Thomas Richter pisze:
> AJ schrieb:
>> Hi All,
>>
>> I was hoping someone would be able to explain why something I have
>> noticed during some image processing analysis occurs.
>>
>> When I am working with extremely degraded images, and am trying
>> different techniques to improve them, I often measure the PSNR (Peak
>> Signal-to-Noise Ratio) of the degraded image against the original,
>> undamaged image. Often, I find that the techniques that give the best
>> (highest) PSNR dB value are NOT the ones that give the best subjective
>> quality i.e. I chose restoration techniques that give a lower PSNR
>> because they LOOK better.
>>
>> Can anyone explain why this happens?
>
> PSNR relates bad to human vision. You should try different metrics
> with a better correlation to subjective quality.
>
> Recommendations: MS-SSIM (Sheik & Bovik) works quite ok (though not
great),
> VDP (Daly's Visual Difference Predictor) works much better, though it is
> grey-scale only and it is extremely slow.
>
>> Maybe it is something to do with
>> the fact that the image quality of the degraded image is so poor to
>> start with, that using PSNR is not that helpful?
>
> If you want to measure subjective quality, PSNR is not very helpful to
begin with...
>
> So long,
> Thomas
i'd also recommend to have look at the S-CIELab
(http://white.stanford.edu/~brian/scielab/)
It is:
- implemented in Matlab
- reasonably fast (or slow)
- well do***ented
You can use it to compute weighted image quality/distortion in both ways
or WMSE or WSNR.
It computes color difference as CIE deltaE distance (Euclidean in Lab
space, prefiltered with eye CSF functions).
regards
Przemek
P.S
There is also minor bug in computing visual angle but its not critical
and result differences are small. If someone is interested i can send a
"patch".
|
