Plan your first vision inspection station: start small and scale

Most small plants do not fail at vision because the technology is hard. They fail by buying a system before proving a check. The workable path is smaller: one station, one camera, one check — label presence, part presence or a printed code — run beside the line until it holds its numbers, then copied to the next checkpoint. This guide walks that path with the arithmetic attached: field of view, focal length, working distance, pixels per millimetre, exposure and depth of field, each worked on a 100 mm example. Still deciding whether to automate the check at all? Start with when manual visual inspection beats automation. For component trade-offs in general, see machine vision system selection by application requirements; this page is the phased plan for the first station and the ones after it.

The first station is one check, not a system

Buy one station. A station is one camera at one position doing one check, and the first one should be the defect that costs you most per month. The default build is an all-in-one smart camera — sensor, lens, processing and I/O in one housing — plus a diffused ring light where the camera has none, mounted over a fixture point. Set-up happens on a laptop; the camera then runs standalone and hands a pass/fail to the PLC.

The discipline that makes scaling repeatable and keeps cost per station down: prove the check at one point before ordering the second camera. Plants that start with one proven station keep the same part number everywhere. Plants that buy five cameras on day one end up tuning five stations at once and trusting none of them. Scaling is a copy, and copies only come out clean when the original is boring, well lit and documented.

Pick the check that pays for the station

Rank your defects by monthly cost, then cross out the ones that need a special viewing geometry. What is left is usually one of three:

• Presence and orientation. Is the label on, the cap down, the clip seated? Flat-on view, forgiving lighting, the fastest check to make reliable.
• Print and code. Date codes, lot codes, OCR. Same flat view, but resolution-limited: at a 100 mm field of view you want 5 MP to read 3 to 4 mm characters.
• Surface. Scratches, dents and smears on shiny parts. The hardest lighting problem and the one most likely to eat the pilot. Take it as station two or three, unless it is the defect your customer keeps catching.

The commit for station one: label or part presence on a stopped or slow part, viewed square-on. That single choice removes most of the motion, focus and glare problems in the rest of this guide.

Camera: all-in-one first, your own lens when geometry forces it

Two routes cover almost every first station.

Route one first. The MV-SC3050XC-16M-ATN-H is the shape: a 5 MP color vision sensor with a global shutter, a 1/1.45" sensor format, an M12 lens mount and a Fast Ethernet interface. The MV-SC6050C-08M-XBN carries its own multi-color light source — 5 MP, color, global shutter, GigE, an 8 mm focal length, M12 — which removes the light bracket for close work. The LS-SMC-0063 is the minimum version: 1.6 MP at 60 fps, a 16 mm lens, built-in white light (polarized and transparent), GigE. At the small end, the MV-SCA002M-03S-WBN-SR is a 0.3 MP mono sensor with a 3 mm lens and focus set at 240 mm — a fixed-focus presence check and nothing more.

Take route two when no built-in focal length meets your working distance or field of view. The LS-SMC-0017 is an 8.8 MP smart camera with a 1" sensor format, global shutter and C-mount, so the whole machine vision lenses range bolts on. Fast Ethernet carries one station's pass/fail traffic without effort; step up to GigE when several cameras share a switch or you stream images off-board.

Use rule-based tools — area, edge, blob, template match — for presence and print; reach for learned models only when good parts vary naturally. AI vs rule-based smart camera inspection: how to choose draws that line.

The three route-one candidates side by side.

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

For a first label station, start here.

Field of view and focal length: the 100 mm worked example

Field of view is the part plus margin, nothing more; every extra millimetre of view spends pixels you paid for. Worked target: a carton 80 mm wide, held to ±5 mm per side by a guide, needs 90 mm. Round to 100 mm — 25% margin and round numbers.

Two numbers set the optics: sensor width and focal length. A longer focal length narrows the view and lengthens the working distance; a shorter one widens it. The geometry is the same for every camera — the angle-of-view basics summarize the trade, and lens field of view: why sensor size and focal length set it covers why sensor size is half the equation.

Assume a 1/2"-type sensor, 6.4 mm wide; take the real width from the camera's own data sheet when you size your build.

With 6.4 mm over 100 mm, m = 0.064×. You have about 300 mm of room above the fixture, so f = 300 × 0.064 / 1.064 = 18 mm. The stock focal lengths either side, and the working distance each one lands on:

• 16 mm lens → WD = 16 × 1.064 / 0.064 = 266 mm
• 25 mm lens → WD = 25 × 1.064 / 0.064 = 416 mm
• 35 mm lens → WD = 35 × 1.064 / 0.064 = 582 mm

Decision: with 300 mm of bracket room, order a 16 mm lens and set the camera at 266 mm. Under a 500 mm ceiling, a 25 mm class lens at 416 mm does the same job from higher up. Among 35 mm C-mount options, the LS-FFL-0679 fits: 35 mm focal length, C-mount, 2/3" sensor format, rated 6-10 MP, an 8-11 mm image circle. That circle clears a 6.4 mm sensor width with margin, and at 35 mm the working distance runs out to 582 mm — order that class when the station must stand clear of an operator's reach.

Run your own numbers with the lens focal length calculator; the same arithmetic worked on a packaging line is in blister inspection working distance and field of view formula.

Resolution: what 5 MP buys at 100 mm

Pixels per millimetre is field of view divided by sensor pixels. A 5 MP sensor is about 2,500 pixels across, so 100 / 2,500 = 0.04 mm per pixel. The working rule is three pixels across the smallest feature you must see: 3 × 0.04 = 0.12 mm. A 4 mm date-code character spans 100 pixels — comfortable OCR territory. A 0.3 mm glue smear spans about 8 pixels — a clean blob hit.

The 1.6 MP class is about 1,450 pixels across: 100 / 1,450 = 0.07 mm per pixel, and the three-pixel floor rises to about 0.21 mm. That finds a missing cap and reads a large label, and it will not reliably flag a 0.15 mm print void.

Decision: presence and orientation — the 1.6 MP class is enough at this view. Date codes, small print, anything under about 0.2 mm — take 5 MP. Finer still? Shrink the field of view with a longer lens instead of chasing megapixels; the camera resolution calculator does the division for your numbers.

Lighting: one diffused ring light, close to the part

For a flat part or label, one diffused ring light is enough. Match the emitter to the field of view — a 100 mm view wants a 100 mm class ring — and mount it about one emitter diameter away, 100 to 150 mm above the part. That usually means the light rides its own bracket at the part rather than the camera at 266 mm, and that is correct: light placement follows the part, not the lens.

Put the diffuser on for glossy labels and film; specular hot spots die there. If glare survives, the next step is cross polarization — cross polarization machine vision: reduce surface glare shows the two-polarizer set-up. In ring lights the pattern to look for is the LS-RNG-0143: a 100 mm diameter ring with a diffuser, 30° output, 24 V.

When the belt runs, drive the light as a strobe timed to the trigger rather than continuous-on. What integrated drivers change is explained in integrated strobe drivers in machine vision lights.

Motion and depth of field: 100 µs at f/8

On a moving belt, blur is exposure times speed, and the budget is half a pixel. At 0.04 mm per pixel and a 200 mm/s belt (12 m/min): t = 0.5 × 0.04 / 200 = 0.0001 s, so 100 µs. Set the camera to 100 µs and strobe the light to match; at that exposure a continuous light leaves the image dark and the part smeared.

If the station indexes — the part stops at the fixture for even 200 ms — the constraint disappears: 2 ms exposures and continuous light work. Index where the line allows it. It halves the difficulty of a first station.

Depth of field at 0.064× is short, and the f-number sets it. With c as one pixel, 2.6 µm here:

• f/4 → 2 × 4 × 0.0026 × 1.064 / 0.064² = 5.4 mm
• f/8 → 10.8 mm
• f/11 → 14.9 mm

Decision: fixture the part to ±5 mm and set f/8 — the 10.8 mm of depth clears it. Buy missing depth with guides and stops, not with the aperture: past roughly f/11 on a 2.6 µm pixel, the diffraction disk softens the image faster than the extra depth helps. Aperture f-number and depth of field in machine vision cameras and airy disk size and the diffraction limit cover that trade in depth.

Prove it: a two-week pilot with seeded defects

Order one camera, one light, one bracket, one cable — then hold the second station until this list is done.

1. Build the defect set. 50 known-good parts, 30 carrying the defect — make them by hand if the line will not supply rejects — and 20 borderline parts: light print, off-centre stops, the worst the guide allows.
2. Write the gate before tuning. Every seeded defect caught, and false rejects at or below 1 in 100 across a 500-part run. Written first, or the target drifts to whatever the camera happens to do.
3. Bench, then line. Stabilize field of view, exposure and thresholds on the bench, bolt the station on the line, re-run the same 100 parts.
4. Log every image for two weeks in shadow mode. The camera records; the reject output stays disconnected. Borderline images are where thresholds get tuned.
5. Attack ambient light on purpose. Overhead banks on and off, dock doors open: if the verdicts move, add a shroud. Ambient light variation in pick and place vision: control it lists the checks.
6. Re-run the seeded set after every change — thresholds, light, mounting, firmware. The 30 defects are your regression test.

If the gate holds two weeks at line speed, order the copy. The same lighting-lens-exposure ordering is argued in stable 2D vision inspection systems.

Scale by copying, not by upgrading

• Same camera everywhere. One spare covers every station, one person's training covers every interface, and the seeded-defect method transfers unchanged.
• One recipe per SKU, selected by input. The inspection job travels with the part number; a changeover is a selector, not a retune.
• One pass/fail log. Every station writes its verdict with an image reference to the same store — that is your traceability record and your regression evidence.
• Vary only what geometry forces. A different field of view changes the lens or the camera; a different face of the part is a new station with the same hardware.

Add stations in the order the defect ranking set when you picked the check. By the third station the work is an afternoon: mount, copy the recipe, run the seeded set, walk away.

When a smart camera is the wrong first station

Four cases break the default build, and each has a better first move:

• Transparent or glass parts. Reflected light fails on them; you need backlight or a dome. Start from how to detect small transparent parts with machine vision.
• Height, bend or coplanarity. A 2D image cannot measure what it flattens. The choice is laid out in 3D vs 2D machine vision system: when to choose 3D.
• Guiding a robot. That is localization with calibration and hand-eye offsets, not inspection: 2D vision pick and place: selecting a planar guidance system.
• Continuous web or parts past about a metre at speed. Area cameras run out of pixels and light together; that is line-scan territory.

None of these says wait. Each says the first station still buys one check, proven then copied — with different hardware.

Checklist before the first purchase order

Optics and mechanics:

• Field of view = widest part + stop tolerance + 20-30% margin, written as one number
• Working distance measured on the machine, bracket drawn
• Lens rated sensor format equal to or larger than the camera's
• Part held to ±5 mm of height, or f/11 budgeted with the sharpness cost accepted

Light and image:

• Ring light sized to the field of view, one emitter diameter from the part
• Diffuser on for glossy surfaces; polarizers if glare survives
• Exposure set for line speed — 100 µs at 200 mm/s and 0.04 mm/px — or fixture indexing
• Image logging on from the first part

Electrical and process:

• 24 V at the station for the light; Ethernet run to the camera
• PLC input wired for the fail signal
• Seeded-defect set built before the camera arrives
• Acceptance gate written: seeded defects 100%, false rejects at or below 1%

When the list is done, browse smart cameras or the camera finder, or let the system configurator walk camera, lens and light in order. Send a part photo and the defect you must catch through the contact page and the numbers above can be checked against your real working distance before you order.