k-means Clustering
Locked
 
|
Dear All,
I am using k-means Clustering to measure the blue and red in my images. I use the threshold to move from one cluster to another. In some images the k-means clustering could produce very good matches (I'd say spot on), but in others the results were less desirable. My settings are Number of clusters to: 4 Cluster center tolerance: 0.0001 Enable randomization seed: ticked Randomization seed: 48 The other options were not ticked Could I do anything more to improve the analysis? and... Is it possible to pre-define the red and blue clusters (ie, assign cluster 1 to be blue)? Any help will be appreciated. Thanking you. Richard Han University of Edinburgh |
|
|
It might be helpful to know what sort of images these are.
I haven't worked a lot with k-means for color discrimination, but AFAIK this is just using kmeans clustering based on pixel RGB value distances, so it doesn't really know "blue." You could follow up the kmeans with a little algorithm that looks at the pixels in each cluster and evaluates which cluster is the most red and which is the most blue. Alternatively, you could use other means of segmenting an image into colors, such as HSV colorspace segmentation. toby >Date: Thu, 6 Dec 2007 09:53:04 +0000 >From: Richard Han <[hidden email]> >Subject: k-means Clustering > >Dear All, > >I am using k-means Clustering to measure the blue and red in my >images. I use the threshold to move from one cluster to another. In >some images the k-means clustering could produce very good matches >(I'd say spot on), but in others the results were less desirable. My >settings are > >Number of clusters to: 4 >Cluster center tolerance: 0.0001 >Enable randomization seed: ticked >Randomization seed: 48 >The other options were not ticked > >Could I do anything more to improve the analysis? and... Is it >possible to pre-define the red and blue clusters (ie, assign cluster >1 to be blue)? > >Any help will be appreciated. Thanking you. > >Richard Han >University of Edinburgh Toby C. Cornish, M.D., Ph.D. Pathology Resident Johns Hopkins Medical Institutions [hidden email] |
|
|
In reply to this post by Richard Han-2
Something else I forgot is that you could try kmeans after conversion from RGB colorspace to the LAB or another colorspace (convert to a 3-slice stack in a new colorspace and run kmeans on the stack). Plugins are readily available for that process.
toby >Date: Thu, 6 Dec 2007 09:53:04 +0000 >From: Richard Han <[hidden email]> >Subject: k-means Clustering > >Dear All, > >I am using k-means Clustering to measure the blue and red in my >images. I use the threshold to move from one cluster to another. In >some images the k-means clustering could produce very good matches >(I'd say spot on), but in others the results were less desirable. My >settings are > >Number of clusters to: 4 >Cluster center tolerance: 0.0001 >Enable randomization seed: ticked >Randomization seed: 48 >The other options were not ticked > >Could I do anything more to improve the analysis? and... Is it >possible to pre-define the red and blue clusters (ie, assign cluster >1 to be blue)? > >Any help will be appreciated. Thanking you. > >Richard Han >University of Edinburgh Toby C. Cornish, M.D., Ph.D. Pathology Resident Johns Hopkins Medical Institutions [hidden email] |
|
|
In reply to this post by Richard Han-2
Greetings List, here is an issue that has so far resisted our attempts...
Executive Summary: Can anyone suggest an approach to deal with a vignetting effect that does not begin from the optical center of the image? Details of the experiment * All figures referred to here can be seen at http://www.uoguelph.ca/~gyoung02 Background Aerial imagery from a tethered helium blimp is used to assess nitrogen status of specific horticultural crops. Images are captured using Kodak DCS-460 and -460c digital cameras from a height of approximately 350 feet (107 m). The images are stored in Kodak's proprietary format and imported using proprietary software. While the system has been used in a number of scientific and industrial applications, it was obviously intended for photojournalism or portrait photography purposes. This can cause problems when quantitative measures of radiance are required. Aside from the issue of non-direct measures of red, green, and blue wavelengths caused by the use of a single CCD array (with RGB filters applied in the Bayer pattern), there is the more important (I think) issue of vignetting effects caused by the geometry of camera system. A vignette is where the intensity (brightness) falls off from the image center, causing the edges (and corners especially) to appear darker. This effect becomes most apparent when applying a linear contrast stretch to the imagery (Figure 1 at the webpage above). Research Question The differences in radiance between regions of interest within the imagery are used to determine if a significant change exists. Problem The presence of vignetting effects may introduce trends that are completely independent of the processes under study. It is therefore important that these effects be quantified and removed if possible. This can be achieved through a knowledge of the camera response model and the radial distortion parameters of the lens. If this information is unavailable (as in my case), various algorithms have been developed which can identify the appropriate functions and use them to remove vignetting. One such algorithm was applied to Figure 1, the result of which can be seen in Figure 2. As you can see, the algorithm has done a fairly good job of removing vignetting, but if you look closely you will notice a trend, from left to right across the image, where systematic brightening seems to be occurring. This effect is the result of the model's assumption that vignetting occurs radially from the optical center of the image. However, by examining an image of a barium sulphate test panel taken with the near infrared camera (Figure 3) it can be seen the the vignette effect is not centered. It still appears to be radial, but with the optical center being shifted from where it should be. I hypothesize that this effect is due to a misalignment between the lens apparatus and CCD chip in the camera body. The REAL Question So my question is, can anyone suggest an approach to deal with a vignetting effect that does not begin from the optical center of the image? Regards, Tom Elliot |
|
|
In reply to this post by Richard Han-2
Hi,
I did something similar in the past. The solution was to use the Color Threshold plugin of Gabriel Landini and then to use k-means clustering. Another option is to try with the Color Deconvolution (again Gabriel) and then to apply the k-means. Another option is to implement Gaussian mixture model classifier. Cheers Dimiter Prodanov PS: You can post a link to the images for us to see. |
|
|
In reply to this post by Richard Han-2
Hi Toby,
Thank you so much for your reply. I would like to measure the area of red/blue from RGB images of histological slides. I have tried the Threshold color (HSV) as suggested to me by Jacqui Ross. It seems the Threshold color is very accurate (better than the k-means). If I use RGB split, the red channel actually resemble the blue in the original image?? I guess I will stick with the threshold color (HSV) for measurement. Regards Richard > > Date: Fri, 7 Dec 2007 10:31:32 -0500 > From: Toby Cornish <[hidden email]> > Subject: Re: k-means Clustering > > It might be helpful to know what sort of images these are. > > I haven't worked a lot with k-means for color discrimination, but > AFAIK this is just using kmeans clustering based on pixel RGB value > distances, so it doesn't really know "blue." You could follow up > the kmeans with a little algorithm that looks at the pixels in each > cluster and evaluates which cluster is the most red and which is > the most blue. > > Alternatively, you could use other means of segmenting an image > into colors, such as HSV colorspace segmentation. > > toby > >> Date: Thu, 6 Dec 2007 09:53:04 +0000 >> From: Richard Han <[hidden email]> >> Subject: k-means Clustering >> >> Dear All, >> >> I am using k-means Clustering to measure the blue and red in my >> images. I use the threshold to move from one cluster to another. In >> some images the k-means clustering could produce very good matches >> (I'd say spot on), but in others the results were less desirable. My >> settings are >> >> Number of clusters to: 4 >> Cluster center tolerance: 0.0001 >> Enable randomization seed: ticked >> Randomization seed: 48 >> The other options were not ticked >> >> Could I do anything more to improve the analysis? and... Is it >> possible to pre-define the red and blue clusters (ie, assign cluster >> 1 to be blue)? >> >> Any help will be appreciated. Thanking you. >> >> Richard Han >> University of Edinburgh > > > Toby C. Cornish, M.D., Ph.D. > Pathology Resident > Johns Hopkins Medical Institutions > [hidden email] > > ------------------------------ > > Date: Fri, 7 Dec 2007 10:48:08 -0500 > From: Toby Cornish <[hidden email]> > Subject: Re: k-means Clustering > > Something else I forgot is that you could try kmeans after > conversion from RGB colorspace to the LAB or another colorspace > (convert to a 3-slice stack in a new colorspace and run kmeans on > the stack). Plugins are readily available for that process. > > toby > >> Date: Thu, 6 Dec 2007 09:53:04 +0000 >> From: Richard Han <[hidden email]> >> Subject: k-means Clustering >> >> Dear All, >> >> I am using k-means Clustering to measure the blue and red in my >> images. I use the threshold to move from one cluster to another. In >> some images the k-means clustering could produce very good matches >> (I'd say spot on), but in others the results were less desirable. My >> settings are >> >> Number of clusters to: 4 >> Cluster center tolerance: 0.0001 >> Enable randomization seed: ticked >> Randomization seed: 48 >> The other options were not ticked >> >> Could I do anything more to improve the analysis? and... Is it >> possible to pre-define the red and blue clusters (ie, assign cluster >> 1 to be blue)? >> >> Any help will be appreciated. Thanking you. >> >> Richard Han >> University of Edinburgh > > > Toby C. Cornish, M.D., Ph.D. > Pathology Resident > Johns Hopkins Medical Institutions > [hidden email] > > |
|
|
On Monday 10 December 2007 10:02:41 Richard Han wrote:
> I would like to measure the area of red/blue from RGB images of > histological slides. Histological stains behave as "subtractive" colours and it is very likely that they will colocalise in some places. If that is the case, colour thresholding does not work because the hue of the co-localised stains is neither red nor blue (as it has contributions from both) while the thresholding will be assigning a pixel to one of the 2 classes. Have a look at Ruifrok's paper on colour deconvolution (I wrote a plugin that does that): Ruifrok AC, Johnston DA. Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol. 2001 Aug;23(4):291-9. Regards, G. |
Re: off-centre Vignette correction
|
|
In reply to this post by Thomas Elliot
Hi Tom,
I've been working some on algorithms for anti-vignetting, a.k.a., Lens Shading Correction, (LSC). I'm not sure what you are looking for. To me, the obvious answer to your question is make the origin of your LSC factor a variable. CMOS sensors (what I'm working with) have pixels that are like little wells. (I believe CCD sensors are the same way, but the pixel wells are not so deep.) When light comes in from off center through the color filters at the top of each pixel well, as is the case anywhere off center in the sensor, the light hits the photodiode active area off center, and may not fully illuminate the photodiode. This is an additional contributing factor to image shading. This is worse for wide angle lenses which have the light hitting the sensor at wider angles on the sides. Because the photodiodes may be asymmetrically shaped and asymmetrically positioned at the bottom of the pixel well, the intensity falloff can also be asymmetrical. So, the total falloff may not be what is predicted from analyzing the lens alone. And the sensor may be centered on the optical axis of the lens, and yet you will still get asymmetrical falloff. For my purposes, I've modeled lens falloff well enough (in some cases) as the simple product of parabolic functions in x any y with adjustable centers, x0 and y0. Your needs are obviously different, and so you will need to experiment with fitting different curves. I've found that the ImageJ plugin, ExpressionNT, (http://www.ulfdittmer.com/imagej/expression.html) is very helpful for exploring the impact of different variable gain functions applied to an image. For scientific least squares fit optimization, we have in the past used Excel and MatLab. Currently, we are using the optimization routines built into SciPy (http://www.scipy.org/SciPy) to calibrate our LSC routines. Hope this helps. -- Harry Parker Senior Systems Engineer Digital Imaging Systems, Inc. ----- Original Message ---- From: Thomas Elliot <[hidden email]> To: [hidden email] Sent: Friday, December 7, 2007 11:18:58 AM Subject: off-centre Vignette correction Greetings List, here is an issue that has so far resisted our attempts... Executive Summary: Can anyone suggest an approach to deal with a vignetting effect that does not begin from the optical center of the image? Details of the experiment * All figures referred to here can be seen at http://www.uoguelph.ca/~gyoung02 Background Aerial imagery from a tethered helium blimp is used to assess nitrogen status of specific horticultural crops. Images are captured using Kodak DCS-460 and -460c digital cameras from a height of approximately 350 feet (107 m). The images are stored in Kodak's proprietary format and imported using proprietary software. While the system has been used in a number of scientific and industrial applications, it was obviously intended for photojournalism or portrait photography purposes. This can cause problems when quantitative measures of radiance are required. Aside from the issue of non-direct measures of red, green, and blue wavelengths caused by the use of a single CCD array (with RGB filters applied in the Bayer pattern), there is the more important (I think) issue of vignetting effects caused by the geometry of camera system. A vignette is where the intensity (brightness) falls off from the image center, causing the edges (and corners especially) to appear darker. This effect becomes most apparent when applying a linear contrast stretch to the imagery (Figure 1 at the webpage above). Research Question The differences in radiance between regions of interest within the imagery are used to determine if a significant change exists. Problem The presence of vignetting effects may introduce trends that are completely independent of the processes under study. It is therefore important that these effects be quantified and removed if possible. This can be achieved through a knowledge of the camera response model and the radial distortion parameters of the lens. If this information is unavailable (as in my case), various algorithms have been developed which can identify the appropriate functions and use them to remove vignetting. One such algorithm was applied to Figure 1, the result of which can be seen in Figure 2. As you can see, the algorithm has done a fairly good job of removing vignetting, but if you look closely you will notice a trend, from left to right across the image, where systematic brightening seems to be occurring. This effect is the result of the model's assumption that vignetting occurs radially from the optical center of the image. However, by examining an image of a barium sulphate test panel taken with the near infrared camera (Figure 3) it can be seen the the vignette effect is not centered. It still appears to be radial, but with the optical center being shifted from where it should be. I hypothesize that this effect is due to a misalignment between the lens apparatus and CCD chip in the camera body. The REAL Question So my question is, can anyone suggest an approach to deal with a vignetting effect that does not begin from the optical center of the image? Regards, Tom Elliot ____________________________________________________________________________________ Never miss a thing. Make Yahoo your home page. http://www.yahoo.com/r/hs |
«
Return to ImageJ
|
1 view|%1 views
| Free forum by Nabble | Edit this page |