Book a consultation

Cognex Deep Learning and AI: Edge AI, OneVision and VisionPro Deep Learning

Cognex Edge Learning, Now Edge AI Tools: What They Do and Need

Cognex edge learning, renamed Edge AI tools in Vision Suite 26.1: the four on-camera tools, how few images Cognex says they need, and where they fit.

Vladimir Romanov


Cognex edge learning is the family of AI tools that run on an In-Sight camera, start from a pre-trained model and adapt it with a small set of your own labelled images. Since In-Sight Vision Suite 26.1.0 the documentation calls them Edge AI tools; the release history reads “ViDi EL tools renamed to Edge AI tools”. Older manuals, forum posts and some Cognex pages still say ViDi EL, EL Read or edge learning for the same thing. This page covers what the four tools do, how few images Cognex says they need, and where they fit. It is part of Cognex deep learning and AI, within the Cognex technology hub.

The four Edge AI tools

Cognex’s 10-K says its edge learning systems “use pre-trained models”, which is why a handful of images can be enough to start (Cognex Form 10-K, fiscal 2023). The tools in the In-Sight Vision Suite documentation are:

ToolWhat it answersWhat you give itNotes from Cognex documentation
Anomaly DetectDoes this image differ from what OK looks like?Images labelled OK or NGIn EasyBuilder since 25.1.0; shows an anomaly score and a heat map
ClassifyWhich class is this: OK, NG or a type you define?Images labelled by classDefault classes OK and NG, more can be added; shows a predicted class with confidence and a model health indicator
SegmentWhere are the defect pixels?Labelled example imagesSecond version in 24.1.0; a “Very Small” feature size added in 24.4.0; we did not read the full tool page
ReadWhat do these characters say, where rule-based OCR fails?Character images labelledIntroduced as EL Read in 22.3.0; single-character training added in 25.3.0

In the spreadsheet environment the same tools appear as functions with ViDi in the name, such as ViDiELClassify and ViDiELAnomalyDetect. That is naming history, not a different product.

Dense collection of assorted metal gears, cogs and machine parts.
Training needs examples of good parts and, for most tools, some bad ones, laid out under the same light the line will use.

How few images Cognex says it needs

Every number here is a vendor claim. Cognex stated “as few as five images” for the In-Sight 2800 and ViDi EL Classify at its April 2022 launch. Its April 2024 L38 release gives five to ten labelled images, and the 2026 In-Sight 6900 release gives 10 to 20 for Few Sample Classification, a separate Advanced AI mode. Cognex-linked articles contrast this with hundreds or thousands of images for PC-based deep learning.

The Anomaly Detect documentation is more specific than the marketing: label at least two OK images and one NG image. For a real job, follow the documentation and test the result against real rejects, not against the images you trained on.

Our view. Five images starts a trial. It does not validate a process. How many you need depends on how much your product and your lighting vary, and the way to find out is to hold back a test set of parts the model has not seen.

Where Edge AI tools run

The Vision Suite 24.3.0 release notes say it supports the new In-Sight 3800 “for increased edge learning performance” and that “the training also happens on the vision system without any external graphical accelerator”. The In-Sight 8900 manual lists “Edge Learning tool support”, the 2022 In-Sight 2800 shipped with ViDi EL Classify, and Cognex’s 5 May 2026 In-Sight 3900 release names Edge AI tools alongside Advanced AI tools and rule-based tools. We did not find a full table of which tool runs on which model. Check the exact model and Vision Suite version before you design around Segment or Anomaly Detect.

Training in the Vision Suite emulator on a PC is a separate matter. Cognex’s support documentation, as we read it, uses a cloud credential file or a dongle for Edge AI in the emulator, and the 26.2.0 release adds a warning when the dongle is lost mid-training. That means an engineer without the licence cannot prepare a job offline.

An Edge AI inspection job, step by step Five boxes in a row. Acquire the image under stable lighting. Locate the part with a pattern, edge or blob tool, which creates a fixture. Run an Edge AI tool, Anomaly Detect, Classify, Segment or Read, on the fixtured region. Produce a result: OK or NG, a score or a class. Send the result to the PLC. A note below says the Anomaly Detect tool needs the upstream fixture so every image has the same perspective. 1. Acquire trigger, light, exposure 2. Locate pattern, edge or blob sets fixture 3. Edge AI tool Anomaly, Classify, Segment, Read 4. Result OK or NG, score, class, confidence 5. To the PLC reject, count, log the image Cognex documentation: Anomaly Detect needs an upstream fixture for a consistent perspective The AI tool judges a stable, located region. It does not replace the steps before it.
An Edge AI tool sits in the middle of an ordinary job. The steps before it decide whether it sees the same picture every time.

A first training pass, as Cognex documents Anomaly Detect

The documented workflow for Anomaly Detect gives a sense of the effort. Other tools differ in detail.

  1. Fixture the part with an upstream tool so every image has the same perspective.
  2. Acquire sample images, live or loaded into the filmstrip, and label each OK or NG.
  3. Draw the region of interest, remembering it cannot be resized afterwards.
  4. Set the feature size, trading speed against resolution.
  5. Pick a threshold mode: Prioritize OK, Balanced or Prioritize NG.
  6. Read the heat map on test images, and check that it highlights the defect and not a shadow or a label edge.

Illustrative scenario. A packing line seals film over trays. A blob tool with a threshold catches dark specks on a white surface well, but wrinkles, product in the seal and print smears vary every shift. A team might keep the rule-based tool for specks, add an Anomaly Detect tool on the seal region, and review a week of rejected images with quality before deciding where the threshold sits. The point is the sequence: rules first, then AI for what the rules cannot describe.

Where Edge AI tools fit

They fit variable defects that are hard to write as rules: cosmetic marks, contamination, surface variation, and sorting parts into classes. They fit lines where a PC and GPU at each station is not wanted, and parts that change often enough that retraining on the device is attractive.

They fit poorly for dimensional gauging, for reading clean codes and text that rule-based tools already handle, and for any station where the image is not repeatable. For the wider decision, see the decision tree on the deep learning page.

For a PC-based alternative with more control over the model, see VisionPro Deep Learning. The cameras themselves are covered under In-Sight.

Where it gets hard

The training set is the product. Good images from one shift and one light condition will not cover the plant. Collect across shifts, lots and changeovers, and keep a held-out set.

Region of interest. The Anomaly Detect documentation notes the region of interest cannot be resized after you click OK, and a loose region of interest on a Read job invites trouble. Set it deliberately, and keep it tight.

Feature size. Anomaly Detect has a feature size control: larger is faster at lower resolution, smaller is more accurate but slower. Cycle time is a real constraint at line speed, so test on the target camera.

Trigger rate. A Cognex release note says the trigger interval must be at least 1.5 times the job execution time or images are missed. AI tools lengthen the job, so recheck.

Thresholds. Anomaly Detect offers Prioritize OK, Balanced and Prioritize NG. That is a business decision, because it trades false rejects against escapes. Make it with quality, not alone at the laptop.

Change control. Vision Suite 25.2.0 added saving jobs while online, and 24.3.0 added audit logging. Use both, and keep a dated backup of every job that goes live. For the data side, see image archiving and data and, for where results land in a controller, EtherNet/IP and Studio 5000.

We have no commercial relationship with Cognex and do not resell its products. See Edge and I/O and Edge solutions for industrial automation for how we think about this layer.

Frequently asked questions

What is Cognex edge learning?

Cognex edge learning is the company's name for AI inspection tools that start from a pre-trained model, adapt to a small set of your labelled images and run on the camera itself. In In-Sight Vision Suite 26.1.0 the documentation renamed the ViDi EL tools to Edge AI tools. Cognex's 10-K describes edge learning systems as using pre-trained models.

Which Cognex cameras run Edge AI tools?

Cognex documentation and releases tie them to the In-Sight 3800, 8900 and 3900, with edge learning also shipped on the 2800, the SnAPP sensor and the L38 3D system. We did not find a complete per-model tool matrix. Confirm that the tool you need, such as Segment or Anomaly Detect, is on the exact model and Vision Suite version you plan to buy.

How many images does Cognex Edge AI need?

Cognex states as few as five images for the In-Sight 2800 and five to ten labelled images for the L38. The Anomaly Detect documentation asks for at least two OK images and one NG image. Those are minimums to start, not validation evidence. Production use needs enough examples to cover the variation your line actually produces.

Do Edge AI tools need a PC or GPU to train?

Cognex's Vision Suite 24.3.0 release notes say that, on the 3800, training also happens on the vision system without any external graphical accelerator. Training in the Vision Suite emulator on a PC needs a Cognex cloud licence file or dongle. Whether your camera trains live or only runs a PC-trained model is worth testing on the actual unit.

Is Edge AI better than rule-based tools?

Neither is better in general. In our view rule-based tools win for measurement and clean codes, and Edge AI wins where defects vary and are hard to describe as rules. Cognex ships both in the same camera, and the usual design is a rule-based tool to locate the part and an Edge AI tool to judge it.

References and further reading

  1. In-Sight Vision Suite EasyBuilder release history · Cognex documentationThe 24.3.0 on-camera entry, the Anomaly Detect tool in 25.1.0 and the 26.1.0 rename.
  2. Anomaly Detect tool (EasyBuilder) · Cognex documentationThe training workflow, minimum images, threshold options and ROI behaviour.
  3. Classify tool (EasyBuilder) · Cognex documentationLabelling by class, the predicted class and the model health indicator.
  4. Cognex In-Sight 2800 combines deep learning and traditional vision · Cognex (via PR Newswire India)The 'as few as five images' claim for ViDi EL Classify.
  5. Cognex Corporation Form 10-K for fiscal 2023 · U.S. Securities and Exchange CommissionCognex's description of edge learning systems as using pre-trained models.
  6. Cognex launches the In-Sight 3800 vision system · Cognex (via PR Newswire India)The April 2023 launch release.

← Deep Learning and AI

Have a project that needs technical ownership?

Send us the situation. If it is not something we should take on, we will say so.

The Plant-Side Brief

The weekly brief for the people who own a plant’s control systems.

One email every Wednesday for engineering and operations leaders responsible for controls, SCADA, MES and OT networks. Skim the top in two minutes; read the Deep Dive when you have fifteen.

  • One modernization problem, worked through to cost, risk and a schedule you can defend. The Deep Dive.
  • End-of-support dates before they reach your plant Windows, servers, controllers and HMI software. Lifecycle Watch.
  • A checklist your engineers can run that week not a framework to admire. Field Kit.
  • Every number sourced, no vendor agenda we are paid by plants, never by vendors.