Visual Quality Assessment of Composite Images: A Compression-Oriented Database and Measurement

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Composite images (CIs) are increasingly prevalent due to generative AI.
• Lossy compression significantly degrades CI visual quality, impacting applications.
• Objective quality assessment for compressed CIs is crucial but underdeveloped.

Purpose of the Study:

• To establish a comprehensive database for composite image quality assessment (CIQA) under compression.
• To investigate the perceptual impact of various compression distortions on CIs.
• To develop an effective no-reference CIQA method for CIs.

Main Methods:

• Created the ciCIQA database with 3,000 CIs, 30 distortions, and 6 codecs.
• Conducted large-scale subjective experiments to gather quality scores.
• Proposed a novel multi-masked no-reference CIQA method (mmCIQA) with specialized modules.

Main Results:

• The ciCIQA database is one of the earliest and most comprehensive for CI compression.
• Insights into the first five just noticeable difference (JND) points of compression effects were gained.
• The proposed mmCIQA method significantly outperformed 17 existing approaches in quality assessment.

Conclusions:

• The ciCIQA database and mmCIQA method advance objective quality assessment for composite images.
• The findings facilitate better perceptual compression strategies for CIs.
• The developed resources are publicly available to support further research.