Identify crossing lines in image

I am trying to read images of magnetograms (Original image) by binarising the image and using bwboundaries, with the signal curves of interest being divided into segments. I would like to avoid burning connecting lines into the image as suggested here, so I am searching for adjacent boundaries and changing their labels to be the same (simplified script with simplified example image). In non-quiet days, like in the example, some signal curves cross the baselines (straight lines) or other curves, causing them to be marked as one boundary - is there a way of detecting cross-overs in bwboundaries() and separating the boundary into parts? It does not necessarily have to be using bwboundaries(), it is just the function I started with.

Answer 1

[Image Analyst]

mask = bwskel(mask);
crossingPoints = bwmorph(mask, 'branchpoints');

[Vangelis Vladimirov]

Thank you for the quick answer. I assume one way is to identify branch points, then cut the skeletonised curves at them and try to match the different segments using their local gradients. This is complicated a bit by the skeletonisation which sometimes causes the crossing points to be separated into two (shown in attached image). An additional complication in the magnetogram images is that the crossing points have to be selected among other, additional points along the curves (usually at local extrema) which are also marked as branching points. I saw a similar discussion in this thread, where you suggest ways of removing spurs - I am currently testing different options on how to do this. Are there any other MATLAB packages or tools which can deal with crossing curves, or is bwmorph the easiest way?

[Image Analyst]

What would you like to have? Get rid of the short segments? Move the branchpoints to the middle of a segment if the segment is too short and reconnect the old branchpoint?

Why do you need branchpoints in the first place? What will you do if you have them?

[Vangelis Vladimirov]

I am identifying the curves on the images by using bwboundaries. The curves however are not continuous, and instead of burning a connecting line or dilating the image, I am matching segments which start near an area where another one ends.

In the first attached image, I show the changes of labels I am able to achieve on a test image I drew, which consists of three broken curves. By using the end points (min/max(boundary(:,2)) of each segment, I am able to match the segments which I think to be belonging to one curve (label 1) even if there are breaks and jumps (either introduced by the binarisation or not present in the original image). The second curve, however, crosses a baseline (or another curve), which causes both the part I want to join with the second curve (red, label 2) and the part which should be part of the baseline (blue, label 6) to appear as one boundary. I would like to separate them, so I have something like in the second image I attach instead, where I can match what I think is part of the continuation of the (red, label 2) to itself and the start of the baseline (green, 6) to its other part (blue, 6). I don't mind if the branch points (where (teal, 2) and (green,6) cross in the second image) are included in both resulting curves, either of them or in neither of them (i.e. teal or/and green curves get broken into 4/3 separate pieces).

[Image Analyst]

Could you show the original gray scale image, and the segmented, binary image from which you called bwboundaries() on? Maybe we can just avoid having line 6 in there in the first place.

Do you want to do edge linking? See attached demos.

[Vangelis Vladimirov]

The figure being processed in my previous comment is only an exagerrated example of what I want to do (initial image is label_and_crossing.png). I also attach a minimal working script which generates and shows the binary and skeletonised images of the example figure and the Original_image.jpg I want to process. Line 6 is supposed to look like the baseline curves in Original_image.jpg, which are used to calculate the recorded fields, field = baseline + the distance of the recorded curve to that baseline times the scale (field/mm).

[Image Analyst]

I'd suggest you get rid of the axes right away by summing the image horizontally

verticalProfile = sum(grayImage, 2);

and then erasing peaks that fall into the certain, known zones where the axes lie. Then maybe some morphology to thicken the line, like imopen() or imerode(). Then perhaps edge linking. Then scan the image column by column looking for dark places.

There are several graph digitizing apps in the File Exchange if you don't want to build one yourself.