Resources/Packaging Vision Inspection: From Visible Defects to Owned Actions
Operational Excellence

Packaging Vision Inspection: From Visible Defects to Owned Actions

Plan packaging inspection around visible features, package-specific validation and named owners. Use worked hold and maintenance records to connect each finding to an action.

12 min read
By Monitory team

Use computer vision to check visible packaging features, retain the image with its lot and recipe, and route pass, reject and unknown outcomes to named owners. A seal-area image can support a quality hold or process investigation. An integrity decision needs a validated relationship to the package-specific reference test.

For example, a camera may show a wrinkle crossing a pouch seal band while leaving the strength of that seal unresolved. Define those as separate requirements before choosing a sensor. The guides to lighting and surface selection, defect size and camera resolution and thresholds and validation cover the station design; this article applies that work to packaging.

Write the Requirement Before You Buy the Camera

Start with a sentence a technician can read off a station placard. Not "detect seal defects." Something closer to: flag pouches where product, powder or liquid residue is visible inside the declared seal band, where a wrinkle crosses the full width of the seal band, or where the closure feature is absent.

Each of those is an inspectable visible feature. Each maps to a geometry you can express in pixels once you know the seal band width. Take that width from the packaging drawing or the sealing specification your packaging engineer maintains. Do not invent it. The tolerance belongs to the product, not to the camera.

A published industrial case on thermoforming food packages is useful here because the authors designed image acquisition and the sealing and closure checks around the specific features of that package, including the complication of printed and transparent materials [1]. That is why a 2021 case still has a place beside newer work: the measurement question was defined before the method.

Patterns worth watching for when the requirement is written loosely:

  • Defect class drift. The model is trained on whichever rejects happened to be available, so "seal defect" expands to include print smear and film haze.
  • Unstated geometry. The seal band width and minimum feature size are never recorded, so nobody can check whether the optics resolve what the station claims to inspect.
  • Borrowed conclusions. A visible-condition check gets reported as an integrity check, and the gap surfaces during an audit or a customer complaint.
  • Unowned outcomes. The station produces a reject signal with no named person responsible for confirming it.

Write the requirement, have the packaging engineer and the quality manager agree the wording, and keep it at the station.

Package featureImaging candidate to evaluateBlind spotHold / review owner
Pouch seal band appearanceDiffuse backlight or coaxial top light; test edge contrast on your filmDoes not measure integrity; a clean-looking seal may be cold or channeledQuality technician raises QA hold; packaging engineer reviews
Product or residue inside seal bandTransmitted light for clear film; reflected light for exposed residuePrinted or opaque film can hide intrusion; tinted product can read as shadowQuality technician; line lead reviews fill timing
Wrinkle crossing seal bandLow-angle grazing light to raise surface relief; verify on samplesShallow channel leaks may produce no visible relief; glare can read as a wrinkleQuality technician; maintenance request if the pattern recurs
Missing closure (zipper, cap, spout)Top-down presence check with fixed framingA partially engaged closure can appear present; reflective caps can saturateLine lead confirms; maintenance request if the applicator repeats
Tray packout arrangementOverhead area or count check with controlled shadowShrink film glare can read as present product; stacked items occludeLine lead; QA hold if the count stays unresolved
Seal thermal signatureInfrared imaging with controlled timing; validate acceptance bandNot an integrity test; ambient and dwell variation shift the signaturePackaging engineer defines acceptance; quality owns disposition

Every imaging entry in that table is a candidate to test on your own film and product, not a working recipe. Whether a feature appears bright or dark depends on the film, the print coverage, the product and the geometry. Run the samples before writing the recipe.

The Transparent Pouch Problem

Transparent film lets a camera see several layers at once. The inspection must distinguish the relevant seal feature from product behind the film and reflections on its surface.

Transmitted light through clear film tends to show product silhouette and film wrinkle at the same time, with limited separation between them. The 2025 MVTec AD 2 dataset was built around conditions of this kind: transparent and overlapping objects, wide normal variation, very small defects and test lighting the model never saw during training [2]. The point to carry across is not a performance figure. It is that these scenarios needed their own benchmark, which means a station design has to earn its result on your samples rather than assume it.

Hypothetical example. A plant runs a clear stand-up pouch with a granular product. Fill timing shifts and a thin band of granules lands across the inner edge of the seal area. Under a single top light, the granules read as a faint texture similar to acceptable pouches where product simply sits behind the seal. With transmitted light and a framing window locked to the declared seal band, the intrusion presents as a dark interruption inside a bright band. Same package, same model family, different physics.

This is why lighting and optics come before model selection. If you cannot state in one sentence the optical mechanism that makes your feature brighter or darker than its background, you do not yet have an inspection. The same applies to resolution. Take the minimum feature size from the drawing, then confirm the lens and sensor resolve it with pixels in reserve.

To use a visible check in a package integrity decision, establish and re-check its relationship to the package-specific reference test. Keep the visual result and the reference-test result as separate fields so quality can assess that relationship.

Separate the vocabulary on the placard

Hypothetical example. A shrink-wrapped multipack shows a bright white region in the overhead image where one unit should be. Two different causes can produce similar pixels: a specular glare highlight from the film, or a genuinely missing unit with the white conveyor showing through. A station that treats both as a missing-item reject can send the team after an applicator fault that is not there. A station that adds a second framing angle and routes the unresolved case to line-lead confirmation rather than an automatic reject gives the team a chance to find that film tension changed. Write "glare" and "missing unit" as separate outcomes on the placard, each with its own owner.

One Reject Is Not a Work Order

Letting every vision reject create a CMMS work order erodes the maintenance queue quickly. The planner ends up with single-package events carrying no diagnosis, and the queue stops being read.

Rejects and maintenance requests are different records with different owners. A reject is a quality disposition about a specific package. A maintenance request is a statement about equipment condition, raised when a reviewed pattern in rejects points at a machine cause. Product hold and release belongs to quality. Maintenance owns station and machine repair, not product release.

The structure to aim for: the vision station writes an inspection event, quality decides disposition, and a separate reviewed rule decides when a run of events becomes a maintenance request against the sealer or the applicator.

Worked QA hold record (hypothetical)

  • Record type: QA hold
  • Trigger: visible product inside declared seal band, pouch SKU family A
  • Station: pouch inspection, lane 2
  • Evidence attached: reject frame, reference acceptable frame, framing window overlay
  • Confirmation step: quality technician inspects the retained package
  • Confirmation outcome: reject confirmed
  • Disposition: hold affected sequence back to the last confirmed acceptable check
  • Scope decision: packaging engineer defines the hold boundary with quality
  • Follow-on: no package-level action; event added to the seal-area intrusion tally for the run

Worked sealer maintenance request record (hypothetical)

  • Record type: maintenance request, not a package disposition
  • Reason: seal-area intrusion tally for lane 2 crossed the review threshold agreed with the packaging engineer across consecutive runs
  • Suspected failure modes to investigate: fill timing drift upstream, seal bar dwell variation, jaw alignment
  • Requested work: diagnose first. Measure fill valve timing against setpoint, inspect seal bar face and jaw alignment, compare dwell setpoint to the sealing specification, then propose corrective work based on what is found
  • Evidence attached: tally trend, representative reject frames, station drift log extract
  • Requested by: quality technician with line lead concurrence
  • Owner: maintenance planner schedules against the lane 2 asset record
  • Closeout requirement: state what was found, what was adjusted and the verification check run afterward

Note that the requested work starts with diagnosis rather than a prescribed part swap. A request that says "replace the seal bar" closes out with a replaced part and no learning. A request that says "measure, inspect, compare, then propose" produces history you can analyse later, which is the point of a work order closeout standard.

When the Station Does Not Know

Design a third outcome alongside pass and reject: unknown.

Unknown is what you have when the trigger fired but no frame arrived, when a frame arrived with the framing window clearly misaligned, when illumination fell outside its monitored band, or when motion blur prevents the required feature from being assessed. Treat each of these as hold, never as pass. A missing frame is not evidence of an acceptable package.

Hypothetical outcome record for a lost frame:

  • Outcome: unknown. Trigger logged with no matching frame
  • Package status: affected units held, routed to quality for disposition
  • Station status: continuing acquisition is acceptable only while containment and unit identity tracking are demonstrably working. If either is in doubt, stop release from the station
  • Escalation: when unknown events exceed the agreed per-run count, the station goes to hold and quality is notified
  • Follow-on: trigger sensor and cabling diagnosed under the station drift procedure

A no-read on a code or a closure check follows the same rule. The affected units stay held until they are verified by another method. Cleaning the lens and continuing does not clear the units already captured.

Research on hybrid automated optical inspection for mixed-SKU bottling lines is instructive on structure: a small experimental study combined learned region selection with deterministic checks and kept a record per bottle for household chemical containers [3]. The structural lesson is what transfers. Learned components choose where to look, deterministic checks decide pass, reject or unknown against stated geometry, and every unit gets a retrievable record.

Deterministic checks also give a technician something concrete to dispute. A stated geometric rule plus a retained frame keeps that conversation short. A bare model score does not. Setting and re-checking those decision points is covered in the notes on thresholds and validation.

Make the Inspection Event Worth Keeping

An inspection result that lives only in a reject counter has little value two weeks later. The context has to travel with it.

GS1's explanation of EPCIS is a useful design pattern even for teams that never implement the standard: events capture what, when, where and the business context, and version 2 extends events to carry sensor and status data [4]. Borrow the shape. Your inspection event should answer what was inspected, when, at which station and lane, under which recipe and lighting state, and which business step was in progress.

The GS1 traceability material adds a second useful split between critical tracking events and key data elements, where the event records that something happened and the data elements describe the specifics [5]. At a packaging station, the event is "seal area inspected at lane 2 pack step." The key data elements are the SKU family, the station recipe, the framing window, the illumination state, the outcome and the person who confirmed it.

Fields worth insisting on for a packaging inspection event:

  • Station and lane identity, so a multi-lane plant can separate causes instead of averaging them
  • Recipe and framing window in force, because a changeover that loads the wrong recipe is an invisible failure otherwise
  • Illumination state, read from the monitored band rather than assumed from a schedule
  • Outcome in the three-value vocabulary: pass, reject, unknown
  • Human confirmation result, including the case where a reject is not confirmed
  • Linked requirement wording, so a later reader knows what the station actually claimed

Those fields are what make drift visible. If lane 2 produces more unknowns after a lighting module is replaced, the record shows it. A daily reject count does not.

Frequently asked questions

Can vision confirm a pouch seal is intact?

No. Vision inspects visible surface condition inside a declared seal band. Channel leaks and cold seals can be invisible at the surface. Integrity remains the job of the tested integrity method defined by your packaging engineer and quality system.

Is infrared imaging an integrity test?

Treat it as a thermal signature indicator under controlled conditions, not an integrity result. Ambient temperature, dwell time and film variation shift the signature. Have your packaging engineer define the acceptable range and who dispositions units outside it.

Should every reject create a CMMS work order?

No. A reject is a quality disposition. Raise a maintenance request when a reviewed pattern points at a machine cause, with evidence attached and diagnosis as the first requested step.

What happens when the trigger fires but no frame arrives?

Record the outcome as unknown and hold the affected units for quality disposition. Count unknowns per run and send the station to hold when the agreed count is crossed. A missing frame is never a pass.

How do we decide whether the camera can see the feature at all?

Take the minimum feature size from the packaging drawing, then work the resolution math before anything else. Resolution and defect size and lighting selection are prerequisites to choosing between rule-based, supervised and unsupervised methods, not follow-ups.

Can a height or fill-level check be done with a single camera?

Sometimes. A validated 2D side view with fixed framing and a reference edge can serve, as can a dedicated gauge. The requirement is validation against your specification, not a particular sensor type.

Start With One Station and One Placard

For the wrinkle crossing the seal band, retain the visible finding, route it to quality, and keep any integrity test linked to the same package record.

Pick your highest-volume packaging station and write its requirement on one page: the visible feature, the geometry taken from the drawing, the three outcomes and the named owner for each. Have the packaging engineer and the quality manager agree the wording. Use that review to identify assumptions about visibility, acceptable variation and release authority.

Track unknown rate per run and lane alongside reject rate, confirmed escapes and false rejects. If unknowns climb, investigate triggers, lighting, framing, product presentation and input records. The guide to station drift and maintenance signals suggests station checks to include.

To work through a specific packaging feature and its decision record, discuss the use case with us.

References

[1] "Deep learning for the quality control of thermoforming food packages". https://www.nature.com/articles/s41598-021-01254-x

[2] "The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection". https://arxiv.org/abs/2503.21622

[3] "Vision-based automated optical inspection for real-time quality control: hybrid single-camera deep-learning AOI for mixed-SKU bottling lines". https://link.springer.com/article/10.1007/s00170-026-18704-3

[4] "What is Electronic Product Code Information Services (EPCIS)?". https://support.gs1.org/support/solutions/articles/43000755325-what-is-electronic-product-code-information-services-epcis-

[5] "What is the GS1 Traceability Standard?". https://support.gs1.org/support/solutions/articles/43000734535-what-is-the-gs1-traceability-standard-

Ready to put this into practice?

See how Monitory helps manufacturing teams implement these strategies.