How many models does a four-branch pipeline need?

Fewer than four. The example here uses one item detector shared by three branches, plus a label defect detector, a character detector, a character classifier and a scratch detector. Models are added per detection step, so branches that share upstream work do not each need their own.

Does one pipeline mean one pass or fail for the whole part?

No. Each branch ends in its own result, so you can see that the label is present but misprinted, or that the surface is scratched while the label is fine. How those verdicts combine into a disposition is a decision you define, not something the pipeline forces.

What happens to the other branches if one model underperforms?

Branches are independent below the point where they diverge, so a weak scratch detector does not affect OCR. The exception is a shared upstream model: if the item detector misses the label, every branch downstream of it loses its input, which is why it is validated first.

Can we add a fifth check later without rebuilding?

Yes. Adding a check means annotating that failure mode, training the model it needs and attaching a new branch to the router surface or to an existing detector. The branches already in production are untouched.

Does running several checks slow the line down?

The image is captured once and shared, and the shared upstream detector runs once for the branches that depend on it, so the cost is far below running four separate inspections. Whether the total fits your takt time depends on the models and hardware, and should be measured on validation images before release.

Do we need a data science team to build this?

No. Annotation, training and pipeline assembly are done in the ReliVision UI, so the process and quality engineers who know the failure modes can build and maintain the branches themselves.

How to Catch Multiple Defects with a Single AI Pipeline

One branching AI pipeline that checks label presence, label defects, printed characters and surface condition — in a single pass, without adding a station for every new check.

One pipeline, four checkpoints
Single pass — no extra cameras or headcount
One model per detection step, shared across branches

Why it matters

Why build a multi-point inspection solution?

A single part rarely has just one thing that can go wrong. A single pipeline should be enough to catch all of them.

One part, many failure modes

A label can be missing, damaged, misprinted or unreadable, and the surface around it scratched. Each is a different inspection problem.

Single-purpose stations don’t scale

Building a separate station or model for every defect type multiplies hardware, integration work and maintenance.

One pipeline, every checkpoint

A multi-point pipeline checks all of them in a single pass without adding extra cameras, stations or headcount.

The core idea

The multi-point inspection architecture

One router surface forks the same image into every detection branch — each branch chains its own models and ends in its own result.

The branch diagram: a router surface forking into the item detector, label defect detector, character detector chain and scratch detector.

Step 1

Annotate your defects

Label every failure mode you want the pipeline to catch — presence, defects, characters and surface issues — across a representative image set.

Scratch, OCR and label annotations captured on the same representative image set.

Step 2

Train your models

One model per detection step, not per branch: branches share upstream models and chain their own.

Watch loss and mean average precision during training, and evaluate each model on held-back images before wiring it into the pipeline.

Why the order matters

A weak upstream detector limits every branch downstream of it. If OCR results look poor, check the item detector before retraining the classifier.

The training view with model loss and mean average precision curves for training and validation.

Step 3

Build the pipeline

Connect your trained models to a single router surface, chain multi-stage branches and define a clear end result for each.

Then run the pipeline on validation images and confirm each end result fires independently of the others.

The ReliVision pipeline editor with the node menu open and a branch being connected.

See it in action

The pipeline in ReliVision

The same router-and-branches structure, built and running as a live multi-point inspection pipeline in ReliVision.

The complete multi-point pipeline running in the ReliVision pipeline editor.

Start building your multi-point pipeline today

Bring one part with more than one way of failing, and the branches follow from the failure modes you already know about. Request a demo to start your journey.

Request a demo

Walk through the router-and-branches architecture on your own parts and failure modes.

Live Demo

Knowledge Hub