Custom vision models: are your defect photos enough?

Can the camera see what the inspector sees?

Suppose your quality team wants to catch fine scratches on painted housings. An inspector can turn a part under a lamp to examine a mark. The model gets an image from the camera. If glare hides the scratch, that image does not contain the same information the inspector gets from handling the part.

Start by reviewing the photos with someone who knows the defects. Can they see the mark? Can they distinguish a defect from an acceptable finish? If the inspector needs another view to decide, first look at the lighting, optics, or part position. Better images may resolve a problem that would otherwise become an expensive AI project.

When the image is consistent, compare a fixed image check with an existing machine vision product as well. A custom model becomes a reasonable option when you can explain what those methods miss or why they send too many good parts back for inspection.

How many parts are behind the photo count?

An archive may contain many scratch photos, all showing one damaged housing. The framing changes, but the scratch and surface belong to the same physical part. A model that flags every image correctly has yet to show that it can find scratches on other housings.

Ask how many separate parts were photographed before judging the results. It also helps to know which production lots they came from and how the images were captured. Those details distinguish recognizing a familiar case from handling the next parts that arrive. A purchasing decision needs results on parts that were not used to adapt the model.

Rare defects make that judgment harder. If you have only a few parts with confirmed scratches, finding all of them tells you relatively little about the next one. A large archive of good products does not resolve that uncertainty. It can help you assess false alarms while leaving missed rare defects largely untested. You may need to postpone custom adaptation or narrow the task to flagging suspicious areas for an inspector to review.

Naming a defect or flagging an unusual area?

Classification assigns an image to a defined category, such as a particular type of scratch. To check that ability, you need examples where the quality team has confirmed that defect. If inspectors disagree about the same mark, they first need to agree on what counts as the correct answer.

Anomaly screening asks whether part of an image differs from the appearance of a normal surface. It can point an inspector toward an area to examine. The flag alone does not identify a defect category or establish that the part is defective. This approach may be worth exploring when defects are rare, provided ordinary variations in appearance do not overwhelm the team with alerts.

Lighting changes also matter in research on these methods. MVTec AD 2, a benchmark for anomaly detection, includes test images captured under lighting conditions that may be absent from the training images. A result on that benchmark does not establish performance on your housings. Your parts need a separate evaluation that covers the lighting changes at your inspection station.

Does it make inspection easier?

A model can find more scratches while also flagging more good parts. The quality manager needs to see both effects. An overall accuracy percentage reveals little when most images show products without defects.

When comparing methods, review the confirmed defects each one missed and the good parts it flagged. Check how much time inspectors spend resolving those alerts. A setting that picks up smaller marks may generate more repeat inspections. Reducing alerts may mean missing more defects. The quality team decides which tradeoff is useful for this task.

Compare the options on the same parts, then look at results by production lot and lighting condition. If a custom model helps only under one lamp setting, you want to know before purchasing it. If a simpler method delivers similar detection results and requires similar review effort, a custom model needs another concrete benefit to justify maintaining it.

When to discuss adapting a model

One option is supervised fine-tuning: further training an existing model on examples labeled with the correct result. For cosmetic inspection, it is worth considering when the defect is visible in the image, confirmed examples come from different physical parts, and simpler options still struggle with the task.

You can discuss image evaluation and possible model adaptation with Syntalith through its custom AI applications service. Start with selected photos of defective and good parts, plus a short description of what inspectors find difficult today. Point out any images that show the same part. You do not need to organize the entire archive first: the initial discussion should help establish whether the images support a comparison of methods or whether you first need clearer images or more examples of the defect. That gives you a basis for agreeing on further work. See the pricing page for pricing information.

For the broader workflow around inspection results, see the article on quality control with computer vision.