Humankind was waiting for it long enough: We detect scratches on chips.
Our challenge is simple: There shall not be any faulty chips!
We detect if a chip is faulty and add a bounding box to the image. Everything is displayed in a user-friendly dashboard, also showing some overall statistics and allowing you to upload your own images that are fed to our model.
The main challenges we ran into:
- Limited training data
- Raw training data
- Unlabeled training data
- Building a user-friendly UI
We used Microsoft Azure's Custom Vision to detect the scratches on the images. For training, we applied various preprocessing techniques to ensure the generalizability of the limited training sample such as:
- Adding rotation
- Flipping
- Changing the gamma parameters
- Changing the brightness
- Adding noise
- Affine transformation
Also, we manually drew more scratches on some chips allowing even more general predictions.
Our model is very accurate and also has a high recall. (both over 90%). Also, our prototype is coded from beginning to end - and works!
Microsoft Azure's Custom Vision actually delivers decent results, despite the little customizability.
It would be awesome to implement the model into the real-time image feed of chip production.