K-NN Classsification on images

i1=dicomread('1.dcm');
i2=dicomread('2.dcm');
i3=dicomread('3.dcm');
i4=dicomread('4.dcm');
Sample=[i1;
        i2];
Training=[i3;
          i4];
Group=['1';
       '2'];
k=2;
Class = knnclassify(Sample, Training, Group, k);
disp(Class);

I get an error - The length of GROUP must equal the number of rows in TRAINING.

The size of both the images are same.

It works when I'm using co-ordinates instead of i1, i2 and so on.

I know I should be using fitcknn, but it should work as well. Any inputs please?

[Walter Roberson]

http://www.mathworks.co.uk/matlabcentral/answers/43073-a-guide-to-tags

Answer 1

[Arthur Goldsipe]

Hi,

Your input arguments Training and Group must have the same number of rows. I see that Group has only 2 rows, but Training probably has many more rows that than. Training consists of tow matrices stacked on top of each other. These matrices represent images as M-by-N matrices. So if all your images (i1, i2, i3, and i4) are all of size M-by-N, then Sample and Training have 2M rows. If your intention is to treat each image as a single entity, then you could construct Sample and Training as follows:

Sample   = [i1(:) i2(:)]';
Training = [i3(:) i4(:)]';

That said, I don't think this classification approach is very useful if you only have one training example from each class.

-Arthur

[prahlad h]

Thanks for your reply. I understood your point regarding the number of rows. My images are 512x512. So, for 2 such images, I have 512x2 rows and only 2 in the Group.

However, I'm not very familiar with the syntax used in matlab. Could you explain what's exactly happening in this code

Sample   = [i1(:) i2(:)]';
Training = [i3(:) i4(:)]';

This seems to be working as I don't get the error anymore, but what exactly do you mean by treating each image as a single entity?

Just to give some background - These are gabor filtered CT scan images of lungs. I'm implementing an algorithm which uses K-NN classification on these images.

[Arthur Goldsipe]

i1(:) reshapes the 512x512 matrix into a (512x1) column vector. The square brackets concatenate . And ' is the transpose operator. So Sample ends up being a 2x512 matrix.

In this case, what I mean by a "single entity" is a "single row" in the matrix, because that's how knnclassify identifies what numbers belong to the same observation.

[prahlad h]

Sorry for the noob question but what do you mean by a column vector? When you're reshaping 512x512 matrix into a 512x1 column vector, is there any loss in the image?

Also, a single row, does that mean that all the pixels are not compared when knnclassify is used?

It's working fine but I just wanted to know for my understanding!

Answer 2

[Image Analyst]

I know you've already accepted an answer, but you might want to check out my knn demo.

[prahlad h]

Sure. I'll look at it. It would help me with understanding it a little better. I've been through your other demos on segmentation and thresholding. They are amazing!