Hi Jeremy, If you are analysing the quality of your denoise strategy, you might also want to analyse the fourier spectrum of your images for each specific frequency pattern, independently. In that case, you typically need three images: your "noisy image", your "denoised image" and a "ground truth image", and you can check how the signal of your denoised image is improved compared to how the noisy image signal is behaving. There is also a question of what sort of noise you are dealing with, if your noise is symmetrically distributed (e.g. Gaussian), you might consider fourier shell correlation (3D) or fourier disk correlation (2D). I can see there are fiji plugins to compute FSC/FRC (e.g. this https://imagej.net/plugins/fourier-ring-correlation).
Cheers, Mauro On Tue, Apr 1, 2025 at 4:47 AM Jeremy Adler <[email protected]> wrote: > Hi Michael, > > You seem to be rather missing my point. > > A decent sized image has maybe 4 million pixels, > Any measurement that employs the peak pixel value is not necessarily very > descriptive of the whole or even part of an image. > And may be very poor indication of the overall image's quality - relative > to the noise, the reason for making the measurement. > > In the most extreme case there might be one intense pixel with the > remaining pixels indistinguishable from background. > In this instance the (peak signal)/noise or the peak(signal/noise) is > misleading, the image is essentially noise. > And assuming the image does contain an area of interest rather than a > single pixel of interest, a better image is required. > > The point is that the SN maybe a very misleading measurement. > This is why I asked for suggestions for measurements that better > illustrate the distribution of intensities of an image relative to the > level of noise. > > Jeremy > > > > > > > > > > > > > > > > > > > -----Original Message----- > From: Michael Schmid <[email protected]> > Sent: Monday, March 31, 2025 2:43 PM > To: [email protected] > Subject: Re: Median 3D filter > > Hi Jeremy, > > this might be a misunderstanding. > It is > (peak signal)/noise, not > peak(signal/noise). > https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio > > (From a grammar standpoint, the hyphens should be Peak-signal to noise, > but then there is no room for adding a the word "ratio", which would > require hyphens again; unless one has two levels of hyphens, > Peak-signal–to–noise ratio.) > > > BTW, what came to my mind after replying: > If Nataliya wants to get rid of salt/pepper noise, the most suitable > filter would be Process>Noise>Remove Outliers. > > For other sources of noise, one can then run some (edge-preserving) > smoothing filter in a second step. > > > Michael > ________________________________________________________________ > On 31.03.25 13:14, Jeremy Adler wrote: > > Signal to noise ratio is meaningful when receiving morse code but when > based on the value of the single most intense pixel in a 6000x6000 image > maybe far less informative. > > . > > > > Any suggestions for a measurement of image quality based on a larger > and therefore more representative population of pixels. > > > > Jeremy Adler > > > > > > =============================================== > > B i o V i s P l a t f o r m of Uppsala University > > Light & EM microscopy / FlowCytometry & Cell > > Sorting / Image Analysis =============================================== > > Jeremy Adler PhD - Senior research engineer > > Light, Confocal Microscopy, Image Analysis > > E-mail: [email protected] > > 0739 188170 > > www.uu.se/biovis > > > > Dag Hammarskjölds v 20 > > 751 85 UPPSALA, SWEDEN > > http://biovis.uu.se/ > > =============================================== > > > > > > > > > > > > > > -----Original Message----- > > From: Michael Schmid <[email protected]> > > Sent: Monday, March 31, 2025 12:28 PM > > To: [email protected] > > Subject: Re: Median 3D filter > > > > Hi Nataliya, > > > > just see the formula for PSNR: > > https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio > > > > You can get the mean squared error (MSE) by calculating the square of > > the difference image (which must be a 32-bit image), > > Process>Math>Square > > > > Then take the mean of the squared deviation, see Analyze>Measure and > > Analyze>Set Measurements... > > > > You can also get the maximum of the original (the signal) with > > Analyze>Measure. > > > > If you use a macro, Analyze>Measure would be replaced with the mean and > max values obtained via > > getStatistics(area, mean, min, max); > > > > > > Michael > > ________________________________________________________________ > > On 31.03.25 10:59, Наталія Тулякова wrote: > >> Hello, Michael. > >> I have implemented my own non-adaptive nonlinear (robust) filtering > >> algorithm as a plugin. > >> I use some noise-free images (such as "Barbara", "Lena" and > >> artificial RSA > >> (radio) images). I use the ImageJ command "Process - Noise" and > >> simulate salt and pepper and Gaussian noise with different variance. > >> I found the plugin for Poisson noise. But I can only calculate the > >> total difference between the pure test image and the image obtained > >> after the denoising algorithm. I use the commands "Process - Image > >> Calculator" and "Analyze - Measure". I mean is it possible by means > >> of ImageJ to calculate PSNR? Maybe there is a plugin for this? > >> Thanks. > >> With best regards, Nataliya > >> > >> > >> пн, 31 бер. 2025 р. о 11:19 Michael Schmid <[email protected]> > пише: > >> > >>> Dear Natalya, > >>> > >>> for calculating the signal-to-noise ratio, you need to know the > >>> "ground truth", i.e., what the image would be without noise. > >>> There are essentially two approaches for this: > >>> > >>> (1) Take a noise free image and add synthetic noise (preferably, > >>> with the same characteristics as the noise of real images; for > >>> photos with digital sensors the noise should be shot noise (Poisson > >>> statistics) + readout noise (roughly Gaussian) + dark current > >>> variations (best taken from a dark field with the same sensor). > >>> > >>> (2) As a reference image, you can use the average of many images of > >>> the same object (of course, without any lateral displacements due to > >>> vibrations, etc.). By averaging, much of the noise will cancel out. > >>> If dark current plays a role, you should subtract an average many > >>> dark fields. Best do these operations in 32-bit mode, since it may > >>> result in slightly negative values in low-intensity regions. > >>> When subtracting this low-noise result from the noisy single > >>> exposure, not that there will be a (roughly) constant offset, which > >>> should be removed before evaluating the noise. > >>> > >>> [A dark field is a photo with no light intensity reaching the > >>> sensor, but the same exposure time. You can use a black lens cap.] > >>> > >>> > >>> Best, > >>> > >>> Michael > >>> ________________________________________________________________ > >>> On 29.03.25 16:46, Наталія Тулякова wrote: > >>>> Hi, Michael. > >>>> > >>>> Thank you very much for the answer. > >>>> > >>>> I have developed nonlinear filtering algorithms and implemented > >>>> them as plugins, taking one of the freely available plugins > >>>> described in the package ImageJ as a prototype. I want to compare > >>>> the filter de-noising efficiency for some of the test 2D gray-scale > >>>> images. But I can only calculate the difference between the test > >>>> and filtered images using Image Calculator, and to obtain the total > >>>> mean value (Analyze - Measure). Is it possible to obtain the filter > >>>> efficiency estimation, for example, by > >>> means > >>>> of “Peak Signal-to-Noise Ratio”, using ImageJ program? > >>>> > >>>> Yours sincerely, Nataliya > >>>> > >>>> вт, 25 бер. 2025 р. о 13:30 Michael Schmid > >>>> <[email protected]> > >>> пише: > >>>> > >>>>> Hi Nataliya, > >>>>> > >>>>> both, Median and Median 3D use a circular support (circular Kernel). > >>>>> I noticed that Median 3D (and the other 3D filters) use a slightly > >>>>> different definition of the radius. > >>>>> For the 3D filters, sometimes one needs to add a small number like > >>>>> 0.5 to the radius of the 2D filters. > >>>>> For a 2D image (no stack), Median 3D with radius=2.5 in x&y and > >>>>> the "usual" (2D) Median with radius=2 do exactly the same (except > >>>>> near the edges, see below). > >>>>> At some radius values. the behavior is the same for the 2D and 3D > >>>>> filters (e.g. radius=10.5 and 14.5), so there is no simple rule. > >>>>> I think that eventually the 3D filters should be modified to use > >>>>> the same definition of the radius as the 2D filters, the one also > >>>>> used for > >>>>> Process>Filters>Show circular masks. > >>>>> > >>>>> The remaining difference is the handling of the edge pixels. > >>>>> The 3D filters consider the out-of image pixels as nonexistent. > >>>>> Thus, when calculating the median near the edge, the 3D median > >>>>> uses fewer pixels. The 2D filters (the "usual" Median, Mean, etc.) > >>>>> assume that the out-of-image pixels are the same as the nearest > >>>>> edge pixel, and the mean, median, etc. is always calculated over > >>>>> the same number of pixels (except for float images with NaN = Not a > Number). > >>>>> The latter convention (assuming repeated edge pixels) is the usual > >>>>> convention in ImageJ, also for Gaussian Blur, and the > >>>>> Process>Binary functions. > >>>>> > >>>>> Hope this helps, > >>>>> > >>>>> Michael > >>>>> ________________________________________________________________ > >>>>> > >>>>> > >>>>> On 22.03.25 12:32, Наталія Тулякова wrote: > >>>>>> Dear colleagues. > >>>>>> I am interested in filtering 2D images. The ImageJ program has > >>>>>> two > >>> median > >>>>>> filters in the "Process-Filters" menu item. I have not found any > >>>>>> documentation explaining how "Median 3D" works. Median 3D > >>>>>> provides > >>> better > >>>>>> results than Median. What is the difference between them for 2D > >>>>>> image processing? Does "Median 3D" use a square window while > >>>>>> "Median" uses a circular window? > >>>>>> > >>>>>> With best regards, Nataliya > >>>>>> > >>>>>> -- > >>>>>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html > >>>>> > >>>>> -- > >>>>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html > >>>>> > >>>> > >>>> -- > >>>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html > >>> > >>> -- > >>> ImageJ mailing list: http://imagej.nih.gov/ij/list.html > >>> > >> > >> -- > >> ImageJ mailing list: http://imagej.nih.gov/ij/list.html > > > > -- > > ImageJ mailing list: http://imagej.nih.gov/ij/list.html > > > > VARNING: Klicka inte på länkar och öppna inte bilagor om du inte känner > igen avsändaren och vet att innehållet är säkert. > > CAUTION: Do not click on links or open attachments unless you recognise > the sender and know the content is safe. > > > > > > > > > > > > > > > > > > > > När du har kontakt med oss på Uppsala universitet med e-post så > > innebär det att vi behandlar dina personuppgifter. För att läsa mer om > > hur vi gör det kan du läsa här: > > http://www.uu.se/om-uu/dataskydd-personuppgifter/ > > > > E-mailing Uppsala University means that we will process your personal > > data. For more information on how this is performed, please read here: > > http://www.uu.se/en/about-uu/data-protection-policy > > > > -- > > ImageJ mailing list: http://imagej.nih.gov/ij/list.html > > -- > ImageJ mailing list: http://imagej.nih.gov/ij/list.html > > VARNING: Klicka inte på länkar och öppna inte bilagor om du inte känner > igen avsändaren och vet att innehållet är säkert. > CAUTION: Do not click on links or open attachments unless you recognise > the sender and know the content is safe. > > > > > > > > > > När du har kontakt med oss på Uppsala universitet med e-post så innebär > det att vi behandlar dina personuppgifter. För att läsa mer om hur vi gör > det kan du läsa här: http://www.uu.se/om-uu/dataskydd-personuppgifter/ > > E-mailing Uppsala University means that we will process your personal > data. For more information on how this is performed, please read here: > http://www.uu.se/en/about-uu/data-protection-policy > > -- > ImageJ mailing list: http://imagej.nih.gov/ij/list.html > -- ImageJ mailing list: http://imagej.nih.gov/ij/list.html
