1 2 Previous Next 40 Replies Latest reply on Jun 18, 2009 6:03 AM by Bill_Janes

    ACR 5.4b: Is this zero noise reduction?

    dorin_nicolaescu Level 5

      Just downloaded a sample Nikon D5000 raw from DPreview. It seems to me that the blotchy noise pattern points to some pretty heavy baseline noise reduction has been applied. Note the NR sliders are set to 0.

       

      So is it RAW enough or does ACR apply some serious NR behind the scene?

      Is it Nikon or is it Adobe?


      noise3200.png

        • 1. Re: ACR 5.4b: Is this zero noise reduction?
          Panoholic Level 2

          Following is by no way a substitute for an answer if and how much noise reduction is done by ACR despite turned down specifications.

           

          1. The D5000's top real ISO (i.e. supported by a specific analog gain) is 1600. ISO 3200 and 6400 are numerical derivatives of ISO 1600, equivalent to an exposure adjustment in raw processing, except that those ISOs reduce the dynamic range by one respectively by two stops (in other words: never use them when recording raw data).

           

          2. The top effective ISO (i.e. the highest ISO, which reduces the noise compared to the lower one) is probably 800 (this is based on a very cursory look, and I have not seen any 1/3 stop ISO shot raw files).

           

          Thus this shot with ISO 3200 is heavily underexposed. The red on the dark blue shorts of the bicyclist is at the end of the eights stop of the dynamic range (-7.61 EV), but as it is in reality an ISO 1600 shot, it is at the end of the ninth stop; this is very dark at ISO 1600 with this camera. Accordingly, the noise is serious. See the following captures showing the red, blue and the composite image:

           

          http://www.panopeeper.com/Nikon/NikonD5000_ISO3200_Noise-7.61EV_Red.GIF

          http://www.panopeeper.com/Nikon/NikonD5000_ISO3200_Noise-6.31EV_Blue.GIF

          http://www.panopeeper.com/Nikon/NikonD5000_ISO3200_Noise_Composite.GIF

           

          Now lo and behold: the blotches are there in the raw data.

          1 person found this helpful
          • 2. Re: ACR 5.4b: Is this zero noise reduction?
            witkacy Level 1

            Never checked this myself, but according to some (for example http://forums.adobe.com/message/1213697#1213697), ACR is applying additional noise reduction at higher ISOs even though sliders are set to zero. If that is still true (the report above is about ACR 4.1) then the dpreview tests are flawed. regards m.

            • 3. Re: ACR 5.4b: Is this zero noise reduction?
              Ramón G Castañeda Level 4

              Just a line to thank you for your consistently helpful input and to remind myself of your new user ID, Gabor.  

              • 4. Re: ACR 5.4b: Is this zero noise reduction?
                Jeff Schewe Level 5

                witkacy wrote:

                 

                ACR is applying additional noise reduction at higher ISOs even though sliders are set to zero.

                 

                That thread was way before fine-tuning done in ACR 4.3 & 4.4 which reduced much of the demosiacing noise reduction component.

                 

                Pretty much all demosiacing algorithms have components of noise reduction in the interpolation. If you didn't, the demosiacing would prolly make noise worse. How much noise reduction is a matter of taste and usefulness. Looking at an image at 100$ zoom is not telling you much of anything other than what the pixel relationships are at 2 to 4 times the likely final print size. So inspecting an image at 100% is useful for seeing those relationships (such as you need for setting sharpening and the amount of noise reduction) but what you are seeing is scifi. Computer displays are low resolution devices...way low.

                 

                Say you process you image out at 300PPI. If your display has a resolution of about 100PPI (such as my 3090 NEC whose resolution is 2560 x 1600 and is just about 25" wide), looking at an image at 100% means that image is 3 times the actual size. So, yes, whatever you are seeing may indeed look like "crap" but it's crap that likely won't print and it's also crap that will likely be eliminated by downsampling.

                 

                About the only time the image will ever look like that is if you do a 100% crop for posting (such as here in the forums). Many people then tend to start wringing their hands because well, yes, at 100 or higher zoom the image looks less good. They also tend to confuse noise with micro-detail (which it ain't) and think noisier looking images have more detail than noise reduced images. Again largely because they don't know what they are looking at.

                 

                Until the time comes that computer displays start to get into the 200DPI range (heck my iPhone has 160DPI resolution) you're gonna have to learn that what you see ain't what you're gonna get on a print. Color managements is far easier than detail management which is what sharpening and noise reduction are all about.

                 

                The only way to tell what an image will look like printed is, to well, print it.

                • 5. Re: ACR 5.4b: Is this zero noise reduction?
                  witkacy Level 1

                  Jeff Schewe wrote:

                  (..)

                   

                  That thread was way before fine-tuning done in ACR 4.3 & 4.4 which reduced much of the demosiacing noise reduction component.

                   

                  (..)

                   

                  So is ACR still doing additional noise reduction based on ISO marker in the metadata with sliders set to zero (just less so)?

                   

                  Your points against useless pixel peeping/real world printing are all valid, I agree wholeheartedly. I also understand that ACR is popular software and have to have pleasing defalts, not the real-ugly ones. But they are just that  - defaults and would I seek 'truer' image interpretation I could change them, exposing what's closer to what camera 'really saw'. Printing extreme enlargements or converting heavily underexposed original to b/w are just two situations I would like to have maximum control over noise reduction. Examples for the review site would be another :\

                   

                  Regards

                  m.

                  • 6. Re: ACR 5.4b: Is this zero noise reduction?
                    Panoholic Level 2

                    Thank you Ramón, it gives a good feeling to know, that someone finds the explanations and demonstrations useful, thus the effort was not in vain.

                    • 7. Re: ACR 5.4b: Is this zero noise reduction?
                      Panoholic Level 2

                      There are two aspects (at least) to this issue.

                       

                      One is, that some noise reduction is done at the best at this stage. Back to the original example: I posted a composite color crop from a noisy patch (not demosaiced, nor color space converted); the blotchiness is very obvious. I repeat it here for easy comparison:

                       

                      http://www.panopeeper.com/Nikon/NikonD5000_ISO3200_Noise_Composite.GIF

                       

                      This version is after WB application. With the actual WB setting of this image, the red channel gets multiplied by 1.9 and the blue by 1.18 (if the greens remain unchanged). This means, that the anyway very noisy red pixel values become amplified compared to the green and even to the blue.

                       

                      The same patch without any white balance application does not look so bad:

                       

                      http://www.panopeeper.com/Nikon/NikonD5000_ISO3200_Noise_Composite_NoWB.GIF

                       

                      It is obvious, that the WB application makes the noise worse. Plus, the noise gets spread (the red pixels are infecting the neighbours through demosaicing and color space conversion).

                       

                      Thus noise reduction at this stage is inherently better than later.

                       

                      The other aspect is, that users may reasonably expect disabling the noise reduction for whatever specific reasons, for example the user will apply selective noise reducation. Another example is the case of the Canon 5D2. ACR is no match for the pattern noise, but certain third party product (products?) can eliminate it quote well. So, why not allowing the user to make this choice?

                       

                      This paternalizing attitude of Adobe sucks. This is the same as with the nonsensical automatic exposure adjustment, which causes lots of blown results (literally blown). Although this can be countered by an oposite adjustment of the "exposure", most users have no idea of it, because Adobe chose not even to indicate, that such an adjustment takes place.

                      • 8. Re: ACR 5.4b: Is this zero noise reduction?
                        xbytor2 Level 4

                        Your efforts are very much appreciated. Though some of the finer points may slip by me at times, your posts do help.

                        • 9. Re: ACR 5.4b: Is this zero noise reduction?
                          Jeff Schewe Level 5

                          Panoholic wrote:

                           

                          This paternalizing attitude of Adobe sucks.

                           

                          Just to be perfectly and absolutely CLEAR, my "patronizing" attitude is mine alone and does not represent the attitude of Adobe which is, if anything bent over backwards in an attempt to do the best thing possible for the end user. I speak for myself not for Adobe...

                           

                          The fact is, most people have no *******' clue what they are looking at when they pixel peep at 100 or 200%, do you dispute that Panoholic? The fact that ACR is doing "some" noise reduction in the demosiacing can be considered a good thing, correct? Do you dispute that? The fact that the demosiacing of false ISO images (amplified way beyond the "real" ISO) look like crap is whose fault?

                           

                          Is it Adobe's or is it the camera makers that take lower real ISO and add a false amplification to it that results is really bad noise signatures that essentially match vastly underexposed images that would suck regardless of the raw processing software used...

                           

                          It does nobody any good to pitch this kinda crap in the forums...you would be better served sticking to the facts rather than trying to insert your own self serving jabs at the Camera Raw engineers bud..

                          • 10. Re: ACR 5.4b: Is this zero noise reduction?
                            Jeff Schewe Level 5

                            witkacy wrote:

                             

                            So is ACR still doing additional noise reduction based on ISO marker in the metadata with sliders set to zero (just less so)?

                             

                             

                            Yes...or more rightly it is doing a base level noise reduction as a component of the demosiacing process...all demosiacing interpolations will have some degree of noise reduction, it's part of the process of demosiacing.

                            • 11. Re: ACR 5.4b: Is this zero noise reduction?
                              Panoholic Level 2
                              my "patronizing" attitude is mine alone and does not represent the attitude of Adobe


                              Jeff, if I had meant you or your statements, I would have said so. The patronizing attitude is that ACR is making decisions on behalf of the user, not only without asking him/her, but not even indicating this.

                               

                              The fact that ACR is doing "some" noise reduction in the demosiacing can be considered a good thing, correct?

                               

                              The fact that ACR is doing noise reduction if the user wants that is a good thing. The fact, that ACR is doing some noise reduction even though the user does not want that is not a good thing.

                               

                              The fact that the demosiacing of false ISO images (amplified way beyond the "real" ISO) look like crap is whose fault?

                               

                              This is really irrelevant in the present context. I am preaching on several forums repeatedly to avoid useless ISOs with raw recording (not only the fake ones, but the ineffective ones as well); the success rate is unknown. However, if and when someone decides to use ISO 25600 (which is four stops past the last useful step with the Canon 5D2, as an example), then he gets what he is asking for, but it is not the task of any raw processor to forcefully "correct" that.

                               

                              It does nobody any good to pitch this kinda crap in the forums...you would be better served sticking to the facts rather than trying to insert your own self serving jabs at the Camera Raw engineers bud

                               

                              You are invited to point out any incorrect statements in my posts.

                               

                              it is doing a base level noise reduction as a component of the demosiacing process...all demosiacing interpolations will have some degree of noise reduction, it's part of the process of demosiacing

                               

                              If done, it is part of the demosaicing process. However, it is not correct that it has to be done, nor that all demosaicing processes always perform noise reduction.

                              • 12. Re: ACR 5.4b: Is this zero noise reduction?
                                Bill_Janes Level 2

                                Panoholic wrote:


                                 

                                If done, it is part of the demosaicing process. However, it is not correct that it has to be done, nor that all demosaicing processes always perform noise reduction.

                                 

                                Here is another way of looking at the problem. I took an image of a uniform flat field with the Nikon D3 using lossless NEF compression at a bit depth of 14 and ISO of 3200. I then rendered the file into ProPhotoRGB at a bit depth of 8 using ACR with noise reduction turned off and with a linear tone curve, separated out the green channel and determined the standard deviation in the central portion of the image. The SD was 1.30.

                                 

                                Using Rawnalize, the standard deviation of green with a gamma 1.8 tone curve was 2.27. This indicates that noise reduction was performed by ACR.

                                 

                                greenResults.png

                                • 13. Re: ACR 5.4b: Is this zero noise reduction?
                                  sandy_mc Level 3

                                  Ahh, no, it just indicates that the same amount of physical noise 

                                  quantized to 8-bits has a lower SD that if its quantized to 14-bits. 

                                  Which is the expected answer.

                                   

                                  Sandy

                                  • 14. Re: ACR 5.4b: Is this zero noise reduction?
                                    Panoholic Level 2
                                    I took an image of a uniform flat field with the Nikon D3 using lossless NEF compression at a bit depth of 14 and ISO of 3200. I then rendered the file into ProPhotoRGB at a bit depth of 8 using ACR with noise reduction turned off and with a linear tone curve, separated out the green channel and determined the standard deviation in the central portion of the image

                                     

                                    Bill, there is a more straightforward way to prove, that NR takes place even it if is set to zero (though the proof is useful only for high ISOs):

                                     

                                    1. convert the raw file with high ISO in DNG,

                                     

                                    2. create a copy of it and change the ISO in a low value,

                                     

                                    3. convert both together with the same settings, black zero, NR zero, and look at the difference in PS.

                                     

                                    I don't know what "high ISO" is in specific cases. The limit is 3200 for the Canon 5D2, but I guess that this depends on the camera.

                                     

                                    If you want to do this and have a hex editor, I explain you how to do it - or you upload the images and I do it for you.

                                     

                                    Btw, I see that the text overflowed in the Rawnalyze dialog. If you press Alt and minus, the font size gets reduced (with plus it gets increased; there are five sizes). That's what I and many others are missing from the DNG converter dialog.

                                    • 15. Re: ACR 5.4b: Is this zero noise reduction?
                                      Bill_Janes Level 2

                                      andy_mc wrote:

                                       

                                      Ahh, no, it just indicates that the same amount of physical noise 

                                      quantized to 8-bits has a lower SD that if its quantized to 14-bits. 

                                      Which is the expected answer.

                                       

                                      Sandy

                                      Not at all. I merely used 8 bit because that is what Photoshop uses for histograms.

                                       

                                      Here is another experiment. I rendered into 16 bit ProPhotoRGB using ACR and then converted to 16 bit linear_RIMM_-RGB_v4.icc using the method outlined on the ICC site. I also rendered into a linear 16 bit RGB using Iris and separated out the green channel for analysis.

                                       

                                      http://http://www.color.org/scene-referred.xalter

                                       

                                      The results can be viewed with ImageJ. The Photoshop-ACR conversion has less noise. The Iris demosaicing has more noise than PS but less than the raw file. Whenever you perform demosaicing you necessarily average with surrounding pixels and will reduce noise. The results are slightly different, becuause I was using only the green 1 channel with Iris and the areas selected for analysis may be slightly different.

                                       

                                      Linear16_Bit_results.png

                                      • 16. Re: ACR 5.4b: Is this zero noise reduction?
                                        Bill_Janes Level 2

                                        https://download.yousendit.com/UmNKcHBKMHdZY1IzZUE9PQ

                                         

                                        Gabor,

                                         

                                        Thanks for your input. Above is a link to the file. I would be interested in your findings.

                                         

                                        Bill

                                        • 17. Re: ACR 5.4b: Is this zero noise reduction?
                                          Panoholic Level 2

                                          Bill,

                                           

                                          I converted the image in DNG and modified it; you can download that from http://www.panopeeper.com/Download/BillJanes_Adobe_ISO03200to200_0111.DNG

                                           

                                          However, this is a too-well exposed shot (nothing is below the third stop of the DR), there is not much the NR can do. Still, I loaded them; the mean difference is only 0.57, it can be seen only for example by curves pulled to the minimum.

                                           

                                          This shows that there IS a difference, but the effect in a really noisy image is clearly visible without aggrevation.

                                           

                                          Gabor

                                          • 18. Re: ACR 5.4b: Is this zero noise reduction?
                                            sandy_mc Level 3

                                            Let me repeat back your own data to you:

                                             

                                            8-bit SD : 1.3

                                            14-bit SD : 2.27

                                            16-bit SD : 90-144

                                             

                                            AKA, comparing the SD's of data at different bit depths is apples to 

                                            oranges.

                                             

                                            Sandy

                                            • 19. Re: ACR 5.4b: Is this zero noise reduction?
                                              Bill_Janes Level 2

                                              sandy_mc wrote:

                                               

                                              Let me repeat back your own data to you:

                                               

                                              8-bit SD : 1.3

                                              14-bit SD : 2.27

                                              16-bit SD : 90-144

                                               

                                              AKA, comparing the SD's of data at different bit depths is apples to 

                                              oranges.

                                               

                                              Sandy

                                               

                                              Sandy,

                                               

                                              You have misinterpreted the data and apparently do not understand how bit depths can be normalized. The SD's of 1.7 and 2.27 are from ACR and Rawnalize respectively at a bit depth of 8 in a gamma 1.8 space. No conflict there--apples to apples. If you look at the Rawnalize SD for the raw green channel, it is 184 with a gamma of 1.0. The SD calculated from IRIS is 175, which is good agreement.

                                               

                                              The mean raw green value reported by Rawnalize is 8050 in the 14 bit raw file and 175.8 in an 8 bit gamma 1.8 file. To calculate the 8 bit gamma 1.8 value from the raw value, one normalizes the 14 bit value by dividing by 16384 to obtain a normalized value of 0.47133. For the gamma correction, one raises this value to the power of 1/1.8 (0.47133 ^ (1/1.8)) obtain 0.6738. To convert to 8 bit notation, one multiplies by 255 to get a value of 172 (close but not exactly the same as the Rawnalize value of 175.8. Gabor may be applying a linear segment for the shadows as is done with sRGB. For the standard deviation, one would have to apply a similar correction.

                                               

                                              With Iris, the raw data values are dumped into a 16 bit space but they are not scaled to 16 bit. In other words, if the 14 bit raw file is saturated at 16383, the value dumped into Iris is 16383 and not 65535. If you want 16 bit values, you have to add one, multiply by 4 and then subtract 1. In normal work, one merely multiplies by 4. One can compare different bit depths by normalizing them. In Photoshop, one can have the info palette report 8 bit or 16 bit data regardless of whether one is working in an 8 bit or 16 bit space.

                                               

                                              Bill

                                              • 20. Re: ACR 5.4b: Is this zero noise reduction?
                                                Panoholic Level 2
                                                The mean raw green value reported by Rawnalize is 8050 in the 14 bit raw file and 175.8 in an 8 bit gamma 1.8 file. To calculate the 8 bit gamma 1.8 value from the raw value, one normalizes the 14 bit value by dividing by 16384 to obtain a normalized value of 0.47133. For the gamma correction, one raises this value to the power of 1/1.8 (0.47133 ^ (1/1.8)) obtain 0.6738. To convert to 8 bit notation, one multiplies by 255 to get a value of 172 (close but not exactly the same as the Rawnalize value of 175.8. Gabor may be applying a linear segment for the shadows as is done with sRGB. For the standard deviation, one would have to apply a similar correction

                                                 

                                                Bill,

                                                 

                                                there is no linear ramp with the Gamma 1.8 mapping of Rawnalyze's implementation (nor in the Adobe RGB implementation). The discrepance is due to the fact, that the green channel is clipping at (or rather around) 15750, unlike the red and blue, which clip at 16383. However, ACR can not account for the different clipping levels, for the DNG specification allows only for one WhiteLevel for mosaic-type data (this is one of the DNG design errors).

                                                 

                                                Gabor

                                                • 21. Re: ACR 5.4b: Is this zero noise reduction?
                                                  sandy_mc Level 3

                                                  Bill,

                                                   

                                                  With all due respect, please stop posting new and different numbers 

                                                  with every post you make. Without going through your entire post, in 

                                                  in the first paragraph, why would you believe that Rawanalyse, when 

                                                  analyzing what you yourself say is 14 bit data, will normalize SD's to 

                                                  8-bit?? Why not 10-bit? Or 16-bit?

                                                   

                                                  If you really believe that Rawanalyze does as you suggest, just please 

                                                  be so good as to give everybody a reference on the Rawanalyse website 

                                                  where the author states that SD's are normalized to 8-bit......

                                                   

                                                  Sandy

                                                  • 22. Re: ACR 5.4b: Is this zero noise reduction?
                                                    Bill_Janes Level 2

                                                    Panoholic wrote:


                                                     

                                                    Bill,

                                                     

                                                    there is no linear ramp with the Gamma 1.8 mapping of Rawnalyze's implementation (nor in the Adobe RGB implementation). The discrepance is due to the fact, that the green channel is clipping at (or rather around) 15750, unlike the red and blue, which clip at 16383. However, ACR can not account for the different clipping levels, for the DNG specification allows only for one WhiteLevel for mosaic-type data (this is one of the DNG design errors).

                                                     

                                                    Gabor

                                                     

                                                    Gabor,

                                                     

                                                    Thanks for the explanation. The spec for aRGB does not include a linear ramp, but the implementatin does allow a slope limit not be greater than 1/32.

                                                    A slope limit of 1/32 affects 8-bit integer values 1 to 14. Refer to Anex C of the spec.

                                                     

                                                    http://www.adobe.com/digitalimag/pdfs/AdobeRGB1998.pdf

                                                     

                                                    In a recent post, Chris Cox confirmed that Adobe uses a linear ramp for both ProPhotoRGB and aRGB.

                                                     

                                                    http://forums.adobe.com/message/1657509#1657509

                                                    • 23. Re: ACR 5.4b: Is this zero noise reduction?
                                                      Panoholic Level 2
                                                      The spec for aRGB does not include a linear ramp, but the implementatin does allow a slope limit not be greater than 1/32

                                                       

                                                      Yes, I know that; I wanted to implement that too, but it is not a clean solution: a linear slope added to the original power function is not a tangent of the power function (it cuts through the power function's curve). The sRGB specification takes care of this by shifting the power function:

                                                       

                                                      f(x) = 12.92x                           if x <= 0.0031308
                                                      f(x) = 1.055 * (x ^ (1/2.4)) - 0.055    if x >  0.0031308


                                                      Instead of providing for a clean solution (difficult without having fixed the slope), Adobe adopted a "dirty" solution, apparent in the DNG converter: there is a "gradual transition" between the linear segment and the power function. Undocumented, of course (the DNG source code is not a documentation of ACR). I wonder who ever tried to invert it beside Adobe.

                                                       

                                                      On the other hand, I admit that this is a rather theoretical issue, except when someone wants to invert the transformation, and that does not occur in Rawnalyze. I make a note to adopt it, but only the 1/32 segment, without the funky transition.

                                                       

                                                      Gabor

                                                      • 24. Re: ACR 5.4b: Is this zero noise reduction?
                                                        Bill_Janes Level 2

                                                        sandy_mc wrote:

                                                         

                                                        Bill,

                                                         

                                                        With all due respect, please stop posting new and different numbers 

                                                        with every post you make. Without going through your entire post, in 

                                                        in the first paragraph, why would you believe that Rawanalyse, when 

                                                        analyzing what you yourself say is 14 bit data, will normalize SD's to 

                                                        8-bit?? Why not 10-bit? Or 16-bit?

                                                         

                                                        If you really believe that Rawanalyze does as you suggest, just please 

                                                        be so good as to give everybody a reference on the Rawanalyse website 

                                                        where the author states that SD's are normalized to 8-bit......

                                                         

                                                        Sandy

                                                         

                                                        For the 1.8 gamma results, the values are 8 bit (0..255), so the SDs would be in the same units. For the raw data (0..16383) the results are in ADUs (analog to digital units). Saturation is near 1^14 = 16384, which inidcates a bit depth of 14 for this camera. Gabor is the author if Rawnalize, and he is participating in this thread, so why don't you ask him directly?

                                                         

                                                        Bill

                                                        • 25. Re: ACR 5.4b: Is this zero noise reduction?
                                                          Panoholic Level 2
                                                          please  be so good as to give everybody a reference on the Rawanalyse website  where the author states that SD's are normalized to 8-bit......

                                                           

                                                          The Rawnalyze documentation is at http://www.cryptobola.com/PhotoBola/RawnalyzeGuide.htm, a general description with version and download information is at http://www.cryptobola.com/PhotoBola/Rawnalyze.htm

                                                           

                                                          The data displayed directly over the image relates to the selection, which is marked by an orangey rectangle. The first row relates to the raw pixel values, per raw channel; the range depends on the camera model and often on the ISO, and sometimes on selection by the user, like with the newer Nikons of the mid and high range, which allow for selecting between 12bit and 14bit raw data. The second row relates to the mapped values, R, G and B, always in the range from 0 to 255. The actual mapping is selected in the "Mapping function" box; some of them are extravagant, but useful when analyzing certain issues (they are substituting for contrast setting).

                                                           

                                                          The users' guide does not describe the fifth value in the group of raw values, which is the average pixel intensity relative to the clipping level, expressed in stops; this is the  "position in the dynamic range", useful when measuring the noise.

                                                           

                                                          Gabor

                                                          • 26. Re: ACR 5.4b: Is this zero noise reduction?
                                                            sandy_mc Level 3

                                                            Gabor,

                                                             

                                                            Thanks. You may have answered this indirectly when you talked about 

                                                            what on different rows, but could you directly say whether when 

                                                            looking at a 14-bit NEF file as in the example previously in this 

                                                            thread, the SD values are normalized to 8-bit, and if so, how the 

                                                            normalization is done?

                                                             

                                                            Thanks,

                                                             

                                                            Sandy

                                                            • 27. Re: ACR 5.4b: Is this zero noise reduction?
                                                              Panoholic Level 2
                                                              when  looking at a 14-bit NEF file as in the example previously in this  thread, the SD values are normalized to 8-bit, and if so, how the  normalization is done?

                                                               

                                                              Sandy,

                                                               

                                                              the numbers in the top row of the statistics on the selection are reflecting the original pixel values. I.e. the standard deviation is much greater with 14bit than with 12bit. The relative intensity is different: it does not change with the bit depth (or not much), because the clipping level changes more or less proportionally when the bit depth is changed from 12bit to 14bit. (This relates to those Nikons, which offer a selection between 12bit and 14bit.)

                                                               

                                                              The numbers in the second row relate to the mapped values; as Bill described above, they are projected on the range 0.0-1.0, and the mapping function(s) are applied. The result is again in the range 0.0-1.0, and that will be projected over the range 0-255. The standard deviation in that row relates to the result of the above process.

                                                               

                                                              Note, that the range of "operative pixel values" is seldom the full numerical range of the bit depth: the clipping level is more often than not under the maximum numerical value, and some sort of black level correction applies with most cameras. Thus the mapping can not always be tracked easily.

                                                               

                                                              Gabor

                                                              • 28. Re: ACR 5.4b: Is this zero noise reduction?
                                                                sandy_mc Level 3

                                                                Gabor,

                                                                 

                                                                Ok, I'm with you - so given that the SD's are of mapped rather than 

                                                                totally raw values, while it is correct that they apply to a range of 

                                                                0-255, you could not compare those SDs to the SD's from any other 

                                                                program, even if that program was also looking at raw values, unless 

                                                                you knew that the mapping function that the other program uses was the 

                                                                same as Rawanalyse uses. Which would be difficult given the very 

                                                                different processing pipelines.

                                                                 

                                                                BTW, are the SD's computed on floating point values in the range 0-255 

                                                                (e.g., 15.3), or integer values (e.g., 15)? - that will also make a 

                                                                difference to the computed SD. Or perhaps you do it against the 0-1 

                                                                values?

                                                                 

                                                                Thanks,

                                                                 

                                                                Sandy

                                                                • 29. Re: ACR 5.4b: Is this zero noise reduction?
                                                                  Panoholic Level 2

                                                                  while it is correct that they apply to a range of 0-255, you could not compare those SDs to the SD's from any other 

                                                                  program, even if that program was also looking at raw values, unless you knew that the mapping function that the other program uses was the same as Rawanalyse uses. Which would be difficult given the very different processing pipelines

                                                                   

                                                                  I don't see here a serious problem. The difference is only the lack of linear ramp in the Rawnalyze mapping, if one selects in ACR "linear tone curve".

                                                                   

                                                                  However, there is a more serious issue: the color space conversion. Rawnalyze presents the pixel values as they are, as if they were RGB values (this has nothing to do with the non-linear mapping function), while ACR's color space conversion is not avoidable, at least to my knowledge.

                                                                   

                                                                  I tried to coax ACR into not transforming the colors by replacing the matrixes in the DNG file by unity matrixes, but it was not a success. The response is one of those useful messages of ACR

                                                                   

                                                                  Could not complete your request because the file-format module cannot parse the file

                                                                   

                                                                  which is, of course, nonsense.

                                                                   

                                                                  are the SD's computed on floating point values in the range 0-255 (e.g., 15.3), or integer values (e.g., 15)?


                                                                  They are calculated using the displayed values, i.e. the integer RGB values. I don't see any point in calculating it based on intermediate values; the result could be strange. For example when selecting only a few pixels, their RGB values can be identical despite having slightly different raw values, thus the min, max and average would be identical but the standard deviation would be greater than 0.

                                                                  • 31. Re: ACR 5.4b: Is this zero noise reduction?
                                                                    Ramón G Castañeda Level 4

                                                                    Sandy,

                                                                     

                                                                    Your latest post is totally blank.   You must have attempted to post by email.

                                                                    • 32. Re: ACR 5.4b: Is this zero noise reduction?
                                                                      sandy_mc Level 3

                                                                      Umm, I hate this forum software.

                                                                       

                                                                      What my post was, was a suggestion to Gabor that the reason why ACR is thowing up with the unity matrixes may be because they are the same  rather than being unity. So what might work is a file with only a single matrix. With two identical matrixes, ACR can't interpolate a new matrix, which may be making it unhappy. But this is just a suggestion based on the way the DNG SDK works - I haven't tried it on ACR.

                                                                       

                                                                      Sandy

                                                                      • 33. Re: ACR 5.4b: Is this zero noise reduction?
                                                                        MadManChan2000 Adobe Employee

                                                                        What my post was, was a suggestion to Gabor that the reason why ACR is thowing

                                                                        up with the unity matrixes may be because they are the same  rather than being

                                                                        unity. So what might work is a file with only a single matrix. With two

                                                                        identical matrixes, ACR can't interpolate a new matrix, which may be making it

                                                                        unhappy. But this is just a suggestion based on the way the DNG SDK works - I

                                                                        haven't tried it on ACR.

                                                                         

                                                                        I am not sure which matrix/matrices Gabor tried to update, but I'm guessing

                                                                        it's the forward matrices, which map white-balanced linear camera

                                                                        coordinates to XYZ relative to D50. You cannot set these to identity

                                                                        matrices, because these matrices are required to map (1,1,1) to the D50

                                                                        white point.

                                                                         

                                                                        Eric

                                                                        • 34. Re: ACR 5.4b: Is this zero noise reduction?
                                                                          dorin_nicolaescu Level 5

                                                                          Eric,

                                                                           

                                                                          what about my original question? Or, at least, give a hint - is there a correct answer in the thread?

                                                                           

                                                                          Thanks.

                                                                          • 35. Re: ACR 5.4b: Is this zero noise reduction?
                                                                            Panoholic Level 2
                                                                            You cannot set these to identity matrices, because these matrices are required to map (1,1,1) to the D50 white point

                                                                             

                                                                            Thanks Eric, that was useful. I guess I would have to replace the original forward matrix with the inverse of the XYZ to target color space conversion in order to neutralize the conversion (but I don't see any reason to make that effort).

                                                                            • 36. Re: ACR 5.4b: Is this zero noise reduction?
                                                                              Bill_Janes Level 2

                                                                              dorin_nicolaescu wrote:

                                                                               

                                                                              Eric,

                                                                               

                                                                              what about my original question? Or, at least, give a hint - is there a correct answer in the thread?

                                                                               

                                                                              Thanks.

                                                                               

                                                                              No, there is noise reduction in ACR as a part of the demosaicing process. For information about the use of Fourier transforms in evaluating noise please see the post by Emil Martinec:

                                                                               

                                                                              http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p4.html

                                                                               

                                                                              Here is a Fourier transform of the green1 channel from my Nikon D3 at ISO 3200 (left). The frequency distribution throughout is uniform, indicative of no filtering. On the right is the green channel from an ACR rendering with noise reduction set to zero. Loss at high frequencies is present, indicative of filtering.

                                                                               

                                                                              fft_123_composite.png

                                                                              • 37. Re: ACR 5.4b: Is this zero noise reduction?
                                                                                MadManChan2000 Adobe Employee

                                                                                Bill, are you including all green pixels from the ACR rendering? (Remember,

                                                                                half of them have been synthesized ...)

                                                                                 

                                                                                 

                                                                                Here is a Fourier transform of the green1 channel from my Nikon D3 at ISO 3200

                                                                                (left). The frequency distribution throughout is uniform, indicative of no

                                                                                filtering. On the right is the green channel from an ACR rendering with noise

                                                                                reduction set to zero. Loss at high frequencies is present, indicative of

                                                                                filtering.

                                                                                 

                                                                                • 38. Re: ACR 5.4b: Is this zero noise reduction?
                                                                                  Bill_Janes Level 2

                                                                                  MadManChan2000 wrote:

                                                                                   

                                                                                  Bill, are you including all green pixels from the ACR rendering? (Remember,

                                                                                  half of them have been synthesized ...)

                                                                                   

                                                                                   

                                                                                  Eric,

                                                                                   

                                                                                  Yes, I analyzed the green channel of the conversion. The label Green 1 for the ACR conversion is misleading. Does the fact that half are synthesized affect the analysis?

                                                                                  • 39. Re: ACR 5.4b: Is this zero noise reduction?
                                                                                    MadManChan2000 Adobe Employee

                                                                                    >> Bill, are you including all green pixels from the ACR rendering? (Remember,

                                                                                    >> half of them have been synthesized ...)

                                                                                     

                                                                                    Yes, I analyzed the green channel of the conversion. The label Green 1 for the

                                                                                    ACR conversion is misleading. Does the fact that half are synthesized affect

                                                                                    the analysis?

                                                                                     

                                                                                    You will usually get different results depending on what subset of green

                                                                                    pixels you study. For example, you may find that green pixel values at

                                                                                    red/blue pixel locations have different frequency characteristics and lower

                                                                                    variance simply because they have been interpolated (i.e., those pixel

                                                                                    locations only contain red/blue pixel data in the original raw image, and

                                                                                    hence they are missing green pixel data, so the resulting green pixel values

                                                                                    at those pixel locations that you see in the final image have been

                                                                                    synthesized -- presumably by interpolating from existing nearby pixel data).

                                                                                    The exact characteristics would depend on the actual demosaic method, of

                                                                                    course.

                                                                                    1 2 Previous Next