Low-light image enhancement is an important problem in image processing, especially in applications like surveillance, mobile photography, and computer vision systems. Images captured in dark environments often suffer from low brightness, noise, and loss of important details, which reduces their overall quality and usability. In this work, we propose a perceptually guided U-Net based approach for enhancing low-light images while maintaining natural color consistency. The model is trained using paired low-light and normal-light images from the LOL-v2 dataset, including both real and synthetic samples. Our approach focuses on improving brightness and visibility without overexposure, while preserving structural details such as edges and textures. To achieve better visual quality, the model uses a combination of Mean Squared Error (MSE) and Structural Similarity Index (SSIM) loss functions. This helps in maintaining both pixel-level accuracy and perceptual quality. The proposed method also performs region-aware enhancement, ensuring that darker areas receive stronger correction while well-lit regions remain balanced. Experimental results show that the model performs better than traditional methods like histogram equalization, CLAHE, and Retinex in terms of PSNR and SSIM. The enhanced images appear more natural, with reduced noise and improved clarity. The model is lightweight and efficient, making it suitable for real-time applications and deployment on edge devices.
Garg et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: