First I start by identifying the vertical position of each question number. You can see that the answer boxes on the right for each question line up nicely with the bulleted question number on the left.
img = Import["https://i.sstatic.net/DjGwH.png"];
questionNumbers =
ImageTake[img, {40, 243}, {1, 30}] // RemoveBackground
Then I find the centroids of each of the components in questionNumbers, and grab only the vertical positions. I then define a Nearest Function to calculate which question number a given vertical coordinate is closest to:
numberCentroids =
ComponentMeasurements[questionNumbers, "Centroid"] // Values;
numberVerticalPos = numberCentroids[[All, 2]];
returnNearestNum := First@Nearest[numberVerticalPos -> "Index", #] &
Now let's look at only the boxes part of the image:
boxes = ImageTake[img, {40, 243}, {900, 1100}]
I just determined by eye where each box column lies horizontally using the coordinates tool. It looks like they lie at horizontal image positions {40,70,100,130,160}. I then define another NearestFunction to return the box column number a horizontal coordinate is closest to:
horizontalComponents = Range[40, 160, 30];
returnNearestBoxColumn :=
First@Nearest[horizontalComponents -> "Index", #] &
Let's now Binarize and ColorNegate our box image. We can delete small, erratic components too while we're at it:
bin = Binarize[boxes] // ColorNegate;
bin = DeleteSmallComponents[bin, 30]
We need to now find a sufficient ComponentMeasurement that splits our components into two groupings: one of boxes with checks, and the other of boxes without checks.
A good candidate is "Rectangularity". The definition of "Rectangularity" (in the ComponentMeasurements documentation) is the fraction of pixels within the minimal bounding box. We expect a filled in box (filled with an X, a check mark, anything) to have a higher fraction of pixels within its bounding box compared to an empty box. If we histogram the Rectangularity of the components in bin, we indeed see it splits the components into two separable groupings:
ComponentMeasurements[bin, "Rectangularity"] // Values // Histogram
It looks like a Rectangularity of 0.3 is a good threshold to split our components by. The ones above this threshold should be the filled in boxes.
Let's use SelectComponents to find the components that have Rectangularity greater than 0.3:
filledBoxes = SelectComponents[bin, #Rectangularity > 0.3 &]
And now we can get the corresponding question number and box column number for each filled-in box by grabbing the Centroids of the components in filledBoxes and using our Nearest functions:
xPos = ComponentMeasurements[filledBoxes, "Centroid"] // Values;
xQuestionNumber = returnNearestNum /@ xPos[[All, 2]];
xBoxNumber = returnNearestBoxColumn /@ xPos[[All, 1]];
Thread[{xQuestionNumber, xBoxNumber}]
{{1, 2}, {2, 4}, {3, 4}, {4, 2}}
The xQuestionNumber part isn't really necessary if we know there will only be one box filled per question, but it is a nice check as it confirms that we are getting an answer for each question.
Hope this helps. Please validate this with other examples to make sure the parameters are robust.
Add-on As an alternative for using Rectangularity, it appears DeleteSmallComponents[bin, 200] works quite well, while also being simple. This makes sense that DeleteSmallComponents works, because filled in boxes should consist of more elements compared to empty boxes. My only concern is that if a box is filled with a small X that doesn't touch the edges, it may be treated as its own component (separate from the box edges) and fall below the DeleteSmallComponents threshold.
