Mailman 3 December 2013 - scikit-image

Re: Better results with Canny/Hough for circular particles
by Johannes Schönberger 14 Dec '13

14 Dec '13

Am 14.12.2013 um 00:09 schrieb Adam Hughes <hughesadam87(a)gmail.com>: > I'm trying multiple methods for isolating nanoparticles for a paper in which the purpose will be to contextualize a lot of available image processing workflows best suited for sizing, separating and counting nanoparticles. The idea is that imaging nanoparticles is becomes a widely standard necessity, and so I want to present an overview of the availble methods and workflows for best processing images, as well best practices for obtaining high quality images in the lab via SEM/AFM. > > I've already covered approaches via auto thresholding, with optionally applying watersheding to find particle boundaries. OK, segmentation would have been my first suggestion. > IMO, Canny, Hough, Sobel etc... give another approach; whereby, one finds and deals directly with the particle boundaries/outlines. The advantage of this approach is that it lends finer control over the particle boundaries, which are often hazy in SEM images. I think that there are some use cases for this approach. For example, I've observed that canny edge detection gives a great fit to particles that have a halo effect, and so I'm pretty confident that, if we could fill all of the edges, that we'd be sizing the particles very nicely. It's easier to assess the "goodness of fit" looking at just the edges, but obviously it introduces more work. I at least had hoped to present a bit of this edge --> filter ---> fill approach in the paper to complement the threshold approaches to segmentation. > > I tried the morphological operators (open, close, dilation, edge), but they are a bit restrictive. I'd need something that has a variable structure parameter, as Dan mentioned sort of is builtin to canny (ie, it's smart about how it closes the regions). I'll probably just leave it at "the input parameters are crucial for getting closed regions" instead of focusing on how to connect the regions after the fact. > > PS, is anyone familiar with a function in scikit or ndimage to give the perimeter/outline of a filled region? skimage.measure.regionprops, skimage.measure.perimeter

2 1

Restoration module
by Stéfan van der Walt 14 Dec '13

14 Dec '13

Hi everyone, Just a heads-up on a new scikit-image submodule: ``skimage.restoration``. Thanks to François Orieux, we are fortunate to have the following deconvolution algorithms implemented: [1] `wiener` [2] `unsupervised_wiener` [3] `richardson_lucy` I'd pay specific attention to Francois' own algorithm [2] -- no more painful parameter tweaking to get a good Wiener-Hunt deconvolution! Thanks also to everyone on the team who worked hard to review and shape this PR. Stéfan

1 0

Re: New examples
by Stéfan van der Walt 14 Dec '13

14 Dec '13

Hey, Adam On Sat, Dec 14, 2013 at 2:42 AM, Adam Hughes <hughesadam87(a)gmail.com> wrote: > Perhaps put any videos like this in the gallery as well? We could also add links to StackOverflow posts. Would you mind being our curator, and send a pull request either on the website repo or on the docs to make this happen? Thank you! Stéfan

1 0

New examples
by Adam Hughes 14 Dec '13

14 Dec '13

Hey guys, The new examples in the gallery look great, and are quite helpful! I just watched a tutorial from Stefan today (PyData 2012), which was really great: http://vimeo.com/53065496 Perhaps put any videos like this in the gallery as well?

1 0

Re: Better results with Canny/Hough for circular particles
by Adam Hughes 13 Dec '13

13 Dec '13

Hi Dan, Thanks for the quick reply. I think that I can get better results if I tweak the parameters. The threshold parameter intuitively makes sense, but I'll have to read a bit to get familiar with sigma and the algorithm in general. Thanks for the explanation; it really helped. I will try out the erosion as well. PS, do you have any feelings towards the applicability of circular hough to my image? On Friday, December 13, 2013 3:03:30 PM UTC-5, Dan Farmer wrote: > > Hi Adam, > > This can be the worst part of image processing, but I'm curious how > much you played with the parameters to Canny? You probably know this, > but canny already tries to close gaps (hysteresis thresholding). What > you want to do is try to lower the low_threshold parameter (values > above the high threshold value get initially labeled as edges, then it > looks for pixels that are connected to edge pixels and whose value is > > low_threshold to link the edges). > > An easy/basic way to get rid of small fragments would be to start with > morphological erosion. > > -Dan > > On Fri, Dec 13, 2013 at 11:47 AM, Adam Hughes <hughes...(a)gmail.com<javascript:>> > wrote: > > Hi, > > > > I have several images of circular particles (see attached for an > example) > > and I've been experimenting with automatic routines to find edges. > > > > I've found that with Canny, I can get really nice edges, but the edges > are > > not always connected. Thus, when I do fill-binary, many of my particles > are > > not painted in due to slight breaks in the border returned by canny. Is > > there an ideal way to fix this, either by connecting "almost" connected > > canny edges? Additionally, what is the best way to filter out small > > fragments and/or non-circular edges? > > > > I've attached an image of the canny outlines; you can see that I > obviously > > want to get rid some of the regions that aren't associated with any > > particles. PS, the coloring of the outlines are based on the brightness > of > > the image at that point underneath it, which has been hidden. (Would be > > happy to share the function if anyone wants it). > > > > Lastly, I tried adapting the circular hough transform example: > > > > > http://scikit-image.org/docs/dev/auto_examples/plot_circular_elliptical_hou… > > > > But struggled with setting it up, due to a naive understanding of the > > algorithm. Given that my image has thousands of particles, but I know > > roughly the size distribution, would the circular hough transform be > useful > > to me? > > > > Thanks > > > > > > -- > > You received this message because you are subscribed to the Google > Groups > > "scikit-image" group. > > To unsubscribe from this group and stop receiving emails from it, send > an > > email to scikit-image...(a)googlegroups.com <javascript:>. > > For more options, visit https://groups.google.com/groups/opt_out. >

2 2

Re: Better results with Canny/Hough for circular particles
by Dan Farmer 13 Dec '13

13 Dec '13

Hi Adam, This can be the worst part of image processing, but I'm curious how much you played with the parameters to Canny? You probably know this, but canny already tries to close gaps (hysteresis thresholding). What you want to do is try to lower the low_threshold parameter (values above the high threshold value get initially labeled as edges, then it looks for pixels that are connected to edge pixels and whose value is > low_threshold to link the edges). An easy/basic way to get rid of small fragments would be to start with morphological erosion. -Dan On Fri, Dec 13, 2013 at 11:47 AM, Adam Hughes <hughesadam87(a)gmail.com> wrote: > Hi, > > I have several images of circular particles (see attached for an example) > and I've been experimenting with automatic routines to find edges. > > I've found that with Canny, I can get really nice edges, but the edges are > not always connected. Thus, when I do fill-binary, many of my particles are > not painted in due to slight breaks in the border returned by canny. Is > there an ideal way to fix this, either by connecting "almost" connected > canny edges? Additionally, what is the best way to filter out small > fragments and/or non-circular edges? > > I've attached an image of the canny outlines; you can see that I obviously > want to get rid some of the regions that aren't associated with any > particles. PS, the coloring of the outlines are based on the brightness of > the image at that point underneath it, which has been hidden. (Would be > happy to share the function if anyone wants it). > > Lastly, I tried adapting the circular hough transform example: > > http://scikit-image.org/docs/dev/auto_examples/plot_circular_elliptical_hou… > > But struggled with setting it up, due to a naive understanding of the > algorithm. Given that my image has thousands of particles, but I know > roughly the size distribution, would the circular hough transform be useful > to me? > > Thanks > > > -- > You received this message because you are subscribed to the Google Groups > "scikit-image" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to scikit-image+unsubscribe(a)googlegroups.com. > For more options, visit https://groups.google.com/groups/opt_out.

1 0

Better results with Canny/Hough for circular particles
by Adam Hughes 13 Dec '13

13 Dec '13

Hi, I have several images of circular particles (see attached for an example) and I've been experimenting with automatic routines to find edges. I've found that with Canny, I can get really nice edges, but the edges are not always connected. Thus, when I do fill-binary, many of my particles are not painted in due to slight breaks in the border returned by canny. Is there an ideal way to fix this, either by connecting "almost" connected canny edges? Additionally, what is the best way to filter out small fragments and/or non-circular edges? I've attached an image of the canny outlines; you can see that I obviously want to get rid some of the regions that aren't associated with any particles. PS, the coloring of the outlines are based on the brightness of the image at that point underneath it, which has been hidden. (Would be happy to share the function if anyone wants it). Lastly, I tried adapting the circular hough transform example: http://scikit-image.org/docs/dev/auto_examples/plot_circular_elliptical_hou… But struggled with setting it up, due to a naive understanding of the algorithm. Given that my image has thousands of particles, but I know roughly the size distribution, would the circular hough transform be useful to me? Thanks

1 0

Re: Pattern drawing library ideas
by Adam Hughes 09 Dec '13

09 Dec '13

Thanks for the advice guys! Josh, that makes sense, thanks. I will implement that suggestion, and thanks for explaining rr and cc. Stefan, thanks! I will still share some example notebooks down the road, but keep the library separate from skimage. If there are any new particle types that might make sense to be added to skimage.draw, then i'd be happy to share at that time. On Friday, December 6, 2013 7:07:57 PM UTC-5, Adam Hughes wrote: > > Hi everyone, > > My colleague Evelyn and I have been using scikit image's draw utilities to > generate test data for modeling particle distributions on optical fibers > (ie the SEM images I sent a few weeks ago). We are doing this to test some > new models for nanoparticle cluster fitting that we've developed. > > In working on this, we believe we've come up with a general framework for > drawing 2D particle ensembles in skimage. One could envision using it to > do something like: > > Step 0: Generate a blank canvas of resolution 1024 x 768 > Step 1: Add circles of average radius 5 pixels until 30% of the image > is covered. > - Arrange these randomly vs. equally spaced on a grid > etc... > Step 2: Add an "cluster" of particles to this image. Paint only the > clusters red. > Step 3: Run an optimization that maximizes the inter-particle spacing. > Step 4: Make all the particles smaller than a certain area green. > Step 5: Try watershedding > etc... > > By generating the sample images in skimage, we wouldn't have to break out > of the python workflow to make our test data. > > We believe that we've come up with some abstractions that would allow for > such an toolset. This includes defining a abstract class for different > particle shapes, and then wrapping skimage.draw() functions as a class > method. Before we get invested in this, I had a few questions: > > 1. Is this something that, if executed well, would be of interest to > incorporate into scikit image? If so, I will start working on it as a > branch; otherwise, we'll just use skimage as a dependency. I'd image it > would either be a submodule of skimage.draw(), something like > skimage.draw.ensemble() or draw.psuedodata()... > > 2. Is anyone aware of a pre-existing tooset/library (preferrably in > Python) that's built for this? And if so, is that library compatible to > skimage? > > 3. When a user runs skimage.draw.circle(), it returns *rr* and *cc*, > what are these? Is cc the "chain code"? > > One caveat: our design plan is object oriented. We thought that the best > way to have an image with several particles would require a *Particle*class to add enough metadata to the returns of scikit.draw() so that > particles could indivually be tracked, isolated and manipulated. The > ensemble would be created on a *Canvas* class (better name? *TestImage*?), > which is responsible for storing an ensemble of Particles, as well as all > the drawing and organization of the ensemble. For example, a circle would > have attributes X, Y, R, which are then passed to a draw() method that > called skimage.circle(). In this way, one could track particle positions, > manipulate and redraw(). Would something like this clash with skimage's > basic design paradigms? If so we, maybe it would best to keep this toolkit > out of skimage. > > PS, is anyone else working on this, or interested? > > Thanks >

1 0

exposure.equalize_adapthist() problem
by Simon Filhol 09 Dec '13

09 Dec '13

Hi, after reinstalling, updating the skimage package version 0.9.x , I cannot get the function equalize_adapthist() to work like the example on the scikit-image webpage. from skimage import exposure, data img = data.moon() img_adapteq = exposure.equalize_adapthist(img, clip_limit=0.03) the following error is returned in the console: >>> img_adapteq = exposure.equalize_adapthist(img, clip_limit=0.03) Traceback (most recent call last): File "<stdin>", line 1, in <module> AttributeError: 'module' object has no attribute 'equalize_adapthist' Would you have any insight into which problem I am running? Thanks

2 1

Re: match_template confidence > 3 ?
by Johannes Schönberger 09 Dec '13

09 Dec '13

See https://github.com/scikit-image/scikit-image/pull/847. Can you please test if this fixes your issues? Am 08.12.2013 um 03:22 schrieb Stéfan van der Walt <stefan(a)sun.ac.za>: > On Sat, Dec 7, 2013 at 9:02 AM, Tony Yu <tsyu80(a)gmail.com> wrote: >> Unfortunately, I'm not sure what the problem is here. You're correct in >> thinking that values shouldn't be outside of [-1, 1], but as your code >> demonstrates, there are apparently inputs that break the current >> implementation. I've played around with isolating a small test image, but I >> haven't had much luck. > > I've filed an issue with a minimal code snippet here: > > https://github.com/scikit-image/scikit-image/issues/845 > > Stéfan > > -- > You received this message because you are subscribed to the Google Groups "scikit-image" group. > To unsubscribe from this group and stop receiving emails from it, send an email to scikit-image+unsubscribe(a)googlegroups.com. > For more options, visit https://groups.google.com/groups/opt_out.

1 0