[ImageJ-devel] IJ2 work/Histogram

[Curtis Rueden]

Hi Aivar,


> The mapping of values to histogram bins counts values that are less than
> the min in the first bin and values that are greater than the max in the
> last bin.  *Is this really desirable behavior?*

*
*The most flexible way of doing it would be to bin such values to special
"under" and "over" bins that are not part of the histogram proper, but
whose counts can be requested via the histogram API. In this way,
downstream code can decide how to consume it. Sound good?


> I haven't actually worked with any histograms of integral valued images
> but I thought they should work the same way as floating point, that is you
> could count a range of values in a single histogram bin.  *Is there any
> reason bins and values should always be 1:1 for integral valued images?*


Good point. We should *not* assume 1:1. Using the histogram code to
resample the lifetime bins would be valuable, I think. This would always be
limited to a "nearest neighbor" style resampling of course, which I think
is for the best.

That said, we need to make sure such flexibility does not complexify the
common 1:1 case. The API should provide shortcuts for that.


> If we specify min/max as centers some values slightly below minimum and
> slightly above maximum will map to the first and last histogram bins
> respectively.  *Do we need to specify min/max as bin center values?*


>From a mathematical standpoint, the two are totally interchangeable—it is
just a pain to convert between them. So the question comes down to which is
more intuitive. Personally I agree with you and find the "min as absolute
min" / "max as absolute max" more intuitive, rather than them being bin
centers. So let's go with that unless someone else has an objection. (At
worst, if there is disagreement, we can always add multiple API signatures
and/or a state flag to toggle the behavior.)

Regards,
Curtis


On Fri, Aug 10, 2012 at 6:13 PM, Aivar Grislis <grislis at wisc.edu> wrote:

>  >The mapping of values to histogram bins counts values that are less
> than the min in the first bin and values that are greater than the max in
> the last bin.  *Is this really desirable behavior? * In my usage of
> histograms I specifically don't want to count under- and over-range
> values.  Similarly when using a LUT I like under- and over-range values to
> appear black.  Of course if we need it work both ways we can provide a
> switch.
>
> I totally forgot, actually thats how IJ1 works: it does assign under-range
> values to the first color and over-range values to the last color.
>
> For my purposes, I had to kludge up the first and last colors of the LUT
> to be black and fiddle with the min/max values to make it all work.
>
>
> On 8/10/12 5:11 PM, Aivar Grislis wrote:
>
> Here is a discussion of changes I propose for dealing with histograms in
> IJ2.  Sorry it's so length, let me know if something is unclear and please
> comment if you have an interest in histograms; I do have some specific
> questions for the group (*in bold italics*):
>
>
> 1)  First of all I would create some helper methods like the following
> (could just be static methods somewhere)...
>
> // maps a value to a bin index, given the range min...max and the number
> of bins
> int valueToBin(int bins, double min, double max, double value);
>
> // returns array of edge values
> double[] getEdgeValues(int bins, double min, double max);
>
> // returns array of center values
> double[] getCenterValues(int bins, double min, double max);
>
> ...and equivalent set for integer values.
>
> Note that these methods would be used for both computing histograms and
> assigning colors from a LUT.  It's nice to use exactly the same code for
> both.  You can characterize a histogram bin or LUT color by the starting or
> edge value or by the center value, I've seen it both ways so I included
> both flavors.
>
> Having these common methods would also avoid bugs:
>
> i) One bug I saw in several places is to compute (value - min) / (max -
> min) and get a value from 0.0 to 1.0, but multiply that by 255.  The
> problem with that is the only way you get bin 255 is if value == min.
> Ideally each bin should represent a range of values; if the data set is
> uniformly distributed the size of the range of values per bin would be
> about the same.
>
> ii) Another bug was to compute width = max - min + 1.  That's certainly
> the way you do it for integer arithmetic but it doesn't work in floating
> point!
>
> I'd rather fix these bugs by refactoring to call the common helper methods.
>
>
> 2)  Regarding the "ImageStatistics" class idea that Curtis mentioned:
>
> For example it might include methods:
>
> // request certain things ahead of time
> doMinMax();
> doHistogram(int bins);
> doHistogram(int bins, double min, double max);
> doMean();
>
> // do minimal number of passes through the image
> process();
>
> // get accumulated results
> double[] getMinMax();
> long[] getHistogram();
> double getMean;
>
> Here if you don't specify a min/max for doHistogram the code has to take
> an initial pass to get min/max then another to build the histogram using
> that min/max.
>
>
> 3) Discussion of  net.imglib2.script.analysis.Histogram:
>
> Both of these ImgLib2 histogram methods combine a single pass through the
> image with generating histogram statistics.
>
> This code exhibits bug (i).
>
> I'm not sure that this handles integral values as well as floating point.
>
> The mapping of values to histogram bins counts values that are less than
> the min in the first bin and values that are greater than the max in the
> last bin.  *Is this really desirable behavior? * In my usage of
> histograms I specifically don't want to count under- and over-range
> values.  Similarly when using a LUT I like under- and over-range values to
> appear black.  Of course if we need it work both ways we can provide a
> switch.
>
> The output of this code is actually a JFreeChart image, there's no way to
> get the raw histogram count array.
>
> FYI, uses TreeMap<Double, Long> as an internal representation of the
> histogram count array.  Note that this combines what I am calling the edge
> values with the histogram counts.
>
>
> 4) Discussion of net.imglib2.algorithm.stats.Histogram:
>
> This has a HistogramBinMapper base class with RealBinMapper and
> IntBinMapper so it's meant to handle real and integral values.  The mapper
> is passed in as a parameter; I think it could be inferred from the cursor
> type T.
>
> Here the RealBinMapper class exhibits bug (ii) when calculating binWidth
> in the constructor.
>
> IntBinMapper doesn't let you specify a bin count, but assumes numBins =
> max - min + 1.  In other words, each bin will track a given integer value.
> I haven't actually worked with any histograms of integral valued images but
> I thought they should work the same way as floating point, that is you
> could count a range of values in a single histogram bin.  *Is there any
> reason bins and values should always be 1:1 for integral valued images?*
> This is so in IJ1 for 8-bit and RGB images.
>
> You pass in the minimum and maximum values when you create a
> HistogramBinMapper, but these are described as bin center values.  My
> concept of the minimum and maximum values has been the minimum is the
> lowest value that maps to the first histogram bin and the maximum is the
> highest value that maps to the last histogram bin.  If we specify min/max
> as centers some values slightly below minimum and slightly above maximum
> will map to the first and last histogram bins respectively.  *Do we need
> to specify min/max as bin center values?*
>
>
> Thanks!
> Aivar
>
>
> On 8/10/12 2:44 PM, Curtis Rueden wrote:
>
> Hi all,
>
>  Aivar & I also discussed histograms, and our plan for those is a bit
> more nebulous. Aivar is going to investigate improving Larry's histogram
> API in net.imglib2.algorithms.stats.
>
>  The general idea is that computing min/max, computing histograms, and
> computing other statistics are intrinsically related, and for performance
> reasons it is nice to minimize the number of passes through the image when
> computing these. So we would like to create a unified ImageStatistics class
> (sound familiar? ;-) that can compute one or more of these in an
> intelligent way, then return the results on demand.
>
>  None of this is reconciled with the net.imglib2.script.analysis package
> (Histogram, ImgMax, ImgMean, etc.), but perhaps in the future the
> net.imglib2.script stuff can take more direct advantage of it.
>
>  Regards,
> Curtis
>
>
> On Tue, Aug 7, 2012 at 4:45 PM, Aivar Grislis <grislis at wisc.edu> wrote:
>
>> I'd also be interested in working on the histogram code.  Mostly the code
>> that generates histogram data, also their display, perhaps the histogram
>> tool itself.
>>
>> SLIM Plugin uses histograms in a very dynamic way, as they are updated
>> periodically during the fit process.  You can change the minimum and
>> maximum values interactively.  So I might be an interested consumer of any
>> common histogram code.
>>
>> (I looked at the histogram code in the AutoContrast plugin,
>> net.imglib2.algorithm.stats.Histogram, and
>> net.imglib2.script.analysis.Histogram and found bugs in all of them when
>> binning up the histogram values.)
>>
>> Aivar
>>
>>
>
>
>
> _______________________________________________
> ImageJ-devel mailing listImageJ-devel at imagej.nethttp://imagej.net/mailman/listinfo/imagej-devel
>
>
>
> _______________________________________________
> ImageJ-devel mailing list
> ImageJ-devel at imagej.net
> http://imagej.net/mailman/listinfo/imagej-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://imagej.net/pipermail/imagej-devel/attachments/20120917/f151cf66/attachment.html>