Segmentation of transparent droplets with oriented illumination

Dear all,

I am currently working on a university project over summer in which we are interested in identifying and tracking the growth of oil droplets from porous media. I have attached a typical picture from my experiments;

the camera I use is a Logitech C920 Webcam (maybe not the best, I know…). The bright spots surrounded by dark edges are oil droplets (typical examples of which are indicated by green enclosures on the uploaded image) which progressively grow from the medium and then reach a steady size. There is also an air bubble (which I have indicated by a red enclosure), but it should be ignored since it’s irrelevant to the experiment. I am now trying to isolate the droplets from the picture and measure their size; I have tried several simple pre-processing steps, using ImageJ, such as subtracting the background, running “Find Maxima” to extract the edge of the droplets, and then fitting an ellipse/circle to the edge using MatLab, but I admit that I am not an expert regarding image analysis (I am a Chemical Engineer by formation, and this is one of the first times I have to perform image analysis).

One of the problems I am running into is that my setup uses a light source which is not directly above the substrate, hence when the droplets are illuminated it seems like a certain portion of their edge appears dark while the other is bright (I understand that when looking at the uploaded picture this might not appear obvious, since it seems like the entire edge is dark; however, the lower portion of the edge is actually the shadow of the droplet, not the true physical edge). This is quite annoying since I cannot accurately identify the droplets on the image, and when I do, they appear as ellipsoids, when in reality they should be spherica

I have been looking around the web for a solution to this problem for weeks now, but I haven’t found anything conclusive. I would therefore be grateful if someone more experienced than me in image analysis could give me some indications or solution. Thank you very much in advance for your time and help.

Nathan

Answer 1

[Image Analyst]

It's a little hard for me to know what dark circles are the thing(s) you want and what's just clutter/noise. I think traditional image analysis techniques will not work well here and I'd suggest you just go with manual tracing (easy to do and to program up) or deep learning (automatic but hard to program up). If the image analysis is not a major part of your dissertation, and you just need to measure the things via any method that works just so you can get stuff done, I think I'd go for the manual approach. See attached demo.

[Nath Ravoisin]

Hi Image Analyst,

First, thank you so much for your quick answer! To address your point, the things I'm interested in are the bright "blobs" surrounded by dark edges; blobs which do not contain a bright cores are just noise. To remove any future confusion I have uploaded a new image where I have enclosed the things I am actually interested in in green:

I completely agree with you that manual tracing is probably the easiest way to go here, however the goal of the project is to track the growth of the droplets over time; unfortunately this means that I'd have to spend a LOT of time tracing everything manually for each frame, and that's only for one experiment... I've also read about deep learning and convolutional neural networks, but as I understand it these typically require thousands of labelled images to train, which is something I do not have. I will try the demo that you have uploaded and see what I can get out of it, but of course I'd appreciate if you could share any knowledge that you have regarding deep learning (e.g. maybe they don't actually require that much training images but are just tough to code). Anyways thank you again for your help, it is much appreciated.

Nathan