Enhancing Descriptive Image Quality Assessment With a Large-Scale Multi-Modal Dataset

Area of Science:

• Computer Vision
• Artificial Intelligence
• Natural Language Processing

Background:

• Vision Language Models (VLMs) are advancing Image Quality Assessment (IQA) for linguistic descriptions.
• Current VLM-based IQA methods are limited by narrow task focus, small datasets, and suboptimal performance.

Purpose of the Study:

• To develop a more practical and comprehensive VLM-based IQA model.
• To address limitations in existing IQA datasets and task diversity.

Main Methods:

• Introduced the Enhanced Depicted image Quality Assessment (EDQA) model with a multi-functional paradigm (assessment, comparison, brief/detailed responses, full/non-reference).
• Developed a ground-truth-informed dataset construction approach, creating the large-scale EDQA-495K dataset (495K images).
• Retained image resolution during training and incorporated a confidence score for response filtering.

Main Results:

• EDQA significantly outperforms traditional and VLM-based IQA methods on distortion identification, instant rating, and reasoning.
• Demonstrated superior performance in real-world applications like assessing web-downloaded and model-processed images.
• The EDQA-495K dataset provides a comprehensive, large-scale, high-quality resource for IQA research.

Conclusions:

• EDQA represents a significant advancement in VLM-based IQA, offering enhanced versatility and performance.
• The developed dataset and model address key limitations in the field, paving the way for practical IQA solutions.
• Open-sourced codes, datasets, and model weights facilitate further research and development.