How to read a subplot as an image and find correlation for two images?

Hi, I am writing a code for image segmentation where it takes multiple images as input and the output is as follow for each images.

The relevent code is:

figure(k);
subplot(1,2,1);
imshow(imageArray,[]);
title('Image');
subplot(1,2,2);
imshow(imageArray,[]);
hold on
for i=1:length(B)
 plot(B{i}(:,2),B{i}(:,1), 'y' ,'linewidth',1.45);
 end
title('Segmented Tumor');
hold off;

Where B is defined in the code. My goal is to find the correlation of two images (Chosen by the user), and correlation value is shown for the second subplot of the selected images, that is "Segmented Tumor" . How can I do it?

[DGM]

I don't understand what you're actually asking. Are you trying to find the correlation coefficient between two different images? That would make sense, but your question and code seems to be asking how to find correlation between an image and the same image with a plot drawn on it -- which doesn't seem like it would be useful to know.

Either way, you can use corr2() to find the correlation coefficient, but I'd recommend against trying to operate on plotted copies of images. Imshow() resizes and does nearest-neighbor interpolation on images. You'd likely be operating on a severely degraded copy of your data. Plots should generally be considered for viewing purposes only.

[Tawsif Mostafiz]

Sorry for the confusion. The full code takes multiple images as input. As you can see, it shows the image itself and the segmented image in two subplot for every images. Which means the code works as follows: 1. I select a folder. 2. It reads all the images in that folder. 3. Produces the figure above for every images. Now my goal is to amend the code in such a way, that the user can select two images, (or two images will be given as input), their segmented version( which is the subplot(122)) will be used in correlation. I have tried to use the corr2 function, but it seems lime it only takes images as input.

[DGM]

But just to be clear, you're trying to find the correlation coeff between an image and the image with the plot line drawn on it, right?

[Tawsif Mostafiz]

No actually. The image that I have uploaded here shows only one output. The code I am writing can take multiple images as input and shows output as this image shows. Which means if I take three input images, there will be three output figures similar to the image shown above. Now, my goal is for example: I take two images as input. Then as the output comes for both, I want to take the second subplot for both images, that is the segmented tumor. And then find the correlation between them. Not an image with and without the plot, but between two different image with plots drawn on. Thanks again for your interest.

Answer 1

[DGM]

This is one way:

% example images and points
% points to plot [x y], one point per row
inpict1 = imread('cameraman.tif');
inpict2 = fliplr(inpict1);
pointlist1 = [80 80; 70 140; 100 190; 200 150; 160 70; 130 80];
pointlist2 = [80 40; 40 160; 130 150; 160 210; 250 160; 100 50];

% create line mask
hl = images.roi.Polyline;
hl.Position = pointlist1;
plm = createMask(hl,inpict1);
% apply mask
markedpict1 = inpict1;
markedpict1(plm) = 255;

% create/apply new line mask
hl.Position = pointlist2;
plm = createMask(hl,inpict2);
markedpict2 = inpict2;
markedpict2(plm) = 255;

% find corr
corr2(markedpict1,markedpict2)

This does not use the images as altered by imshow. The lines are drawn directly on a copy of the original image. Of course, there are considerations.

The lines that are drawn are approximately 1px wide with no antialiasing. As far as I know, there are no practical tools in base or IP toolboxes for drawing antialiased lines or lines of variable width. There may be some on the FEX, but I'd have to further question the utility. As the correlation coefficient becomes a function of line properties, what exactly does it mean conceptually?

Is it possible to extract the image directly anyway? Sure. Use getframe(), but the images will be whatever size they are rendered on the screen. They'll also likely be RGB, which corr2() can't handle, so you'll have to convert them to grayscale.

% do it the lossy way
inpict = imread('cameraman.tif'); % same image, same points

imshow(inpict)
S = getframe; 
f1 = rgb2gray(S.cdata); % captured plot

Again, why should the correlation coefficient be a function of the line color? Since the plot geometry depends on the figure geometry, the correlation coefficient is also dependent on how big the figure is. What technical meaning does that have?

Is it possible to force them to be rendered exactly one pixel per pixel? Yes, but it's not as trivial as the imshow docs make it sound. Of course, none of this will work unless the images can actually fit within the figure at 1:1 resolution.

% demonstrate forced 1:1 image display using imshow

% create perfect 1px fine checkerboard
x = mod(1:250,2);
y = mod(1:210,2);
inpict = xor(x,y');

s = size(inpict);

figure
set(gcf,'position',[0 0 640 480])

% may need to use tight borders so there's enough room for 1:1 display
options = {'border','tight'};
h1 = subplot(1,2,1,'units','pixels');
h1.Position(1:2) = h1.Position(1:2) + h1.Position(3:4)/2 - s([2 1])/2;
h1.Position(3:4) = s([2 1]);
imshow(inpict,options{:});

h2 = subplot(1,2,2,'units','pixels');
h2.Position(1:2) = h2.Position(1:2) + h2.Position(3:4)/2 - s([2 1])/2;
h2.Position(3:4) = s([2 1]);
imshow(inpict,options{:});

Again, I'd try to avoid ever using captured plots for any analytical purposes.

[Tawsif Mostafiz]

Thanks @DGM!