Mark an inspection image unusable when it lacks the evidence required to make the agreed decision. Give the person reviewing it a specific reason and an action. Record unreadable images separately from clear examples with a visible defect.
Tie each usability rule to evidence
Ask which feature and reference make the inspection possible. A label can be fully visible while its placement reference is cropped out. A fastener may look absent if the mount blocks its recess. One task might accept a small motion blur across irrelevant background. Another might require a crisp edge exactly where the blur sits. Avoid a generic score that treats these images alike.
For a fictional closure check, suppose the top-view image is meant to show one sealing tab and its adjoining opening. Record “tab and opening in view” as the usable condition. An image that contains the tab but cuts off the opening does not satisfy the requirement. If close-up review reveals both but leaves alignment uncertain, a person can adjudicate it using the rubric. The initial reason should say which uncertainty is present.
| Image state | Example | Action |
|---|---|---|
| Usable and clear | Tab and opening both visible under the usual view | Apply the release rubric |
| Usable and questionable | Both edges visible but tab sits near the tolerance | Route to the quality reviewer |
| Unusable | Opening hidden by the fixture | Request a clear image or use the approved manual check |
| Unmatched | The item ID is absent | Resolve the identity before attaching a disposition |
Define the reason a shift can act on
Use reasons such as “lower edge clipped,” “small slot obscured,” “duplicate image of previous unit,” or “insufficient focus at the target.” Preserve details in the record rather than assigning an operator a long list of reasons to choose from. Ask the operator how a reason would change the next action. Group two reasons only when they lead to the same practical response and the extra specificity offers no diagnostic benefit.
Retain the original rejected capture when your image policy allows it. Link it to the unit and the retry image. A repeated capture is better measured as a failure and successful retry than as one clean result. If an alternate view is used routinely, test how much time it takes and whether it actually reveals the missing reference. Note the setup before and after if a lamp or station adjustment caused the new image to become readable.
Measure unusable captures without hiding failure
Start with all eligible original captures. Report how many were usable immediately, usable after recapture, routed to a person for uncertainty, or still unresolved at the end of the run. Keep reference-defect counts for unusable frames when an independent measurement establishes them. A software result can be evaluated on the readable images, while the process must still handle cases that fail the input gate.
A fictional set of 300 packaging images includes fifteen unusable first views. Twelve can be resolved by capture before the next step and three go to an in-person check. That is a 5% unusable-first-view rate for the set. If one scored only the final resolved images, that workload would disappear. Use the slow cases to test whether the retry procedure is usable during a shift change.
Review one example of every excluded reason
Compare the capture and result with someone who runs the station. Check whether another controlled view can resolve the exclusion. If a wrong-image error comes from a unit-ID issue, fix the association before adjusting an image-quality threshold. If a missing edge is caused by camera placement, gather example views before changing the acceptance rule. Give the next reviewer the original examples and expected actions, with the date the rule was agreed.
Use the image usability rubric for the first review and the camera view guide when you find that a necessary feature is routinely obscured. For one vendor’s acquisition controls, consult Cognex’s advanced acquisition settings documentation.
Write the usability test in the language of the work
Name the feature and the evidence a human needs before deciding whether a picture can be used. On a label placement check, that can mean the outside edge of the label and a visible reference on the carton face. On a part presence check it can mean the mounting point and enough surrounding geometry to see whether a part obscures it. A picture may be sharp and correctly exposed across most of the frame while leaving one required edge hidden. Judge the image for the decision it has to support.
Determine the full unit of evidence. If one appearance depends on a top and a side image, define which question each view answers. Then decide what should happen when one image does not arrive. Sometimes the first view can establish enough evidence, while the second is only an optional fallback. In a check that requires both views, the missing one leaves the unit undecided. Record that relationship in the rubric and test the status the operator would see for each missing view.
Use a short, specific convention for your reason codes. “Missing lower label edge” points a collector toward adjusting the frame or finding a second view. “Low quality” tells a person very little. Avoid assigning too many codes that staff cannot distinguish in practice. The raw reason can describe what the user can see. A separate field can identify the likely cause for later investigation. For example, “target outside frame” describes evidence. “Fixture guide slipped” is a cause that can be checked.
Give the code an action. A carton present at the station might be recaptured using an approved alternate view. A unit already moved into a hold area may need a physical inspection. An evaluation image with no recoverable identity might remain unresolved. Do a short trial with an operator and a quality lead who have both used the process. They should agree on the information they need for the next step and how their decision attaches to a unit. Use the image usability rubric to record the minimum evidence and review actions.
Compare blur where the feature is smallest
Start with a reliable control under the same imaging path. A still reference can show whether the lens, focus or image transformation lost detail. A unit at line speed shows blur the reference may miss. Judge the target area rather than the edges with the most obvious contrast. A barcode may remain sharp enough for a person to read its size and location while a low contrast label edge has softened too far to locate near its tolerance boundary.
Inspect the direction of the softness. A smearing pattern aligned with motion suggests an acquisition problem that changes with speed or exposure. A soft region across a stationary sample can point to focus, depth, lens cleanliness, or preprocessing. Use matched frames to test each candidate cause. If images are resized for review, open the original and exported image at consistent scale. A resize can remove usable detail between capture and labeling. Preserve the exact derived file to explain what each reviewer saw.
Imagine a fictional assembly check on a small clip. The camera shows clear markings on the tray and a thin edge where the clip should be visible. On a moving batch the edge grows indistinct over a wider region. The tray markings still look crisp because they are large and high contrast. The quality lead reviews three images on the same clip and sees the edge in both still images but has no evidence in one moving capture. In the pilot, record the moving capture as unusable for clip presence. Preserve the other images as attainable evidence and ask whether an approved change to capture timing or light helps the moving case.
Check around near boundaries. A larger defect may remain observable with slight softness while small differences cannot be interpreted. Use a borderline reference that actually matters to the process. Compare repeat images of its lowest-contrast material under the conditions you intend to approve. Ask a second reviewer to mark the observed edge and compare whether their measurement lies within the decision margin. If the visible feature varies by position, treat the range as part of the usability rule. Keep the real procedure within the tolerances the site can repeat.
Separate framing faults from true absence
Write a positive definition of a clear field of view. If a required fastener is missing, show the expected fastening point sufficiently clearly to support that finding. An empty-looking crop can be cropped above or below the fastening point, show the wrong side, or be the wrong unit. Ask which feature anchors the location and how the image tells you it was possible to see it. Store a normal example and an authentic missing-part example from that same viewpoint.
A fictional tray with two compartments illustrates the difference. The overhead camera reliably sees the farther compartment, but a closer divider obscures the bottom of the near one when the tray sits against the back guide. A person looks at the image and sees one insert. One interpretation is that the near insert is missing. Another is that the nearer compartment is hidden. A side image from the same unit can resolve the latter in the right setup. Test a known empty near compartment and a known filled one at the expected extremes. Use this pair to define when the original view is usable for the near compartment.
Also review the reference when a box moves. A carton that is rotated as a whole can leave the label clear while cropping the edge needed to establish its relative position. A position worksheet may still accept a numerical input generated from that crop. The calculation is valid only when the physical reference is visible and the permitted geometry applies. Add a usability test before accepting the coordinate reading. Have the reviewer verify both datum and label edge rather than judging the position from an isolated label image.
Set a separate route for wrong view. That reason can be recognized and acted on without guessing about image quality. A correct label photographed on the wrong side for the desired inspection may be readable by itself and irrelevant for the required check. Use a target pose reference where possible, then identify the expected station and sequence. Label the true capture fault for training or audit and preserve the valid product outcome when someone obtains the needed view.
Read exposure and clipping in the region that matters
Find the significant boundary first. A hotspot over smooth blank board may have little consequence for a check focused on an intact black label. A small patch that hides the only evidence of a torn edge can be decisive. Compare the captured original with known correct and failing cases in a paired fixture. If the same patch moves with a light change, record whether the feature becomes visible and whether that setup makes other areas harder to review.
Exposure should support both the evidence and the capture cadence. A site might choose a brighter test setting for a stationary image that will be too slow under live movement. The photo could then have enough image brightness to see the edge in a paused sample, while blurring it on actual output. Mark the current settings, light, position and motion state. Add an example of the acceptable and failed appearances at operating speed. The lighting guide works through matching images by unit and condition.
Observe how dark materials behave. A black clip in a black recess may need a view and illumination that reveal the outline. If the part appears dark in every sample, a broad brightness rule rejecting all dark regions may increase workload without recovering detail. Decide whether there is a stable contrast or shape in the expected region. Compare the same clip in accepted and absent conditions. If the outline cannot be seen reliably even after an approved setup change, route that product condition to the supported evaluation method.
Protect unusable examples for evaluation. If all clipped examples are dropped before checking performance, the resulting report may describe a well-behaved subset of inputs and conceal the workload for the production line. Preserve counts by cause and shift. Determine what the check would do on the clipped examples and what the operator actually did. The model evaluation can exclude images without a settled visual outcome from its four-cell report while the workflow evaluation still accounts for them.
Verify unit identity before interpreting the picture
Make the capture and the package identity retrievable by someone on the receiving shift. A clear image attached to the wrong carton is unusable for deciding that carton’s disposition. In a production system, pairing issues can happen at a changed trigger, a partial file upload, a duplicate save, or between cameras with different timing. Test by using a short sequence of distinct known specimens and preserving the order. Check that each result points to the right physical unit and the correct picture.
For example, consider a fictional line that uses an overhead image and a side shot of the same open kit. Two trays arrive closer together after a restart. The system pairs the side image of the following tray with the overhead image of the first because both carry timestamps rounded to the same second. Both images are sharp and each looks sensible in isolation. A person trying to decide the first tray receives conflicting evidence. The team needs an independent pairing key or an explicit unresolved state when pairing cannot be verified. A specification about focus and clipping alone misses the failure.
Assess remote review of a recapture. A retry may generate another file from the same unit and a second device. Preserve a reference to both and note which image superseded the earlier observation. A reviewer needs to understand the sequence and identify whether the original unusable frame happened before or after a change at the bench. That can explain an apparent difference where the first image has a blocked edge and the second shows a new condition. Use the capture metadata guide to define the expected retrieval fields.
Treat interrupted evidence as a separate reason from “wrong identity confirmed.” You might have a known capture and an absent unit identifier, or a pair of images with conflicting identifiers. Record what can be recovered and who can settle it. If the identity remains uncertain, keep the images for investigating the station without including them in an independent test with unit-level reference outcomes. Ask the process owner to establish which physical units remain on hold.
Rehearse recapture without losing the first observation
Test the recapture action with a person on the line and a known difficult unit. The person should see why the first view was unusable, what to change using the approved process, and where to put the next frame. Record the original and repeats under the same unit ID and distinguish time and camera pose. Use the actual time to examine, retrieve or recapture during the trial. Report that time separately from any fainter effect of the accuracy result.
When recapture will be common for a condition, check where a unit waits. A packaging line with a fixed pace may need a hold tray at the next safe step. A bench operator may be able to take a second image immediately. If the next action requires another operator, have them find the unit and both frames during a realistic handoff. Include the case where the raw image is unavailable or a worker moves it unexpectedly. Rely on the site’s approved method for physical handling and containment.
Define a point at which another capture has little value. Four consecutive unreadable images with a blocked edge may show a view problem that needs a different route. Use a reason-specific recapture limit chosen by the operations owner. An alternate station, physical measurement, or separately documented manual procedure may finish the case. Track the time until resolution and preserve the attempted recaptures. Reassigning the same unit repeatedly can obscure the true human workload.
Separate recapture evidence from downstream pilot decisions. The observation that an operator recaptured an image is evidence the workflow can recover some cases. Its success rate depends on which units receive the recovery path. Show the number that became usable and the number that required other action. If a second image adds evidence but cannot show the defect rule, retain an unresolved decision until a person can inspect the physical item or another appropriate record. Bring these outcomes to the uncertain image routing guide.
Measure availability and outcomes at the appropriate level
Start by counting attempts, unique units, and evaluated outcomes separately. In a fictional shift, 120 units enter the check, and exactly 120 initially captured images arrive. Twelve first images are unusable. Eight of those units get one additional capture, and five then become usable. If you simply count acceptable frame files, the denominator is now 128 rather than 120. The process figures that matter include twelve initially unusable units, five recovered units and seven that needed another decision path. Report first-try and final visibility separately.
Show the relationship of the unusable group to other check metrics. A pilot may report its matrix on the 113 units with settled automated outcomes using the fictional shift above. If seven units also had reference outcomes after manual work, you can report how the overall process handled them while preserving the separation. Specify which denominator applies in each statement. For instance, “false alarms among decided readable units” answers a narrower question than “manual interventions among units that came to the station.” Both matter to the quality owner who evaluates scope.
Avoid silently removing an inconvenient group from a performance comparison. If glossy material has greater capture loss, a comparison restricted to usable frames can make glossy and matte look equally good among readable cases. Report which cases were missing, their outcome where available, and the production consequence. If your evaluation intentionally includes unusable crops from a challenge set, keep their expansion separate from any estimate of routine frequency. Bring the station test and the reference labels into the same decisions log.
Disaggregate by meaningful condition. A sample with several hundred images from the large carton and only five from the tall carton provides very different certainty about each format. If the only failure happens on tall stock, ask whether that format deserves focus before ordering more frames. When a category has little data, save clear and failed examples and schedule a focused follow-up. Summarize the next evidence you need and the interim action for the condition. Use error reporting by condition to preserve the difference between an observed unusable image and a possible cause.
Teach the difference with intentionally difficult examples
Keep a small set of examples that practitioners can use without a trainer. Include a textbook usable frame, an obvious framing failure, and ambiguous examples whose correct outcome depends on the decision. Pair an image with a tiny motion smear to one with a similar visual effect caused by a softened printed border. Give each example a reference outcome and a reason tied to the region. Add an original and one crop if the tool displays a reduced view.
Have two people apply the usability rubric independently before they compare notes. Ask each to identify the physical feature and show where the image supports it. Look for recurring disagreements over the same reason. If one person labels an unfamiliar package as unreadable and another as admissible, review the format list and adjust the job instructions. If both agree on what to check but see different boundaries, evaluate whether a second view or a more precise reading rule is needed.
Use the disagreements to improve the rubric. Suppose the team originally said an image is unusable if “any part of the label is blocked.” A fixture hides the upper margin in ordinary frames, but the check only uses the two lower corners and a visible carton datum. That broad rule could reject an otherwise defensible view. If the quality owner agrees, define the set of points required for the decision and update the version. Recheck the known cases near the physical tolerance and the next independent sample. Preserve the previous wording and observations so earlier reports remain interpretable.
Record a qualification example for each important failure type. It should show the feature the route is intended to protect, the expected review action, and how somebody resolves it. Do a short rehearsal after a camera move, a crop update, or an expected material change. Have a receiving reviewer explain the reason and result using the saved evidence alone. A practical rubric is successful when someone can act on a hard case and the next person can verify what they did.
Decide where an automatic usability check helps
Start with failures that can be detected in a simple and repeatable way. An expected frame has to exist and have the correct dimensions and view identity. A missing image or absent unit identifier can cause a workflow exception before the visual model runs. More subjective conditions, such as a small area of low contrast on a reflective feature, may need a visual test with meaningful examples. Separate these gates so an operator knows whether to look for an image, recapture, or inspect the item directly.
Test an automated blur flag on sound and soft images from each format. A statistic based on edge strength can be sensitive to material texture and contrast. Plain board may produce fewer crisp edges than a printed field even when it is captured correctly. A rule fitted to a busy background may routinely reject the quieter format. Compare corresponding crops with the actual target in both groups. Document the situations the metric can handle and the supported action when it does not give enough evidence.
Where a station uses a bright-region count to identify clipping, evaluate it with small critical features. Imagine an image where most of a white carton face remains visible, but a two millimeter highlight blocks the sole datum used for tolerance judgement. The share of bright pixels over the whole frame hardly changes. A usability check that inspects the reference region can be more informative. Evaluate it using an acceptable white feature that legitimately reflects light and an unreadable area on the target region. Store the region definition and station configuration so later maintainers can repeat the comparison.
Keep the people reviewing exceptions informed about the reasons automation actually provides. A flag caused by a frame width check should lead to verifying the export and source path. A flag caused by a missing reference requires locating another view or explaining why the region is absent. If the reason is an observed model score only, the person should be able to identify the image features and the permitted closeout action. Avoid giving a person a risk category they cannot translate into a useful diagnosis. Test each failure path through the same interface and work steps it will use when live.
Check that variants keep the same meaning when the crop changes
A team can make a useful crop for one orientation and accidentally apply it to a format that needs a different region. Before selecting a crop, compare images that are representative of the small, large and rotated units. Draw each target and reference on the original image. Confirm that every required point stays inside the crop for all approved routine positions. Keep a small margin that reflects how positioning really varies. Verify the displayed crop and any separate one used for automatic inspection.
Run the check after ordinary operating changes. A recipe might alter the export dimensions while preserving the position of an object in the frame. A mount change could preserve dimensions and shift the location. A saved crop can look exactly as expected on a common product and exclude a datum on a taller variant. Use one known boundary example on each affected format and compare it with its earlier image. If the feature moved relative to the crop, identify whether the image setup or permitted position changed, and re-evaluate the affected condition.
Store a rule or configuration revision with each raw image. When reviewing an old result, recreate the same crop if needed. A modern crop may show an edge only because the pipeline now includes more of the original. Reviewing it as if the old check had seen it can turn an earlier evidence gap into an apparent model error. Keep the historical observation separate from a proposed replay. Record the output of the replay with its own version and explicitly link it to the input used.
Use earlier failure images as intentional regression tests for the image quality route. A new crop should continue to pass usable cases and reject or route cases where evidence is unavailable. Some prior failures can become usable if the camera recorded the missing region outside the old crop. Other failures stay blocked by the fixture or clipped by exposure. Record which examples recovered and why. Then test on fresh units to see whether the difference persists under the currently approved setup.
Report the usability decision in a way the next owner can use
Close with one of three kinds of statements. A usable image can proceed under a stated check because specified features are visible. A recoverable image needs a named new view or action before anyone can interpret it. An unresolved image requires the approved alternative and an owner. The record should include the identity of the unit, original source, rule version and reason. Keep the allowed next action connected to the position of the physical unit when the workflow depends on it.
Use a fictional handoff to calibrate the team. At the end of first shift, two cartons wait for review. The first preserves a valid unit ID and clear label but shows a blocked datum. The second has two readable views and no evidence they belong to the same unit. In the first case, the receiving lead can retrieve and photograph the expected face or make the authorized physical observation. In the second, a sharper picture of one tray cannot solve the relationship. The receiving lead checks sequence, location and identity records, then applies the approved unresolved disposition if the link remains uncertain.
When two categories seem interchangeable, ask what action follows. A routing field succeeds when a person can pick up the work quickly and another person can reconstruct why they did so. If the two categories lead to the same action now but identify two distinct improvement tasks, keep the observed reasons in the analysis. If new categories lead people to debate labels without changing the review or engineering response, simplify the operator-facing fields. Retain raw evidence for the quality team.
Record the time from a failed first image to recovered evidence. If the station needs a recapture in the first thirty seconds to avoid losing the unit, a diagnostic whose notification arrives minutes later serves a different purpose. Use it for identifying patterns with the maintenance lead. Reserve the operational exception for a signal a person can receive and act on while the item is still recoverable.
Before the pilot advances, pick an example each of blur, clipping, occlusion, wrong view and uncertain identity from the approved sample collection. Have a person process them on the normal interface, then check how the failed images were recorded, whether the right next owner received them and what happened to the unit. Make any correction required for someone to complete that action on the next shift. Start with the usability rubric and use the pilot acceptance worksheet to record its result.
