Inspiration
We want to improve ZEISS' ability to ensure zero defect photomasks and thus the chips that improve our lives.
What it does
First, there is a training phase with only image labels (vision/train...). Thereafter, the very small model is used to segment new images. An inference script infers more attributes from the segmentation masks (whether the object lies on the edge, orientation, size).
How we built it
With our combined passion and skills ;)
Challenges we ran into
Exploring off-the shelve image analysis functions took some time until we settled with the final training regime.
Accomplishments that we're proud of
Weakly supervised training on only about 70 images. 1 1/2 minute training time.
10 Mega Pixel per second inference on GPU. Solid Dashboard that is setup to work with many machines.
What we learned
We brainstormed, developed and tested a new form of weakly supervised loss that works well for turning image labels into accurate binary segmentations.
What's next for Apollo PhotoMask Rescue: Defect Detection + Dashboard
This setup is ideal to continually learn with easy to label data. On top of that, predictive maintenance tasks become feasible because it tracks performance over time.
Log in or sign up for Devpost to join the conversation.