Vision-Based Label and Contents Verification

A pack-out camera station exists to stop a wrong label or a wrong inner pack from leaving the line. The check is binary for the operator: the unit is either cleared for seal or it is held. Quality owns the rule set. The camera only applies it.

The label side of the check reads the fields that the customer, carrier, or next process will treat as truth. Typical captures are the 1D or 2D barcode, the human-readable SKU or GTIN, the lot or batch, the expiry or manufacture date when the label carries it, and the quantity or pack-count printed on the outer. Those values are compared to the order line and the expected pack recipe for that SKU, not to a generic “looks like a label” pass.

The contents side of the check confirms that what is in the carton, case, or kit matches that same order line. Depending on the SKU mix, that can mean presence and count of inner cartons, a color or shape signature for a known SKU family, a secondary barcode on a bag or bottle, or a count of units against a taught nest. Mixed-SKU kits need a taught bill of materials and a camera view that can see every pocket. Homogeneous cases can use count plus a single SKU barcode on an inner.

The two checks belong together because a correct inner pack with the previous order’s label, or a correct label over the wrong inners, both fail the same customer. Run them in one station, with one decision, before tape, glue, or stretch wrap commits the unit.

This station is not the same job as inline vision-based defect detection. Defect detection looks for damage, missing components, or process faults on the product. Pack verification looks for identity and completeness against the order. Keep the models, reject reasons, and disposition codes separate so a dented carton is not filed as a lot mismatch.

Trigger the inspection after the packer or robot has placed the last inner and applied the shipping or case label, and before the sealer or case closer. If the label is printed at the station, fire the cameras after print-and-apply confirms a successful apply, not on the print-complete signal alone. A label in the printer is not a label on the pack.

Feed the vision controller the expected identity from the warehouse or manufacturing execution system for that serial, carton ID, or station assignment. Do not teach the camera a “good label” in isolation. The expected barcode, lot, and quantity come from the current order. If the WMS has already allocated a lot, the camera confirms that allocated lot, not any valid lot for the SKU.

Light and fixturing matter more than brand. Specular tape, glossy varnish, and curved bottles all defeat a camera that was commissioned on a flat brown case. Fix the pack in a nest or against a backstop so the label plane and the contents window are repeatable. If operators can rotate a carton 90 degrees, either add a second camera or add a presence sensor that withholds the trigger until orientation is correct.

Vendors such as Cognex, Keyence, and SICK all ship pack-station tools that combine code reading, OCR, and presence. Choose on working distance, code types on your labels, and how the controller exposes pass, fail, and no-read to the PLC. Keep the PLC as the interlock owner. The camera reports a result. The sealer, printer inhibit, and conveyor do not move to seal or discharge unless the PLC has a fresh pass for that unit ID.

When the SKU is regulated or the outer must carry hazard marks, run identity first, then confirm the mark set still matches the classified goods. Pack verification does not replace hazmat regulatory classification. It only confirms that the label and contents still agree with the classification already assigned to that SKU or lot.

A mismatch is not an automatic scrap. The packer still clears mismatches after the station has made the disagreement visible. Typical fails are barcode vs order, lot vs allocation, printed quantity vs inner count, missing inner, extra inner, and label present but unreadable.

Present the fail so the packer can act in one glance: which field failed, expected vs read, and whether the contents view or the label view is at fault. A no-read is not a wrong-lot. Treat it as a separate reason so the first response is clean the window, reseat the pack, or reprint, not open every inner.

The packer’s clearance is a recorded decision, not a bypass button. Require a scan of the packer’s badge or a logged confirm on the HMI, capture a still from both cameras, and write the unit ID, order line, expected values, read values, and reason code to the quality log. If the packer reprints the label, the station must re-inspect the new label and the contents again. Clearing a lot fail by reprinting without a second pass recreates the original risk.

Do not let the packer edit expected lot or quantity on the station to force a pass. If allocation is wrong, the correction belongs in the WMS or MES, then a new expected record is sent to the camera. Station-level edits hide systemic allocation errors and train the line to treat the camera as optional.

Escalation is for repeats and for regulated SKUs. Two consecutive fails on the same SKU, or any fail on a serialized or hazardous SKU, should hold the lane and call quality, not loop the same packer through reprints. The packer remains the person who physically corrects the pack. Quality owns whether the recipe, the print template, or the allocation is wrong.

A camera that is out of calibration is worse than no camera because it stamps a pass on an unproven pack. If calibration, focus, illumination, or trigger health fails, the station goes empty and hold. No unit is sealed. No unit is palletized from that lane. Downstream accumulation may still run if it is physically separate. This cell does not.

Commission a scheduled calibration target: a known label with known codes and a contents nest with known count. The controller compares decode grade, OCR confidence, and pixel position against the last accepted baseline. Fail that check on startup, after a lamp change, after a crash or tool change, and on a timed interval during the shift. Tie the result to the same PLC interlock as a product fail, with a distinct reason so maintenance is not hunting a phantom SKU.

No-trigger, blocked photocell, torn Ethernet, and PLC heartbeat loss are also empty-and-hold. A silent camera is not a pass. If the sealer can still cycle when the vision controller is offline, the interlock is incomplete. Prove the fail-safe on commissioning by powering down the camera and confirming the sealer and discharge stay locked.

When the hold clears, do not release the queue of packs that sat in the nest during the fault. Those units never received a valid inspection. Recirculate or re-present each one. Packs already sealed during a later-discovered calibration drift are a contained quality event: isolate the time window, hold the pallet positions built in that window, and re-inspect or rework before ship. Do not average a calibration fail across a pallet and ship the rest.

A pass means this unit’s label data and visible contents matched the order at the moment of inspection. It does not mean the pallet pattern is legal, the case will cube correctly, or the carrier label will survive the dock. Hand the cleared unit to seal, then to pallet build. Use 3D bin and pallet configuration optimization for orientation, stack stability, and cube. Do not overload the pack camera with pallet geometry.

Keep the inspection image and decode payload with the carton or case ID through palletization. If a customer claims a mislabel, you need the still and the expected vs read record, not only a station counter. Retention follows your quality system and any customer or regulated SKU rules. The image is evidence of the check, not a substitute for the physical pack.