Machine Vision Software Compared: 9 Platforms for 2026

The best machine vision software is the one the person who'll edit your inspection in two years can actually work in. For a software engineer, that's an SDK like MVTec HALCON, Cognex VisionPro or Zebra's Aurora Imaging Library. For a controls engineer or quality tech, it's a configurable tool like MERLIC, Sherlock, Aurora Design Assistant or a Keyence controller. OpenCV belongs in prototypes more often than on the line.

We build robotic cells with machine vision built in, and we've programmed most of these platforms on real stations. No camera vendor pays us, so we'll tell you where each one's annoying. Everything we state about licensing, deep learning and camera support below comes from the vendor's own pages, checked in October 2026, and we say so when something's our judgment instead. If you're still choosing hardware, start with our Cognex vs Keyence comparison. This post is about the software you'll live with.

Key takeaways

Pick software by who maintains the job, not by the demo. Code SDKs reward engineers and punish everyone else.
Most commercial platforms sell a development license once and a runtime license for every station, often on a USB key.
HALCON, MERLIC and Zebra's tools run with a wide range of GenICam cameras. Keyence is a closed set.
OpenCV is free under Apache 2.0, but you'll write the acquisition, PLC handshake, operator screen and validation yourself.
Deep learning is now in every major platform. The differences are in training tools, GPU needs and how it's licensed.

What does machine vision software do?

Machine vision software takes images from an industrial camera, runs inspection tools on them (find the part, measure it, read the code, judge the surface) and sends a pass, fail or coordinate to a PLC or robot fast enough to keep up with the line. The tools matter. The plumbing around them matters just as much.

That plumbing is what separates industrial vision software from an image processing library. A production job needs camera triggering and acquisition, tools that report calibrated units, a handshake with the PLC, recipes for each part number, an operator screen, image logging and some way to lock the job so nobody edits it at 2 a.m. without leaving a trace. Every platform in this post gives you some of that. They differ in how much, and in how you build it.

There are really three ways to build. You configure (pick tools from a menu and set parameters on the image), you draw (a flowchart or point-and-click program), or you write code against an SDK. Camera choice is the other axis, and it's where the closed and open platforms split.

Fig. 01Machine vision software map · how you build vs which cameras

Columns tell you who can maintain the job; rows tell you how locked in you are. Placement is our judgment from vendor documentation and our own commissioning work, and several platforms straddle columns (VisionPro's QuickBuild is graphical, its full power is in .NET code). OpenCV is highlighted because it's the only one where camera support is entirely your job.

How do the main machine vision software platforms compare?

The big split is code versus configuration. VisionPro, HALCON and Aurora Imaging Library are SDKs for engineers who program. MERLIC, Sherlock, Design Assistant, pylon vTools and Keyence's controllers are built so you can set up a job without writing code. OpenCV is a free library, and AI platforms like LandingLens focus on training models from images.

Here's the whole field on one page. We only put a licensing fact in the table if the vendor states it publicly. Where a vendor doesn't publish something, the cell says so rather than guessing.

Machine vision software comparison · vendor-stated facts, October 2026
Platform	How you build	Deep learning	Cameras	Licensing (as published)	Best fit
Cognex VisionPro + VisionPro Deep Learning	QuickBuild graphical setup, then .NET code	Separate VisionPro Deep Learning product; NVIDIA GPU2	Cognex and GigE cameras; camera-count licenses	Each tool needs a license on a Cognex USB security key1	PC stations with custom logic, Cognex shops
MVTec HALCON	HDevelop IDE, then C, C++, C# or Python	Built in (classification, detection, segmentation, anomaly detection, OCR)	Broad GenICam, GigE Vision, USB3 Vision support	Steady and Progress editions; individual quotes3	Hard algorithms, 3D, embedded and Linux builds
MVTec MERLIC	No code, set up on the image	Deep OCR and Deep Counting in base; all DL in Professional4	GenICam GenTL, GigE Vision, USB3 Vision	Lifetime license, one-year upgrade window4	Teams who want HALCON tools without programming
Zebra Aurora Imaging Library	C, C++, C# or Python API5	Included as tools and models	Device interfaces in the API	Free trial and a free Lite subset5	OEM machines, Matrox heritage code
Zebra Aurora Design Assistant	Flowchart IDE plus web operator screen6	Classification, segmentation, object detection steps	Iris GTX smart cameras, or PCs with CoaXPress, GigE Vision or USB3 Vision	USB key for development; runtime by software key or USB key6	Controls engineers, mixed camera fleets
Teledyne Sherlock	No-code, point-and-click7	Models trained in Astrocyte, run on CPU or GPU	Teledyne DALSA plus many third-party GigE GenICam cameras	Licensed to run on any Windows x64 PC7	Teledyne DALSA hardware, plants that want no code
Basler pylon vTools	Graphical recipes in pylon Workbench	AI vTools (classification, detection, segmentation)	Built around Basler's pylon SDK and cameras	Bought per vTool; dongle or PC-locked; 180-day evaluation8	Basler camera users adding a few tools
Keyence CV-X, XG-X, VS	Menu-guided (CV-X, VS) or flowchart (XG-X)	VS series has built-in AI tools	Keyence controllers, cameras and lights	Quoted with the hardware	Fast setup with direct vendor support
OpenCV	C++ or Python code	DNN inference module; train elsewhere	Anything, through a vendor SDK you add	Free, Apache 2.09	Prototypes, research, custom products with a software team
LandingLens and other AI platforms	Label and train in a web app	The whole point	Images from your existing cameras	Subscription; cloud, Docker or LandingEdge deployment10	Cosmetic defects that rules can't describe

Two cautions on reading it. "Built in" deep learning doesn't mean free: most vendors license the AI tools separately from the rule-based ones. And "broad camera support" means the interfaces exist, not that every camera has been tested. Ask for the specific model you plan to use to be on the vendor's list, or test it before you buy ten.

HALCON vs VisionPro: which is better?

Neither is better across the board. HALCON has the broader algorithm library, runs on Windows, Linux and Arm, and works with a very wide range of cameras. VisionPro is the natural choice when the plant already runs Cognex, wants a .NET stack and values Cognex's support network. The skill of the person programming it matters more than either.

Here's how the decision usually goes on the stations we see. HALCON wins when the job's algorithmically hard. Odd 3D geometry, subpixel gauging on a part that won't sit still, an embedded Linux box inside a machine you're going to sell. Its HDevelop environment lets an engineer prototype interactively and then export to C++, C# or Python, and the operator count is enormous. MVTec sells it in two editions, Steady for long-term stability and Progress for quicker access to new features, and says both can run side by side.3 Prices are quoted individually.

VisionPro wins when the plant's already a Cognex plant. Your integrator knows it, your In-Sight cameras' tools behave the same way, and QuickBuild lets someone set up a job graphically before it moves into C# code. Its licensing is the most granular we deal with. Every vision tool needs at least one enabled license on the Cognex USB security key, and GigE camera licenses come in tiers like VisionPro.Cam2 for up to two cameras and VisionPro.CamMulti for unlimited.1 That's fine once you know it. It's a surprise the first time a station won't start because the key's missing a tool someone added in the lab.

Deep learning splits them too. Cognex sells VisionPro Deep Learning as its own product, and version 4.1 supports NVIDIA CUDA GPUs with compute capability 6.1 and up.2 HALCON folds deep learning into the same library and the same HDevelop session. If you'll mix rule-based and AI tools on one image a lot, that integration saves real hours.

Honestly, the tiebreaker we use most is boring. Who's going to support this at 6 a.m. when it faults? If that's an integrator with five VisionPro engineers within driving distance, that beats a theoretically better library nobody nearby can touch.

Which machine vision software needs no programming?

MVTec MERLIC, Teledyne Sherlock, Zebra Aurora Design Assistant, Basler's pylon Workbench with vTools and Keyence's CV-X and VS series all let you build an inspection without writing code. They differ in how far they stretch. Flowchart tools handle branching logic better, and menu-driven tools are faster for standard jobs.

No-code isn't a lesser category. On a lot of lines it's the right one, because the person who'll edit the job is a quality tech who touches it twice a year. What you're buying is guardrails. MERLIC sets parameters directly on the image and its base package comes with a lifetime license and a one-year upgrade window, which we wish more vendors did.4 Sherlock is point-and-click and talks Modbus, EtherNet/IP, PROFINET, OPC UA and MQTT out of the box.7 Design Assistant is a flowchart IDE with a web-based operator screen you edit in the same tool, and it can deploy to a Zebra Iris GTX smart camera or a PC.6

The catch shows up when a job doesn't fit the tool's assumptions. A menu-guided controller is wonderful until you need a measurement that depends on two other measurements and a part-number lookup. Then you're working around the software. Flowchart tools (XG-X, Design Assistant, Sherlock's program tree) bend further before they break. Most of them also let a developer add a custom step in code. MERLIC Professional, for example, runs custom tools built in HALCON's HDevelop.4 That escape hatch is worth checking before you commit.

How is machine vision software licensed?

Most commercial machine vision software uses two kinds of license. A development license opens the full editor for the engineer who builds the job. A runtime license lets one production station run it. Runtime licenses live on a USB dongle or are locked to the PC, and some vendors license each tool separately.

This is where budgets go wrong, because the development seat is a one-time cost and the runtime license is a cost per station, forever. Ten identical stations means ten runtime licenses. Zebra's Design Assistant spells the model out clearly. The development package includes a single-user USB hardware license key, and each runtime is licensed with a software key generated from a lock code that identifies the target system, or with a pre-programmed USB runtime key.6 Basler's vTools work the same way at a smaller scale. You buy runtime licenses per vTool, load them onto a USB dongle you can move between PCs or lock them to one computer, and you get a 180-day evaluation license that can't be renewed.8

Fig. 02Development vs runtime licenses · where the cost scales

The development seat is a one-time cost; runtime licenses scale with every station and spare. This is the common pattern across Cognex, Zebra and Basler as their documentation describes it. MVTec, Teledyne and Keyence publish less detail, so ask them the four questions in the bottom box directly.

Two details bite on real lines. First, spares. A spare vision PC on the shelf is useless if its license is locked to the PC that died, so ask how a license moves and whether that needs the vendor's office to be open. Second, maintenance lapses. Cognex's documentation describes a software update maintenance period programmed onto each security key, which ties the key to VisionPro releases from those years, so an old key won't simply run the newest version. Budget the maintenance renewal or plan to freeze the version, and decide which before the station ships.

Machine vision software licensing models · as published
Platform	What you license	Where it lives
Cognex VisionPro	Each vision tool, plus camera count	Cognex USB security key1
MVTec HALCON	Steady or Progress edition, development or runtime	Quoted per project3
MVTec MERLIC	Package (base or Professional) plus add-ons	Lifetime license4
Zebra Design Assistant	Development package; runtime packages by tool group	USB key, or software key from a lock code6
Teledyne Sherlock	Sherlock on a Windows x64 PC	PC license; included on some Teledyne systems7
Basler pylon vTools	Each vTool, sold singly or in bundles	Movable USB dongle or locked to the PC8
OpenCV	Nothing	Apache 2.0 open source9

Is OpenCV good enough for industrial inspection?

OpenCV's algorithms are good enough for a lot of industrial inspection. The library isn't the problem. What it doesn't give you is everything around the algorithm, like camera acquisition, PLC communication, an operator screen, recipe control, audit logs and a vendor to call. On a production line you'll build and support all of that yourself.

We like OpenCV. It's free under the Apache 2.0 license,9 it's in every computer vision course, and it's the fastest way to find out whether a measurement is even possible on your images. Plenty of good products ship with it inside, built by companies with a real software team.

It's the wrong choice when a plant expects to treat the station like equipment. The usual story goes like this. A capable engineer writes a Python script in two weeks that works on the bench. Then the line needs hardware triggering from an encoder, and the script needs a vendor SDK or a GenICam library to talk to the camera. Then it needs EtherNet/IP to the PLC, a screen operators can use with gloves on, recipes for twelve part numbers and a log the quality team can audit. Six months later the engineer's moved on and nobody else can open the code. That's not OpenCV's fault. It's just what a library is.

Fig. 03OpenCV for industrial inspection · what you still have to build

OpenCV covers the algorithm row well; the rest of the station is on you. Acquisition libraries named here (Aravis, Harvesters) are common open-source GenICam options, not endorsements. Commercial platforms differ in how much of each row they include, so check the PLC and operator-screen rows against your line.

So our rule is simple. Use OpenCV to prove feasibility, and to build products where you own the software for life. For a one-off production station your maintenance staff will inherit, port the logic into a commercial platform before it ships. It costs a license. It saves you the phone call in year two when the only person who understood the script is gone.

Which vision software has the best deep learning?

Every major platform now has deep learning, so the real question is fit. HALCON and Aurora Imaging Library suit engineers who'll tune models in code. VisionPro Deep Learning, Teledyne's Astrocyte, Keyence's VS and MERLIC Professional train from labeled images with little or no code. Cloud platforms like LandingLens add team labeling and version control.

The differences that matter on a line are three. Where training happens, what hardware the model needs at runtime, and how the AI tool is licensed. Sherlock runs models trained in Teledyne DALSA's Astrocyte on CPU or GPU.7 Zebra's Design Assistant has classification and object detection steps in a separately licensed classification package.6 LandingLens lets you run in the cloud, in a Docker container or on its LandingEdge app, and only the Docker and LandingEdge options keep inspecting when the plant's internet goes down.10 For an inline reject decision, that last point isn't optional.

We'd pick the AI platform last, not first. Decide whether you need deep learning at all, which our guide on deep learning vs rule-based vision walks through, and then pick the tool that sits in the platform you already chose. Bolting a separate AI stack onto a station adds a second vendor, a second license and a second thing to validate. Our post on AI visual inspection on a real line covers how to pilot one honestly.

Does machine vision software lock you into one camera brand?

Some does. Keyence software runs on Keyence controllers and cameras. Sherlock, VisionPro and pylon vTools center on their own vendor's cameras but support third-party GigE Vision or GenICam cameras to varying degrees. HALCON, MERLIC and Zebra's tools support a broad range of GenICam cameras, and OpenCV works with anything you can write a driver for.

The standards doing the work here are GenICam, GigE Vision and USB3 Vision. MERLIC, for example, uses HALCON's acquisition interfaces for GenICam GenTL, GigE Vision and USB3 Vision. Sherlock supports many third-party GigE cameras through GenICam but not third-party USB cameras directly.7 Design Assistant runs on any PC with CoaXPress, GigE Vision or USB3 Vision cameras.6

Is lock-in bad? Not always. A closed Keyence system means one vendor owns every problem, and their sales engineers will often come set it up. That's worth real money to a plant without a vision engineer. Lock-in hurts when you need a camera the vendor doesn't make, like a specific line-scan or SWIR camera, or when a model goes end-of-life and the replacement forces a re-validation. If you expect to standardize across many stations, an open platform protects you more. Our smart camera vs PC-based vision comparison goes deeper on that architecture choice, and Cognex vs Zebra compares two of the bigger open-ish ecosystems.

How do you choose machine vision software?

Start with who will maintain the inspection, then check that the platform supports your cameras, your PLC protocol and the deep learning you need, and price the runtime licenses for every station and spare. Finally, run your hardest real parts through a trial license before you commit.

Fig. 04Choosing machine vision software · start with the person

Choose machine vision software by its future owner, then by its features. This is the order we use when we scope a station. The examples in each box are typical, not exclusive; Sherlock, for instance, sits comfortably in either of the first two columns.

Once you've got a shortlist of two, here's the test that actually tells them apart.

Collect your worst real parts, including the borderline ones your inspectors argue about, and a few dozen good ones from different lots.
Get a trial license for each platform. Basler gives 180 days on vTools,8 and most others offer time-limited trials on request.
Have the person who'll own the job build it, with the vendor watching but not driving.
Wire it to a PLC, even a bench one, and time the full trigger-to-result cycle.
Make a change your plant will really make, like a new part number, and see how long it takes and who had to help.

Step five is where most decisions get made. Image quality still sets the ceiling for any software, so get the optics right first with our machine vision lighting guide and the camera resolution calculator. And price the whole station, not just licenses. Our breakdown of machine vision system cost shows where software sits in the stack, and the Cognex and Keyence alternatives post covers the hardware side of the same vendors. If you'd rather see these platforms working on real applications first, the machine vision library has worked examples from surface defect inspection to vision-guided robotics.

How we compared

Licensing, deep learning and camera facts come from each vendor's own product pages and documentation, checked in October 2026, and they're cited inline. Placements in the figures and the "best fit" column are our judgment from commissioning stations on these platforms. Vendors change editions often, so confirm the current license terms in writing before you buy.

Frequently asked questions

Is HALCON free?

No. HALCON is commercial software, and MVTec quotes prices individually based on edition, license type and project. You can get a free trial to evaluate it on your own images. If you want something free to experiment with, OpenCV is open source, and Zebra offers a free Lite subset of its Aurora Imaging Library alongside a trial of the full version.

What programming language is used for machine vision?

Mostly C++, C# and Python. HALCON has its own HDevelop language and exports to C, C++, C# and Python. VisionPro is built on .NET, so C# is the usual choice. Zebra's Aurora Imaging Library offers C, C++, C# and Python. OpenCV is C++ at its core with excellent Python bindings. No-code tools hide the language entirely.

Can you use Cognex VisionPro with non-Cognex cameras?

Usually, yes, for GigE Vision cameras. VisionPro sells camera-count licenses for GigE cameras, such as up to two cameras or unlimited, and supports GigE Vision devices as well as Cognex frame grabbers. Support for a specific third-party model can vary, so test the exact camera with your VisionPro version before you order a batch of them.

Do you need a dongle to run machine vision software?

Often, but not always. Cognex VisionPro uses a USB security key. Basler's vTools and Zebra's Design Assistant can use a USB dongle or a license locked to the PC. Sherlock licenses run on a Windows PC, and OpenCV needs no license at all. A movable dongle makes swapping in a spare PC much easier, which is why we like having one per station.

Is MERLIC or HALCON better for beginners?

MERLIC. It's MVTec's no-code product, built so people can set up inspections directly on the image without programming, and it shares HALCON's underlying tools. HALCON is the better choice when you have an engineer who'll write code and needs algorithms or deployment targets MERLIC doesn't cover. Many teams start in MERLIC and add HALCON-built custom tools later.

Can machine vision software run on Linux?

Some can. HALCON runs on Windows and Linux, including Arm-based Linux boards. LandingLens's LandingEdge runs on Linux or Windows, and its Docker option runs anywhere Docker does. OpenCV runs on almost anything. Sherlock is Windows x64 only, and Zebra's Design Assistant IDE needs 64-bit Windows. Check the runtime platform, not just the development one.

How much does machine vision software cost?

It ranges from free (OpenCV) to several thousand dollars per seat, and most vendors don't publish list prices. Expect a one-time development license plus a runtime license for each station, with deep learning often priced separately. Basler sells individual vTools, which makes small jobs cheap to license. Ask for a quote that covers every station and spare.

Machine vision software compared: pick the platform your team can still edit in year two.