Hi, everyone,
I am a newbie on R. Now I want to do image pixel classification by random
forest. But I has not a clear understanding on random forest. Here is some
question:
As for an image, for example its size is 512x512 and has only one variable
-- gray level. The histogram of the image looks like mixture Gaussian Model,
say Gauss distribution (u1,sigma1), (u2,sigma2),(u3,sigma3). And a image
classified by K-means or EM algorithm, so the class label image is also
512x512 and has 0, 1, 2 value.
I read the binary image data as follows:
datafile <- file("bone.img","rb")
img <-
readBin(datafile,size=2,what="integer",n=512*512,signed=FALSE)
img <- as.matrix(img)
close(datafile)
labelfile <- file(label.img","rb")
label <-
readBin(labelfile,size=2,what="integer",n=512*512,signed=FALSE)
label <- as.matrix(label)
close(labelfile)
img_and_label <- c(img,label) // binds the image data and class label
img_and_label <- as.matrix(img_and_label)
img_and_label <- array(img_and_label, dim=c(262144,2))
Random Forest need a class label like "Species" in the iris. I do not
know
how
to set a class label like "Species" to the img. So I run the command
as
follows:
set.seed(166)
rf <- randomForest(img_and_label[,2],data=image_and_label,importance=TRUE,
proximity=TRUE)
which outputs:
Error in if (n == 0) stop("data (x) has 0 rows") :
argument is of length zero
Could anyone tell what is wrong and how can do the RF?
Secondly, if there is an new image , say img3 (dimension is 512x512,too),
how can I
use the former result to classifify the new image?
Thirdly, whether or not random forest be used well if there is only one
variable, say pixel
gray level, or three variables, such as red, green, blue color component to
an true color
image?
Thank you very much!
Best,
Fucang
=======================================Fucang Jia, Ph.D student
Institute of Computing Technology, Chinese Academy of Sciences
Post.Box 2704
Beijing, 100080
P.R.China
E-mail:fcjia at ict.ac.cn
It looks like image_and_label has only 2 columns, so when you take
img_and_label[,2] you have a vector left. Even if that weren't the case,
you're going to need to pass in both the gray scale points and labels,
presumably in a data frame. You've created a character matrix below, so
you're just passing in a character vector of labels.
You'll probably want something like
rf <- randomForest(label~image,data=image_and_label,importance=TRUE,
proximity=TRUE),
assuming that image_and_label is a data frame with elements image and label.
For the second question, see the documentation for the predict method for
random forests; for the third, the answer is yes, random forests can be used
with multiple variables.
There is an introduction to the random forests package in volume 2, issue 3
of the R newsletter (available in the documentation section of cran).
Hope this helps,
Matt Wiener
-----Original Message-----
From: r-help-bounces at stat.math.ethz.ch
[mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Fucang Jia
Sent: Monday, November 24, 2003 10:31 AM
To: r-help at stat.math.ethz.ch
Subject: [R] Questions on Random Forest
Hi, everyone,
I am a newbie on R. Now I want to do image pixel classification by random
forest. But I has not a clear understanding on random forest. Here is some
question:
As for an image, for example its size is 512x512 and has only one variable
-- gray level. The histogram of the image looks like mixture Gaussian Model,
say Gauss distribution (u1,sigma1), (u2,sigma2),(u3,sigma3). And a image
classified by K-means or EM algorithm, so the class label image is also
512x512 and has 0, 1, 2 value.
I read the binary image data as follows:
datafile <- file("bone.img","rb")
img <-
readBin(datafile,size=2,what="integer",n=512*512,signed=FALSE)
img <- as.matrix(img)
close(datafile)
labelfile <- file(label.img","rb")
label <-
readBin(labelfile,size=2,what="integer",n=512*512,signed=FALSE)
label <- as.matrix(label)
close(labelfile)
img_and_label <- c(img,label) // binds the image data and class label
img_and_label <- as.matrix(img_and_label)
img_and_label <- array(img_and_label, dim=c(262144,2))
Random Forest need a class label like "Species" in the iris. I do not
know
how
to set a class label like "Species" to the img. So I run the command
as
follows:
set.seed(166)
rf <- randomForest(img_and_label[,2],data=image_and_label,importance=TRUE,
proximity=TRUE)
which outputs:
Error in if (n == 0) stop("data (x) has 0 rows") :
argument is of length zero
Could anyone tell what is wrong and how can do the RF?
Secondly, if there is an new image , say img3 (dimension is 512x512,too),
how can I
use the former result to classifify the new image?
Thirdly, whether or not random forest be used well if there is only one
variable, say pixel
gray level, or three variables, such as red, green, blue color component to
an true color
image?
Thank you very much!
Best,
Fucang
=======================================Fucang Jia, Ph.D student
Institute of Computing Technology, Chinese Academy of Sciences
Post.Box 2704
Beijing, 100080
P.R.China
E-mail:fcjia at ict.ac.cn
______________________________________________
R-help at stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
It's not clear to me what you want to do, but if I understand your problem somewhat, I don't see how randomForest would be relevant. Sounds like you are doing the following: o Read in a 512x512 image with pixel intensities. o You somehow fit a 3-component normal mixture model to the intensity data, and have labels for which component the pixels belong to. o You want to be able to "fit" (or "predict") other images to the 3-component mixture model you have; i.e., create the "label" data given an image. If that's about right, I don't see why you would need some learning algorithm such as randomForest. You should be able to compute the likelihood that a pixel belong to each of the 3 components in the mixture model, based on the fitted parameters of that model. The simplest I can think of, ignoring the mixing proportions, is to simply compute the absolute Z scores of a pixel with respect to the three components: z1 abs((x-u1)/sigma1), z2 = abs((x-u2)/sigma2), z2 = abs((x-u3)/sigma3), and assign the pixel to the component with the largest absolute z-score. HTH, Andy> From: r-help-bounces at stat.math.ethz.ch > [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Fucang Jia > > Hi, everyone, > > I am a newbie on R. Now I want to do image pixel > classification by random > forest. But I has not a clear understanding on random forest. > Here is some > question: > > As for an image, for example its size is 512x512 and has only > one variable > -- gray level. The histogram of the image looks like mixture > Gaussian Model, > say Gauss distribution (u1,sigma1), (u2,sigma2),(u3,sigma3). > And a image > classified by K-means or EM algorithm, so the class label > image is also > 512x512 and has 0, 1, 2 value. > > I read the binary image data as follows: > > datafile <- file("bone.img","rb") > img <- readBin(datafile,size=2,what="integer",n=512*512,signed=FALSE) > img <- as.matrix(img) > close(datafile) > > labelfile <- file(label.img","rb") > label <- > readBin(labelfile,size=2,what="integer",n=512*512,signed=FALSE) > label <- as.matrix(label) > close(labelfile) > > img_and_label <- c(img,label) // binds the image data and class label > img_and_label <- as.matrix(img_and_label) > img_and_label <- array(img_and_label, dim=c(262144,2)) > > > Random Forest need a class label like "Species" in the iris. > I do not know > how > to set a class label like "Species" to the img. So I run the > command as > follows: > > set.seed(166) > rf <- > randomForest(img_and_label[,2],data=image_and_label,importance=TRUE, > proximity=TRUE) > > which outputs: > > Error in if (n == 0) stop("data (x) has 0 rows") : > argument is of length zero > > Could anyone tell what is wrong and how can do the RF? > > Secondly, if there is an new image , say img3 (dimension is > 512x512,too), > how can I > use the former result to classifify the new image? > > Thirdly, whether or not random forest be used well if there > is only one > variable, say pixel > gray level, or three variables, such as red, green, blue > color component to > an true color > image? > > Thank you very much! > > Best, > > Fucang > > =======================================> Fucang Jia, Ph.D student > Institute of Computing Technology, Chinese Academy of Sciences > Post.Box 2704 > Beijing, 100080 > P.R.China > E-mail:fcjia at ict.ac.cn > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://www.stat.math.ethz.ch/mailman/listinfo/r-help >