MagicBox

Software Architecture MagicBox

Inspiration

Manufacturers face the challenge that no real-time information on the current product quality is available to react short term on up stream and down stream changes. We want to address this need and provide transparency on, for example, cost impact of waste, delays, losses and errors in production lines. We help to ensure consistent quality in every product (=screens) in a highly dynamic manufacturing environment.

With the help of Accenture and the provided Azure Account we're ready to hack!

What it does

Our solution provides real-time information about the quality outcome of a product. In this case we take a look at defect detection on surfaces of screens. In this manner we propose a cascading low-level image preprocessing strategy to get the best results out of a lightweight Deep Learning algorithm, which can provide fast and accurate scratch detection performance. Most importantly we will show you all our achievements in a real-life demonstration of our implemented dashboard.

How we built it

The Data provided from Accenture was labeled with the tool "LabelMe" (see given credits in our Slides). We used Azure services to train and validate our DL-Model. For preprocessing and image augmentation we used the Visual Studio IDE and the python language. The dashboard was created with the frontend framework React. Javascript has been used to create the clean UI.

Challenges we ran into

We also address the challenge of a small dataset in our solution by giving insights in our Augmented Data Concept and Image Preprocessing Strategy to overcome the challenge and train an even more efficient DL-Model.

Accomplishments that we're proud of

Meeting in person to get to know each other and enjoing every minute of the hackathon. We started as a collection of strangers and got out of the HackaTUM as friends.

What we learned

If your in extreme coding flow you wont get any sleep.

What's next for We Wont Byte

For the future the possibility of generating more data additional to the result of the screen quality should be considered. This can help to detect which parameters are most important to get high quality products and which are critical. All in all this would provide a digital twin of the manufacturing process and create the possibility to understand the process in more depth as well as react to changes before the product loses in quality and prevent defect products rather than just detecting this cases.

With this information the Dashboard will be improved and additional features provided.