Skip to main content
AI // Computer vision

Cameras that notice the scratch before your customer does.

Vision models detect defects, count items, read labels, and check assembly steps, then record the result in your quality workflow automatically.

inference.tracelive

{

"summary"
: "Good lighting, good labels, then a model. In that order.",
"use_cases"
: "Defects · counting · OCR",
"deployment"
: "Edge devices or cloud",
"integration"
: "Quality checks in Odoo",
"reviewed_by"
: "a human"

}

Where vision helps

Defect detection

Scratches, dents, misprints, and missing parts flagged on the line.

Counting & verification

Pallets, cartons, and items counted from images.

Label reading

Lot numbers, expiry dates, and serials read and checked against the ERP.

Safety monitoring

Protective equipment checks and restricted-area alerts.

What a vision pilot needs

  • Consistent camera position and lighting
  • Labeled examples of good and defective items
  • A clear decision: stop the line, flag, or log
  • A way to record results in the quality system

Why visual inspection is a good fit for AI

Visual inspection is one of the most repetitive jobs in any factory or warehouse. An inspector looks at the same kind of part hundreds of times a shift, searching for scratches, dents, missing components, or a misprinted label. Attention drops as the shift goes on, and defects that slip through are found by customers instead.

Computer vision models are well suited to this work. Given a fixed camera, consistent lighting, and enough examples of good and bad items, a model can check every unit at line speed and never get tired. It does not replace the inspector's judgment; it handles the routine checks and passes uncertain cases to a person.

Most vision projects succeed or fail before any model is trained. Camera placement, lighting, and how defects are labeled matter more than the choice of algorithm. We spend the first weeks of every project getting those right.

Approaches we choose between

ApproachWhat it doesBest for
ClassificationLabels a whole image as good, defective, or a defect typeSingle parts photographed one at a time
Object detectionFinds and boxes each item or defect in an imageCounting, missing components, multiple items
SegmentationMarks the exact pixels of a defect or regionMeasuring defect size or area
Anomaly detectionLearns what normal looks like and flags anything elseRare defects with few examples
OCRReads printed text, codes, and numbersLabels, lot codes, expiry dates, serials

From camera to quality record

Capture

Industrial or IP cameras triggered by sensors or line timing, with fixed lighting.

Inference at the edge

Models run on a small computer near the line, so decisions take milliseconds and work offline.

Decision

Pass, fail, or review, mapped to an action such as a reject gate, a light, or an operator alert.

Record

Results, images of failures, and confidence stored against the production order.

Review

Quality teams see defect rates by line, shift, and product, with example images.

Pilot plan

  1. Step 01

    Site visit

    Week 1

    Look at the line, the parts, the defects that matter, and where a camera can go.

  2. Step 02

    Capture setup

    Week 1–3

    Install a camera and lighting, and collect images under real conditions.

  3. Step 03

    Labeling

    Week 2–4

    Label good and defective examples with your quality team's definitions.

  4. Step 04

    Model and thresholds

    Week 4–6

    Train, test on held-out images, and choose thresholds with the quality team.

  5. Step 05

    Shadow run

    Week 6–8

    Run alongside human inspection and compare decisions before acting on them.

Pilot scope

Included

  • One line or station, one product family
  • Camera and lighting recommendations
  • Model training, evaluation, and thresholds
  • Integration with quality records

Decided later

  • Rollout to other lines
  • Automatic rejection without review
  • Additional defect types

What makes a vision use case viable

  • The defect is visible to a person looking at the same image
  • Items appear in a consistent position and orientation
  • Examples of the defect exist or can be collected
  • Someone owns the decision when the model is unsure

Common pitfalls we help you avoid

The most common pitfall is starting with the model instead of the setup. Inconsistent lighting, reflections, and varying camera angles cause more failures than any algorithm choice, so we fix the capture first.

The second is unclear defect definitions. If two inspectors would disagree about whether an item passes, the model will be inconsistent too, so we agree examples and rules with the quality team before labeling.

The third is deciding automatic rejection too early. Shadow runs alongside human inspection show where the model is reliable and where uncertain items still need a person.

What we measure

MeasureWhy it matters
Detection rateShare of real defects the model catches
False reject rateGood items wrongly flagged, which costs time and material
ThroughputInspections per minute compared with line speed
Review volumeUncertain items sent to inspectors

Who does what

We handle

  • Camera and lighting recommendations
  • Model training, evaluation, and edge deployment
  • Integration with quality records

Your team handles

  • Access to the line and sample parts
  • Defect definitions and labeling review
  • Decisions on rejection rules

A worked example

Think of a packaging line where a misprinted lot code or a missing seal occasionally reaches customers. A camera above the conveyor photographs each pack, a model checks the seal and reads the printed code, and the result is compared with the production order in the ERP.

Packs that fail are diverted for inspection, and each failure is saved with its image against the batch. The quality team sees defect rates by shift and can trace problems to a specific machine setting.

Before you start

  • A sample of good and defective items
  • Space and power for a camera near the line
  • Agreed defect definitions
  • A decision on what happens to rejected items

What success looks like

Success is fewer defects reaching customers, inspectors focusing on the uncertain cases, and a quality record with images that explains every rejection.

Governing vision decisions

Vision models, like other AI, make mistakes with a measurable rate. NIST's AI Risk Management Framework, released in January 2023 for voluntary use, provides a structure for managing those risks across an AI system's lifeSource [1]. For inspection systems, that translates into concrete practices: measure error rates on your own images, set thresholds by defect severity, and route uncertain cases to people.

We keep every image, result, and override, so quality teams can review decisions and the model can be retrained on real corrections. A pilot on images from your actual line, lighting, and cameras is the only reliable way to judge viability.

Frequently asked questions

How many images do we need?
It varies by task. Pilots often start with a few hundred labeled examples per class and grow from there.
Can it run without internet?
Yes. Edge devices run models locally and sync results.
What about privacy for people on camera?
We design for the minimum capture needed and follow your policies on retention and blurring.
Does it replace human inspectors?
It handles repetitive checks and escalates uncertain items to inspectors.
Can results go into Odoo Quality?
Yes, as quality check records linked to the production order.
Do we need special cameras?
Not always. Some tasks work with good IP cameras; fast lines or tiny defects need industrial cameras and lenses. The site visit decides.
How fast can it inspect?
Edge inference typically keeps up with line speed for single-item inspection. We test throughput during the pilot.
What if defects are very rare?
Anomaly detection learns from good items only, which helps when bad examples are scarce.
Can it read dot-matrix or engraved codes?
Often yes, with the right lighting and camera angle. We test sample items early.
Who maintains the model?
We hand over monitoring and retraining procedures, or support it under a maintenance plan.
Can one camera inspect several products?
Yes, if the product is identified first, for example from a barcode, so the right model and rules are applied.
What happens if the line changes?
Changes in product, lighting, or position may need new images and retraining. Monitoring shows when accuracy drifts.
Can it work with existing cameras?
Sometimes. We test images from your current cameras before recommending new hardware.
Can results trigger an alert to the operator?
Yes. A light, a screen message, or a notification can alert the operator the moment an item fails or needs review.
How accurate does a vision system need to be?
It depends on the cost of a missed defect versus a false alarm. We measure both on your images and set thresholds accordinglySource [1].
Can people override the system?
Always. Overrides are recorded and used to improve the model.

Sources

  1. [1]AI Risk Management Framework (opens in a new tab)NIST, released 26 January 2023