I am not going to Montreal

I am not presenting my work at the CIGI-QUALITA 2019 conference, in Montreal. This is because of a really soon to come newborn. The associate paper can be found here. This paper is a teaser of my onwriting PhD. I show that a simple webcam can be use with Deep Learning to detect complex defect patterns on injection molded parts.

Example of a good and a defective part, captured just after molding with a simple webcam

The software architecture is a REST API endpoint, with an GPU powered inference server and a client system which integrates hardware sensors. When the production line start, I have zero training data. Then it grows slowly, as the human expert annotate some parts. This is why transfer learning is necessary. I tried many different architecture, and the most robust one is VGG16.

Transfer learning for quality control with VGG16

The embedded system runs on a low power ARM. It is responsible for the robust part detection. Fusion of multiple sensors was needed for a robust detection: sonar, Time Of Flight and camera frame analysis.

Necessity of robust part detection

Finally, I designed a simple Flask UI, so the user can train the Deep Learning model as the production line goes. Every ten new annotate parts, the neural netorks model is re-trained.

Deep Quality Control User Interface

To explain the view of the neural networks, guided backpropagation is used. The saliency image shows that the model is sensible to the part geometry and to defects.

Model explained: guided backpropagation saliency image

I am actually doing industrial scale trials of the system at IPC Technical Center and Plastic Omnium.