Edges in wheat leaves

I am trying to find the edges in these leaves. However due to shadow and other nuances normal edge detectors do not work very well.

I have tried a laplacian filter however it does not give back the results I hoped for. I would like to separate overlapping leaves as well though I do not need it to be 100% foolproof. The following code gives me a good idea of the edges however I still need to refine it more.

img=imread("path_to_image");
h=fspecial("laplacian",[3000000 0],10.0);
B=imfilter(img,h);
C=imsubtract(img,B);
imshow(B);
figure;imshow(C);

[Image Analyst]

Why do you need the edges? What is your overall goal? For example to find the amount of green growth in the scene? So you think you need to count the blades of grass? If so, a simple way to get the area fraction of green would be just as good, or even better, and much, much simpler.

[David David]

find out the number of leaves there are and using probability find an estimate for how many plants there are. It is not necessary to separate the background leaves.

Answer 1

[Image Analyst]

With this field of view, you cannot know, because there are plants outside the field of view whose leaves are in the field of view, and conversely there are plants inside the scene whose leaves are outside the scene. Plus there may be plants in the scene that are mostly or totally obscured by leaves and thus not visible at all.

Plus with such a small field of view, your statistics will be very inaccurate. You can't just take that one small field and extrapolate the answer out to a whole farm many acres in size.

If you back up and take a larger scene when the leaves are not so big that they blend into neighboring plants, then you'd have a better chance at counting plants.

I still think the color fraction of green is your best bet. It will correlate very well with the number of plants. First you'd build a calibration table where you manually count the number of plants and then find its green fraction. Do this at several time points if you need to, or want to, take snapshots at different times. Then for a given time, and given green fraction, you just look up the number of plants from your calibration lookup table.

Try a search of VisionBib to get papers on crop analysis.

[David David]

That was actually my objective(sampling).

The farms where I am doing this work are very homogeneous.

I've managed to get this image however I do not know how to join the borders(the gray islands).

[David David]

Also I don't fully understand what you mean with green fraction. The overlapping renders this impossible if it is what I think.

[Image Analyst]

If the white represents the green in your image, you simply do

areaFraction = sum(greenMask(:)) / numel(greenMask);

Multiply by 100 if you'd prefer "percent area" instead of a fraction between 0 and 1.

[David David]

I don't think you comprehend what my objective is. I want to find the leaves which are more visible(those in the foreground) and try to pick them out. I can't use your method as overlapping gives me a lower green percent than that I would have hoped for.

[Image Analyst]

I don't know what you want to do because you just reveal it in parts over several posts. And there are contradictory statements, like "It is not necessary to separate the background leaves." and "I want to find the leaves which are more visible(those in the foreground) and try to pick them out." The best I can get from your overall objective is "find an estimate for how many plants there are." I think the best way to do that is when the leaves are not overlapping, when the plants are small. Sure, some may die so it won't be totally accurate for all times, but I think that estimate may serve you well for whatever your overall goal is (which I still don't know). For example, counting the number of plants is not really your goal. Let's say you knew that number. Let's say I told you there were 40 plants in that image? So what? Why care about that? I think your overall goal may really be something like you want to determine which strain or supplier of wheat seeds gives you the best germination/sprouting rate. Or which fertilizer is most effective. Or something like that. And the number of plants is a good indicator of that overall goal - but the number of foreground plants is not a goal in itself. But I claim that percentage green would also be a good indicator of that and would track well with the number of plants (which you might have to determine manually rather than by image analysis).

That said, if you want to create a distance map to figure out what leaves are in the foreground or background, you can use a 3-D camera system, of which there are several.

[David David]

It doesn't really matter what I want to do with the information. I will use an image from an area(1/2 m^2) and try to determine the amount of plants there are by tracking the leave down to its roots. I don't need to find more than one leaf per plant. The image which I showed at the beginning is only an approximation of the image that I will use but I consider it good enough however if you deem this problem impossible I will not persist.

[Image Analyst]

I deem it impossible with the image you have supplied. I have 40 years of image analysis experience and do it as a full time job, and sometimes what you want or think you want is just not attainable so you have to try to measure other things that will still let you move ahead with your research. That's why I'm suggesting alternative approaches that will probably give you the information you really need.