Book a consultation

Cognex

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

Cognex deep learning explained: ViDi and SUALAB history, Edge AI, OneVision and VisionPro Deep Learning, when AI is the wrong answer, and how it fails.

Vladimir Romanov


Cognex deep learning is a family of AI inspection tools that share a lineage but run in different places: Edge AI tools on the camera, newer Advanced AI tools on the 2026 In-Sight hardware, and VisionPro Deep Learning on a PC with a GPU. Cognex bought the technology in two deals, ViDi Systems in 2017 and SUALAB in 2019, and spent 2022 to 2026 moving it from a GPU workstation onto cameras and into a cloud platform. This page explains the line, how it differs from rule-based inspection, when it is the wrong answer and how it fails in plants. It sits in the Cognex technology hub, and the company behind it is profiled in Cognex Corporation.

Close-up of a dark blue circuit board with gold pads and a chip.
The model runs on a processor module. Where that module sits, in the camera, a controller or a PC, is the first selection question.

Cognex deep learning, from ViDi to Edge AI

Cognex entered deep learning by acquisition. The ViDi Systems purchase closed on 4 April 2017 (Vision Systems Design report of the Cognex announcement), and the SUALAB purchase closed on 16 October 2019 for about $195 million in total, of which $171 million was cash at closing and $24 million was deferred to the fourth anniversary and tied to key staff staying (Cognex Form 10-K, fiscal 2019). Cognex stated in that filing that, combined with the ViDi intellectual property, it believed it was “a leading provider of deep learning-based industrial vision software”. The deal terms are covered on the acquisitions page.

DateEvent
4 April 2017ViDi Systems (Switzerland) acquired
16 October 2019SUALAB (Seoul) acquired; its SuaKIT software became the High Detail mode
20 April 2020In-Sight D900 announced; Cognex’s release headline calls it the first industrial smart camera powered by deep learning, and trade press reports that ViDi applications for it were trained on a GPU-enabled Windows PC
28 July 2021Cognex announces VisionPro Deep Learning 2.0 together with VisionPro 10.0
April 2022In-Sight 2800 with ViDi EL Classify running on the camera
April 2023In-Sight 3800 launched; Vision Suite 24.3.0 later ran edge learning on the camera without an external graphics accelerator (release date not found)
June 2025OneVision cloud platform in beta
27 August 2025VisionPro Deep Learning 4.1.0 dated in Cognex’s release file
28 April to 13 May 2026In-Sight 6900 (28 April), In-Sight 3900 (5 May) and OneVision general availability (13 May)
Vision Suite 26.1.0“ViDi EL tools renamed to Edge AI tools” (release date not found)

The pattern is that training moved toward the device. In 2020 the camera ran the model and a PC trained it. By Vision Suite 24.3.0 the camera did both.

The Cognex deep learning line, as of November 2026

There are four things to tell apart, and Cognex’s own names for them have changed.

Edge AI tools are the on-camera tools formerly called ViDi EL or edge learning: Anomaly Detect, Classify, Segment and Read. Cognex says they start from pre-trained models and adapt to your images. See Edge AI tools.

Advanced AI tools appear on the In-Sight 3900 and 6900. The 6900 release describes “Few Sample Classification”, which it calls transformer-based and trainable from as few as 10 to 20 images, and pixel-level segmentation. Cognex has not published tool-level documentation for them in the sources we could read, so we do not describe their behaviour.

VisionPro Deep Learning is the PC and GPU product, the direct successor to VisionPro ViDi per Cognex’s release notes. See VisionPro Deep Learning.

OneVision is a cloud platform for building, validating, versioning and deploying models, not a camera. See OneVision.

TierWhere it runsHow it is trainedImages, as Cognex or Cognex-linked sources stateWhere it fits, in our view
Rule-based tools (In-Sight, VisionPro)Camera or PCAn engineer sets tool parametersNot applicableMeasurement, location, known codes and text
Edge AI toolsOn the In-Sight cameraPre-trained model adapted with your labelled imagesAs few as five (2800, 2022); five to ten (L38, 2024)Variable defects and sorting on a fast line without a PC
Advanced AI toolsIn-Sight 3900 and 6900Transformer-based per the 6900 release10 to 20 (Few Sample Classification)Harder cases at the edge; details unpublished
VisionPro Deep LearningPC with a supported GPUTrained on the PCCognex-linked articles say hundreds or thousands; no official figure foundFine or numerous defect types, and PC-based vision

Every image count above is a vendor claim. None has been independently benchmarked in the sources we read. Treat them as a starting point for a trial on your parts.

How deep learning differs from rule-based inspection

A rule-based tool is an explicit instruction: find this edge, measure this gap, count blobs above this size, match this shape. An engineer sets it, and when it fails you can usually say why. A deep learning tool is a model trained from labelled images. It is strong where a defect or a product varies in ways that are hard to specify, such as natural products, cosmetic marks and damaged codes. It is weaker where you need a repeatable measurement or an explanation.

It needs representative examples, including bad ones. The model only knows the variation it has seen, so a validation set and change control for retraining are part of the job.

Hybrid is normal. Cognex ships both kinds in the same camera, and the In-Sight Vision Suite documentation says Anomaly Detect needs an upstream tool to fixture the part so every image has the same perspective. In practice a pattern or edge tool locates the part and the AI tool judges it.

When Cognex deep learning is the wrong answer

Decision tree: rule-based tools or Cognex AI tools Four questions in order. First: is the pass or fail rule a measurement or a known code? If yes, use rule-based tools. If no, second: is the image stable, with fixed lighting, fixturing and position? If no, fix lighting and fixturing first. If yes, third: can you collect good and bad examples? If no, collect samples before choosing a tool. If yes, fourth: are the defects fine or numerous, and is a PC with a GPU acceptable? If yes, consider VisionPro Deep Learning. If no, use Edge AI tools on the camera. A regulated plant locks and validates the model version before any of them. Is the rule a measurement or a known code? Is the image stable: fixed light, fixture, position? Can you collect good and bad examples? Fine or numerous defect types, and a PC with a GPU is acceptable? No Yes Yes No Rule-based tools Yes Fix lighting and fixturing first No Collect samples before choosing a tool No VisionPro Deep Learning consider, and trial against Edge AI Yes Edge AI tools on the camera with a fixture tool upstream Regulated plant: lock the model version and validate it before any AI tool goes live.
Most "AI or rules?" decisions are settled by the first two questions. A bad image or a measurement is not an AI problem.

It is the wrong answer for measurement. If the pass criterion is a dimension, a gap or an angle with a tolerance, a rule-based gauging tool gives a number you can calibrate and defend. A model gives a score.

It is usually the wrong answer for clean standard codes and text. In our view rule-based readers handle well-printed codes, and Cognex’s Edge AI Read tool is aimed at hard text, so reach for AI on a code only when rule-based reading has already failed on your marks.

It will not fix an unstable image. In our view the most common root cause behind a “bad AI model” is light that moves with shifts and ambient sources, a loose lens or a part that sits differently in the fixture. A model trained on those images learns the instability.

It is wrong when you cannot get samples. A new product with no history of defects can still start with anomaly detection, but plan for more validation, not less.

How Cognex deep learning fails

Failure modeWhat you seeWhat we would do
A defect type never seen in trainingPasses that should fail, with no warningKeep a reject challenge set and add new defects to it as they appear
Drift after a change to light, lens, product colour or packA rise in false rejects or escapes after changeoverRe-verify the model after any optics or material change
Too few defect samplesModel overfits and misses variantsCollect real rejects over time, treat early results as provisional
Inconsistent labellingThe model learns the disagreementOne written defect catalogue, one reviewer on borderline images
Model updated without change controlBehaviour changes with no recordVersion and lock models like any other recipe

Samples. Cognex’s Anomaly Detect documentation asks for at least two OK images and one NG image. Follow the documentation for the job in front of you and test with real rejects.

Labelling. In our view this is the quiet failure. Two operators who label a borderline scratch differently teach the model that the scratch is random.

Validation in regulated plants. GAMP 5 Second Edition (ISPE, published July 2022) is risk-based guidance, not regulation. 21 CFR Part 11 applies where a vision system creates or holds records the regulations cover. The FDA’s computer software assurance final guidance, issued in September 2025 and replaced in February 2026 by a final guidance titled “Computer Software Assurance for Production and Quality Management System Software”, addresses production and quality system software for device makers under the Part 820 quality management system regulation. The FDA page does not mention Part 11. Whether a given Cognex product offers Part 11 features, we could not confirm from public sources, so ask in writing. The Vision Suite release history lists audit logging (24.3.0), a System Validation utility and Job Validation panel (25.1.0) as tools that support the work, not as the work itself. Our practical list: locked and versioned model files, challenge tests with known good and bad parts at qualification, periodic re-verification, and change control when lighting or the lens is replaced.

What we do here

Our team has programmed in-line inspection on Cognex In-Sight cameras at large food and consumer-goods manufacturers, and we integrate Cognex cameras and readers with PLCs, line controls and data systems as part of controls projects. We have no commercial relationship with Cognex and do not resell its products. This page is written from public sources, not from our own deep learning deployments. Where we help:

  • Rule-based or AI? Working through the decision above with your parts and your lighting before you buy a tier.
  • Specification. Writing what the inspection must catch, how it is proved, and what result the PLC receives.
  • PLC handshakes. Passing results, scores and class codes to the controller and handling the “no result” case; see systems integration.
  • Commissioning and troubleshooting. Light, fixturing and trigger problems that look like model problems.
  • Data paths. Where images and results go next; see plant data.

Limits and open questions

  • No independent benchmarks. Image counts and speed claims are Cognex’s.
  • Advanced AI is thinly documented. Tool names and their relation to Edge AI and VisionPro Deep Learning are not published in the sources we read.
  • OneVision pricing and supported cameras after launch are not public. See the OneVision page.
  • Lifecycle. Cognex has not published an end-of-life notice we could find for the In-Sight D900, so we call it older, not discontinued.
  • Dates. We could not find release dates for most Vision Suite versions or VisionPro Deep Learning 4.2.
  • Cognex product pages were not readable by our tools, so everything here rests on documentation, releases and filings.

Deeper

Frequently asked questions

What is Cognex deep learning?

It is Cognex's AI inspection line, built on technology from two acquisitions, ViDi Systems (2017) and SUALAB (2019). As of November 2026 it comes in tiers: Edge AI tools that run on In-Sight cameras, newer Advanced AI tools on the In-Sight 3900 and 6900, the PC and GPU product VisionPro Deep Learning, and OneVision, a cloud platform for managing models.

What is the difference between Cognex Edge AI and VisionPro Deep Learning?

Cognex describes Edge AI as pre-trained models adapted with a few of your images and run on the camera, with no external graphics accelerator. VisionPro Deep Learning trains and runs on a PC with a supported GPU, licensed through a Cognex security dongle. In our view the PC product suits harder, finer defects and the camera tools suit fast lines with fewer parts to buy.

Did ViDi EL become Edge AI?

Yes. The In-Sight Vision Suite 26.1.0 release history reads 'ViDi EL tools renamed to Edge AI tools'. Older documents and forum posts still say ViDi EL, EL Read or Edge Learning for the same tool family. Cognex's 5 May 2026 In-Sight 3900 release already used the Edge AI name.

How many images does Cognex AI need to train?

Cognex states as few as five images for the In-Sight 2800 edge learning tool (2022), five to ten labelled images for the L38 3D system (2024) and ten to twenty for Few Sample Classification on the In-Sight 6900 (2026). These are vendor claims, not benchmarks. The number that matters is how many good and bad examples cover the variation on your line.

When should I not use deep learning for inspection?

Do not use it for pass or fail rules that are a measurement, for clean standard barcodes and text that rule-based tools read reliably, or when lighting and fixturing are unstable. In our view AI on a bad image only hides the problem for a while. Fix the optics first, then decide whether rules or a model fit the defect.

Can Cognex deep learning be validated in a regulated plant?

It can be validated like other computerised systems, but the burden is higher than for fixed rules: you must control training data, lock model versions and challenge-test with known good and bad parts. We could not confirm from public sources that Cognex products meet 21 CFR Part 11 requirements by default, so ask Cognex for that in writing.

References and further reading

  1. In-Sight Vision Suite EasyBuilder release history · Cognex documentationThe 24.3.0 on-camera edge learning entry and the 26.1.0 rename from ViDi EL to Edge AI.
  2. Cognex Deep Learning release notes (version 1.0) · Cognex documentationStates that VisionPro Deep Learning is the direct successor to VisionPro ViDi.
  3. Cognex Launches OneVision (general availability) · Cognex (via AAP)The 13 May 2026 release, including the customer quotes and Cognex's claims.
  4. Cognex launches highest-performance embedded AI vision system powered by Qualcomm · Cognex (via AAP)The In-Sight 3900 release, which names Edge AI tools and Advanced AI tools.
  5. Cognex Launches In-Sight Vision Controller Powered by NVIDIA · Cognex (via PR Newswire)The In-Sight 6900 release, with Few Sample Classification and pixel-level segmentation.
  6. Cognex Acquires Maker of Deep Learning Software for Industrial Machine Vision · Cognex (via Business Wire)The ViDi Systems acquisition, April 2017.
  7. Cognex Corporation Form 10-K for fiscal 2019 · U.S. Securities and Exchange CommissionThe SUALAB purchase terms and Cognex's own description of its deep learning position.
  8. Computer Software Assurance for Production and Quality System Software · U.S. Food and Drug AdministrationThe February 2026 final guidance, which replaced the September 2025 one, on assurance of production and quality management system software for device makers.
  9. GAMP 5 Guide, Second Edition · ISPEPublished July 2022.
  10. Deep learning powered smart camera introduced by Cognex · Vision Systems DesignThe In-Sight D900, 20 April 2020, with GPU PC training.
  11. Cognex VisionPro Software Breaks Down Barriers Between Deep Learning and Traditional Vision · Cognex release, 28 July 2021VisionPro 10.0 and VisionPro Deep Learning 2.0.

← Cognex

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.