Why so many vision software brands, and why pay for any of them?
Walk the floor at Automate and count the machine vision booths. Cognex, Keyence, MVTec, Zebra, SICK, ifm, Banner, Omron, Beckhoff, Basler, Teledyne... every one of them will find your part, read your barcode, and measure your gap. So why are there a dozen of them? And in 2026, why does a vision license cost more than the PC it runs on when OpenCV is free and Claude will write the code for you?
We use most of these tools in our shop. Here's the honest version.
They're not really selling software
The first thing to understand is that almost none of these companies started out selling vision software. The software is the visible part of something else.
Cognex started in the early 80s doing OCR and built PatMax, which is still the pattern-matching tool everything else gets compared to. Matrox started as a frame grabber company...the MIL library existed to sell boards, and now the whole imaging division belongs to Zebra. MVTec spun out of a German research institute and Halcon still feels like it, in a good way. Keyence, SICK, ifm, Banner, and Omron are sensor companies. Vision is one more line in a catalog that already has a rep standing in your lobby. And Beckhoff put vision inside the PLC runtime because that's where Beckhoff puts everything.
So the answer to "why so many brands" is that machine vision was never one market. It's a feature that grew out of five different industries at once, and each one kept its own sales channel, its own hardware, and its own idea of who the user is.
That last part matters most. Some of these tools assume the user is a C++ programmer. Some assume it's a controls engineer who has never compiled anything. Both assumptions are correct...for different plants.
Two families
Everything on that list falls into one of two buckets. Libraries you program, and cameras you configure.
Product | What it really is | Who it's for |
|---|---|---|
Cognex CVL / VisionPro | C++/.NET libraries with the patented tools (PatMax) | Programmers building PC-based systems |
Cognex In-Sight | Smart cameras with a spreadsheet/point-and-click interface | Controls engineers, plant techs |
Keyence CV-X / XG Vision Editor | Menu-driven controller (CV-X) and a flowchart tool (XG) tied to their hardware | Plants that want the rep to set it up |
Matrox MIL (now Zebra Aurora) | Deep PC library plus a flowchart tool (Design Assistant) | Programmers, OEM machine builders |
MVTec Halcon | The research-grade library, HDevelop IDE, best-in-class 3D and deep OCR | Vision specialists, hard problems |
SICK / ifm | Vision as an extension of a sensor catalog, app-based setup | Simple presence/measure jobs, sensor buyers |
Banner | Low-cost smart cameras, dead-simple setup | Error-proofing, sensor-level budgets |
Omron FH / FHV | Vision fused into the Sysmac PLC ecosystem | Plants standardized on Omron controls |
Beckhoff TwinCAT Vision | Vision running inside the PLC real-time kernel | Machine builders already on TwinCAT |
OpenCV | The free library everything academic runs on | Anyone who can program, or has an AI that can |
Same math, different packaging. A blob tool is a blob tool. What you're choosing is who programs it, what hardware it's married to, and who answers the phone.
And here's roughly what the choices cost, per camera point, before anyone bolts anything to a machine:
So why pay?
Because the algorithm is maybe twenty percent of what you're buying. This is the part that's easy to miss if you've never had a vision system fail on a running production line.
What the license actually buys:
The patented tools really are better at the ugly stuff. PatMax finding a part at 40% occlusion under bad lighting is not the same as OpenCV template matching. The gap has narrowed a lot, but it's not zero.
A deterministic runtime on hardware the vendor controls. The smart camera fires in the same number of milliseconds every trigger, forever. It doesn't take a Windows update. Nothing else is running on it.
An interface a maintenance tech can open. When a job needs a threshold nudged at 2am, somebody who has never seen Python is going to do it, from a GUI, with the line waiting.
A paper trail. In pharma and med device, validation wants versioned firmware, a vendor of record, and documentation that a regulator has seen a hundred times before. "We pip-installed it" is a hard sentence in an audit.
Obsolescence management. That camera model will still be purchasable, with the same mounting and the same firmware lineage, when the machine is eight years old.
The 2am test
We call it the 2am test. When the line is down and product is backing up, who fixes it? Here's what 2am actually looks like, from systems we've been called in on over the years:
A power blip reboots the vision PC and the two USB cameras come back enumerated in the opposite order. The script runs fine. It's just inspecting station 2 with station 1's camera, and rejecting everything.
IT pushes a Windows update on a Tuesday night. The PC reboots mid-shift and comes back with a different .NET runtime, and the inspection app doesn't.
Summer hits. Somebody props open a bay door and afternoon sun rakes across the conveyor. False rejects spike at 2pm, disappear by 4, and come back the next day. Nobody suspects the door for a week.
Your part supplier quietly changes plating vendors. Nothing looks different to your eye. The match scores drift down for a week and then fall off a cliff on night shift.
And the classic...the engineer who wrote the script took another job 18 months ago, the source is on a laptop nobody can find, and the only documentation is the script.
None of these are exotic. Every one of them is a normal Tuesday somewhere. Now run the same failures on the paid path: there's a spare In-Sight on the shelf, a tech swaps it, loads the job file from the backup, and the line is running in twenty minutes. There's a vendor hotline that has seen your failure mode a hundred times. And if an integrator like us built the cell, we've got your job files archived and can remote in before you've finished your coffee.
That's what the license and the integrator actually cost out to. You're paying to not own the problem alone.
The license is the small slice anyway
Here's the part the brochure wars hide. On an installed system, the software license everybody comparison-shops is around fifteen percent of the bill:
The other 85% is physics and iron. And it's where inspections actually live or die:
Lighting. Pass/fail is decided before a single pixel hits the sensor. A scratch that's invisible under a ring light glows under dark field. The best algorithm on earth can't recover contrast the light never made.
Lens selection. Working distance, field of view, and how many pixels land on your smallest defect. Get that math wrong and no software tier fixes it.
Mounting and fixturing. A camera that vibrates, grows with the afternoon heat, or gets bumped by a forklift doesn't hold a calibration. Repeatability is a steel problem.
Electrical integration. Triggering off the encoder, part-in-place sensing, the PLC handshake, reject confirmation, cabling that survives washdown and cable tracks. The camera is one device on a machine full of them.
That's the part of the problem we sell at Saber. We're glad to work in whatever software your plant standardized on...Cognex shop, Keyence shop, Halcon on a PC, doesn't matter to us. The demo happens on a screen, but the light, the lens, the steel, and the wire decide whether it works in month eighteen. Twenty-five years of our systems are still running because of the 85%, not the 15%.
Where OpenCV plus Claude honestly wins
And it does win, more than the vendors would like. If the job is offline analysis, a lab tool, a prototype, an R&D fixture, a one-off measurement study, or anything that isn't gating product at line speed...the free path is legitimate now. It was always technically legitimate. What changed is the labor.
The expensive part of OpenCV was never the library, it was the six weeks of a good engineer's time. An AI that writes competent OpenCV in an afternoon just collapsed that cost. The bottom of the vision market, the simple presence-check and measure-a-width jobs, is going to feel that. Honestly, some of it already does.
But notice what the AI didn't collapse. It didn't make your USB camera enumerate in the same order after a reboot. It didn't give your maintenance crew a GUI, and it didn't sign your validation package. The code got cheap. The system around the code didn't. OpenCV is free the way a puppy is free.
Can you actually make money with the free stuff?
Yes. OpenCV is Apache 2.0 licensed. You can use it commercially, ship it in products, charge money, and owe nobody a dime or a line of source code. Plenty of commercial vision products have OpenCV somewhere in their guts.
The licensing trap isn't OpenCV, it's what gets bolted to it. The popular YOLO object detectors from Ultralytics are AGPL-3.0...use one in a commercial product and you either open-source your application or buy their commercial license. Pretrained model weights carry their own licenses too. If a lawyer will ever look at your product, read the license of everything pip installs, not just the headline library.
Is OpenCV plus AI the best free stack? As the chassis, yes. But the free shelf is deeper than most people know, and the best free tool for a specific job is usually something more specific sitting on top of it:
Tool | License | What it's good at |
|---|---|---|
OpenCV | Apache 2.0 | The general toolbox: filtering, matching, measurement, calibration |
scikit-image | BSD | Measurement and analysis in Python, NumPy-native |
Anomalib (Intel) | Apache 2.0 | Deep-learning defect detection trained on good parts only |
Ultralytics YOLO | AGPL-3.0 (paid commercial option) | Object detection — mind the license |
Segment Anything / Grounding DINO | Apache 2.0 | Zero-shot "find the thing" from foundation models |
Tesseract | Apache 2.0 | OCR — decent on clean print, weak on dot peen and low contrast |
ZBar / ZXing | LGPL / Apache 2.0 | Barcode reading — fine on clean codes; paid engines still win on damaged ones |
OpenMV | Open hardware, MIT | A real smart-camera form factor for light-duty jobs |
There's also a middle rung most buyers never hear about because it doesn't have a rep in your lobby. MVTec MERLIC, Euresys Open eVision, and Zebra's Aurora Vision Studio (the old Adaptive Vision) are serious tools at a fraction of flagship pricing. And the wave of low-cost cameras out of Hikrobot and Daheng comes with bundled software that's gotten legitimately good.
AI in the camera vs. the other AI
Every one of these vendors has bolted AI onto their product line by now. Cognex bought ViDi and put edge learning in the In-Sight 2800 and 3800. Keyence ships AI tools on CV-X. Halcon has deep learning and a genuinely good Deep OCR. Zebra, SICK, Banner, and Omron all have their versions.
None of this is as new as the brochures make it sound. Twenty years ago we were working with Guy Paillet and Anne Menendez at General Vision, over in Petaluma. Guy had co-invented the ZISC chip with IBM France back in 1993...actual silicon neurons on a die, a neural network you could hold. By the mid-2000s he and Anne had pushed it to the CM1K: 1,024 hardware neurons you could train right at the bench from a handful of part images, recognizing in microseconds, no GPU, no internet, no cloud. That technology later ended up inside Intel's Curie module. Guy and Anne were putting neurons on factory problems when "neural network" still got you funny looks on a plant floor.
And here's the part that belongs in this article. Being a decade ahead on the AI didn't exempt them from the 85%. The neurons still needed somebody to install them on real machines, test them, document them, and integrate them with everything else on the floor. Because Guy and Anne had built on a factory-standard Matrox platform instead of something exotic, they could go find an integrator who already knew it...and that's how they found Saber. The smartest silicon of its decade still showed up at the plant needing lighting, mounting, wiring, and a manual. Some things don't change.
So when a rep tells us AI inspection just got invented, we smile.
What today's in-camera AI actually is: small convolutional networks you train on a few dozen images of your own part, running on a GPU or an in-camera chip, answering one narrow question very fast. Scratch or no scratch. Good weld or bad weld. And they're genuinely useful...surface defects on shiny metal is a problem rule-based tools whiffed on for thirty years, and these tools eat it.
Then there's the other AI. The big foundation models. Claude can look at a photo of your part and tell you in plain English what's wrong with it, why the lighting is causing the false rejects, and what to try next. That's a different animal entirely, and it cannot run 300 parts a minute at 40 milliseconds of latency on a shop floor with no internet connection. It's not supposed to.
So the in-camera AI and the big AI aren't competitors. The narrow model inspects the part. The big model is starting to do everything around the inspection...write the OpenCV, configure the job file, read the shift's reject images and tell you why lane 3 drifted, draft the FAT procedure. The vendors are adding AI inside their products. The bigger shift is the AI standing outside all of them, using them like tools.
That second thing is where we think this is actually heading. Guy and Anne had the right idea in 1993 — the surprise is which direction the intelligence ended up arriving from.
The short answer
Why so many brands? Because vision grew out of frame grabbers, sensor catalogs, PLCs, and research labs all at once, and every lineage kept its channel.
Why pay? You're not paying for the algorithm. You're paying to not own the problem alone at 2am...and the license was only 15% of the system anyway.
And why learn OpenCV plus AI anyway? Because for everything that isn't a production gate, it just became the cheapest engineer you'll ever hire.
The 85% — lighting, lens selection, mounting, and electrical integration — is our day job at Saber. We'll work in whatever vision software your plant runs. If you're staring at this decision for a real machine, [get in touch](https://www.sabereng.com/#contact) — we've probably shipped your problem before.
Comments