded inputs to high-quality outputs. Convolutional architectures in particular have demonstrated strong results in tasks like demonising, super resolution, and low-light
© 2025, IRJET
|
Impact Factor value: 8.226
Article [2] "Low-light image enhancement by deep learning network for improved illumination map" by M. Wang, J. Li, and C. Zhang in 2023: This work proposes a deep learning network that enhances illumination maps to improve lowlight images. By employing depth wise separable convolutions inspired by Mobile Net, the model balances enhanced feature extraction and computational efficiency. The network effectively brightens dark areas while preserving texture details, suitable for real-time applications.
|
ISO 9001:2008 Certified Journal
|
Page 684