Choosing a machine vision system standard for your line

Here is the choice, up front: put a vision sensor or smart camera at every station that decides pass/fail on its own, put one vision controller behind four or more cameras that share a decision or an image archive, and standardize the whole plant on GigE Vision cameras under GenICam — that is the standard that keeps cameras and hosts interchangeable. For your three products the committed answer is per-station smart cameras on the plastic and aluminum tube lines, and one 4-port controller with four 2.5 MP GigE cameras on the tin-can line: at 480 cans per minute each camera draws 20 MB/s on a port you plan at 100 MB/s. You inspect plastic tubes, aluminum tubes and tin cans with boxes from several vendors, and every new line adds another integration to maintain; choosing one standard on purpose ends that. Those two choices — where images are processed, and which interface ecosystem the cameras speak — outlast any brand, because a GigE Vision, GenICam-compliant camera from anyone talks to a compliant host. Below are the decision rules, the bandwidth arithmetic that binds them, the worked example behind the can-line answer, and the decision for each of your products. If the project is not scoped yet, start with scoping a machine vision system before you call an integrator; for the basics, what a machine vision system is and how to choose one covers the ground.

The decision first

Commit to the architecture before opening a catalog. If every station can accept or reject on its own, and nobody programs vision on site, put a vision sensor or smart camera at each station. If four or more cameras share processing, one reject decision or one image archive, put them on one vision controller. Plants making tubes and cans usually end up hybrid: sensors on the simple stations, one controller for the high-resolution and cross-line checks — and that is the right answer, not a compromise.

Then commit to the ecosystem before comparing vendors. For line cameras the default is GigE Vision under GenICam; USB3 Vision inside a cabinet; plain Fast Ethernet only for results from fixed-function sensors. The reason is economics: machine vision connectivity standards keep cameras and hosts interchangeable, so an end-of-life camera becomes an evening swap instead of a redesign.

What actually drives the choice

Three numbers pick the architecture. Station count: one to three independent stations favor a device per station; four or more cameras sharing logic favor a controller, because one controller with four camera ports replaces four separate computers. Staffing: a vision sensor is configured by a trained operator, a smart camera is programmed by one trained engineer, a controller-based system wants an engineer or your integrator on call — and as an integrator yourself, you already have that person. Line rate: each camera must capture one frame per part, so the frame rate you need is simply parts per second.

Resolution sets a second floor: put at least 3 pixels across the smallest feature you must find. The LS-CMR-0166 is a 12 MP USB3 camera that prints 4000 × 3000 pixels at 30 fps. Spread a 300 mm field of view over those 4000 pixels and one pixel covers 300 ÷ 4000 = 0.075 mm; a 0.3 mm pinhole then spans 4 pixels, which detects stably. Halve the field of view and one pixel covers 0.038 mm.

Pick the interface ecosystem before the vendor

The link is the contract between camera and host. A GigE Vision link runs 1 Gbit/s — 125 MB/s raw — and you plan on 100 MB/s of image data after protocol overhead; it reaches 100 m on Cat5e/Cat6 and can carry camera power over PoE. USB3 Vision moves more, around 350 MB/s to plan on, but only a few metres without active or optical cabling. Fast Ethernet is 100 Mbit/s — a results link for pass/fail strings, never for full-rate images.

The table turns resolution into rate in the open: a 2.5 MP camera tops out near 40 frames per second on one GigE port — 2,400 parts per minute — while a 20 MP camera on the same link reaches about 5 frames per second, or 300 parts per minute. You buy speed and detail with either link bandwidth or on-board processing. Two deeper comparisons: the most reliable camera interface for production lines and why GigE cameras still win on industrial interfaces.

The three architectures compared

Vision sensors: configured, not programmed

A vision sensor ships fixed checks — presence, edges, code reading — set up in a browser screen and left alone. The MV-SC3050XC-16M-ATN-H is a 5 MP color sensor on Fast Ethernet. The MV-SCA002M-03S-WBN-SR is a 0.3 MP mono unit with a 3 mm lens and a fixed focus distance of 240 mm. Neither streams full images at line rate, which is exactly why the wiring stays small.

Smart cameras: your logic, light included

A smart camera runs your inspection logic where the image is made, with the illuminator built in. The MV-SC6050C-08M-XBN is a 5 MP color smart camera with an 8 mm M12 lens and a multi-color light source, so one part number mounts, lights and decides at one station. When the defect is appearance-based rather than dimensional, read AI versus rule-based smart camera inspection before you commit the station's method.

Controller with area scan cameras: shared processing

A controller-based system centralizes: area scan cameras on GigE or USB3 links, software on a fanless vision controller in the cabinet. The MV-VC3301H-128G66 carries 5+ LAN ports, 3-4 USB 3 ports and 4 built-in light channels with a Core i3 and 8 GB of memory; the MV-VC5525-256G70-NN adds PoE camera ports, 8 light channels and 32 GB. Any industrial mini PC can host the software, but a vision controller adds camera ports, strobe channels and fanless duty in one housing. Wiring cameras to it is standard Ethernet work — connect a GenICam camera to an industrial controller shows the pattern.

For the cameras, settle the interface first — GigE PoE versus USB3 variants — then pick per station by resolution and shutter:

The first is the high-rate line camera; the second is the detail camera for sub-0.1 mm features.

Worked example: four stations on a tin-can line at 480 cans per minute

Inputs, each with its source:

• Line rate: 480 cans per minute — the planning number for this example; substitute your own rate in the same arithmetic.
• Stations: four — end seam, side seam, label print, fill level — grouped to match a controller with 4 camera GigE ports, 1 configuration port, 1 communication port and 4 built-in light channels, the MV-SVC2506-16G60-128G.
• Camera: the MV-CU020-90GC above — 2.5 MP, global shutter, GigE.
• Frame format: one frame per can, 8-bit Bayer, which is 1 byte per pixel.
• Link budget: 1 Gbit/s per port, plan on 100 MB/s usable.

Rule in words: a station fits on one GigE port while frame rate × frame size stays under 100 MB/s, and one frame per part converts that ceiling directly into parts per minute.

Arithmetic, written out:

480 cans/min ÷ 60 = 8 frames/s per camera. 2.5 MP × 1 byte/pixel = 2.5 MB per frame. 8 frames/s × 2.5 MB = 20 MB/s per camera. 20 MB/s ÷ 100 MB/s = 20% of one port — fivefold headroom for exposure overlap and retries. 4 cameras × 20 MB/s = 80 MB/s aggregate, each camera on its own port. Ceiling per camera: 100 MB/s ÷ 2.5 MB = 40 frames/s, and 40 × 60 = 2,400 cans per minute.

Decision: at 480 cans per minute, order one 4-port controller, four 2.5 MP GigE cameras and four lights on the built-in channels; every port runs at a fifth of its budget. Above 2,400 cans per minute at 2.5 MP, that station leaves the shared-controller pattern — give it a smart camera that decides on board and ships results, or a 10GigE link (5GBASE-T bandwidth, cables and distance covers both).

The decision for each product you run

Plastic tubes. Label presence, cap seat and fill level on clear plastic want transmitted and polarized light right at the part, and every check is local — per-station smart cameras win. The LS-SMC-0063 is a 1.6 MP smart camera at 60 fps with a 16 mm lens and built-in white light in polarized and transparent modes. At one frame per tube, 60 fps covers 3,600 tubes per minute, so the camera is never the bottleneck on the line.

Aluminum tubes. Print registration, crimp geometry and color checks stay on smart cameras with a multi-color light source; laquer pinholes and other sub-0.1 mm defects go to one controller-fed 12 MP camera, where the 0.075 mm-per-pixel sizing above applies directly.

Tin cans. Seam and label checks at line rate follow the worked example: four 2.5 MP global-shutter cameras on one 4-port controller. Global shutter is mandatory wherever the part moves.

Checks before you order

1. Compliance in writing: every camera and host you buy claims GigE Vision or USB3 Vision and GenICam. That claim is what makes a camera a commodity you can swap in year five.
2. Count light channels against lights: a controller with 4 built-in channels drives four strobes; a line with eight lights takes the 8-channel model.
3. Add the cameras' wattage and compare the total with the controller's PoE budget before you order PoE cameras.
4. Measure cable runs against the 100 m GigE limit now; a longer run gets a switch or fiber in the plan, not on install day.
5. Buy one spare camera per model and keep it on your shelf. Under a shared standard the swap is a settings file, not a porting job.
6. Draw the reject path early: strobe, trigger and pass/fail lines all land on the controller, so the PLC wiring list stays one page.

When the usual answer is wrong

A smart camera on every station fails when stations must share a decision. A can rejected by two stations should be rejected once, with both images in one archive; cross-station logic and shared evidence are a controller's job.

One central PC for the whole plant fails on reach and blast radius. A 100 m run from the farthest line already needs a switch, and one crash stops every line at once. Segment by line, not by plant.

Maximum resolution everywhere fails on rate. Twenty megapixels on GigE is about 5 frames per second — 300 parts per minute; a camera that misses parts is worse than a camera with coarser pixels. Buy resolution per station.

Fast Ethernet sensors fail you at audit time. They ship results, not images; if a customer disputes a reject, that station needs a camera that can hand over the frame.

A short checklist, in order

• Architecture per station: local decision → device at the station; shared decision or archive → controller.
• Ecosystem per plant: GigE Vision with GenICam as the default, USB3 in cabinets, results-only links for sensors.
• Arithmetic per station: frame rate × frame size against the 100 MB/s per GigE port.
• Staffing named on paper: who configures, who programs, who is on call.
• Spares and standards: one spare per camera model, one configuration environment for everything.
• Cable runs and PoE budgets measured before the purchase order.

Put your own numbers in and the architecture falls out. To move, run the camera finder, walk the system configurator, or for a plant-wide standard across several lines, send us the line rates and station list through the contact page and we will lay out the architecture with you.