[IPOL discuss] [IPOL] new version demo NLmeans

[Jean-Michel Morel]

Dear all,

For the denoising benchmark, I forgot  to  write three crucial points 
that are missing.

1) Noiseless images:
First of all, the benchmark will not be "real" in the sense that (white) 
noise will be added to noiseless images. Thus  a series of noiseless 
images must be proposed to the users. Noiseless images are simply best 
quality  photographs with many  photons (typically outdoor scenes), that 
have  been zoomed in by an 8 factor.

2) RMSE!
Since  in the benchmark we have the "ground truth" (the noiseless 
image), the demo must sytematically give the pair (original \sigma, RMSE 
(original,  denoised).

3) In fact contrarily to intuition denoising in presence of low SNR
is useful because in dark  parts of raw images this is what really happens.
  Thus to go up to $\sigma = 40, 60, 80, 100$  is not absurd.
I am aware that the algorithms may  have to undergo some serious changes 
to cope with such low SNR's. Nevertheless, I invite the denoising algo 
designers to think it over, because this is the real thing.
At the  very least we should go up to \sigma= 40. (SNR simeq 3). It 
might be that the hierarchy of algorithms is changed by low SNR's.

On the other side, we all see that with low SNR's  (\sigma = 2 to  5) 
the gains in RMSE are inexsitent to modest, and this is another point 
that will probably come out as a major challenge (or may be an impossible).

All the best,
Jean-Michel



Guoshen Yu a écrit :
> Hi José Luis, hi Jean-Michel and Toni,
> Thanks very much for asking. I find the interface superb and it would be
> great if our DCT denoising can have a same demo like this. Just two
> remarks/questions. 
> 
> 1. The denoising algorithm uses a noise sigma parameter. Is this parameter
> estimated from the noisy image or is it simply copied from the interface?
> The noise estimation and the denoiser can be considered as two completely
> different algorithms, whereas the precision of noise estimation has a
> direct impact on the denoiser performance. If this demo is to test the
> denoiser, I think it would be better to directly copy the sigma as if it is
> given. Moreover, in most denoising benchmarks, sigma is assumed known. If
> in this demo the noise sigma is estimated, the comparison with the other
> benchmark results will be unfair, in favor of most other benchmark results.
> 
> 
> 2. I think it would be really great if we can have an interface where a
> number of denoising algorithms are simultaneously and systematically tested
> together. For example, by clicking a button, all the algorithms are
> executed and their results on the same image are displayed one next to the
> other. I would be very happy to have the DCT denoiser bundled with any
> denoising algorithm if possible. 
> 
> Jose Luis, thanks again in advance for implementing the DCT denoiser demo!
> 
> Best,
> Guoshen
> 
> On Wed, 09 Mar 2011 17:36:25 +0100, José Luis Lisani
> <joseluis.lisani at uib.es> wrote:
>> The updated version of the NLmeans demo is available:
>>
>> http://demotest.ipol.im/~lisani//bcm_non_local_means_denoising/ 
>> <http://demotest.ipol.im/%7Elisani//bcm_non_local_means_denoising/>
>> (user: demotest, passw. demotest)
>>
>> The demo features:
>> - subimage selection
>> - result images displayed with a zoom in when they are small (less than 
>> 200 pix. in any dimension)
>> - all the images displayed on the same frame
>>
>> It remains a pending issue:
>> the use of 'float' images to prevent the loss of information due to 
>> quantization
>> (Nicolas?)
>>
>>
>> @Guoshen: tell me if you like this kind of demo for your dct_denoising 
>> algorithm
>>
>>
>> Best,
>> José Luis
>