Label Inspection

       Flash animation, 49 MB, ZIP

   After download click index.html 



            70 pages, 37 MB


The interaction of a material irradiated by electro magnetic waves, implies three different effects: reflection, refraction and diffraction.    Light irradiated by a LED most commonly operating infra-red (IR) at a wavelength of ~1μm, or visible red colour (see figure below).   Light source contained in the same package of the sensor, typically a photo-transistor, whose out feed enters an operational amplifiers.   Light irradiated to the label, from where it’ll be later reflected, refracted and diffracted.   Just a small fraction of the light incident to the bottle external sidewall shall be retro-reflected diffused toward the sensor, lying a few millimeters far from the LED source visible in the figure at left side.   Setting, as sensitivity threshold, a value intermediate between the retro-reflected signal when label is present, and the value measured when no label is present, it results possible to discriminate the presence or absence of labels on the container.   

  Commercial photoelectric proximity switch


     Spectral position of the infrared (IR) band adopted by photoscanners devoted to Label Presence Inspection

How Label's Colour influences the Scanning Distance

Refer to the figure below showing the behaviour of a red light emission photosensor in the range of distances in between sensor and label: (50 – 250) mm, the one typically used for label presence inspection.   

Three different re-emissions percentages associated to labels whose colours lie in three different tones:  

  1.      6 %, black, 
  2.    18 %, grey,
  3.    90 %, white.


  Re-emission characteristic curves of the photoscanner above at left side, when illuminating a target with red light, for a range of distances (50 –250) mm.  Visibly, only different distances sensor-label can optimize the detection following the colours' differences

Label Presence inspection at the Conveyor. Semi-transparent labels particularly dark, because of their blue dominant.  We saw ourselves compelled to direct the photoscanner toward the only one white colour detail written in Arab language.   Observe the angle <30º we adopted.   An higher angle increases favourably the Signal-to-Noise relation but it was not possible to use it here, again because of the semi-transparent nature of the labels

By the characteristic curves, it is possible to see that a black colour label, 6 % of re-emission, shall be optimally detected with an incoming signal sufficient to switch the outfeed contact for a scanning distance sensor to label 6 % of 250 mm.    If the material is black colour, 6 % of re-emission, the optimal distance for switching shall be 100 mm and the truly useful range shan’t be wider than 150 mm.   If the material is white, 90 % of re-emission, the switching shall be assured thru all range (50 – 250) mm, optimized around 125 mm.

 Semi-transparent labels.  Their coloured area particularly dark, because of its blue dominant.   We saw ourselves compelled to direct the photoscanner toward the only one white colour detail written in Arab language in an Electronic Inspector operating in a PepsiCo franchisee in Oman, Persian Gulf.   Observe the angle <30º we adopted.    An higher angle increases favourably the Signal-to-Noise relation but it was not possible to use it here, again because of the labels’ semi-transparent nature, implicitly causing low re-emission 

What above is directly hinting to several potential problems, when adopting the cheap label inspection with analog photo-sensor:

  • in the inspector there is no way to adjust the distance sensor - label when changing format and label’s colour;
  • no way to select the colour of each one format in the Bottling area: it is typically a Marketing Dept. who chooses their colour;
  • the labels are curved and moving, rather than flat as those used by the tests' base for the re-emission graphic above.  Implying that in the Electronic Inspectors there is time available only for a few measurements.   As a consequence, just a few values of the random variable shall be registered in conditions far from what the definitions of a “good measurement” in Physics.   

Yet these factors alone imply that the signal outcoming by the same photoelectric switch in the same electronic inspector, may be extremely wide for a clear label or extremely short for a dark.   Due to the fact that when signal duration is < 2 ms, these (signals) are interpreted as: “No label” signals by the inspectors, we have to be ready to understand the root causes for sometimes impressive false rejects. 

How many sensors to use ? 

They exists 3 types of label-container geometrical relations, requiring different amounts of photosensors.   The amount of sensors in the Table at right side assumes labels fully reflective and non-flagging.  

If semi-reflective, or partially transparent like in the figure below, their amounts have to be double than those marked in the table for cases: (360º,  < 180º) say in an amount of 4.   

 Sensors/Label/Container relation for one fully reflective non-flagging label.  If more than one label has to be detected, the amount marked in the column “Sensors” has to be proportionally increased

To apply four photosensors rather than one, increases only (1 - 5)% the total price of the electronic inspector, due to the low cost of the sensors.   This, when of the highest Quality photoscanners on the Market, cost ~230 € each (~300 $), and average cost being 40 €.   

As photoscanners of the “highest Quality on the Market”, we are meaning those (like, e.g. Omron’s model E3S-CD16): 

   Semi-transparent labels on PET bottles

In presence of semi-transparent labels, nor the visible couple of photo-sensors were not enough to prevent false rejects   

  • specifically designed for the transparent and semi-transparent labels used by the Beverage Bottling Companies;
  • operating with fuzzy-logic;
  • including a microprocessor, and not just an operational amplifier;
  • using polarised light, rather than non-polarised;
  • stainless steel case (IP65), and not plastic;
  • NPN and PNP, and not only NPN;
  • with glass front window, and not plastic easy to scratches;
  • with output short-cut protection;
  • preventing mutual interferences with other photoscanners;
  • designed 15 years ago, and not nearly 40 years ago.

The two figures below at right side, show an example where 2 photosensors (of good Quality) were not enough to detect semi-transparent labels for PET bottles filled with water.  

Detection ratio resulted a poor 80 % and false reject 0.01 % for flagging-labels.    Flagging-labels being the most frequent (> 90 %) in between all possible types of defect, and also the most difficult to detect too.   When choosing a standalone in-the-Conveyor label inspector on semi-transparent labels, it is a must to order the Flagging-Label option totaling 4 photo-sensors.   Flags nearly always are oriented downward by gravity.   Meaning that the change on reflection, the Signal, is observed in the optimal condition only directing the light beam exclusively to the upper side of the labels.

How to adjust sensitivity

The sensitivity to adjust by mean of the sensor trimmer, has to be kept close to 135°, say one-half of its 270º range, to optimise linearity of the internal amplifier.   The angle of the photo-sensor, with respect to the label, has to be in the range (15–35)º to prevent sensor over exposure.   

All these conditions can be summed implying an adjustment of the three parameters defining the sensitivity, so that:

actual value of the measurement  <  upper limit for measurement value  < length of the measur. window

lower limit for measur. value  <  length of the measurement window  <  actual value of the measurement

     Label inspection parameterization can only be “tailored” to the specific application during commissioning

Light beam has to be directed always to the top side of the labels.  Figure at left side shows how a camera on top of a bottle with a flagging label.   The erroneously applied label, folded and protruding out of the sidewall contour for approximately 4 mm, is easily detected by a camera.

 A label bottle with a label visibly flagging, as seen by a camera over its top

What is a “good Label inspection” ?

The correlated Label-Photoscanner State

  To establish a relation between a label (“Label”) and a Photoscanner (“Photoscanner”), both differentially related with the Environment (“Environment1”, Environment2”), is necessary Time.   Time to transform the previous state, in which all possible kinds of correlation of the Photoscanner coexist, in a following state in which the Photoscanner is “aware” to be correlated to a label, because having recorded eigenvalues for the eigenfunction ΦiS1 describing a cap.  Quantomechanical explanation of the measurement process, unaffected by the circularities implicit in the classic explanations.  The vertexes represent interferences

As we saw with plenty of details elsewhere in this web site, as an example we’ll try to determine if a bottle is labelled by mean of a common Photoscanner.   Photoscanner including a LED irradiating the container's sidewall or neck, where the label is expected to be, later receiving and processing just a fraction of the reflected diffused waves.  Whatever physical system, label included, is represented by its wave function or state vectors.    The physical meaning of the state vector becomes apparent when making a measurement.   Then the state of the system assumes one of the eigenstates, with probability given by the Born rule, and the result of the measurement is the corresponding eigenvalue.

Until eventual switching its digital output if the diffused retro-reflected waves’ intensity is over a threshold level pre-set during commissioning.

To definetely perceive a correlated Label-Photoscanner State, they are necessary:

  1. Timeto transform the previous state, in which all possible kinds of correlation of the Photoscanner coexist, in a following state in which the Photoscanner is “aware” to be correlated to a Label, because having recorded eigenvalues for the eigenfunction ΦiS1 describing a Label.  The correlation between the two systems, Photoscanner and Label, is progressively established during interaction and proportional to the natural logarithm ( ln t ) of the interaction time t.   An ideal correlation, corresponding to a maximised information of the Photoscanner about the Label, can only be reached allowing an infinite time. The fact we cannot wait for an infinite time causes the measurements’ fluctuations, a synonimous of the spectrum of the eigenvalues, resulting in the Electronic Inspector's false positives (false rejects).  Time, for what ?    Time to transform the previous state, in which all possible kinds of correlation (themselves, superpositions) of the Photoscanner coexist, in a following state in which the Photoscanner is aware to be correlated to a Label, because having recorded eigenvalues for the eigenfunction ΦiS1 describing a Label.   

 To detect a Label they are necessary Time and a kind of interaction which do not reduce the information in the marginal distribution of the Label.  All times bottle and label share the same material nature like this Russian plastic bottle and label, the application is more difficult and are necessary labels of high re-emission (   Baltika OAO, CARLSBERG Group/2014) 

  1. Interaction between the systems such that the information in the marginal distribution of the object inspected is never decreased.   In a probability distribution deriving by two random variables, we remember that marginal distribution is where we are only interested in one of them.   Otherwise, we’d have forcedly reduced the sample space where one of the random variables is derived and then, we could not have any more repeatability of the following measurements.    As an example, this should be the case if we’d erroneously try to use a beam of high energy neutrons, rather than LED's low energy photons, to interact with the Label.   The neutrons should modify the molecular structure of the Label, modifying its eigenstates and then the eigenvalues we expected to derive by the measurement.

What precedes is the unique scientific background existing for the fact that also double, triple or quadruple photoscanners, and also if of the best Quality (costing ~300 $ each), can result insufficient to reach satisfactory results.   Satisfactory in terms of defects' rejection ratio with minimum false positives.  Be wary of proposals for systems which could later leave only you and your Company with the nightmare to have the product store, filled with yet palletised flagging-, malpositioned- or missing-labels. 

Probability Multiplication Theorem 

A different, extremely cheaper approach in the figure below.   Here, four independent label inspections check the wrap-around label position by four different places.   Photoscanners’ independence being the key to satisfactory performances, without extra-costs, operative complexity and expensive spare parts, always associated to camera-systems.

  Flagging label inspection, by mean of four independent inspections, adopting the cheapest general-purpose photo-scanners.  Their outfeeds take advantage of the Probability Multiplication Theorem.  Application on PET bottles labelled with semi-transparent labels

Such a design profits of the implications of the Probability Multiplication theorem. Applied  to an Electronic Inspector (a Binary Classifier) composed by: 

  • a set of N inspections, each one identified by the suffix-i 
  • whose individual rejects converge to a single Rejector, 

the probability that each one of the N inspections fails to detect a single defect is reduced to:

In our case, the probability P1,2,3,4  of joint occurrence of four events 1, 2, 3, 4,  any collection of outcomes of an experiment:

  • whose respective probabilities are P1, P2, P3 and P4;
  • shaped following the normal (gaussian) distribution;
  • reciprocally independent;

is the product of their individual probabilities:

                              P1,2,3,4    =    P1  *  P2  *  P3  *  P4

Finally, after considering that each one of the N inspections has:  

                                                                    0 < detection ratio < 1

it results immediately evident an impressive reduction on the amount of defective containers falsely considered “correct” (False Negatives), proceeding to the Market.  

Case Study

5 Binary Classifiers in a single Label Presence Inspector

As an example, we’ll consider the case of = 5 (five) inspections, all of them featuring the same individual False Positive (false reject) ratio FPi < 0.01 %.    

Five independent Binary Classifiers, independently acting on a single rejector, whose individual:

  • accuracies; 
  • reproducibilities; 
  • specificities; 
  • sensitivities;

are such that the same property of an object is classified with Probability of False Negatives, say the undetected nor rejected but truly “defective” bottles:

                                                          10.00 %

Defects’ Detection Ratio

Then, the joint probability representing the Risk R that the five homogeneous systems (inspecting different projections of one and the same physical property, different sample spaces of the same state space) simultaneously fail to recognise one and the same defect, shall be limited to:

R1,2,3,4,5   =   Πi=1 Ri   =   0.001 %



  • that no more than 1 (one) non-labelled bottle in a row of 100000 non-labelled consecutive bottles out feeding by the Labeller Machine, shall not be detected defective by the Electronic Inspector, thus passing to the Market,
  • detection ratio >99.998 %. 

 The detection of the flagging labels represents an area where the Probability Multiplication Theorem aids the Food and Beverage Quality Control.  Greatly enhancing defects’ rejection and reducing the false rejects.  In the figure, the Yattala Brewery, Fosters Group and CUB, Australia, part of SAB MIller Plc (  Tom Parker/OneRedEye/2012)

False Rejects Ratio

We can look also on the other side, to the expected average value of the False Positives (false rejects) deriving by the superposition of the random variables measured by a Binary Classifier.    We have immediately different possible approaches, corresponding to our own prudential choice.  Choice referred to the level of confidence we attribute to the classification enacted by each individual Binary Classifier.   

For the value expected by the sum (or, superposition) of N independent random variables, following the confidence we attribute to the outcomes of our measurements, we’ll have at least three kinds of estimation:

  1. most prudential 
  2. prudential 
  3. less prudential.

  1. Imagine that we’d decide to adopt the most prudential definition of sum of N independent random variables, an Electronic Inspector (a Binary Classifier) composed by a set of N inspections whose individual rejects Ri converge to a single Rejector (or, ejector), each one identified by the suffix-i.     In this case, the probability that each one of the N homogeneous inspections simultaneously and erroneously classify “defective” (False Positive) their physical measurements of a container  has an upper bound which is their sum: 

          and, for i = 1, 2, …, 5 we’ll have:

                                     FPexpected %   < 0.05 %    

  1. Adopting the prudential definition of superposition of N independent random variables, we’ll have a correspondingly best estimation of the expectation value for the total False Positives (false rejects) of the entire system of Binary Classifiers, as:

          which let us estimate the Total False Positives ratio as:

                                   FPexpected %   =  0.022 %  

  1. Adopting the less prudential definition of superposition of N independent random variables, we’ll have a correspondingly estimation of the expectation value for the total False Positives (false rejects) of the entire system of Binary Classifiers, as:

         implying an estimation of the Total False Positives ratio:

                                  FPexpected %   =  0.01 %  


Well before to mislabel 100 000 consecutive bottles, the Rejects Accumulation Table of the Electronic Inspector, should be fully occupied by rejects.   The sensor of the Conveyor Control Cabinet managing this part of the Beverage Bottling Line, set in the Rejects Accumulation Table should control an automatic stop of the Labeller Machine.   

The straightforward application of the frequentist definition of Probability, silently assuming an infinite capability to accumulate rejected bottles, is countered by a reality where the:

  1. limited accumulation space, typically (80 - 300) standstill bottles, on the Rejects' Accumulation Table,
  2. existence of a sensor to detect that status and controlling Labeller Machine stop.

Then, it is better to get out of that classic definition of Probability and modernise our idea of expectation of the future Events to the relatively modern Bayesian viewpoint.   The expected Risk that a single non-labelled bottle pass to the Market, when consecutively inspected by five independent gaussian measurement systems, results still much lower than 0.001 %. 

Case Study 

Redundant inspection of Flagging-Labels 

Is it reasonable to expect these excellent values for the losses (false rejects <0.022 %) and for the defects’ detection ratio (>99.998 %) ?  

Yes.   We are using as a base for the following calculations a value for the Risk R = 10 % that a non-labelled bottles (equivalent to grossly-malpositioned label) is a False Negative.  Then truly defective but not detected.   The reality is that this 10 % is an excessively prudent one: missing labels are typically (2-sigma, 95 % of the cases) yet detected, leaving a Risk limited to <5 %.   Banally, we preferred to use the most prudent possible 10 % to account for all possible kinds of labels.   Thus including also the most difficult kinds existing: semi-transparents or transparents. 

Is it technically feasible the solution in the scenario presented above ?   

The answer is positive.  Since 2008 it is being applied also in some particularly fast and multiformat Nestle’ Waters® Bottling Lines, treating the wrap-around labels, like those shown in a figure here at right side.  Applied in more than one factory to Flagging-Label inspection over PET-containers moving at speed > 1.7 m/s.    We remember here that the coupling of a plastic bottle with a modern and typically-plastic wrap-around label, has plenty of esthetic advantages but is ill-fated when coming to the Electronic Inspection of the label.

  NESTLE’ WATERS® AQUAREL.  Yet in 2008 its production started to be deeply controlled for the presence of flagging labels, applying maximum redundancy to the inspection (   NESTLE’ WATERS®, 2014)

  S. PELLEGRINO® mineral water in  a glass bottle with paper label (   NESTLE’ WATERS®, 2014)


It provides lower Signal-to-Noise values to the infeed of the photoscanner’s phototransistor, quite banally because the optic properties (refraction, reflection and diffraction) of the plastic bottles result not particularly different by those of the labels.   Remaining in the sector of the mineral waters’ bottling, think to the traditional application of a paper label over a glass bottle.

Refer now to the figure below.  It shows the digital I/O card 0 (CS 1 and CS2) in one of the world’s most common models of Electronic Inspector.       


   I/O card 0, CS 1 and CS2 in the world’s most common model of Electronic Inspector has yet 8 digital inputs ready to apply label inspection photoscanners outputs, and its upgrade is then limited to a software enabling the display of the menus related to the additional photoscanners, fixing and photoscanners themselves

Within green coloured rectangles the 8 digital inputs ready to apply label inspection photoscanners outputs.  In other terms, and with reference to one of thsie problems permanently affecting the beverage Bottling Lines, to prevent flagging labels, 90% of all of the type of defects getting out of Labeller Machines, from reaching the Market, does not require complex camera based systems.  Hardware is there, waiting to be fully recognized in its usefulness.

What above is presenting a column of bold facts; facts where maximum Quality is obtained the lean-way jointly with maximum economy in: 

  1. an Electronic Inspector of the best Quality existing in the World, costing total 30000$ (or, 25000 €),
  2. equipped with five photoscanners of the best Quality costing <260 $/each (<230 €/each) and easily available, 
  3. checking Label Presence at-the-Conveyor
  4. costs in the end an amount which is four times less than a camera based final inspection system (~100000 €),
  5. reaching withiut any effort inspection performances on the Label Presence inspection close to those reached for the same kind of defect (missing or grossly-misplaced label) by the camera-equipped systems,
  6. with a lower False Reject ratio, then lower losses on Production,
  7. does not need the Advanced Training of the Electronics Maintenance Dept. technicians of the Bottling Plant, vital to assure constancy of the performances of the camera-equipped models, curbing the natural tendency of the camera-equipped models to increase their False Rejects' ratio,
    1. presents a Total Cost of Operation (TCO, futher informations here) much lower than the camera-equipped models: 
      1. all changeovers can easily be made by the Electronics Maintenance Dept. technicians of the Bottling Plant,
      2. there are not inspection protective panes nor strobo flashers to replace,
      3. proper maintenance, keeping apart rare cases of hardware fault, can always be assured by the Bottler’s Staff.

Links to the other pages:

This website has no affiliation with, endorsement, sponsorship, or support of Heuft Systemtechnik GmbH, MingJia Packaging Inspection Tech Co., Pressco Technology Inc., miho Inspektionsysteme GmbH, Krones AG, KHS GmbH, Bbull Technology, Industrial Dynamics Co., FT System srl, Cognex Co., ICS Inex Inspection Systems, Mettler-Toledo Inc., Logics & Controls srl, Symplex Vision Systems GmbH, Teledyne Dalsa Inc., Microscan Systems Inc., Andor Technology plc, Newton Research Labs Inc., Basler AG, Datalogic SpA, Sidel AG, Matrox Electronics Systems Ltd. 

                                                                                                                      .                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         .                   
TRUSTe Certified Privacy Seal

.                 .                                                                                         .                                                                                              .                                                                                                                                 66                                                                                 Webutation

                                                                                                                                                                                                                                                                                                                                                          .                                                                                                                         PRICE LIST       REFERENCES      SERVICES      CONTACT     MULTIMEDIA       TECHNOLOGIES       DEVELOPER       TERMS                                                                                                                                                                                                                                                                                                                         .                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           
TRUSTe Privacy Policy Privacy Policy
Site protected by 6Scan