My work focuses on image quality enhancement where qualitative comparisons (for ex using PSNR and SSIM distortion metrics) are not possible, although the majority of reviewers request it ...
So I plan to rely on survey results where people will judge visual image quality, and I wonder if this study (or vote) is sufficient in a research paper.
What de you think ?
The best assessment of visual quality is by running a perceptual quality assessment by a number of human subjects who give scores between 1 for very bad and 5 for very good. Then you get a Mean Opinion Score from all scores of an image from the different human evaluators. Usually 4 can be ok, 4.5 might mean very good, etc...
The best assessment of visual quality is by running a perceptual quality assessment by a number of human subjects who give scores between 1 for very bad and 5 for very good. Then you get a Mean Opinion Score from all scores of an image from the different human evaluators. Usually 4 can be ok, 4.5 might mean very good, etc...
You can look at unsupervised image quality metrics.
See this paper
Unsupervised feature learning framework for no-reference image quality assessment
Dear Kawther Aarizou
Image Visual Quality or ( Imperceptibility) can be measured by either quality measurements or quantity measurement. The one you are looking for is quality measurements called ( Humane visual systems or HVS), this measures is less immunity than the other one( quantity measure) which has the highest powerful robustness.
Please try to work with MSE, PSNR, NCC. I can help you well is you like.
regards
Qualitative comparisons by people is quite subjective and usually cannot be relied on seriously unless the population size is very big. Research methods usually prefer objective measurements which can be replicated without ambiguity.
Metrics such as PSNR and SSIM have exact, closed-form mathematical definitions so in that regard they are quantitative rather than qualitative metrics. Your proposition of primarily relying on survey results of subjective judgment would only contribute which images are perceived as 'high quality' and which are not, and that would not be easily generalizable. These are all highly subjective. Assuming you can get enough sample responses that would be statistically reliable, you are still only gaining knowledge regarding people's subjective perceptions about the image.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images