APSNR: Artifact Peak Signal-to-Noise Ratio for Image Quality Assessment

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Image processing pipelines introduce artifacts that degrade quality and mislead vision tasks.
• Current image quality assessment (IQA) models use rigid fidelity criteria, failing to differentiate artifacts from enhancements.

Purpose of the Study:

• To develop a novel IQA metric that selectively penalizes image artifacts while allowing legitimate enhancements.
• To address the limitations of existing IQA models in distinguishing harmful artifacts from beneficial image processing operations.

Main Methods:

• Introduced the artifact peak signal-to-noise ratio (APSNR), a new IQA metric with a selective fidelity criterion.
• APSNR identifies artifacts by detecting pixels violating an artifact-free intensity mapping.
• Calculates PSNR exclusively on artifact-corrupted regions.

Main Results:

• APSNR demonstrates strong correlation with human perception of artifacts.
• The metric remains robust to beneficial image enhancements.
• Extensive experiments validate the effectiveness of APSNR.

Conclusions:

• APSNR offers a more nuanced evaluation of image processing algorithms compared to traditional IQA metrics.
• Provides a principled tool for benchmarking artifact suppression techniques.
• Enables better assessment of image quality in the presence of both artifacts and enhancements.