Perceptual Quality Assessment of Low-Light Enhanced Images: A Multi-Annotated Subjective Dataset and a Multimodal...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement algorithms (LIEAs) improve visibility but often introduce distortions in enhanced low-light images (ELIs).
• Existing quality assessment methods for ELIs are limited due to a lack of comprehensive datasets and failure to exploit multi-modal information.
• Current methods often focus on single visual modalities, neglecting crucial structural information across attributes.

Purpose of the Study:

• To systematically study the subjective and objective quality assessment of ELIs.
• To address the lack of publicly available datasets for ELI quality assessment.
• To develop a novel method for predicting ELI quality that leverages multi-modal information.

Main Methods:

• Construction of the first Multi-annotated and multi-modal Low-light image Enhancement quality dataset (MLE) with 1,000 ELIs.
• Subjective studies to gather multiple attribute annotations, quality scores, and textual descriptions for ELIs.
• Proposal of an Attribute-guided Vision-Language Graph Reasoning Network (AVGR-Net) for ELI quality prediction.

Main Results:

• The MLE dataset provides a reliable resource for ELI quality assessment research.
• AVGR-Net effectively integrates multi-attribute visual and textual information using cross-modal graph reasoning.
• Experimental results demonstrate the superior performance of AVGR-Net over existing state-of-the-art methods.

Conclusions:

• The developed MLE dataset and AVGR-Net offer significant advancements in low-light image quality assessment.
• The proposed AVGR-Net demonstrates robust performance in predicting ELI quality by utilizing multi-modal data.
• This work provides a foundation for future research in objective quality assessment of enhanced low-light images.