Low‐light image enhancement is one of the fundamental challenges in computer vision, aiming to improve brightness, contrast, and color balance under insufficient illumination. In this work, we present a novel entropy–fidelity and deep white‐balance (EF–WB) framework that integrates information‐theoretic optimization with deep learning‐based color correction to achieve perceptually consistent and structurally reliable enhancement. Operating in the Lab color space, the proposed entropy–fidelity module maximizes entropy while preserving pixel fidelity to avoid overexposure and maintain a valid intensity range. The deep WB module re‐renders illumination and adjusts color temperature through a learned mapping to produce natural and visually balanced outputs. Extensive experiments were conducted on four benchmark datasets: LOL, LOLv2‐Synthetic, LIME, and DICM, demonstrating the effectiveness of the proposed framework. On the LOL dataset, EF–WB achieved a PSNR of 12.02 dB and SSIM of 0.3526, outperforming traditional methods such as CLAHE (10.44 dB) and LR3M (8.75 dB). Furthermore, it obtained a PSNR of 7.56 dB and SSIM of 0.1676 on the LOLv2‐synthetic dataset, indicating stable performance across synthetic environments. In nonreference evaluations, EF–WB recorded lower NIQE (3.39) and PIQE (35.28) scores, surpassing CDAN, Lighten Diffusion, and Dark IR. With low computational complexity and no requirement for paired supervision, EF–WB shows strong potential for real‐time low‐light image enhancement in practical applications.
Shahbaz et al. (Thu,) studied this question.