What is a machine vision system and how to choose one

A machine vision system answers a production question with an image: is this part present, is it the right one, is the label straight, where is the hole for the robot. Understanding what a machine vision system is comes down to five building blocks and one decision: where the image processing runs. This guide walks through each block with the numbers that size it, then compares smart cameras, vision controllers and PC-based systems so you can pick the right architecture before you pick a model.

The five building blocks

Every system, from a 0.3 MP presence check to a 16-camera robot-guiding cell, has the same structure:

1. Camera and sensor - converts light into pixels. Sensor format, resolution and shutter type set the ceiling on what the system can see.
2. Lens - frames the field of view and projects it onto the sensor. The lens must cover the sensor format and resolve enough detail.
3. Lighting - makes the feature you care about visible. Poor lighting can defeat an otherwise suitable camera and inspection tool.
4. Processing hardware - a smart camera's on-board processor, a vision controller, or an industrial PC that runs the image processing.
5. Software - turns pixels into results: tools for measurement, code reading, matching or classification, plus the I/O and protocol handling that reports results to a PLC or robot.

The wiring between blocks matters as much as the blocks. Trigger, strobe and result lines must be timed so the image is captured at the right moment; the camera trigger source and GenICam line settings article covers the common failure points on that path.

Cameras: resolution and shutter

Size resolution from the feature, not the part. As an initial detection estimate, allow 3 to 5 pixels across the smallest feature. Measurement requirements need a separate accuracy and uncertainty assessment:

Worked example: a 100 mm field of view with a 0.5 mm feature spans 100 / 0.5 = 200 feature widths. Multiplying by 3 to 5 pixels per feature gives 600 to 1,000 image columns. If instead you measure the feature's width to 0.05 mm, the scale changes and you should work through the camera resolution calculator below and the guidance in choosing camera resolution for dimensional inspection.

Resolution is not the only filter. Global shutter avoids rolling-shutter distortion on moving parts; rolling shutter suits static scenes. Both still require exposure times short enough to limit motion blur. Sensor format must match the lens image circle, and the interface (Fast Ethernet, GigE, USB3) affects cable length and bandwidth. For a deeper walk through the sensor terms, see industrial camera specifications: a buyer's guide.

If the inspection is simple, a vision sensor puts camera and processing in one housing. The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter and an M12 lens mount, connecting over Fast Ethernet.

A GigE smart camera is another option when planning the acquisition interface. The LS-SMC-0063 is a 1.6 MP smart camera running at 60 fps with a 16 mm lens and white light, described with polarized and transparent-light options. Test the available illumination on sample parts before deciding whether separate lighting is needed.

Where resolution must rise, the MV-SC5060XM-08M-WBN is a 6 MP mono GigE smart camera with an 8 mm focal length and an M12 mount in the SC5000X series. Note it uses a rolling shutter, so check motion-related distortion before selecting it for moving parts.

Browse the full range in the smart cameras category or narrow it in the camera finder.

Lenses: one formula decides it

The lens choice follows from the field of view, the working distance and the sensor. Magnification is sensor width divided by field of view. A thin-lens estimate gives a starting focal length; actual working distance references and lens construction affect the final choice:

Example: a 1/2" sensor with an active width of approximately 6.4 mm, a 120 mm field of view and a 300 mm working distance gives m = 6.4/120 = 0.053. Then f = 300 x 0.053 / 1.053, roughly 15 mm, making 16 mm a nearby standard focal length to evaluate. Confirm the actual sensor width rather than relying on the optical-format label alone.

Run your own numbers with the focal length formula above, then check that the lens image circle covers your sensor format; the article on lens sensor format explains why a lens too small for the sensor darkens the corners. Step-by-step selection is in how to choose a machine vision lens focal length, and the lenses category lists the available mounts.

Lighting: where most systems are won or lost

The camera cannot see a feature the lighting does not reveal. Match the technique to the surface: ring or bar lights for general contrast, coaxial light for flat shiny surfaces, dome or cloudy-day illumination for curved and wrinkled parts, backlight for dimensions. Four rules keep it simple:

• Choose the lighting geometry before the light model; a good light in the wrong geometry still fails.
• Consider synchronized strobing when motion requires a short illumination pulse. Stay within the light and controller's documented pulse limits.
• Match wavelength to material contrast and sensor sensitivity. Shorter wavelengths can resolve finer optical detail, but surface response may matter more.
• Shield ambient light; changing overhead illumination can wash out the contrast established during setup.

The lights finder sorts by geometry, color and shape; use the AI versus rule-based smart camera inspection guide to compare inspection approaches.

Processing hardware: three architectures

This is the decision that shapes the project. The three options differ in where the processing runs and who maintains it:

• Smart camera / vision sensor: processing on the device, results over Ethernet or I/O. No separate processing PC is needed, although configuration tools, licenses and update requirements vary by model.
• Vision controller: a compact industrial computer purpose-built for vision. Depending on the model, it may combine camera ports, light outputs and machine I/O. Confirm the required software and industrial protocols before ordering.
• PC-based system: a standard industrial PC with your own acquisition and processing code, maximum flexibility for custom algorithms, deep learning pipelines and databases.

The MV-VC2140-128G40-1T is a vision controller in the VC2100 series with an N-series processor, 8 GB of memory, 128 GB SSD plus a 1 TB HDD, 3-4 LAN ports and 4 built-in light channels. Check camera bandwidth, processing time and light-output compatibility before assigning cameras to it.

Another option, the MV-VC3301H-128G66 in the VC3000 series, has a Core i3 processor, 8 GB memory and 128 GB SSD, with 5+ LAN ports, 4 built-in light channels, PCIe expansion and an included VM6200Pro license.

A good middle position: build each inspection station as a smart camera where one camera suffices, and use a controller where three or more cameras or coordinated lighting make separate devices awkward. If you lean toward a PC and your own code, read camera acquisition software compatibility: integration options first - SDK and driver support is the real cost of that route. The vision controllers category shows port counts and light-channel counts per series, and the system configurator helps organize the component selection.

Which architecture to choose

An independent take on the same decision, written from the controls side, is this guide to selecting a vision system for parts inspection.

Checks before you order

• Field of view and working distance measured on the real part, not from the drawing.
• Required pixels per feature computed, and the camera resolution chosen from that.
• Lens image circle covers the sensor diagonal; the mount and flange distance match.
• Shutter type and exposure time suit part motion; global shutter alone does not remove motion blur.
• Lighting geometry tested on sample parts under production ambient light.
• Result interface agreed with the PLC or robot side: discrete I/O, an industrial Ethernet protocol, serial or a file.
• Mounting and focus locks specified; vibration will move an unsecured focus ring.

When the usual answer is wrong

A single high-resolution camera is often the wrong answer to a multi-part inspection. Two cameras, each with a tight field of view and dedicated lighting, can outperform one camera with a wide field of view and compromises everywhere: 3 pixels per feature provides more sampling than 1.5. Likewise, if the feature only shows under a specific geometry, no amount of camera resolution recovers it - fix the lighting first.

If the vision result must request an ordinary process stop, plan the diagnostics path early; mapping vision diagnostics to PLC status and alarms shows how to surface camera faults where operators can act on them.

Short checklist

• Define the decision the system must make, in one sentence.
• Compute resolution from feature size; compute focal length from field of view and working distance.
• Choose lighting geometry on real parts, with the lights you can actually buy.
• Pick the architecture: one camera and simple checks - smart camera; several cameras and shared lighting - controller; custom code - PC.
• Agree the result interface with the machine builder before commissioning.

Start with the smart cameras or vision controllers categories, or work through the system configurator step by step; our team is available for a review of your application before you order.