t024484 Posted September 27, 2009 Share #1 Posted September 27, 2009 Advertisement (gone after registration) Could someone please provide me with 3 DNG's taken at 640, 1250 and 2500 ISO of a white wall, not overexposed, but medium bright, to perform a 2D FFT noise analysis. Quality of the picture is fully unimportant, more important is that all 3 pictures are the same. This will show if Leica is using a in camera noise reduction algorithm at higher ISO values. Hans Link to post Share on other sites More sharing options...
Advertisement Posted September 27, 2009 Posted September 27, 2009 Hi t024484, Take a look here 2D FFT Analysis at ISO 2500. I'm sure you'll find what you were looking for!
ulrikft Posted September 27, 2009 Share #2 Posted September 27, 2009 Could someone please provide me with 3 DNG's taken at 640, 1250 and 2500 ISO of a white wall, not overexposed, but medium bright, to perform a 2D FFT noise analysis.Quality of the picture is fully unimportant, more important is that all 3 pictures are the same. This will show if Leica is using a in camera noise reduction algorithm at higher ISO values. Hans I'll try to fix that tonight. Link to post Share on other sites More sharing options...
scott kirkpatrick Posted September 27, 2009 Share #3 Posted September 27, 2009 I don't have an M9 to help out, but to clarify, you need this one with a little texture, and in focus, so that the fft has something to match up, right? Or do you want the wall out of focus, so that the only variations are noise? scott Link to post Share on other sites More sharing options...
t024484 Posted September 28, 2009 Author Share #4 Posted September 28, 2009 I don't have an M9 to help out, but to clarify, you need this one with a little texture, and in focus, so that the fft has something to match up, right? Or do you want the wall out of focus, so that the only variations are noise? scott Maybe it is a good idea to have both, focussed and unfocussed. We do not know yet how the M9 is acting upon noise. Hans Link to post Share on other sites More sharing options...
scott kirkpatrick Posted September 28, 2009 Share #5 Posted September 28, 2009 Once you have done a 2D FFT of a fairly uniform image, how will you analyze the spatial frequencies that you get from that? And what is your model of the noise to look for? scott Link to post Share on other sites More sharing options...
t024484 Posted September 28, 2009 Author Share #6 Posted September 28, 2009 Once you have done a 2D FFT of a fairly uniform image, how will you analyze the spatial frequencies that you get from that? And what is your model of the noise to look for? scott Good question. A bit of trial and error I guess. First thing to do is to compare the spatial frequency picture for the different ISO values. Second step is to compare it to the M8. If some filtering takes places, it will most likely have its effect to the higher frequencies. From what I have seen from the M8, it is better to start with pictures that are not too bright. In this case the S/N is worse, and noise can be better analysed. Hans Link to post Share on other sites More sharing options...
5541Q Posted September 28, 2009 Share #7 Posted September 28, 2009 Advertisement (gone after registration) hans - i have sent you an email with a download link to three DNG files. christoph Link to post Share on other sites More sharing options...
t024484 Posted September 28, 2009 Author Share #8 Posted September 28, 2009 As a first impression, it seems that ISO 2500 is indeed applying some form of noise reduction. In ascending order you see the 2Dimensional Fourier Transformation of pictures of a white wall taken at resp. ISO 604, 1250 and 2500. The centre of the image is the white wall, the further from the centre, the higher the frequency. At ISO 2500, the edges are much darker as with ISO 640 and 1250, meaning that a low pass filter has been applied to the data, filtering the higher frequencies. Hans Welcome, dear visitor! As registered member you'd see an image here… Simply register for free here – We are always happy to welcome new members! Link to post Share on other sites Simply register for free here – We are always happy to welcome new members! ' data-webShareUrl='https://www.l-camera-forum.com/topic/98259-2d-fft-analysis-at-iso-2500/?do=findComment&comment=1053364'>More sharing options...
scott kirkpatrick Posted September 28, 2009 Share #9 Posted September 28, 2009 Can you show us the result you get for ISO 160, as a baseline? scott Link to post Share on other sites More sharing options...
robertwright Posted September 28, 2009 Share #10 Posted September 28, 2009 I thought that wrt to cmos vs. ccd, cmos has a lot of hi frequency noise whereas ccd tends to have a lot of low frequency noise. why is leica filtering the hi frequency noise-this is not going to help banding-am I correct? and banding is what we get a lot of at 2500. the hi frequency noise tends to look more like grain so it is not a problem. at least this is what I understand. Sean Reid guessed that 1250 and 2500 were being filtered by looking at the file write speed. If you could do this transformation on all ISO's you could see where the filtering kicks in-but it could be that all ISO's are filtered too, just the strength is different. thanks for the analysis. Link to post Share on other sites More sharing options...
sandymc Posted September 28, 2009 Share #11 Posted September 28, 2009 Hans, Interesting analysis!! I see the analysis is from JPEGs - are those in camera JPGs, or generated from a raw developer? If a raw developer, what settings? Regards, Sandy Link to post Share on other sites More sharing options...
carlmuck Posted September 28, 2009 Share #12 Posted September 28, 2009 Hans, Interesting analysis!! I see the analysis is from JPEGs - are those in camera JPGs, or generated from a raw developer? If a raw developer, what settings? Regards, Sandy Sandy, I believe that what Hans did was perform the FFT on the dng file, and then plotted it and produced jpegs of the plotted data. I don't think he calculated across the jpg, but I could be wrong... c. Link to post Share on other sites More sharing options...
carlmuck Posted September 28, 2009 Share #13 Posted September 28, 2009 Sandy, I believe that what Hans did was perform the FFT on the dng file, and then plotted it and produced jpegs of the plotted data. I don't think he calculated across the jpg, but I could be wrong... c. Looks like I'm wrong. If Hans is using ImageJ (which is what it looks like) it doesn't grok DNG files, so he had to convert them. Your questions remain... c. Link to post Share on other sites More sharing options...
t024484 Posted September 29, 2009 Author Share #14 Posted September 29, 2009 Hans, Interesting analysis!! I see the analysis is from JPEGs - are those in camera JPGs, or generated from a raw developer? If a raw developer, what settings? Regards, Sandy Hi Sandy, I started with DNG's, converted them to Jpeg in LR2.5 without sharpening and without noise reduction. The analysis is not final. I suspect the results of ImageJ, and I am still working on this to have more evidence. One of the problems is that with Higher ISO, the level of the (shot) noise content gets higher as it is at Lower ISO, making the energy level in the frequency spectrum different. When transforming a crop of the picture, the main frequency content, being the white wall, causes a Sin(x)/x fallof in the frequency spectrum, hiding partly or even completely the real noise, depending on the energy level of the noise. So a conclusion on the base of the pictures above is unsafe, What I have to do, ist to modify the crop with a cosine or some other window, making it possible to look much deeper into the noise. Hans Link to post Share on other sites More sharing options...
t024484 Posted September 29, 2009 Author Share #15 Posted September 29, 2009 I thought that wrt to cmos vs. ccd, cmos has a lot of hi frequency noise whereas ccd tends to have a lot of low frequency noise.why is leica filtering the hi frequency noise-this is not going to help banding-am I correct? and banding is what we get a lot of at 2500. the hi frequency noise tends to look more like grain so it is not a problem. at least this is what I understand. Sean Reid guessed that 1250 and 2500 were being filtered by looking at the file write speed. If you could do this transformation on all ISO's you could see where the filtering kicks in-but it could be that all ISO's are filtered too, just the strength is different. thanks for the analysis. At higher ISO values, almost all noise come from Shot noise which is dictated by nature, and cannot be influenced. The use of CCD or CMOS does not change this. Only the pixel size has influence on the shot noise. So there is no issue of low or high frequency noise that I am aware off. The extra noise that CMOS sensors could have is because of larger differences in pixel sensitivities, but this is mostly compensated after exposure with a correction matrix. Filtering the high frequency noise, makes the surrounding pixels more equal and effectively filters noise. The price to pay is that the resolution decreases, but at High ISO this is the tradeoff one has to make. Hans Link to post Share on other sites More sharing options...
tgray Posted September 29, 2009 Share #16 Posted September 29, 2009 Wouldn't you want to use Tiff for analysis like this? JPG has compression, even if it is very little at the higher quality settings. The image is broken down in to 8x8 blocks and some kind of transform is applied to those blocks before storage. We've all see compression artifacts on low quality JPGs, but even high quality ones could affect the results of this kind of analysis, I would think. Link to post Share on other sites More sharing options...
robertwright Posted September 29, 2009 Share #17 Posted September 29, 2009 At higher ISO values, almost all noise come from Shot noise which is dictated by nature, and cannot be influenced. The use of CCD or CMOS does not change this. Only the pixel size has influence on the shot noise. So there is no issue of low or high frequency noise that I am aware off.The extra noise that CMOS sensors could have is because of larger differences in pixel sensitivities, but this is mostly compensated after exposure with a correction matrix. Filtering the high frequency noise, makes the surrounding pixels more equal and effectively filters noise. The price to pay is that the resolution decreases, but at High ISO this is the tradeoff one has to make. Hans thanks for the response. I was also thinking of total system noise not just shot noise. CCD's filter off the chip and so you can have shielding issues, timing issues etc. the banding at 2500 in the M8 I would guess is a timing issue between the various components. CMOS chips I thought delivered filtered digital information at the chip level, no further degradation can occur. What noise there is is mostly shot noise. ? Link to post Share on other sites More sharing options...
carlmuck Posted September 29, 2009 Share #18 Posted September 29, 2009 Hi Sandy, I started with DNG's, converted them to Jpeg in LR2.5 without sharpening and without noise reduction. The analysis is not final. I suspect the results of ImageJ, and I am still working on this to have more evidence. One of the problems is that with Higher ISO, the level of the (shot) noise content gets higher as it is at Lower ISO, making the energy level in the frequency spectrum different. When transforming a crop of the picture, the main frequency content, being the white wall, causes a Sin(x)/x fallof in the frequency spectrum, hiding partly or even completely the real noise, depending on the energy level of the noise. So a conclusion on the base of the pictures above is unsafe, What I have to do, ist to modify the crop with a cosine or some other window, making it possible to look much deeper into the noise. Hans Hans, If you're trying to look at the sensor noise, would't a dark image (shoot a lens cap) be better? I did a quick experiment with my M8, equal exposures (2s at 160, 1s at 320, 1/2 at 640, 1/4 at 1250, and 1/8 at 2500) of a lenscap, converted to TIFF and then calculated (using ImageJ) their FFT. The FFTs are progressively lighter (indicating less signal at lower ISO and more signal (noise) at higher, correct?) (ISO 160 FFT image followed by ISO 2500 image.) (ISO 160, 2 s., image of lenscap) (ISO 2500 1/8 s., image of lenscap) c. Link to post Share on other sites More sharing options...
sandymc Posted September 29, 2009 Share #19 Posted September 29, 2009 FYI, I just did a very quick and dirty 1-D FFT analysis on the actual raw data in the center row of some black field images at ISO 2500, ISO 1250 and ISO 200; the attached chart shows the graphs on a log scale, with each FFT adjusted for gain, and trendlines for each. If there was noise filtering going on, what I'd expect to see is the graph going down as it moves to the right. Basically, in 1-D anyway, I see no evidence of noise filtering - there's just random shot type noise. Not really conclusive as its only a 1-D FFT, and on black field image data rather than midtones, but interesting nevertheless. Sandy Welcome, dear visitor! As registered member you'd see an image here… Simply register for free here – We are always happy to welcome new members! Link to post Share on other sites Simply register for free here – We are always happy to welcome new members! ' data-webShareUrl='https://www.l-camera-forum.com/topic/98259-2d-fft-analysis-at-iso-2500/?do=findComment&comment=1054729'>More sharing options...
t024484 Posted September 29, 2009 Author Share #20 Posted September 29, 2009 Sandy, What you did is basically correct, but a number of things can be commented on: 1) You probably used pictures taken with the lenscap on, because if it had been a picture of a white or gray surface, you would have seen a peak in the middle of the FFT. 2) Noise is much better to analyse, when adding 10 to 20 pictures taken of the same object in a sequence with everything else the same . After adding, you should divide the sum by 10 or 20. As an alternative, you could add 10 to 20 different lines of one and the same picture, lying quite a bit apart, to prevent any correlation. What you will get then after transformation is a peak in the middle, and a smooth noise pattern on both sides. instead of the wild peaks going up and down as in your current picture. 3) to prevent sidebands from the main peak in the spectrum, the time sequence should be multiplied by a window, like a cosine or a Hanning window. This can be done after having added the whole sequence, and before calculating the FFT. If you perform like this, you should see the noise rising relative to the main peak, when ISO goes up, and at the same time will it be possible to see if the noise spectrum is filtered at higher ISO values. At this very moment, I do not have the time to do this, but the outcome will be absolutely reliable. What I did with ImageJ is very unsatisfactory, and sometimes even hard to reproduce. Hans Link to post Share on other sites More sharing options...
Recommended Posts
Archived
This topic is now archived and is closed to further replies.