Hard-to-Catch Molding Defects: What AI Vision Inspection Does Differently

Inspection in injection molding is not a new subject. Vision cameras are already on most lines. Yet in many plants a person still does the final check. Why?

Where rule-based systems are weak

Conventional vision inspection runs on rules: “if brightness in this region falls below a threshold, reject”; “if a spot exceeds this size, reject.”

That works well for unambiguous defects. The problem is the ambiguous ones.

• Fine scratches that appear and disappear with the lighting angle
• Color or gloss variance sitting on the edge of the normal range
• Correct geometry but different surface texture
• Contaminants that appear in a different place each time

Trying to write rules for these creates a dilemma. Tighten the threshold and good parts are rejected, driving over-rejection up. Loosen it and real defects escape. People end up re-inspecting, and inspection headcount becomes the bottleneck.

How deep learning handles it

With deep learning, nobody writes the rules. The model learns what normal looks like from the images themselves.

If lighting or angle shifts slightly and that variation was present in the training data, the model treats it as within normal range — exactly the cases that had to be handled as rule exceptions before.

Training on good parts alone

The question we hear most often is whether this works when there are hardly any defect images.

It does. Anomaly detection trains only on images of normal parts. The model learns the distribution of normal and flags whatever falls outside it.

The advantages are clear:

• No waiting months to accumulate defect samples
• Defect types never seen before still register as “not normal”
• The more stable the line, the richer the training data

There are limits. It will not classify the type of defect. If you need statistics by defect type, a classification model has to follow.

Keeping line speed

A common worry is that AI inspection will be slow. Sending a heavy model’s inference to a server does make matching line speed difficult, because of the round trip.

The answer is edge inference: a lightweight model on an edge device beside the camera, deciding on the spot. Without the network hop, latency drops, and because images never leave the site, the security burden shrinks too.

Making a model lighter trades against accuracy, so it cannot shrink indefinitely. Set the target takt time first, then pick the most accurate model that fits inside it.

Detection alone is not enough

Everything so far is about finding defects well. But however good the inspection, the defect rate itself does not fall. You are only filtering out what was already made.

To reduce the rate you have to look at the conditions that produced it: injection pressure, mold temperature, cooling time, material lot, seasonal humidity. Analyzing those variables against actual inspection results shows which combinations raise defect rates.

That requires inspection data to be connected to process data. If the inspection system and the equipment each accumulate data separately, the analysis is impossible. This is why the data connection structure should be settled at the same time vision inspection is introduced.

Evaluating accuracy

“99% accurate” means less than it sounds. On a line with a 1% defect rate, calling everything good scores 99%.

Two things actually matter:

• Escapes (missed defects) — the cost of a defect reaching the customer
• Over-rejection (good parts called bad) — the cost of scrapping sound product

The two generally trade off. Which to avoid more depends on the product and the customer’s requirements. The inspection threshold has to be set on that judgment — it is closer to a business decision than a technical one.

Order of adoption

1. Review current inspection data — what images are stored, and how many
2. Validate on one product and one line — start narrow
3. Agree escape and over-rejection targets — production and quality together
4. Deploy to edge and verify speed — confirm it fits the takt time
5. Connect process data — tie inspection results to process variables
6. Extend to other lines — including a per-line retraining procedure

---

IMPIX’s V² is a deep learning vision inspection solution supporting anomaly detection trained on normal images and lightweight edge models. Quality prediction and process optimization based on XGBoost and genetic algorithms are provided alongside it, so the path runs from detection through to prevention.