Automating manual visual inspection: when one click works

The demo promises one click: aim, shoot, verdict. On a production line the useful question is narrower. Is the defect you catch by eye a stable visual feature a sensor can resolve, and can the part be presented the same way every time? When both hold, the check converts — and on a running line it usually becomes continuous, so the click disappears entirely. This page gives you the decision test, the pixel rule that sizes the camera before anything is bought, a worked example around a 0.5 mm scratch, and the cases where eyes stay in the loop.

Where the click actually goes

Software takes over a manual inspection in three shapes, and in practice they are stages rather than alternatives.

• Operator assist. The operator captures with a click, the software judges the image. The fastest way to pilot, because part handling stays as it is.
• Triggered in-line. A part-present signal fires the camera; operators see only rejects. No clicks at all — this is the end state for nearly every fixtured check.
• Continuous. Every part is imaged as it passes and judged in a continuous camera acquisition loop. Worth it once parts flow faster than an operator can fixture them.

The decision: pilot with a click, deploy without one. A project that stops at operator-assist saves little labour; the consistency gain alone carries it, and routing the verdict to a reject pusher or robot removes the operator from the judgement entirely.

The feasibility test: which defects convert

Split the defect list into three classes.

Converts cleanly:

• Presence, absence, position. Missing screw, cap not seated, label skewed, seal ring in place. Binary geometry with contrast is exactly what vision sensors are built for.
• Surface flaws. Scratches, dents, stains, discolouration, burrs — as long as lighting makes them contrast with the surrounding finish. The recipe changes with material: see lighting for metal surface defect inspection and angle and wavelength choices for plastic parts.
• Dimensional features in the image. Hole diameter, thread presence, slot width. Tight tolerances turn this into metrology — choosing camera resolution for dimensional inspection covers the stricter math.

Does not convert:

• Anything hidden or functional. Leak rate, torque, insulation continuity, flow through a port: no image contains the measurement.
• Genuinely rare defects. A class that appears a few times a year gives you neither a training set nor a stable feature list on day one.
• Designed-in appearance variation. Leather, natural stone, weld seams — rule sets drown in false rejects and the project becomes an AI project (adding AI surface defect inspection to existing cameras).

Write the list down with the exact criterion the operator applies today. Failed pilots usually fail here first: poor finish is a feeling, not yet a visual feature.

The pixel rule that sizes the camera

Resolution closes on paper before any hardware is ordered. Two thresholds, both standard in resolution planning, decide it:

• 4 pixels across the narrowest dimension of the smallest rejectable defect — the working minimum for detection. The standard planning example, a 0.5 mm defect, is sized at four pixels across for a first cut.
• 10 to 20 pixels across the same defect when it must be classified or measured, because the extra pixels let software tell a scratch from a shadow.

The chain of formulas:

formula
pixel size = field of view / sensor pixels across the field
pixels on defect = defect width / pixel size
field of view = part width + placement margin

The field of view carries the part plus its positioning spread — budget 10-20% of part width in a tooled fixture, more on a loose conveyor.

Run your own numbers:

If the count lands below 4, the fix is optics, not software: shorten the field of view, tighten the fixture, or step up sensor resolution. The same sizing method drives choosing resolution for defect detection, and the deployment order — lighting, lens and exposure before algorithms — decides whether the pixels you bought stay usable on the line.

Worked example: a 0.5 mm scratch on a 40 mm cap

Inputs for a typical closure-inspection case — rerun the chain with your own numbers:

• Smallest rejectable defect: 0.5 mm wide scratch, the planning-example size and a realistic polish flaw on a moulded cap.
• Field of view: 44 mm — the 40 mm part plus about 10% placement margin in a tooled nest.
• Candidate sensors: a 1.6 MP vision-sensor array (1440 pixels across) and a 5 MP array (2448 × 2048), the two common formats.

Rule in words: divide the field of view by the pixel count to get mm per pixel, then divide the defect width by that.

Arithmetic:

• 1.6 MP: 44 mm / 1440 px = 0.0306 mm/px. Scratch: 0.5 mm / 0.0306 mm/px = 16.4 px.
• 5 MP: 44 mm / 2448 px = 0.0180 mm/px. Scratch: 0.5 mm / 0.0180 mm/px = 27.8 px.
• Detection floor for this field: 44 mm × 4 / 0.5 mm = 352 px across, so even a 640-pixel sensor resolves the scratch (44 / 640 = 0.069 mm/px → 7.3 px) — detection only, not classification.

Result and decision: at 16 px across the scratch, the 1.6 MP array clears the 10 px classification bar, so a vision sensor closes this check. The 5 MP array earns its keep when the same field must also catch a 0.2 mm flaw: 0.2 / 0.0306 = 6.5 px on the 1.6 MP array (detection only) against 11.1 px on the 5 MP array (classification). Below 4 px per defect no algorithm rescues the image — change the optics or the field.

Cycle time is not the constraint: one part per second needs one judged image per second, and vision sensors in this class read tens of frames per second.

From click to no click: trigger, light, fixture

An unattended check holds only if all three of these are fixed:

1. Trigger. A photoelectric sensor or PLC part-present signal starts the exposure, so every image catches the part at the same spot. Soft-triggering on image content reintroduces the variability you just removed.
2. Fixed light. An LED ring or dome strobed only during the exposure freezes motion and shuts out ambient swings. Shroud the station: sunlight crossing a conveyor is the classic false-reject generator (controlling ambient light variation).
3. Repeatable presentation. A nest or fixture, or locate-then-inspect software that finds the part and inspects relative to it. The 10-20% field-of-view margin in the pixel rule is what absorbs this tolerance.

If the part moves during exposure, a global shutter beats rolling; in a static nest either works.

Where eyes stay in the loop

Automation takes the routine judgement, not the whole quality role:

• Hidden and functional checks stay manual or move to dedicated test equipment — torque, leak and continuity never had an image to begin with.
• Audits replace sampling. An in-line camera judges 100% of parts, so sampling discipline moves to auditing the auditor: formal switching rules such as those under BS 6001 govern when an audit tightens on evidence of trouble and relaxes as parts pass — see BS 6001 switching rules for automated sampling inspection.
• Novel defects are found by people first. For the first months, review the reject bin daily; the system only knows the defects you taught it.
• Set the bar before the pilot. Agree the escape tolerance, then measure the camera call-by-call against the inspector. Machine consistency is a property of fixed lighting and fixture, and it is measured, never assumed.

Choosing the hardware tier

Hardware comes in three tiers — vision sensors and smart cameras that judge on the device, and area scan cameras feeding a PC — and the check picks the tier.

For a single presence or surface check, a vision sensor is the shortest path: the HK-TD032-12-WBN pairs a 1.6 MP mono sensor with a global shutter and an M12 mount.

When the same station needs colour or extra resolution margin, the MV-SC3050XC-16M-ATN-H keeps the all-in-one form as a 5 MP color vision sensor with a global shutter, Fast Ethernet output and an M12 mount.

For several features per part or an on-device AI model, a smart camera keeps the judgement at the station: the MV-SC6050C-08M-XBN is a 5 MP color smart camera with a global shutter, an 8 mm M12 lens, a built-in multi-color light source and a GigE interface.

When checks multiply across stations or processing grows heavy, split sensing from computing and choose the lens freely on a C-mount: the MV-CU020-90GC is a 2.5 MP color camera with a global shutter over GigE.

Fine defects across a wide field pull the other lever, resolution: the MV-CS200-10UM is a 20 MP mono camera on a 1-inch sensor with USB3 output, C-mount.

Checks before you fund the pilot

Five checks separate funded projects from shelfware:

1. Golden parts, real defects. Collect 20 good parts and 20 with each rejectable defect from the line — CAD mockups do not carry the surface texture the algorithm will actually see.
2. Measure the spread. Sit 50 parts in the nest and log the actual position scatter; recompute the field of view from the measurement, not the drawing.
3. Trial the light at the station, ambient included. A recipe that works in a dark lab can fail under shop skylights.
4. Run blind for two weeks. The camera judges silently alongside the inspector; log agreement call by call against the escape tolerance you set beforehand.
5. Plan the falsing investigation. When good parts start rejecting, the cause is optics drift, light ageing or part drift — camera fault isolation for repeatable unit-specific failures gives the isolation sequence.

When the usual answer is wrong

Three cases break the standard path:

• The defect is a depth, not a brightness. A dent or a step hides from flat lighting; a 3D laser line profiler reads the geometry directly instead of inferring it from shading.
• The dataset is tiny and appearance varies. Deep-learning classifiers want hundreds of examples per class; with a handful, start rule-based or anomaly-based — the rule-based versus AI choice is a sequencing decision, not a fashion one.
• The check is really a test. If the operator presses, flexes or listens today, the inspection is functional: budget dedicated test equipment, not a camera.

If the feasibility test passes, size the check with the worked example, then shortlist hardware in the camera finder or assemble the stack in the camera configurator — and if a review of your golden parts would settle the decision, talk to the applications team.