How to compare 2 Images?

Hi all,

I need to compare 2 multiple choice paper images against each other. One sample with the correct answers and the other that I need to compare and calculate the score. I tried extracting the matrix value for one of the answer on the sample and did a normxcorr on the the other image. However, the program return value at another value. Can someone please help advise where did I go wrong? Or do I need to use another function to compare?

Answer 1

[Image Analyst]

Not sure what you did. It should have found the "filled in" circles, so I don't know what went wrong.

Of course you don't even need normxcorr2 for something like filled-in circles - just simply have a template array of masks and look in each mask to see if it's dark or bright. Why did you use normxcorr2? It seems like an overly complicated way of doing it. If you do it that way, even when you have bright spots where all the filled circles are, then you still have to match up their centroid locations with the template - you can't simply just threshold, and run regionprops - you won't have the "question number" - to - "centroid location" correspondence that you need to determine which question went with which blob. Hence my suggestion to use an array of ROI masks as a template, where you do know the correspondence. The only reason I can see for using normxcorr2 is in the registration step where you're looking for the corners or crosses at the outer corners to define the coordinate system so you can register the image. If you really want to use normxcorr2, can you upload the original image and the normxcorr2 output image?

[Yao]

Hi sir, thank you for your reply, just to clarify, I am comparing a hand written multiple choice answer image and hence I did not fill in the circles and compare the mask. I was thinking of extracting the answer for each question out from the template image and comparing it with the image for that particular question to see if they match.

[Image Analyst]

I suggest you use circles. Handwritten OCR is not easy at all. It's very very challenging.

[Yao]

Hi Sir, my project requires me to use handwritten OCR as the choice for answering the MCQ questions. Am I able to specify the ROI in the template and compare it with the target image and calculate if the they are similar?

[Image Analyst]

It's time for you to upload an image so we can stop wasting time guessing. But for now, I guess that you're unfortunately going to have to go to a lot more work because you're going to have to use OCR for handwriting. See the link I gave. There is just way too much variability in how people hand draw letters to do something like what might work for perfect computer fonts. For example someone might write A, someone might write a, and someone might write A without closing the cross bar so that you can't depend on the Euler number (number of hole). Someone might write E, e, or E in several parts because the horizontal lines are not connected to the upright line. With standard techniques, you have a chance only if all your respondents have perfect handwriting in the same font and the font size is uniform - so basically you have no chance at all with standard techniques because you'll never be able to assume that.

[Yao]

Hi sir,

Attached the first image is the sample image and the second image is the image I would like to compare. The criteria is set that I will only be comparing the letters in upper case (A, B, C, D, E). For the moment, I can only assume that all the font and font size is uniform. Thanks for your help.

http://www.flickr.com/photos/88195404@N02/8063185534/in/photostream

http://www.flickr.com/photos/88195404@N02/8063186910/in/photostream

[Image Analyst]

If the font and font size is uniform, then just calculate the area within each box. Each letter would have a different area. First you need to find the boxes, then look inside and find the character. Then measure its area, like using the techniques in my Image Segmentation Tutorial in my File Exchange.

[Matt J]

OK, and what preliminary steps are you able to do successfully so far?

Can you find the boxes and their orientations? It should be pretty easy. Apply imfill(...,'holes') to fill them in and then search for them using regionprops(..,'Area','PixelIdxList').

Can you assume that the letters are not tilted relative to the box boundaries? If so, finding them using template matching should be fairly easy as well, once you know the box positions and orientations.

[Yao]

Thanks Sir, that's what I was trying out earlier. However, I noticed that the area from the Sample and the area from the target is different. I tried finding the difference in the 2 areas and see how much they differ. Is there a way in which I can set a rule that says if the difference is too great then they are not equal?

[Image Analyst]

Yes. You can try some and see. Then the one that works best is maxDifference. So you use code like

if thisDifference > maxDifference
  % They're different letters
else
  % They are the same letter.
end

[Yao]

Thanks a lot sir for your help :)

[Yao]

Hi Sir,

I have another question, I used regionprops and I am able to locate all the boxes of interest.

How do I find the area of each box so that I can compare? This is what I have gotten so far: http://www.flickr.com/photos/88195404@N02/8077403755/in/photostream

Iprops=regionprops(Ilabel,'all');

NumProp = size(Iprops,1);

allPropArea = [Iprops.Area];

getProp = (3200 < allPropArea) &(allPropArea <3800);

findProp= find(getProp);

keepProp=ismember(Ilabel, findProp);

[Image Analyst]

filteredAreas = allPropArea(findProp);

[Yao]

Hi Sir,

Thank you for your help.

For this command, will it show the value of areas of all the boxes of interest or do I need to write another command to show the area.

And also how do I compare them individually? I have tried this code, but I need to execute them individually for 20 times. Is there a shorter way to do it?

if DiffQ1 > MaxDiff

Results = Results + 0

else

Results = Results + 1

end;

to

if DiffQ20 > MaxDiff

Results = Results + 0

else

Results = Results + 1

end;

[Image Analyst]

filteredAreas has all the areas that satisfy your constraint: (3200 < allPropArea) &(allPropArea <3800)

I don't know what compare them individually means. I don't know what you are trying to do with those if statements and why you're executing them 20 times. I don't know what MaxDiff is, or what DiffQ1 is. Difference from what to what??? And what is Results???

[Yao]

Hi Sir,

In this image there are 5 regions that i need to compare the area, which I have used regionprops to get the region of interest and the area.

I need to compare the area for the regions to see if they match with the output image. How do I compare them? Do I comparing them region by region?

Answer 2

[Matt J]

Could you use a 3rd reference image in which all of the answers in the questionnaire are simultaneously selected?

Then you can segment that image to see where all the answer choices are.

Then you can add this reference image to the image of any normal questionnaire (assuming they are co-registered) so that the answers chosen in the normal questionnaire will create a higher intensity region than the other answers.

Finally, search the summed image for the higher intensity answer choices.

[Yao]

Hi Matt, thanks for the reply, I already have a reference questionnaire which I am using to compare with. The image I am comparing is a handwritten questionnaire which I have tried to segment and regionprop the boundary to determine where I should be comparing the image. I am not sure if I am using the correct function to compare both images.

[Matt J]

Hi Matt, thanks for the reply, I already have a reference questionnaire which I am using to compare with.

Yes, but my suggestion was that you don't use that one. If you can make your own template image, A, in which all the circles are filled in, and sum that with the questionnaire image , B, that you are scanning, then

C=A+B

will be an image where the chosen answers in B are easy to spot. Chosen answers will be large peaks in C and unchosen answers will be small peaks in C.

You can repeat the same thing with your sample image, S, of correct answers

C=A+S

to automatically determine which answers are the correct ones.

[Yao]

Hi Matt, my projects requires me to use handwritten format as the choice for answering the questionnaire, I am not sure if I can use your method, I would assume for your method you are referring to using OMR as the choice of questionnaire?

[Matt J]

Yes, I was assuming OMR when I Answered.

Answer 3

[Hiren Maniyar]

hello all,

my name is hiren maniyar. i am working on OMR reorganization so i want help from all of you anyone tell me how to compare circles of two OMR sheet.i have all ready done detection of circles just i want to compare that circles. i there ant algorithm for that or any method are there then please tell me.

[Image Analyst]

Start a new question and insert the two sheet pictures with the green and brown frame icon. We need to know if the two sheets are the same blank form, or different ones. Basically you can use a template of predefined circle locations and look in there. Align the image to the template if you don't have precise control over where the form is in your image.