How to sort and get the coordinates of specific blobs

I've already attached some image to give you an idea of what i'm trying to do. You see, this is a scanned picture. I'm kinda new when it comes to this kind of stuff, so i don't know much yet

i put four blue dots on the 4 corners of the paper so i could use them as control points(im trying to align this image with another one using the cp2tform function). I've tried to extract the coordinates of the blobs manually through MS paint and it worked. Theres no problem with the control points for the base exam paper since it will be the same for each computation, while for the moving image, it would be different for each scanned image. Please help me find a way to sort hose four blobs from the rest and extract their coordinates. Thanks guys.

Answer 1

[Image Analyst]

First I'd straighten the image by finding the centerline, getting it's angle (orientation from regionprops()) and them calling imrotate.

Then take the red channel, which will give the most contrast, then invert the image (to make lines white and background dark). To find the centerline, first call imdilate (to make sure it's connected and has no breaks or gap), then threshold, label with bwlabel(), call

measurements = regionprops(labeledImage, 'MajorAxisLength', 'Orientation');

asking for MajorAxisLength, and look for the blob with the largest MajorAxisLength and Orientation. Then use imrotate, passing in the orientation, to straighten the image.

straightenedImage = imrotate(grayImage, orientationAngle);

I wouldn't bother with cp2tform. That's just more complicated than it needs to be.

See my Image Segmentation Tutorial if you need an example: http://www.mathworks.com/matlabcentral/fileexchange/?term=authorid%3A31862

What might be better than to get all 50 centroids in one shot would be to get a sub image for each question. This would involve straightening the image (described above) then getting the vertical profile by calling

verticalProfile = mean(grayImage, 2);

and by thresholding that profile identifying where each line starts and stops. Then have a loop where you extract the rows for the question, threshold, call regionprops. Do that for both the left half of the band of rows you extracted, and for the right half.

It doesn't look too hard - I could probably do it within an hour. I mean, to give the letter selected for each question. Come back with a good attempt at the coding if you have trouble.

[Teddy valladolid]

thank you mr. image analyst, im kinda new in programming and i jst want to make it clear that im comparing two images. And actually, i already have a code that compares the answer from the BASE paper (fixed) to the exam paper (moving) by verticalprofile. My problem now is, the images are not aligned with each other. knowing that the Base paper is fixed, i tried 4 putting control points (the 4 blue dots near the corners of the paper) on paper. Now whenever i scan the exam paper (moving image) sometimes i get images that need to be rotated to be aligned with the base paper. Im only intrested with the centroids of those 4 blobs. Once i get those 4 coordinates i'll use them as inputpoints for the cp2tform function. Ive also tried the first part in you comment where i should find the maximum value of the MajorAxisLength and orientation, but i just dont know what to do after using the

"measurements = regionprops(labeledImage, 'MajorAxisLength', 'Orientation')" function

like i said im only new, and what i know is still so little. Thank you in advance.

[Image Analyst]

I think your approach is a LOT more complicated but if you want to try that, be my guest. You'd need to first find your landmarks, then define some desired location for them. Then make up the transform, then apply it.

If you want to continue with my approach:

% Extract red channel
redChannel = rgbImage(:,:,1);
% Dilate it
redChannel = imdilate(redChannel, true(5));
% Threshold it
binaryImage = redChannel < 128; % or whatever value works.
% Label it
labeledImage = bwlabel(binaryImage);
measurements = regionprops(labeledImage, 'MajorAxisLength', 'Orientation');
% Get all the major axes lengths into one array.
allAxisLengths = [measurements.MajorAxisLength];
% Sort them in descending order.  
[sortedLengths, sortIndexes] = sort(allAxisLengths, 'Decend');
% Now the long center line ID# will be index #1
indexOfCenterLine = sortIndexes(1);
% Get the angle of that line:
angle = measurements(indexOfCenterLine).Orientation;
% Rotate image by that angle to straighten it:
straightImage = imrotate(grayImage, -angle); % Or maybe +angle (not sure).

Hopefully you came up with something very much like that when you did it on your own. That's just off the top of my head and may need some debugging or tweaking. Good luck.

[Image Analyst]

I saw your "Answer" below. So the result is the same but it used a different (automatic) approach than you had in mind (a manual approach). If you want to do it manually with cpselect(), instead of my automatic method, that's fine. So since you've answered yourself, I guess the project is done. If you need any more help, speak up.

[Teddy valladolid]

no, what i meant was the output of your code was pretty much the same with my first trial but then it gave me an image rotated in about 37degrees and thats way off the base image. i tried to change the sign of the angle but still the image it produced cant be compared to my base image. And about my post below, yes I did mine manually and im finding another approachthat would give me the blob centroid coordinates(automatically).

[Teddy valladolid]

i just wanna give an idea of how i did it manually so i posted it below..

[Image Analyst]

Teddy, you're just not doing the correct commands to do it, probably because you're just a beginner. If you follow my instructions, it could do it. That's what I did in the code below. Note how the centerline is perfectly straight:

Here is the code, attached below in blue text as test3.m. Now all you have to do is to crop off the top, then get the vertical profile and chop the image into 25 bands then go down looking for filled spots in both the left half (for questions 1-25) and the right half for questions 26 - 50.

[Teddy valladolid]

thanks a lot! it worked! now there's just one more thing, how can i transform/make the rotated image in such a way that itll follow the xdata and ydata of my Base image. their correction angle is more or less the same now, but the boxes are off by (x,y) i just want it to fit with the base image when use imshowpair or when i multiply the images when they are in the BW/logial format.

[Image Analyst]

I've read this 4 times and I don't know what you're talking about. What are xdata and ydata? What boxes? Why are you using imshowpair()? Why are you multiplying images? Like I said, simply look in the areas where you'll find the boxed and find which one, if any, are filled. You know the length of the line and the two endpoint locations so you know the rows where the questions start and stop and know how many questions are in between so just look in those bands like I explained earlier.

Answer 2

[Teddy valladolid]

Yup, I've tried your codes above and the result was pretty much the same. It rotated the whole image by an angle and its kinda diffrent from what i have in mind. I just cant really explain it well with words so i figured out that id just show you what i did. you can see that i've put the control points manually. I used the cpselect tool for that, and now im trying to make a code which will provide me the coordinates of the Control Point Markers in the Student Sheet image to use it as the input_points for my cp2tform since i've already assigned fixed values for my base_points. I really appreciate that you are helping me mr. image analyst.

>> i=imread('base.jpg');

>> i2=imread('exam.jpg');

>> t=cp2tform([500 169;477 771;35 77;10 754],[483 158;481 751;18 83;12 758],'affine');

>> info = imfinfo('base.jpg');

>> registered = imtransform(i2,t,'Xdata',[1 info.Width], 'Ydata',[1 info.Height]);

>> figure,imshowpair(i,registered)