Optical Sorting of Frozen Vegetables: Defects, Blind Spots and False Rejects

Optical sorting can remove selected colour defects, misshapen pieces and foreign material from frozen vegetables when the installed sensors, product presentation and reject system can distinguish and separate those targets. It does not follow that every unwanted object is detectable, or that every object detected reaches the reject outlet. The useful purchasing question is which defined defects were removed from this product, under which conditions, and how much acceptable food was lost with them.

A photograph of a sorter cannot answer that question. Neither can a report that simply lists the weight rejected. A useful demonstration examines the incoming material and both outgoing streams, with an agreed definition of acceptable product. It also records the actual product setting and operating conditions so the result can be interpreted later.

We would begin with a small set of priority defects and ask for evidence against each one. The aim is to establish what the proposed supply process can demonstrate, where uncertainty remains, and what additional checks the product requires. The examples below are proposed ways to structure that discussion, rather than a standardised validation procedure.

Define the defect before asking about the camera

“Optically sorted” describes a processing step. It does not define the finished specification. A dark spot on a carrot, an unusually short bean segment and a leaf fragment present three different questions. The first concerns an appearance boundary, the second may concern size or shape, and the third concerns unwanted material. A single setting or removal percentage should not combine them without showing the separate results.

Start with physical examples or sufficiently clear photographs of each target, together with the reason for rejecting it. Define the boundary between normal crop variation and a defect that affects the intended product. For instance, a small change in green tone may be acceptable in a soup ingredient but outside a buyer’s reference for a visibly uniform vegetable side dish. That commercial distinction should be resolved before the machine is adjusted.

Keep critical foreign-material requirements separate from appearance preferences. The fact that a material is visually distinctive in a sample does not establish that its removal has been validated for every size, orientation or product condition. A food-safety control must be assessed by the responsible technical team within the site’s hazard controls. This article’s inspection examples do not establish a safe limit for hazardous material.

Also agree what counts as acceptable food. If the operator calls every piece in the reject stream defective, there is no independent basis for measuring unnecessary loss. Keep an approved reference covering normal variation in colour, shape and surface condition. When a borderline piece appears during review, record the uncertainty instead of changing the definition to make the demonstration look successful.

Separate seeing, deciding and rejecting

Three events must work together. First, the sensing system obtains a useful signal from the object. Second, the selected product setting classifies that signal. Third, a physical action separates the object from the accepted stream. A failure at any one of these stages can leave a target in the finished product, but the corrective action will differ.

The word camera covers several configurations. TOMRA’s vegetable processing overview describes cameras combined with different lighting and sensing technologies, including infrared, as well as multiple viewing angles. These can provide different kinds of contrast. An ordinary colour image is therefore an incomplete description of a machine that combines several sensing methods. Equally, the presence of an infrared sensor does not establish a universal ability to find hidden defects.

Key’s COMPASS description gives a useful manufacturer example of sensing combined with object recognition, product settings and pneumatic rejection. Those are separate capabilities to confirm for the relevant installation. Request the actual model, installed sensor options, product recipe and sorting position. A brochure for an optional configuration is not evidence that those options were fitted or active during the supplied product’s run.

When a target reaches the accepted outlet, investigate whether it was visible, whether it was classified for rejection and whether the reject action worked. Available event records may help, but the physical sample still matters. Conversely, acceptable pieces can reach the reject outlet because they were classified as defects or because they accompanied a nearby target during diversion. Both cost product; they call for different adjustments.

Do not infer all of this from the sound of air valves or a counter on a screen. Depending on the system, a count may represent detected objects, classification events or reject commands. Ask what the displayed quantity means before comparing it with a hand count or a weighed reject sample. Those numbers may legitimately have different denominators.

Map the blind spots in the actual product stream

A surface must be available to a relevant sensor before its appearance can contribute to a decision. Consider a blemish facing the camera, the same blemish facing away, and the same surface hidden under another piece. Those are different viewing conditions even though the target itself is unchanged. Additional views may improve coverage, but the actual arrangement and product path determine which surfaces are presented.

Product distribution therefore belongs in the trial record. Record whether pieces are separated or overlapping, whether the flow is steady or surging, and whether clumps enter the inspection area. Average tonnes per hour cannot describe all these local conditions. A short, evenly spread demonstration may present an easier inspection task than a crowded section of the same production stream.

Contrast creates another boundary. A bright foreign object among pale kernels is a different test from material close to the product’s visible colour. Some sensor configurations use information beyond visible colour to distinguish materials; that capability needs to be demonstrated with the relevant target. A successful test on one conspicuous material should not become a general statement about all plastics, all vegetable matter or all contaminants.

Surface condition also deserves a record. Compare the demonstration material with the product actually being supplied: frozen or unfrozen, individually separated or attached, and with the surface appearance present at the intended sorting stage. If frost or adhering material changes what the sensor sees, testing a cleaner or differently conditioned sample may answer the wrong question. Do not deliberately change product temperature just to reproduce an image; use the site’s controlled handling procedure.

Finally, identify the sorting position. A check before freezing and one near packing inspect different points in the process. TOMRA’s vegetable processing overview describes several possible positions, including checks after the freezer. A result applies to material presented at that point. It cannot establish that nothing was introduced or changed during later handling.

Build a challenge matrix around the agreed specification

A useful challenge matrix connects a target with an observation and a decision. It should be small enough to investigate properly. Choose the defects that matter for the product and application, then describe the relevant variations in size, surface, orientation or contrast. Avoid adding a long list of targets merely to make the document look comprehensive.

Use naturally occurring, identified defects where appropriate, or a documented test method approved by the responsible technical team. Do not introduce hazardous materials into saleable production to create an informal demonstration. Any controlled challenge requires suitable isolation, accounting, recovery and disposition under the site’s procedure. A buyer’s request for evidence should respect that process.

Record the challenge quantity and what was recovered. If a known set of target pieces is used, distinguish recovered targets in reject, recovered targets in accept, and any targets not accounted for. Missing pieces make the interpretation incomplete; they should not be counted automatically as successfully rejected. Likewise, multiple fragments from one broken target should not silently become several successful removals.

For a natural production observation, describe the incoming sample and the sampling interval honestly. Without a known input target quantity, the trial may show residual defects and reject composition but may not support a target-removal percentage. That narrower result can still be useful. Report the measurement that the available evidence supports.

Check acceptable product in the reject stream

The reject outlet contains information about both quality control and yield. Separate material that meets the agreed reject definition from food that still meets the acceptable reference. Keep uncertain pieces in their own recorded category until the criterion is resolved. Photograph representative examples beside their classifications so later reviewers can see what was counted.

For procurement discussions, it helps to distinguish two meanings of false reject. A classifier may identify a sound piece as defective. Separately, sound pieces may be physically diverted with a neighbouring target. The buyer usually needs to know the total acceptable-product loss in the reject stream; the operator also needs the cause to improve it. Call the measured quantity what it is rather than assuming a machine’s internal false-reject statistic measures the same thing.

Consider an illustrative 100 kg input producing 98 kg accepted material and 2 kg reject, with no other loss in this simplified example. Hand review finds 0.5 kg of specified targets and 1.5 kg of acceptable vegetables in the reject. Acceptable product lost to reject is 1.5% of the original input. Acceptable product also represents 75% of the reject stream. Both figures are arithmetically correct, but they answer different questions.

Neither percentage establishes the removal rate for the target. That requires knowing how much target material entered, or another justified method for estimating it, and accounting for the target left in accepted output. Nor is 1.5% the loss relative to all acceptable product entering the sorter unless that acceptable input quantity is known. Write the denominator beside every percentage.

A low reject weight is not automatically a better result. It may reflect a cleaner input, a less demanding specification or a setting that misses more targets. A high reject weight may reflect poor incoming quality rather than excessive rejection of sound product. Compare like inputs and examine both quality and loss before deciding which setting is preferable.

Read both outlets before accepting a setting

For each condition under comparison, retain identified observations from input, accepted output and reject. Match the collection windows and handling method as closely as the production arrangement allows. Record any recirculation, additional sorting or manual removal, because those steps change the relationship between the streams. A simple one-pass accounting example cannot describe an unrecorded multi-pass process.

Read accepted output against the agreed finished-product criteria, not just against the previous machine setting. Read reject against the same acceptable-product reference. The most useful comparison shows residual target findings alongside acceptable-product diversion. If one improves while the other worsens, the record exposes a real trade-off that procurement and technical teams can discuss.

Keep results for materially different targets separate. A combined removal figure can look strong when easy, numerous targets dominate it while a less frequent difficult target remains. A buyer concerned about a particular type of vegetable matter needs the result for that material. The combined figure may be retained for context, but it should not conceal the individual findings.

Zero findings in an examined sample mean that no target was observed in that sample under that review method. They do not establish that the complete lot contains none. State the sample quantity, how it was selected, what the examiner looked for and whether the review could identify the target reliably. The broader release decision belongs to the agreed inspection and control programme.

A short repeated comparison can be more informative than one unusually favourable run. If performance changes noticeably between observations under apparently similar conditions, investigate the variation before settling on a single representative figure. The right number of repetitions and the acceptance rules depend on the target, intended decision and responsible technical review; this guide does not prescribe universal values.

Retest when the product or operating conditions change

A stored product recipe is a useful starting point, not a permanent approval for every crop or cut. A change from a single vegetable to a medley alters the range of colours and shapes that should be accepted. Smaller cuts alter presentation and the relevant defect reference. A new crop may bring normal appearance variation that was missing from the original demonstration.

Before transferring a setting, identify which assumptions have changed. Review the target definition, normal product reference, stream condition, installed sensing arrangement and reject setup. The Key brochure’s broad list of IQF applications should not be read as evidence that one unchanged recipe suits every item or mixture on that list.

Operational changes also merit attention. Cleaning, maintenance, sensor or lighting work, reject-system work and software or recipe changes may affect the relationship between the original evidence and the present run. The site’s procedure should define checks and responsibilities. For a buyer, the practical request is the date and reason for the latest relevant verification, rather than an unsupported statement that the sorter is always accurate.

If a complaint reveals a target that was absent from the original challenge, add the new evidence to the investigation. Preserve its identity and the product context. Do not immediately increase sensitivity without checking whether the issue is visibility, classification, physical rejection or later introduction. An adjustment that raises acceptable-product loss may still fail to address the actual cause.

Turn a demonstration into a usable supply record

A usable record identifies the product, cut, normal reference and target defects; the sorting stage and relevant equipment configuration; the selected recipe and trial conditions; and the results from both outlets. It includes photographs that explain the classifications, the quantities and denominators used, and any target recovery or sampling limitations. These details make the result reviewable without pretending it guarantees every future pack.

Separate what was directly demonstrated from what was inferred or remains to be confirmed. For example, the record may show acceptable-product loss under one tested condition while leaving target-removal efficiency unquantified because the natural incoming target quantity was unknown. That is a clear, usable limitation. Replacing it with an unqualified efficiency claim makes the record less useful.

When comparing supply options, request equivalent definitions before comparing percentages. One report may measure acceptable mass lost per input mass, another the acceptable proportion of reject, and another a machine-generated event statistic. Align the quantities or keep them explicitly separate. A purchasing decision should reflect the finished-product requirement and the strength of the evidence, alongside the commercial effect of sorting loss.

References

• Key Technology: COMPASS for fresh and IQF corn, peas and green beans — example equipment description.
• Bühler: SORTEX F PolarVision — sensing, settings and reject-system features.
• TOMRA: vegetable processing spotlight — possible sorting positions and product demonstrations.
• Key Technology: COMPASS IQF vegetables brochure — configuration and application context.