Use Cases

Automated vision applications for quality assurance and NDT. Each playbook below is a worked example: a real dataset, the annotation approach, the AI blocks used, the training settings and the pipeline that produced the result — so you can follow the same path on your own images.

Animated overview of the ReliVision platform, from data curation and training through to shop-floor inspection

One label per part

Classification

The simplest case: the whole image gets one verdict. You annotate with the whole-image tool, train a single classification block, and the model returns a class per part rather than a location.

Classification of OK / NOK bottles

Sort bottles into defected (NOK) and no-defect (OK) classes with a single classification block, using whole-image annotation.

Rotors presented for inspection, the subject of the OK / NOK rotor classification playbook

Classification of OK / NOK rotors

The same whole-image approach applied to rotors: two classes, defected and no-defect, trained on a small balanced sample set.

Where, not just whether

Object detection

When the position of the thing matters — a character to read, a scratch to locate, a threat object to flag — you annotate regions with rectangles or polygons and train a detection block. One model can carry several classes at once.

A pseudo-coloured X-ray image of airport baggage

Detection of threat objects in X-ray inspections

Detect hammers, scissors, guns and knives in pseudo-coloured X-ray baggage images for airport security — four classes, one model.

The surface of an electronic device inspected for cosmetic scratches

Detection of cosmetic defects on an electronic surface

Detect scratches on router images in two classes, “scratch_normal” and “scratch_wide”, so severity comes out of the model.

An aerial photograph of solar panel arrays

Detection of solar panels from aerial photographs

Detect solar panels in aerial photographs — the same detection workflow applied to overhead imagery rather than a station.

Shape, not a box

Semantic segmentation

Some defects have no sensible bounding box — a crack wanders, a glue smear spreads. Polygon annotation and a segmentation block trace the actual outline, which is also what you want when the extent of the defect decides the verdict.

A ceramic tile surface inspected for cracks and contamination

Detection of defects on tiles

Segment cracks, dye, oil and glue on ceramic tiles as a single defect class, without discriminating the type.

Two models in series

Two-stage pipelines

When you need to find something and then say what kind it is, one model does not do both well. A detector locates the region, its crop is handed to a classifier, and the two are wired together in the Pipeline Editor — no code involved.

A metal plate with an imprinted sequence of alphanumeric characters

Detection and classification of characters on metal plates

Detect the imprinted alphanumeric characters “A”, “F” and “3” on metal plates, annotated with rectangles.

Silicon wafers being inspected for defects

Detection and classification of defected silicon wafers

A detector finds each wafer, then a classifier labels it “Good”, “Small Defect” or “Big Defect” from the cropped region.

Metal nuts presented for defect inspection

Detection and classification of defects on metal nuts

A defect detector feeds a three-class classifier that decides whether the defect is bent, colour or scratch.

Build your own

Where to next

Every playbook above was built with the same tools, in the same order. The rest of the Knowledge Hub is how you do it on your own parts.

Knowledge Hub