Blind Image Quality Assessment by Gaussian Mixture Distribution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional image quality assessment (IQA) methods primarily use mean opinion scores (MOS) and overlook the potential of distribution of opinion scores (DOS).
• Existing IQA databases often lack comprehensive DOS data, limiting the development of advanced IQA models.
• Gaussian mixture distribution (GMD) shows promise in accurately modeling the DOS of image quality.

Purpose of the Study:

• To propose a novel blind IQA method that utilizes GMD to learn image quality distributions.
• To improve the prediction accuracy of both MOS and DOS in image quality assessment.
• To address the scarcity of DOS data by introducing a pseudo DOS generation strategy.

Main Methods:

• A blind IQA method integrating a visual feature learning module (Swin Transformer, CLIP) and a GMD learning module (mixture density network).
• The method learns GMD-based image quality, where the GMD mean predicts MOS.
• Auxiliary training using DOS and a pseudo DOS generation strategy for enhanced applicability.

Main Results:

• The proposed method accurately models DOS using GMD on SJTU IQSD and KonIQ-10K databases.
• The method demonstrates superior performance in predicting both MOS and DOS compared to state-of-the-art IQA methods.
• Pseudo DOS generation significantly improves the method's applicability.

Conclusions:

• Learning GMD-based image quality offers a more comprehensive approach to IQA than relying solely on MOS.
• The proposed method advances IQA by effectively utilizing DOS information, even with limited data.
• This work facilitates deeper research into the distribution of opinion scores in image quality assessment.