Where does rule-based inspection struggle?
In classic image processing, measurements and thresholds are written for each defect. On products with variable surface texture, these rules either miss the defect or reject good products.
What does the model learn?
A deep learning model derives the rules from examples. Trained on images of good products, the model learns the product’s normal state and marks parts that deviate from it. With this method, there is no need to collect separate samples for each defect type.
Why the company’s own data matters
An off-the-shelf model doesn’t know the product, the lighting conditions or the production line in question. When the model is trained on images taken from the line, it learns the actual production conditions. Results should also be measured on labeled parts of the same product.
