- H2O Hydrogen Torch trains image classification, object detection and segmentation models on your own labeled inspection images, without writing deep learning code.
- A defect classifier can sort each image into the defect classes your quality team already uses, such as scratches, crazing, pitted surface or rolled-in scale.
- Inspection prioritization ranks lots, batches or units by predicted defect risk from process and test data, built with H2O Driverless AI.
- Models run in deployment options that keep images and process data inside the plant.
Send inspectors to the parts that matter
Manual visual inspection is slow and tiring, and the same defect can be called differently by two inspectors on two shifts. Sampling helps with volume, but a sample can miss the lot that matters.
Rule-based vision systems help on a stable line. They struggle when the product mix changes, when lighting or surface finish varies, or when a new defect type appears that no rule describes, so you end up trading line speed against the risk of an escape.
How visual defect detection works
H2O can be used to train defect models on your images, score new images on the line, and combine them with process data to decide what to inspect first.
Step 1: collect and label your inspection images
Start with the images your cameras already capture, labeled with the defect classes your team uses today, including images with no defect.
Step 2: train defect models with Hydrogen Torch
H2O Hydrogen Torch trains computer vision models with transfer learning from pretrained networks, covering image classification, object detection, and semantic and instance segmentation for defects where size and shape matter.
Step 3: score images on the line
Hydrogen Torch models package for H2O MLOps, which deploys them as scoring endpoints, tracks versions and watches for model drift when lighting, materials or products change.
Step 4: rank what to inspect first
H2O Driverless AI trains a classification model on process and test data: machine settings, sensor readings, supplier lot and electrical test results. Each lot or unit gets a defect-risk score with reason codes, and image scores can join the same model as features.
Step 5: build the inspection list
Agents can combine the image results and risk scores into an inspection list for the next shift and draft a summary of defect trends by line and product; your quality engineer approves the list and any hold.
Where the same pattern applies
In semiconductor fabs, the same pattern can predict wafer failures from electrical test data, so wafers most likely to fail go to review first, with reason codes showing engineers which settings or tests drove each score.