Cameras catch surface defects before shipping

The challenge

What was going wrong

1. 1

   Collect real examples

   We worked with quality staff to photograph good and bad parts under real factory light. Inspectors labeled the defects that matter most.

2. 2

   Train and test

   We trained deep learning models and tested them against parts the models had never seen, adjusting until misses were rare.

3. 3

   Fit it to the line

   We placed cameras and lighting at one station and ran the model on a small computer beside the line, so no part waits on the network.

4. 4

   Monitor and retrain

   We set up a pipeline to track model quality over time and retrain the model when new part types or lighting changes appear on the line.

• Inline camera station

  Fixed cameras and controlled lighting capture every part as it passes.

• Defect detection model

  Finds and outlines scratches, dents and surface marks in real time.

• Operator display

  Shows pass or fail with the flawed area highlighted on a simple screen.

• Model upkeep pipeline

  Tracks accuracy and retrains models when parts or conditions change.

• Fewer flawed parts shipped

  Surface defects are now flagged at the line instead of being found by customers.

• Consistent inspection

  The check performs the same way on every shift, in every light condition.

• Inspectors freed for harder work

  Staff review flagged parts and investigate causes instead of staring at every item.

• Python Readable language for business logic, data and automation.
• PyTorch Framework for building and training neural networks.
• OpenCV Image and video processing library for vision tasks.
• ONNX Portable model format for fast, flexible deployment.
• MLflow Tracks experiments and manages model versions.
• Docker Packages software so it runs the same everywhere.