Qualitative and Quantitative Quality Assessment of Low-Light Enhanced Images: A Dataset and Benchmark Metric

ICLR 2026 Conference Submission13285 Authors

Keywords: Image quality assessment; low-light image enhancement; qualitative quality assessment; vision-language model

Abstract: Low-light image enhancement (LLIE) improves visibility and restores details in challenging lighting conditions. It is crucial to fairly evaluate LLIE methods to foster the development of more effective models. However, quality assessment of low-light enhanced images proves to be as challenging as the enhancement itself. From a quantitative perspective, full-reference image quality assessment (FR-IQA) metrics (e.g., PSNR and SSIM) are commonly employed to assess the perceptual quality of enhanced images. However, they are not suitable when a pristine reference image is unavailable, which is often the case in real-world applications. From a qualitative perspective, the absence of a standardized and reproducible evaluation pipeline makes it extremely difficult to ensure fair comparisons across different studies. To confront these challenges, we present the Low-light Image Distortions and Quality (LIDQ) dataset, featuring both overall quality scores and distortion distribution annotations collected through formal subjective testing. Leveraging LIDQ, we propose a no-reference Low-light Enhanced Image Quantitative and Qualitative Quality Assessment (LIQ$^3$A) method that not only estimates perceptual quality without requiring a reference, but also provides qualitative assessments of enhancement-induced distortions. Experiments show that LIQ$^3$A aligns closely with human perception while accurately identifying distortion patterns. We anticipate that the proposed dataset and metric will facilitate future advances in low-light image enhancement by providing reliable evaluation feedback.

Primary Area: datasets and benchmarks

Submission Number: 13285