thermo2geo - geometry from thermography in injection molding

I presented a work on the use of thermography to infere the final geometry of an injection molding plastic part at the IEEE ICIT2018 conference, in Lyon, France.
I used the pix2pix GAN to model the relation between thermography and the heights map.
The UNet autoencoder architecture was used.

UNet autoencoder architecture

Model training

This architecture is robust, but a bit heavy to train. Also training requires fine hyperparameters tuning because achieving convergence is hard.

Noisy loss evolution during training

You can see the network learning the translation between thermography and geometry:

Learning thermography to geometry translation

Evaluation the geometric reconstruction

Then with Thomas Lacombe, we evaluated the geometric recontruction accuracy using a discrete modal decomposition. Results are impressive and valid the method.

Modal spectrum comparison to validated the geometric reconstruction

The next step will be to explore the latent space of the auto-encoder. Also, with some optimization, inference could be run on the production line, which could help for quality control of complex geometries.