R help - Oct 2004 - fidelity of generated raster images (R and perl)

Hi:

Goal: use R to turn a matrix of 1's and 0's
          into a corresponding image (e.g. png)
          of black and white pixels.

Why R: Yes, I can do this more efficiently and precisely
            with a perl module like Image::PBM.  Been there,
            done that many times, etc.  (Just humor me.
            I'm trying to do this with R for a number of reasons.)

Problem: Difficult to get a perfect rasterization.  There can
                be appended or removed pixel columns or pixel rows
                depending on plot region dimensions.  I witness this
                with both R version 1.8.1 and R version 2.0.

print($out "bitmap('/usr/local/mycrow/tmp/out.png', type =
'png256',
height = ".(int($height*68/64)/64).", width = 
".(int($width*69/64)/64).", res = 64, pointsize=0)\n");
print($out 'par(mar=c(0,0,0,0))'."\n");
my $width1 = $width-1;
my $height1 = $height-1;
print($out <<END);
plot.new()
plot.window(c(0,$width1),c(0,$height1))
rect(m[,1], m[,2], m[,1], m[,2], col="black",
border="black")

There are alternatives to rect (plot with type="p", pch=".",
etc)
and I have also tried png() instead of bitmap().  (I do prefer
bitmap so this can run without x11.)

I am guessing that R's internal region calculations are
vector based, which generally makes sense for most statistical
plots.  However, I do have some ideas for R and the presentation
of cellular automata results.

Any tips out there?  (Is it just a matter of height=50px to
overcome the inches default, etc?).....

Regards,
Scott

On Thu, 14 Oct 2004, Scott Harrison wrote:
> Hi:
> 
> Goal: use R to turn a matrix of 1's and 0's
>           into a corresponding image (e.g. png)
>           of black and white pixels.
Maybe use the write.pnm() function in the pixmap package:
> try1 <- matrix(0, nrow=100, ncol=100)
> try1[upper.tri(try1)] <- 1
> write.pnm(pixmapGrey(try1), file="~/tmp/try1.pbm",
type="pbm")
should do it (but as an ASCII file, bits are not so easy to do one by 
one). You need to watch which row order is used if the matrix and image 
row orders are to "look" the same. This is assuming that you don't
actually need to use a graphics device, but have an R matrix object you 
need to represent. 

Another package you could explore is rimage, which looks relevant.

Roger
> 
> Why R: Yes, I can do this more efficiently and precisely
>             with a perl module like Image::PBM.  Been there,
>             done that many times, etc.  (Just humor me.
>             I'm trying to do this with R for a number of reasons.)
> 
> Problem: Difficult to get a perfect rasterization.  There can
>                 be appended or removed pixel columns or pixel rows
>                 depending on plot region dimensions.  I witness this
>                 with both R version 1.8.1 and R version 2.0.
> 
> print($out "bitmap('/usr/local/mycrow/tmp/out.png', type =
'png256',
> height = ".(int($height*68/64)/64).", width = 
> ".(int($width*69/64)/64).", res = 64, pointsize=0)\n");
> print($out 'par(mar=c(0,0,0,0))'."\n");
> my $width1 = $width-1;
> my $height1 = $height-1;
> print($out <<END);
> plot.new()
> plot.window(c(0,$width1),c(0,$height1))
> rect(m[,1], m[,2], m[,1], m[,2], col="black",
border="black")
> 
> There are alternatives to rect (plot with type="p",
pch=".", etc)
> and I have also tried png() instead of bitmap().  (I do prefer
> bitmap so this can run without x11.)
> 
> I am guessing that R's internal region calculations are
> vector based, which generally makes sense for most statistical
> plots.  However, I do have some ideas for R and the presentation
> of cellular automata results.
> 
> Any tips out there?  (Is it just a matter of height=50px to
> overcome the inches default, etc?).....
> 
> Regards,
> Scott
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
> 
-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no

>>>>> "Scott" == Scott Harrison <harris41 at
msu.edu>
>>>>>     on Thu, 14 Oct 2004 11:13:16 -0400 writes:
    Scott> Hi: Goal: use R to turn a matrix of 1's and 0's into
    Scott> a corresponding image (e.g. png) of black and white
    Scott> pixels.

    Scott> Why R: Yes, I can do this more efficiently and
    Scott> precisely with a perl module like Image::PBM.  Been
    Scott> there, done that many times, etc.  (Just humor me.
    Scott> I'm trying to do this with R for a number of
    Scott> reasons.)

    Scott> Problem: Difficult to get a perfect rasterization.
    Scott> There can be appended or removed pixel columns or
    Scott> pixel rows depending on plot region dimensions.  I
    Scott> witness this with both R version 1.8.1 and R version
    Scott> 2.0.

 <....>

    Scott> There are alternatives to rect (plot with type="p",
    Scott> pch=".", etc) and I have also tried png() instead of
    Scott> bitmap().  (I do prefer bitmap so this can run
    Scott> without x11.)

I think the "most typical" alternative is to use  image()
which is probably the most efficient currently --- it is used by
the plot() method of "pixmap" objects in package 'pixmap'
{which
you probably should really look at!}.

Note however that image() is still very inefficient for
high-resolution (already 1000x1000) images since it really fills
a rectangular polygon for each pixel.  This e.g. leads to
horribly large postscript files when plotting such pixmaps.

In some way, this is a known "missing feature" in R's graphics
engines.  However a nice implementation would probably use
properties of the graphics device:

A smartImage() function would ``see'' that for a given output
device, it should only draw pixels instead of rects(); even
smarter for even larger pixmaps (where there are less "pixels"
on the output device than you have in your image) it would even average
("color-smooth") neighboring pixmap pixels into device pixels.


    Scott> I am guessing that R's internal region calculations
    Scott> are vector based, which generally makes sense for
    Scott> most statistical plots.  However, I do have some
    Scott> ideas for R and the presentation of cellular automata
    Scott> results.

    Scott> Any tips out there?  (Is it just a matter of
    Scott> height=50px to overcome the inches default,
    Scott> etc?).....

    Scott> Regards, Scott

Regards,
Martin Maechler