03 · Flash and deformation

A defect you cannot describe with a single rule.

Identifying moulding, geometry and product-edge defects — a classic example of an irregular defect, where classic geometric rules quickly stop being enough.

Flash, burrs and moulding deformation rarely have one fixed shape or size — they vary between batches, moulding tools and even individual cycles. That irregularity is exactly why simple threshold- or template-based approaches fail where a model trained on representative examples handles it better.

In this scenario, the diversity of training data matters most — the model needs to see enough natural process variation to avoid confusing an acceptable deviation with an actual defect. That is why the pilot dataset (up to 500 annotations) deliberately includes examples from different batches where available.

Edges and geometry are also often harder to image cleanly than a flat surface — so assessing lighting and optics is often the first step here, before any model training begins.

When this scenario usually works well

01 The defect is visually apparent, even if hard to describe with a single geometric rule
02 Examples of the defect from different batches or production cycles exist, not just one specimen
03 Lighting shows the edge or geometry of the product without excessive shadows or glare
04 A tolerance threshold is defined — how much deformation is still acceptable

What the pilot report shows

For irregular defects, the false-accept rate matters most — whether the model lets borderline defects through. The qualification report shows this value separately, alongside recall and precision.

How this works in practice

AIVQC Trainer → AIVQC Server → AIVQC Production

FAQ

Flash and deformation

Will the model tell an acceptable process deviation apart from a defect?+

That is exactly what validation measures — the acceptance threshold is agreed with you before the pilot, based on the risk and economics of the process, not a single universal threshold.

What if flash only appears under specific moulding-machine settings?+

It's worth flagging during the initial consultation — it helps us plan the dataset to cover that scenario instead of missing it.

Other applications

Missing component

Checking assembly completeness and the presence of required components — one of the most common enquiries, because a missing component is easy to miss at line speed and easy to show unambiguously in an image.

↗02

Incorrect assembly

Detecting reversed, displaced or incorrectly seated parts — a defect that differs from a missing component in that the part is present, just in the wrong orientation or position.

↗04

Cracks and damage

Visually detecting scratches, cracks and surface damage — a defect where the stakes can be high, since damage may affect both function and visual appearance.

↗05

Labels and print

Verifying the presence, orientation and legibility of markings — a scenario that combines a completeness check (is the label there at all) with a print-quality check (is it legible and correctly placed).

↗06

Contamination

Detecting foreign objects, dirt and surface deviations — inherently the most unpredictable scenario, since contamination can take almost any shape.

↗