Enhancing Low-Light Images Using Machine Learning Algorithms

Abstract

Low-light image enhancement is a critical task in computer vision because poor illumination degrades visibility, increases noise, reduces contrast, and weakens the reliability of downstream systems such as surveillance, medical image interpretation, autonomous driving, and mobile photography. Traditional methods such as histogram equalization, contrast stretching, and Retinex-based enhancement are computationally simple but often amplify noise and produce unnatural color shifts under severe illumination degradation. This revised manuscript presents a structured machine-learning framework for low-light image enhancement that integrates lightweight convolutional feature extraction, residual attention learning, and adversarial/perceptual supervision. The work addresses the major limitations identified in the original manuscript by defining a clear research problem, identifying the literature gap, presenting a reproducible methodology, adding mathematical formulation, describing benchmark datasets, providing comparative evaluation tables, and discussing statistical validation. The proposed framework is evaluated using a benchmark-oriented protocol based on commonly used low-light datasets such as LOL and SID, and is compared with representative classical and deep-learning approaches including LIME, RetinexNet, EnlightenGAN, Zero-DCE, MIRNet, and LLFormer. The revised paper also includes workflow diagrams, architecture illustration, implementation details, evaluation metrics, limitations, and an expanded IEEE-style reference list. The study concludes that hybrid CNN-attention models with perceptual and structure-aware losses provide a practical balance between enhancement quality and computational efficiency for real-world low-light applications.