Comparative Analysis of Attention-Based Deep Learning Architectures for Low-Light Image Enhancement

Kanikicherla, Likhita (2025) Comparative Analysis of Attention-Based Deep Learning Architectures for Low-Light Image Enhancement. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Low-light image enhancement continues to be an essential issue in computer vision, restricting the efficacy of imaging technology in surveillance, phone camera, and autonomous navigation use cases. The comparative efficacy of attention-enabled deep learning structures for low-light image enhancement has been explored in this research, with specific attention given to the comparison between channel attention-augmented U-Net and multi-scale dual attention networks. The fundamental question of whether architecture complexity leads to improved enhancement performance in use cases with limited resources is addressed by the study.

The study employs two different architectures: the cutting-edge Multi-scale Image Restoration Network (MIRNet) with dual attention mechanisms and an improved U-Net with Squeeze-and-Excitation blocks for channel attention. Both models were trained using an augmented dataset that was systematically transformed to increase the number of image pairs from 1,000 to 3,500. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics were used in a thorough evaluation that covered 25 training epochs.

The results show that both structures significantly improve low-light images and achieve the SSIM threshold of 0.7. The U-Net structure worked best at 18.64 dB PSNR and 0.811 SSIM, while the MIRNet structure worked best at 18.50 dB PSNR and 0.797 SSIM. The simpler U-Net structure was more stable, converged faster, and took 34% less time to make guesses than MIRNet. These experimental results show that more complicated architectures do not make things work better. They show that channel attention alone strikes the best balance between speed and quality of enhancement, making it more useful in real life. The study offers empirical evidence supporting architectural simplicity when computational resources are constrained.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Maniganze, Taylou UNSPECIFIED
Uncontrolled Keywords:	Low-light image enhancement; Deep learning architectures; Channel attention mechanisms; Multi-scale dual attention networks; U-Net with Squeeze-and-Excitation; Multi-scale Image Restoration Network (MIRNet); Peak Signal-to-Noise Ratio (PSNR); Structural Similarity Index (SSIM); LOL dataset; Data augmentation; Computer vision; Real-time image processing
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	12 Aug 2026 09:00
Last Modified:	12 Aug 2026 09:00
URI:	https://norma.ncirl.ie/id/eprint/9506

Actions (login required)

View Item