Automating inspection is the bottleneck for inline CT settings. Machine learning is inevitable but challenging: data is scarce and sensitive, custom models are required for each setting, and production systems have strict latency and throughput constraints.

Roughly speaking, there are three attack angles:

1. Make training data acquisition easier.
2. Develop models that require less (annotated) data for training.
3. Make models more computationally efficient.

The goal of DeepSpect is to explore these angles to help lowering the barrier to entry for ML for industrial CT. Check out the latest notes for more details.