I have some images from an experiment with plastics particles and water waves. The goal is to identify the plastics particles automatically. They sometimes overlap, and I don’t need to find individual particles, it is enough to find those pixels that contain plastics.
Since the particles are red, and the background is mostly white or black, I thought I could go for a simple thresholding, saying that pixels are plastics if R > 5*B and R > 0.25, where R and B are the red and blue channels. However the exposure varies quite a bit between experiments, and sometimes within an experiment when part of the surface is covered by water, so my approach doesn’t work very consistently, and sometimes mis-identifies the dark cracks along the sides.
I’m wondering what could be other options. I have limited experience with neural networks, so I’m not sure if that would work (with a reasonable amount of effort). In particular, I don’t think shape is going to be much help, since the particles are close together and partially overlapping with poor contrast between them, but maybe colour is enough?
Example images:
