Low-light image enhancement using autoencoder-based histogram matching

Low-light image enhancement (LLIE) is a fundamental preprocessing task in computer vision which is essential for enhancing the quality and visibility of images that are captured under poor illumination. Traditional LLIE methods have lower computational costs but often lack to maintain natural tone distributions. In contrast, deep learning-based LLIE methods produce high-quality results at the cost of complex computations and significant training resources. This study introduces Autoencoder-Based Histogram Matching (AEHM), a hybrid LLIE framework that combines the effectiveness of the conventional histogram matching method with the representational capability of autoencoders. In AEHM, a pre-trained autoencoder predicts an optimal reference histogram from the low-light input image histogram, which is then used to perform histogram matching. Experiments conducted on multiple benchmark datasets demonstrate that AEHM outperforms traditional methods and delivers performance comparable to deep learning-based methods, while operating at a fraction of their computational costs, as measured by FLOPs. In particular, AEHM yields average improvements of about 2–3.5 dB in PSNR and 8%–12% in SSIM, together with a 30%–45% reduction in LPIPS, demonstrating its effectiveness in enhancing visual quality while preserving structural and perceptual fidelity.