From 2011 to 2013

Some crystalline substances under tension experiences a “flow” of its micro crystals. Empty spaces can form between the crystals. These cracks can be viewed with an electronic microscope. However, to process a sample, scientists use a lot of chemical and mechanical treatments. These treatments leave many artifacts that makes the detection difficult. 

Below is such an image. There are certain features (highlighted with yellow circles) that are important. Other (like the lines highlighted with red arrows) are artifacts that must be ignored.

Example of an image generated by scanning electron microscope. The probe is a piece of steel that underwent some mechanical traction. Yellow circles highlight the features that are important.

Metallurgists and physicians want to investigate the quality and quantity of these cracks/holes in the material. Size, shape, orientation, spatial correlation (eg cracks along a certain line) are very important to characterize the macroscopic behavior of the materials.

I devised a fairly robust Computer Vision pipeline to detect these features, automatically. While there was some ML involved, there was no Deep Learning back then. Should be interesting to try to compare classical pipeline with an end to end DL one. The low data volume (few tens of images) and tedious labeling might not be the best environment for Deep Learning to shine.