Cracking up

Being stuck at home has given me some time to (finally) make progress in learning Python and tinkering around with electronics which, while fun, also has a more serious intent as I am trying to understand AI and computer vision a bit better, especially since the temperature scanner in the office managed to recognize my masked face after a single unmasked picture was taken.

Anyway, to combine the fun with the serious and having a large collection of concrete deterioration on jetties I read-up on image recognition of cracks etc. and found an open source program on GitHub; KrakN.?

Now my pictures are often underdeck, taken from a moving boat at different tides and hence the distance, angle and lighting vary significantly, adding to which there are the different colours of the concrete, different widths of cracks, sizes of delamination and extent of corroded reinforcement to mention some variables. In short I did not expect amazing results of the first trial of a random selection of defects and in that sense I wasn’t disappointed by the amount of mislabeling and false positives but I was a bit surprised to find just about the only thing not labelled in some pictures was … the enormous crack.?

It cracked me up.

But seriously; the other reason to chose this for my tinkering is to get a better ‘feel’ for the system my colleagues in Japan developed to inspect a deck with an unmanned boat, develop a 3D model and identify defects with >>90% accuracy. Mizuno san’s paper on the development won the 2018 De Paepe-Willems award from PIANC and the system is being further developed. I hope to be able to use it here in Singapore (or nearby) at some point and see how its done properly.

Still, not going to give-up completely on my own trials, next I’ll actually pre-sort some of the images and use 2-stage classification.