Prevent product recalls with machine vision inspection points

A recall turns a quality problem into a logistics problem: every shipped unit becomes suspect, and the cost stops being one part's price and starts being retrieval, replacement and lost trust. The fix is placement, not just better cameras. Machine vision prevents product recalls when the inspection points sit at the four stages where a defect gets expensive — incoming parts, in-process assembly, final defect detection, and code or lot verification. This checklist gives you each gate's items and the reason they matter, then sizes one camera end to end with the 3-pixel rule, so you can repeat the arithmetic for your own smallest critical defect. It is written for quality and process engineers planning a line or retrofitting gates into an existing one.

Decide by stage: the four gates that stop a recall

Put one camera gate at each stage below, in this order of value. The same defect costs the part's price at receiving; the part plus the labor built onto it after assembly; and, once shipped, retrieval and replacement across the whole suspect population. Each gate catches what the previous one cannot see, which is why the order matters more than the resolution of any single camera.

Gate 1 and Gate 3 run on area scan cameras, Gate 2 on smart cameras, and Gate 4 on fixed code readers.

Gate 1 checklist: incoming parts and materials

• Inspect at the first unpackaged touchpoint. Packaging hides the surfaces that matter. Photograph the faces that will be assembled, not the box, because transit damage and supplier mix-ups are cheapest to reject before they enter good stock.
• Compare every lot against a golden reference. A stored reference image set for each approved part lets the gate reject a mixed or wrong-revision part before it joins the line. Part identification by dimensions shows the measurement approach.
• Size this gate for lot-level defects. A wrong part and gross damage separate at far lower resolution than the final gate's smallest defect, so a modest sensor with the right optics is enough here; save the pixel budget for Gate 3.
• Log the lot ID with every image. If a supplier lot later appears in a field failure, the archive proves what arrived — the difference between a supplier claim and an internal write-off.

Gate 2 checklist: in-process assembly verification

• Gate every irreversible step. A pressed-in insert, a welded connector, a glued joint: once the next step buries the interface, no downstream camera can check it. Mount the camera immediately after the step that closes it.
• Verify before value is added. Check the insert before overmolding and the connector before potting. The insert inspection approach for injection molding shows a camera-and-light layout that confirms inserts while they are still visible.
• Keep one camera per gate, on-device. A dedicated smart camera carries its own light and optics, so the verdict does not drift when someone moves a fixture. The MV-SC5060XM-08M-WBN pairs a 6 MP mono sensor with an 8 mm M12 lens, which suits fixed-standoff presence and orientation checks:

• Write a binary verdict per unit. Pass or fail plus the failing feature name, logged with a timestamp. SMT assembly verification at 3.3 boards per minute is the same pattern at line speed: one camera, one question, one verdict per station.

Gate 3 checklist: final defect detection

• Start from the smallest critical defect in millimeters. Take the number from the control plan or the customer drawing; it sizes the sensor, the lens and the light. Then apply the 3-pixel rule: the smallest defect must span at least 3 pixels, or its contrast spreads into neighboring pixels and disappears. Camera resolution for defect detection works this rule out in detail.
• Buy contrast before pixels. A scratch on a textured face needs the right light direction before it needs more sensor; lighting for defect detection on webs and parts matches light geometry to the surface.
• Freeze the motion, then judge the image. Keep blur under one pixel: the exposure limit is pixel scale divided by line speed. At 300 mm/s and 0.033 mm per pixel that is about 110 µs, and the strobe pulse must cover the exposure — this strobe sizing method applies to any strobed gate. Moving, strobed parts want a global shutter sensor; for those lines the MV-CH120-11UC combines a 12 MP global shutter with USB3 transfer in a high-speed version.

• Make the verdict physical. A fail that does not remove the part did not happen: wire the camera verdict to a reject mechanism with its own part-removed sensor, and treat a reject that times out as a fault, not a pass.

Gate 4 checklist: code and lot verification

• Read and grade every code, not a sample. The code gate is what shrinks a recall's scope, and a marginal print is a future unread. Date code print quality inspection covers verifying the string and grading the print, so bad prints reprint instead of shipping.
• Check that the code belongs to the product. A perfect read of the wrong label is itself a recall trigger; match the decoded content against the order or recipe active at that station. Barcode reader verification for label inspection is the reference set-up.
• Set a confidence threshold on every decode. A decode below threshold sends the unit to re-check, never silently through. Barcode read confidence explains how to judge a decode.
• Archive code, verdict and timestamp. Mapping a field failure to a production window is what turns a months-long pull into a one-shift pull. Two reader classes cover most code gates:

Worked example: a 0.1 mm scratch on a 131 mm housing face

The gate is final inspection of an anodized aluminum housing, customer-visible face. The control plan calls any scratch 0.1 mm or longer on that face critical (a typical case — substitute your own control-plan value). The face is 131 mm wide and the camera photographs it in one view, part stopped on a fixture.

Inputs: smallest critical defect 0.1 mm, from the control plan; one view of 131 mm field width (face plus fit margin); a 12 MP camera; nominal width of the 1/1.7" format, about 7.4 mm; working distance about 150 mm, chosen to clear the fixture hardware. The LS-CMR-0166 supplies the array for this example, and its data sheet prints the resolution as 4000 × 3000.

The rule in words: a defect needs 3 pixels across it, so the pixel scale must stay at or below one third of the defect size. That sets the widest field the sensor can cover; the lens focal length then follows from the magnification the field demands.

Arithmetic. Pixel scale: 0.1 mm ÷ 3 = 0.0333 mm per pixel. Field limit: 4000 px × 0.0333 mm/px = 133 mm, so the 131 mm field fits with about 1.5% margin. Magnification: m = 7.4 mm ÷ 133 mm = 0.056. Focal length at a 150 mm working distance: f = 150 mm × 0.056 ÷ 1.056 = 7.9 mm, and the nearest stock value is 8 mm. Check with 8 mm: m = 8 ÷ (150 − 8) = 0.0563, field = 7.4 mm ÷ 0.0563 = 131 mm, pixel scale = 131 mm ÷ 4000 px = 0.0328 mm/px, and the scratch spans 0.1 ÷ 0.0328 = 3.05 px — just past the rule, so hold the field at 131 mm and do not widen it.

Decision. With the field at or below 133 mm, a 12 MP camera with an 8 mm lens at 150 mm working distance detects 0.1 mm scratches in one view. If the face is wider, split it into views or add horizontal pixels: the rule needs FOV ÷ 0.0333 mm/px pixels across, and the camera resolution calculator converts your field directly:

The LS-CMR-0166 is a 12 MP USB3 area scan camera with a 4000 × 3000 resolution on a 1/1.7" rolling-shutter sensor. With the part stopped, that rolling shutter photographs cleanly; moving gates take the global-shutter option from Gate 3.

When the usual answer is wrong

• "Add megapixels." Extra pixels deliver only when the optics resolve them and the light separates defect from texture; on a textured part under unchanged light, a bigger sensor yields sharper noise, not more detections. Fix light geometry first, then re-check pixel scale.
• "One camera can see the whole part." Widen the 131 mm field to 300 mm and the same 4000-pixel sensor runs at 300 ÷ 4000 = 0.075 mm per pixel, so the 0.1 mm scratch spans 1.3 px — below the rule. Wide coverage and small defects need multiple views or a staged gate.
• "We already audit a sample." An audit estimates the defect rate; it does not stop the unit that carries the defect. Gates inspect every unit; keep the audit to discover defect classes the recipes do not yet know.
• "AI will clean it up." Models inherit the optics: adding AI surface defect inspection to existing cameras works when the images already contain the defect.

Before you order: five checks

1. Smallest critical defect in mm, per surface — from the control plan; every other number follows from it.
2. Field width and view count per gate — one view per region that must meet the pixel rule.
3. Pixel scale at or below defect ÷ 3 in every view — compute it before quoting optics. Matching sensor, lens and software to a parts-inspection task is walked through in this parts inspection system selection guide.
4. Line speed and shutter type — the exposure budget is pixel scale ÷ speed; strobed motion wants a global shutter.
5. Verdict wiring — reject mechanism, part-removed sensor, and a logged pass/fail per unit.

Match hardware to these numbers in the camera finder or configure a full system; more applications guides cover the individual gates, and the contact page reaches the applications team with your control plan.