No-Reference Screen Content Image Quality Assessment With Unsupervised Domain Adaptation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Human visual system perception is primarily adapted to natural environments.
• Assessing image quality for screen content images (SCIs) is challenging due to differing statistical properties compared to natural images (NIs).
• Existing quality assessment models trained on NIs do not directly transfer to SCIs.

Purpose of the Study:

• To develop the first unsupervised domain adaptation method for no-reference quality assessment of SCIs.
• To leverage subjective ratings from NIs to predict the quality of SCIs.
• To improve the transferability and discriminability of quality assessment models across different image domains.

Main Methods:

• Unsupervised domain adaptation leveraging subjective ratings of NIs for SCI quality assessment.
• A novel quality measure based on improving feature transferability and discriminability.
• Introduction of three complementary losses: center-based loss for classifier rectification, Maximum Mean Discrepancy (MMD) for feature discrepancy minimization, and correlation penalization for feature diversity enhancement.
• Utilizing a light-weight convolutional neural network.

Main Results:

• The proposed method achieves higher performance across various source-target settings.
• Demonstrates effective transfer of quality prediction models from NIs to SCIs.
• The approach successfully enhances feature discriminatory capability and reduces feature discrepancy.
• Achieves higher feature diversity through correlation penalization.

Conclusions:

• The developed method provides a robust solution for no-reference quality assessment of SCIs without subjective evaluations.
• This approach enables quality assessment for novel, application-specific content.
• The findings suggest a promising direction for cross-domain image quality assessment.