30-min study: Discover Fractures on Building Walls using Machine Learning

Recently, an old building in Hong Kong was discovered to have the potential to collapse, as major fractures are found in the external wall of the building.

A fracture found in the external wall of a building in Hong Kong

While I’m daydreaming, I wonder if machine learning can help find these fractures, by supplying enough sample photos. That’s why I initiate this 30-minute project.

The platform I’m using is Teachable Machines by Google. It is a machine learning platform that does not require any programming, at least at this stage to prove my concept is workable.

I collected around 20 photos each, with and without fractures, in Google Images, and supply them to the Teachable Machines platform.

Categorize the sample photos into 2 categories

After a half-minute machine training stage, the system is ready for analyzing photos. I then upload a Google street view of the fractured building.

The fractured building in Google Street View

The Results

The platform told me that the building has a 79% of chance that it has no fracture at all. Oops. That’s probably not true. It is because the street view contains too many things other than the concerned building.

1st round result: wrong answer

Then I try to zoom in on the building and re-supply this new image for analysis.

Street View: Zoomed In to the building

Bam! The result is accurate this time. It told me that the building has a 93% of chance that it has some fractures.

2nd round: correct results. Finally.

Conclusion & Future Implementation

To conclude, Teachable Machines by Google are useful to find out fractures on the external walls in the building, even if I supplied very few sample photos. Well done again, Google.

Teachable Machines in action

For further development, I wish there is a way to walk around a city via Google Maps Street View to find out the fractures automatically. A lot of steps are needed, as in my imagination right now:

1. Automation of walking around a city (with a reasonable boundary) using Google Maps Street View
2. Capture all the buildings along the path (and crop away non-building parts)
3. Upload the building images to the trained model for analysis
4. Send the problematic cases to an administrator / building service department for further inspection (with geolocation / address, which can be obtained in Step 1)

The trained model can be used in TensorFlow for further processing using programming, which has a huge potential to become a useful platform. I hope some building service guys can see this article and research this area.

Okay, my work is done now. 28 minutes including typing this article.

Links

Teachable Machine project link: https://drive.google.com/file/d/1rs5aPllHxNg0xK_O4Pq7HBWL6E51qysj/view?usp=sharing

The problematic building in Google Maps:
https://www.google.com/maps/@22.3076725,114.1672915,3a,27.5y,183.86h,107.38t/data=!3m6!1e1!3m4!1sh6BK035DsYEB-uZikzzzXQ!2e0!7i16384!8i8192