Custom vision models: are your defect photos enough?

A scratch under the lamp and in the photo

The quality department wants to catch fine scratches on painted housings. An inspector can turn the part under a lamp and look closely at the mark, while the model only gets the camera image. If glare hides the scratch, the photo does not give the model the same information. Start by looking at the photos with someone who knows these defects: is the mark visible, and is the difference between a defect and an acceptable surface visible?

Also ask how many physical parts were photographed, from which batches and under what conditions. A thousand photos of ten parts tell you less than a hundred photos of a hundred parts. With rare defects, correctly flagging a few known cases tells you little about the next one.

How much a simpler change gives, and how much training adds

In this count a ready-made vision product gives a good part of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive. An error in this count is a missed defect or an alarm on a good part, because both end in another inspection.

The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We include that cost in the quote together with the training itself.

Recognising a defect or flagging an anomaly

Classification assigns an image to a category, such as a specific type of scratch, and needs examples with a confirmed defect. Anomaly detection points to an area that differs from a correct surface, but does not settle on its own which defect it is. The MVTec AD 2 data set for comparing anomaly detection methods contains test photos under different lighting conditions. Supervised fine-tuning is further training of an existing model on examples with the correct result. The wider process is covered in the article on quality control with computer vision.

Where to start

To begin, selected photos of defective and good parts and a description of what the inspector checks today are enough. We run the comparison on parts not used for training, separately for batches and light settings. We adapt models as part of private LLMs, and we build the app where an employee sees the result next to the source as part of AI apps. For the free analysis, a few examples of an error that keeps coming back and a note on who corrects it today are enough.