Image Defogging Quality Assessment: Real-World Database and Method.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Defogging algorithms are crucial in computer vision but lack robust real-world datasets and effective image quality assessment (IQA) methods.
• Existing research is hindered by the absence of naturally occurring foggy image datasets and simple, reliable IQA metrics for evaluating defogged images.

Purpose of the Study:

• To address the limitations in real-world foggy datasets and IQA methods for image defogging.
• To introduce a new dataset, Multiple Real-World Foggy Image Dataset (MRFID), and a novel IQA metric, FRFSIM.

Main Methods:

• Created MRFID with 200 outdoor scenes, each including clear and multiple-density foggy images captured over a year.
• Processed MRFID images using 16 defogging methods, generating 12,800 defogged images (DFIs).
• Conducted subjective evaluations with 120 subjects to collect Mean Opinion Scores (MOS) and developed the Fog-relevant Feature based SIMilarity index (FRFSIM).

Main Results:

• The FRFSIM metric demonstrated higher consistency with MOS compared to other IQA methods.
• Experimental results validate FRFSIM's suitability for assessing the visual quality of defogged images.
• The MRFID dataset provides a valuable resource for developing and evaluating defogging algorithms.

Conclusions:

• The developed MRFID dataset and FRFSIM metric significantly advance the field of image defogging.
• FRFSIM offers a more reliable and user-friendly approach to evaluating defogged image quality.
• This work provides essential resources for future research in computer vision-based fog removal.