Researchers propose sharpness-aware image quality metrics
New IQA metrics outperform PSNR by 67% in user preference tests
New research introduces a novel class of Sharpness Informed (SI) Image Quality Assessment metrics that properly penalize over-sharpening in image restoration. Using a four-protocol subjective study, the researchers found that images restored using a sharpness-aware composite loss are preferred in 67% of binarized comparisons versus losses that do not explicitly target sharpness. The proposed SI-PSNR metric outperforms all other PSNR variants in correlation statistics on IQA benchmarking datasets, and the study also introduces a novel dataset of images with uniform sharpness increments along with Difference Mean Opinion Scores.
- Researchers from Columbia University and Trinity College Dublin introduced SI-IQA metrics to address oversharpening in AI image restoration
- SI-PSNR outperformed all PSNR variants in benchmarking tests and user preference was 67% higher for sharpness-aware restored images
- Proposed a novel dataset with uniform sharpness increments and DMOS for evaluation
Why It Matters
This work provides measurable improvements in AI image restoration quality, directly impacting applications in photography, medical imaging, and visual effects.