阅读说明：本技术 用于对型坯进行光学检查的装置 (Device for optically inspecting a parison ) 是由 多纳托·拉伊科 西蒙娜·尼格罗 于 2019-12-12 设计创作，主要内容包括：一种用于对型坯(2)进行光学检查的装置(1),包括：发光器(3),其被构造为发出指向位于检查位置(10)处的型坯(2)的光束；探测器(4),其被构造为捕获介于发光器(3)与探测器(4)之间的型坯(2)的图像(10),其中发光器(3)包括发射偏振滤光器(32),其被构造为产生偏振光束,并且其中探测器(4)包括接收偏振滤光器(41),其被构造为接收偏振光束。(Device (1) for optical inspection of parisons (2), comprising: a light emitter (3) configured to emit a light beam directed at the parison (2) at the inspection position (10); a detector (4) configured to capture an image (10) of the parison (2) between the light emitter (3) and the detector (4), wherein the light emitter (3) comprises an emission polarizing filter (32) configured to generate a polarized light beam, and wherein the detector (4) comprises a reception polarizing filter (41) configured to receive the polarized light beam.)

1. Device (1) for optical inspection of parisons (2), comprising:

-a light emitter (3) comprising a light source (31) configured to emit a light beam directed at the parison (2) at an inspection position (10);

-a detector (4) comprising a camera (41) configured to capture an image (20) of the parison (2) at the inspection position (10), wherein the parison (2) is operatively interposed between the light emitter (3) and the detector (4) at the inspection position (10),

characterized in that said light emitter (3) comprises an emission polarizing filter (32) configured to intercept said light beam emitted by said light source (31) and generate a polarized light beam, and in that said detector (4) comprises a reception polarizing filter (42) configured to receive said polarized light beam, so that said parison (2) is operatively interposed between said emission polarizing filter (32) and said reception polarizing filter (42) at said inspection position (10).

2. The device (1) according to claim 1, further comprising a processing unit comprising:

a memory (5) containing a reference data set (51, 52);

a processor (6) programmed to process the image (20) captured by the detector (4) on the basis of the reference data set (51, 52) so as to derive a diagnostic indication (23) relating to a defect of the parison (2).

3. The apparatus (1) according to claim 2, wherein the processing unit is configured to:

-processing the image (20) captured by the detector (4) based on the reference data set (51, 52) to derive values of a plurality of image features (21) from the image (20);

-processing the values of said plurality of image features (21) to derive said diagnostic indication (23) relating to said defect of said parison (2).

4. The device (1) according to claim 3, wherein the processing unit is configured to:

-generating an image (22) reconstructed from values of the plurality of image features (21) and based on the reference data set (51, 52);

-deriving a diagnostic indication (23) related to a defect of the parison (2) by comparing the image (20) captured by the detector (4) with the reconstructed image (22).

5. Device (1) according to claim 4, comprising a self-learning system (7) configured to:

-receiving as input a plurality of said images (20) captured by said detector (4) for a corresponding plurality of said parisons (2);

-processing each image (20) of a plurality of images (20) captured by the detector (4) based on the reference data set (51, 52) to derive for each image (20) corresponding values of the plurality of image features (21) based on a predetermined criterion;

-generating a corresponding reconstructed image (22) for each image (20) of the plurality of images (20) based on the reference data set (51, 52) using corresponding values derived for the plurality of image features (21);

-comparing each image (20) of a plurality of images (20) captured by the detector (4) with the corresponding reconstructed image (22) so as to derive, for each image (20) of the plurality of images (20), a corresponding similarity parameter (24) representative of a similarity between the image (20) captured by the detector (4) and the corresponding reconstructed image (22);

-updating the reference data set (51, 52) for each image (20) of a plurality of images (20) based on the similarity parameter (24) and a predetermined threshold (72) for the similarity parameter (24).

6. The apparatus (1) according to claim 5, wherein the self-learning system (7) is configured to update the plurality of image features (21) based on the similarity parameter (24) and the predetermined threshold (72).

7. The apparatus (1) according to claim 5 or 6, wherein the predetermined criterion comprises a maximum number of image features (21) for the plurality of image features (21).

8. The apparatus (1) according to any one of claims 5-7, wherein the self-learning system (7) comprises a convolutional neural network.

9. The device (1) according to any one of the preceding claims, wherein the emission polarizing filter (32) is a linear filter configured to polarize light in a first polarization direction.

10. The device (1) according to any one of the preceding claims, wherein the receiving polarization filter (42) is a linear filter configured to polarize light in a second polarization direction.

11. A production line (100) for manufacturing containers of thermoplastic material, comprising a moulding machine (101) configured to manufacture parisons (2), characterized in that it comprises a device (1) for optical inspection of parisons (2) according to any one of claims 1-10, wherein said device (1) for optical inspection is operatively positioned downstream of said moulding machine (101).

12. A production line (100) for manufacturing containers of thermoplastic material, comprising a blow-moulding machine (103) configured to receive parisons (2) and blow-mould them in moulds to make said containers, characterized in that it comprises a device (1) for optical inspection of parisons (2) according to any one of claims 1-10, wherein said device (1) for optical inspection is operatively positioned upstream of said blow-moulding machine (103).

13. A method for optical inspection of a parison, comprising the steps of:

-emitting a light beam directed at the parison (2) at the inspection position (10) by means of a light emitter (3) comprising a light source (31);

-capturing an image (20) of the parison (2) at the inspection position (10) by means of a detector (4) comprising a camera (41), wherein the parison (2) is operatively interposed between the light emitter (3) and the detector (4) at the inspection position (10),

it is characterized by also comprising the following steps:

-generating a polarized light beam by intercepting said light beam emitted by said light emitter (3) on an emission polarizing filter (32) interposed between said light source (31) and said parison (2);

-receiving the polarized light beam on a receiving polarizing filter (42) interposed between the parison (2) and the camera (41),

wherein the parison (2) is operatively interposed between the emission polarizing filter (32) and the reception polarizing filter (42) at the inspection position (10).

14. The method according to claim 13, comprising a step of processing the image (20), wherein the processing step comprises the sub-steps of:

- (61) processing the image (20) captured by the detector (4) based on a reference data set (51, 52) so as to derive values of a plurality of image features (21) from the image (20);

- (62) generating an image (22) reconstructed by values of the plurality of image features (21) and based on the reference data set (51, 52);

- (63) deriving a diagnostic indication (23) related to a defect of the parison (2) by comparing the image (20) captured by the detector (4) with the reconstructed image (22).

15. The method according to claim 14, comprising a self-learning step comprising the sub-steps of:

-capturing a plurality of images (20) for a corresponding plurality of parisons (2);

-processing each image (20) of the plurality of images (20) based on the reference data set (51, 52) to derive, for each image (20) of the plurality of images (20), corresponding values of the plurality of image features (21) based on a predetermined criterion;

-generating a corresponding reconstructed image (22) for each image (20) of a plurality of images (20) using the corresponding values of the plurality of image features (21) and based on the reference data set (51, 52);

-comparing each image (20) of a plurality of images (20) with said corresponding reconstructed image (22) and deriving a corresponding similarity parameter (24) representative of the similarity between said image (20) captured by said detector (4) and said corresponding reconstructed image (22);

-updating the reference data set (51, 52) and the plurality of image features (21) based on the similarity parameter (24) and a predetermined threshold (72).

16. Method according to claim 15, wherein the images (20) of a plurality of images (20) captured by the camera (41) in the self-learning step represent a corresponding plurality of flawless parisons (2).

17. Method according to claim 15 or 16, comprising the step of supplying the parisons (2) of a plurality of parisons (2) to the inspection position (10) one at a time and according to a predetermined orientation with respect to the emission polarizing filter (32) and with respect to the reception polarizing filter (42).

18. The method according to any one of claims 13-17, wherein the emission polarizing filter (32) is a linear polarizing filter configured to polarize light in a first polarization direction.

19. Method according to claim 18, wherein the parisons (2) are oriented at the inspection position (10) with respective axes (a) parallel to the first polarization direction.

20. The method according to claim 18 or 19, wherein the receiving polarization filter (42) is a linear filter configured to polarize light in a second polarization direction different from the first polarization direction.

21. A method for processing an image of a parison captured by a probe (4), the method comprising the steps of:

-processing an image (20) captured by the detector (4) based on a reference data set (51, 52) to derive values of a plurality of image features (21) from the image (20);

-generating an image (22) reconstructed by the values of the plurality of image features (21) and based on the reference data set (51, 52);

-deriving a diagnostic indication (23) related to a defect of the parison (2) by comparing the image (20) captured by the camera with the reconstructed image (22).

22. Process for processing an image of a parison according to claim 21, comprising a self-learning step comprising the following sub-steps:

-capturing a plurality of images (20) for a corresponding plurality of parisons (2);

-processing each image (20) of the plurality of images (20) based on the reference data set (51, 52) to derive from each image (20) of the plurality of images (20) corresponding values of the plurality of image features (21) based on a predetermined criterion;

-generating a corresponding reconstructed image (22) for each image (20) of a plurality of images (20) using the corresponding values of the plurality of image features (21) and based on the reference data set (51, 52);

-comparing each image (20) of a plurality of images (20) with said corresponding reconstructed image (22) and deriving a corresponding similarity parameter (24) representative of the similarity between said image (20) captured by said detector (4) and said corresponding reconstructed image (22);

-updating the reference data set (51, 52) and the plurality of image features (21) based on the similarity parameter (24) and a predetermined threshold (72).

23. A computer program comprising operating instructions configured to perform the steps of the method according to claim 21 or 22 when run on a computer.

24. A method for processing an image of an article made of plastic material, said image being captured by a detector (4), said method comprising the steps of:

-processing the image (20) captured by the detector (4) based on a reference data set (51, 52) to derive values of a plurality of image features (21) from the image (20);

-generating an image (22) reconstructed by the values of the plurality of image features (21) and based on the reference data set (51, 52);

-deriving a diagnostic indication (23) related to a defect of the article by comparing the image (20) captured by the detector (4) with the reconstructed image (22).

25. The method according to claim 24, comprising a self-learning step comprising the sub-steps of:

-capturing a plurality of images (20) for a corresponding plurality of said items;

-processing each image (20) of the plurality of images (20) based on the reference data set (51, 52) to derive from each image (20) of the plurality of images (20) corresponding values of the plurality of image features (21) based on a predetermined criterion;

-generating a corresponding reconstructed image (22) for each image (20) of a plurality of images (20) using the corresponding values of the plurality of image features (21) and based on the reference data set (51, 52);

-comparing each image (20) of a plurality of images (20) with said corresponding reconstructed image (22) and deriving a corresponding similarity parameter (24) representative of the similarity between said image (20) captured by said detector (4) and said corresponding reconstructed image (22);

-updating the reference data set (51, 52) and the plurality of image features (21) based on the similarity parameter (24) and a predetermined threshold (72).

26. The method according to claim 25, wherein in the self-learning step the plurality of image features (21) are updated further based on the similarity parameter (24) and the predetermined threshold (72).

27. The method according to claim 25 or 26, wherein the predetermined criterion comprises a maximum number of image features (21) for the plurality of image features (21).

28. The method according to any one of claims 25-27, wherein the self-learning step comprises using a convolutional neural network.

29. The method according to any one of claims 25-28, wherein the image (20) of the plurality of images (20) captured by the detector (41) in the self-learning step represents a corresponding plurality of defect-free items.

30. Method according to any one of claims 24-29, wherein the article is a preform or parison (2).

31. A computer program comprising operating instructions configured to perform the steps of the method according to any one of claims 24-30 when run on a computer.

32. Device (1) for optical inspection of articles made of plastic material, comprising:

-a light emitter (3) comprising a light source (31) configured to emit a light beam directed at the parison (2) at the inspection position (10);

-a detector (4) comprising a camera (41) configured to capture an image (20) of the item located at the inspection position (10), wherein the item is operatively interposed between the light emitter (3) and the detector (4) at the inspection position (10);

-a processing unit comprising a memory (5) containing a reference data set (51, 52) and a processor (6) programmed to process the image (20) captured by the detector (4) based on the reference data set (51, 52) to derive values of a plurality of image features (21) from the captured image (20) and to process the values of the plurality of image features (21) to derive a diagnostic indication (23) related to a defect of the article,

characterized in that the processing unit is configured to generate a reconstructed image (22) from the values of the plurality of image features (21) and on the basis of the reference data set (51, 52), and to derive the diagnostic indication (23) related to a defect of the item (2) by comparing the image (20) captured by the detector (4) with the reconstructed image (22).

33. Device (1) according to claim 32, comprising a self-learning system (7) configured to:

-receiving as input a plurality of images (20) captured by the detector (4) for a corresponding plurality of the items;

-processing each image (20) of a plurality of images (20) captured by the detector (4) based on the reference data set (51, 52) to derive for each image (20) corresponding values of the plurality of image features (21) based on a predetermined criterion;

-generating a corresponding reconstructed image (22) for each image (20) of the plurality of images (20) based on the reference data set (51, 52) using the corresponding values derived for the plurality of image features (21);

-comparing each image (20) of a plurality of images (20) captured by the detector (4) with the corresponding reconstructed image (22), thereby deriving, for each image (20) of the plurality of images (20), a corresponding similarity parameter (24) representing a similarity between the image (20) captured by the detector (4) and the corresponding reconstructed image (22);

-updating the reference data set (51, 52) for each image (20) of a plurality of images (20) based on the similarity parameter (24) and a predetermined threshold (72) for the similarity parameter (24).

34. The apparatus (1) according to claim 33, wherein the self-learning system (7) comprises a convolutional neural network.

35. A production line (100) for manufacturing containers of thermoplastic material, comprising a moulding machine (101) configured to manufacture parisons (2), characterized in that it comprises a device (1) for optical inspection of articles according to any one of claims 32-34, wherein said device (1) for optical inspection is operatively positioned downstream of said moulding machine (101) and wherein said articles are said preforms (2).

Technical Field

The invention relates to a device for optically inspecting a parison.

A production line for manufacturing plastic containers, in particular bottles, generally comprises a moulding machine configured to form parisons by PET (polyethylene terephthalate) and a blow-moulding machine configured to blow-mould the parisons in moulds to make the containers. Some preforms may be defective, such as irregular thickness, holes, bubbles, or foreign objects; these defects must be detected by rapid inspection to quickly remove defective parisons from the line.

Background

From patent document EP2976204B1 a system for optical inspection of parisons is known, which is configured to inspect the parisons as they are conveyed by a conveyor towards a collection container; the system comprises a camera and a light source arranged so that the light source illuminates the parisons from behind and the camera captures an image of each parison from the front.

Other inspection systems are disclosed in the following documents: US2017/129157A1, DE102006047150A1, US2018/311883A 1. One limitation of prior art optical inspection systems is that they limit quality inspection to image features (e.g., bubbles) visible to the camera, but are unable to detect internal defects in the material, such as irregular residual stresses in the polymer chains of PET.

In addition, prior art systems detect defective parisons based on similarity to images of other defective parisons stored in the database. These systems must therefore be initialized with a database containing all the possible defects that need to be detected; however, it is difficult to find a complete and exhaustive database of this type, since the defects are diverse and the defective parisons constitute only a small part of the parisons produced.

Thus, in general, the systems of the prior art have limited reliability in identifying defective parisons.

Disclosure of Invention

It is an object of the present disclosure to provide an apparatus and method for optical inspection of parisons that overcomes the above-mentioned deficiencies of the prior art. It is another object of the present disclosure to provide a method for processing an image of a parison that overcomes the above-mentioned deficiencies of the prior art.

These aims are fully achieved, according to the present disclosure, by a device for optical inspection of parisons, by a method for optical inspection of parisons and by a method for processing images of parisons, as characterized in the appended claims.

More specifically, the present disclosure relates to a device for optical inspection of parisons (or preforms) or other articles made of plastic material (for example, caps or capsules) or other articles made of metal (for example, caps); in this respect, it should be understood that the remainder of the description made with reference to the parison also applies to other articles made of plastic or metal material. The word "parison" is used to denote an intermediate product in a process of manufacturing a plastic container (e.g., a beverage bottle). More specifically, the parison is formed by moulding (typically injection moulding or compression moulding) a plastics material and is expanded by blow moulding in a later stage to form the final container. The parison is made of a plastic material, preferably PET (polyethylene terephthalate).

An apparatus for optically inspecting a parison according to the present disclosure includes a light emitter. The illuminator includes a light source configured to emit a beam of light directed at a parison located at an inspection position.

In one embodiment, the apparatus includes an inspection chamber configured to receive a parison at an inspection location. In another embodiment, the inspection chamber configured to receive the parison at the inspection location is part of a production line that includes the apparatus, among other things.

The apparatus includes a detector. The detector includes a camera. The camera (i.e., detector) is configured to capture an image of the parison at the inspection location.

In one embodiment, the light source is configured to continuously emit a light beam. In one embodiment, the light source is a stroboscope and is configured to emit a light beam at predetermined emission intervals (each emission interval corresponding to the time taken to supply the parisons to the inspection position).

In one embodiment, the camera is configured to capture images at predetermined capture intervals (each capture interval corresponding to the time taken to supply the parisons to the inspection position); if the light source is a stroboscope, the emission interval corresponds to (i.e., is equal to) the capture interval.

At the inspection position, the parison is operatively interposed between the light emitter and the detector. In this way, the light emitter illuminates the parison from a first side (e.g., from the front), while the detector captures an image of the parison from a second side opposite the first side (e.g., from the back); the detector thus captures a backlit image of the parison.

Preferably, the light emitter (or said device) comprises an emission polarizing filter (or a first polarizing filter). The emission polarizing filter is configured to intercept a light beam emitted by the light source. The emission polarizing filter is configured to produce a corresponding polarized light beam from the light beam emitted by the light source.

Preferably, the detector (or said inspection device) comprises a receiving polarisation filter (or a second polarisation filter).

Preferably, in the inspection position, the parison is operatively interposed between the emission and reception polarizing filters.

The receive polarizing filter is configured to receive the polarized light beam. More specifically, the receive polarizing filter is configured to receive the beam polarized by the transmit polarizing filter and refracted by the parison and produce a second polarized beam. The camera thus receives the second polarized beam.

Since the material from which the parison is made (preferably PET) is characterized by birefringence, polarized light impinging thereon (from the first polarizing filter) is refracted according to the refractive index that varies based on stress when the polymer chains therein are oriented and subjected to stress. More specifically, the light hitting the parison is split into two rays oscillating in perpendicular planes, and the second polarizer allows only some parts to pass, brings them into the same plane and forms interference; thus, regions subjected to the same stress will have the same interference and therefore the same color, while regions subjected to different stresses will have different colors. Thus, the camera captures an image showing a color pattern representing the internal stress distribution in the parison.

A first and a second polarizing filter are provided which are arranged in parallel planes. The first polarization filter (or emission filter) provided is a linear polarization filter. In particular, the first polarization filter (or emission filter) is configured to polarize light in a first polarization direction. It is provided that the axis of the parison in the inspection position is oriented parallel to the first polarization direction (of the emission filter). The term "axis of the parison" here means the central axis of symmetry of the parison about which the side wall of the parison extends.

The polarization filter (or receiving filter) provided is a linear polarization filter; the second polarization filter (or receiving filter) is configured to polarize light in a second polarization direction. Preferably, the first and second polarized filters are both linear polarized filters. In one embodiment, the first and second polarization directions are parallel to each other (and the resulting light is white in this case). In one embodiment, the first polarization direction and the second polarization direction are different from each other. In particular, the first and second polarization directions may define an angle ranging between 5 ° and 50 °; for example, the first and second polarization directions may be perpendicular to each other (and the resulting light is black in this case). In another example, the first polarization direction and the second polarization direction may define an angle of 45 °. In another example, the first polarization direction and the second polarization direction may define an angle of 30 °. In another example, the first polarization direction and the second polarization direction may define an angle of 15 °.

In other embodiments, the first polarized filter and/or the second polarized filter are circular polarized filters.

In other embodiments, the first and second polarizing filters are not provided, and the camera thus captures a monochromatic, backlit image (with light and dark regions) of the parison.

The supplied parisons are supplied individually, i.e. one after the other, to an inspection location. For example, the inspection device provided is part of an apparatus (which may itself be the subject of the present disclosure) comprising a conveyor configured to continuously convey parisons along an inspection path (within which an inspection position is defined). The conveyor may include a suction belt defining a plurality of orifices and configured to contact an upper edge of the parisons to create a negative pressure (i.e., a vacuum) in the interior cavity of the parisons. The suction belt is configured to support the parison by the negative pressure. The suction belt is configured to move the parisons along the inspection path and position them one after the other in the inspection position. The light emitter and the emission polarizing filter may be disposed on a first side of the inspection path (i.e., the absorbent strip); the camera and the receive polarizing filter may be disposed on a second side of the inspection path (i.e., the absorbent strip) opposite the first side; the parison, supported by the suction band, is therefore interposed between the emission and reception polarizing filters.

In one embodiment, the apparatus includes a processing unit. The processing unit is connected to the detector.

The processing unit includes a memory containing reference data. The processing unit comprises a processor programmed to process the images captured by the probe on the basis of the reference data, so as to derive a diagnostic indication relating to a defect of the parison.

In one embodiment, the processor may include one or more criteria for identifying defective parisons based on one or more thresholds stored in the memory. For example, the processor may be configured to identify a defective parison in the event that the brightness of the image captured by the camera is above a predetermined threshold (in fact, an image with high brightness represents a parison with a very thin wall section) and/or in the event that the image represents a sharp color transition above a certain threshold.

In a preferred embodiment, the processing unit is configured to process the image captured by the detector (based on the reference data) to derive values of a plurality of image features from the captured image; this process encodes the image code (by means of transforming or compressing the image according to a predetermined algorithm, or obtained by the processing system in a self-learning step). Thus, the processing unit is configured to assign a value to each of the plurality of image features. In one embodiment, encoding the image includes reducing the size of the image (e.g., the plurality of image features may include 500 features).

These image features represent an image. Thus, the processing unit is configured to extract from the image a plurality of values assigned to a corresponding plurality of image features and reduce the image to these representative (or significant) values of the image.

In one embodiment, the processing unit is configured to process the values of the plurality of image features (based on the reference data) in order to derive a diagnostic indication relating to a defect of the parison.

In one embodiment, the processing unit is configured to classify the image based on values of a plurality of image features; for example, for each image feature, the memory may contain one or more typical values for a good (i.e., non-defective) parison and one or more typical values for a defective parison, and the processing unit may be configured to confirm the parison as defective when at least one value of the image feature is reasonably close (above a predetermined threshold) to the representative typical value for the defective parison and/or when a particular combination (e.g., multiplication) of the image features is reasonably close (above a predetermined threshold) to a reference value for the combination that represents the defective parison.

For example, the image characteristics may include whether a particular color or combination of colors is present and/or whether a particular symmetry and/or light intensity is present at a particular point; the processing unit may be configured to confirm the parison as defective when the image exhibits a particular color or combination of colors or has (or does not have) a particular symmetry or light intensity above or below a threshold at a particular point.

In one embodiment, the processing unit is configured to generate an image reconstructed from values of a plurality of image features (and based on the reference data). In one embodiment, the processing unit is configured to derive a diagnostic indication (based on the reference data) relating to a defect of the parison by comparing the image captured by the detector with the reconstructed image.

More specifically, the processing unit is configured to compare the image captured by the detector with the reconstructed image and derive a corresponding similarity parameter representing a similarity between the image captured by the detector and the reconstructed image. The processing unit is configured to compare the similar parameter with a predetermined threshold (which may itself be part of the reference data) and derive a diagnostic indication based on the comparison of the predetermined threshold with the similar parameter. For example, the processing unit may be configured to confirm the parison as good when the similarity parameter is above a particular similarity threshold. Thus, the processing unit is configured to confirm the parison as good when the reconstructed image is sufficiently similar to the initial image captured by the camera.

In fact, the processing unit trains (i.e. encodes) the image on the basis of a good, i.e. defect-free, parison, deriving the values of a plurality of image features and generating therefrom a reconstructed image. The processing unit may be trained by a self-learning system as described below, or it may be a pre-training unit commercially available. If the parison is good, the processing unit is able to correctly process its image and generate a reconstructed image similar to the original image; on the other hand, if the parison is defective, the processing unit based on good parison training cannot correctly reconstruct the image and therefore generates a reconstructed image that is significantly different from the original image.

Thus, the processing unit provided is trained by a sample of a single type (e.g., non-defective type) of item; subsequently, the processing unit will be able to distinguish between at least two types of articles (e.g., defective and non-defective).

In one embodiment, the apparatus comprises a self-learning system.

In one embodiment, the self-learning system is integrated in the processing unit. In one embodiment, the self-learning system is connected to the processing unit. In one embodiment, a self-learning system is connected to the memory.

The self-learning system is configured to receive as input a plurality of images captured by the detector for a corresponding plurality of parisons.

The self-learning system is configured to encode (based on the reference data) each of a plurality of images captured by the detector for a corresponding plurality of parisons to derive corresponding values of a plurality of image features from each of the plurality of images. Preferably, the self-learning system is configured to encode the image based on predetermined criteria (which may itself be part of the reference data).

The self-learning system is configured to generate, for each image of the plurality of images, a corresponding image reconstructed from corresponding values of the plurality of image features (and based on the reference data).

The self-learning system is configured to compare each image of the plurality of images with a corresponding reconstructed image, thereby deriving for each image a similarity parameter indicative of a similarity between the image captured by the detector and the corresponding reconstructed image.

The self-learning system is configured to update the reference data based on the similarity parameter such that the similarity parameter is below a predetermined threshold (in case the similarity parameter is proportional to the similarity between the images; conversely, above the predetermined threshold in case the similarity parameter is inversely proportional to the difference between the images). In one embodiment, the predetermined threshold is itself part of the reference data. Preferably, the self-learning system is configured to (iteratively) update the reference data based on the similarity parameter for each image of the plurality of images.

In a preferred embodiment, the self-learning system is configured to update the plurality of image features (preferably in combination with the reference data) based on the similarity parameter and a predetermined threshold. More specifically, the reference data and/or the plurality of image features are updated such that the similarity parameter is greater than a predetermined threshold for each of the plurality of images (in the case where the similarity parameter is proportional to the similarity between the images).

On the other hand, if the similarity parameter is directly proportional to the difference between the images (i.e. inversely proportional to their similarity), the self-learning system is configured to update the reference data and/or the image features such that the similarity parameter is below a predetermined threshold.

In a preferred embodiment, the predetermined criterion comprises (or is defined by) a maximum number (or a predetermined number) of image features of the plurality of image features. The predetermined criteria ensure that the system does not simply perform the function of identifying the captured image without regard to the imperfections of the parison.

Thus, in a preferred embodiment, the self-learning system is configured to determine both image features and reference data from images captured by the detector.

This system performs particularly well when the captured image represents a good parison; in this way, the system learns how to encode and reconstruct images of good parisons (i.e., the reference data used to encode the image features to be extracted (since they represent good parisons) and the reference data used for reconstruction). Subsequently, when the processing system is to encode and reconstruct a defective parison, it will not succeed and a reconstructed image will be generated that is completely different from the original image.

In another embodiment, the predetermined criteria includes (or is defined by) a plurality of image features (i.e., features from which values are extracted).

In one embodiment, the self-learning system includes a classifier. The classifier is configured to classify the captured image based on parameters such as color, color gradient, color standard deviation near the pixel, average color near the pixel, and the like. In particular, the classifier may be programmed to create a histogram representing the distribution of colors in an image and classify the image based on the symmetry and/or uniformity of the colors in the histogram. In one embodiment, it is provided that the parameters may be preset (i.e., well-defined) in the classifier. For example, the classifier provided is of the "one-class" type. The classifier may be configured to select one or more parameters to be used in the examination process among a plurality of preset parameters (such selection may be performed by the classifier in a learning phase).

In one embodiment, the classifier is a decision branch. In particular, the "one-class" classifier provided is a "random forest" type decision branch. It has been observed that these classifiers can be trained (only) from images of articles without any defects (since the classifiers are of the "single class" type).

In a preferred embodiment, the self-learning system comprises a neural network. In particular, the self-learning system may comprise a convolutional neural network.

"convolutional neural network" refers to a neural network configured to encode an image by deriving values for a plurality of image features through a series of convolution steps alternating with a series of pooling steps. In the convolution step, a convolution filter (the value of which is part of the reference data) is applied to the image (i.e. to a matrix of pixels representing each color of the image), resulting in a converted image; in the pooling step, the size of the converted image is reduced, for example, by maximizing or minimizing or averaging mathematical operations between adjacent pixels. The convolution step and the pooling step are thus used to obtain values for a plurality of image features.

In one embodiment, the neural network is pre-trained to extract (i.e., derive) values for a plurality of image features; in one embodiment, the neural network may comprise known neural networks (or portions thereof) configured to identify objects in the image (e.g., based on the dataset "ImageNet").

In one embodiment, the convolutional neural network is configured to classify an image based on a comparison of values of a plurality of image features with reference values (which form part of the reference data). In one embodiment, the image is classified as representing a good parison or a defective parison; in another embodiment, the image is classified as representing a good parison or a parison with a particular defect. In one embodiment, the neural network includes an "anomaly detection" classifier or a "fully connected" network to classify images based on values of a plurality of image features. For example, for extraction (i.e. deriving the values of the image features), a part of a known network based on a dataset such as "ImageNet" is used, which network comprises convolution and pooling (and is therefore a pre-trained network), and for classification of the images, a "fully-connected" network is used, which is trained by a self-learning system (preferably with examples of good parisons and examples of defective parisons) and is able to distinguish between good and defective parisons based on the values of the image features.

In one embodiment, the learning system includes a type of neural network known as a "generative confrontation network (GAN)"; the generative confrontation network comprises a generator and a discriminator; the generator uses the values of the plurality of image features (obtained from the actual image captured by the detector by the convolution step) to generate a corresponding reconstructed image and transmits it to the discriminator; the discriminator attempts to discriminate whether the image it receives is actual (i.e. captured by the detector) or reconstructed and sends a feedback result to the generator; based on the feedback results from the discriminator, the generator learns to generate reconstructed images that are as similar as possible to the actual images (so that the discriminator will regard them as actual). During training, the GAN preferably receives images of good parisons; thus, when the processing system inspects the parisons using the reference data (and image features) determined by the self-learning system, the discriminator identifies as reconstructed images only those images that are related to the defective parison.

In a preferred embodiment, the self-learning system includes a "self-encoder" convolutional neural network. In one embodiment, the reference data comprises a first reference data set and a second reference data set. The first reference data set is associated with a series of convolution steps (e.g., which includes a plurality of convolution filters); the second reference data set is associated with a series of upsampling steps. In particular, the second reference data set may be associated with a series of deconvolution steps (e.g., which include multiple deconvolution filters).

The autoencoder neural network is configured to extract (i.e., derive) values of a plurality of reference features from the image based on a first set of reference data associated with a series of convolution steps, and reconstruct the image using the values of the plurality of image features based on a second set of reference data associated with a series of deconvolution steps (i.e., generate a reconstructed image).

In the described embodiment comprising a self-encoder network, the learning system is configured to receive as input a plurality of images of good parisons to derive therefrom corresponding values of image features based on a first reference data set; the learning system is configured to generate a plurality of reconstructed images using the values of the image features and based on a second reference data set associated with the deconvolution step; finally, the self-learning system is configured to compare the plurality of reconstructed images with the corresponding plurality of original images and to update the reference data (in particular the first reference data set or the second reference data set or a combination of the first reference data set and the second reference data set) so as to minimize the difference between the original images captured by the detector and the corresponding reconstructed images. More specifically, in one embodiment, the self-learning system is configured to derive a similarity parameter for each image (e.g., calculated by one or a combination of two or more of norm I1, norm I2, "SSIM" structural similarity index, "PSNR" peak signal-to-noise ratio, "HaarpI" Haar wavelet-based perceptual similarity index), and to minimize the similarity parameter (i.e., make it below a certain threshold) if the similarity parameter is proportional to the difference between the images, or conversely, maximize the similarity parameter (i.e., make it above a certain threshold) if the similarity parameter is proportional to the similarity between the images. In the described embodiment comprising a self-encoder network, the learning system is configured to update the filters of the network (and hence the plurality of image features) in conjunction with the first reference data set and the second reference data set, to identify the image features that best represent images of defect-free parisons and to collectively identify the convolution step for obtaining these image features and the deconvolution step for generating a reconstructed image using these image features. Preferably, the self-learning system comprises a network of self-encoders that are constrained by at least one predetermined criterion (i.e., constraint) when encoding the image; for example, the criterion may relate to a maximum number of image features that may be identified for a plurality of image features. Thus, in practice, the encoding operation involves compressing the image.

Thus, in one embodiment, the self-learning system is configured to learn reference data and image characteristics that are subsequently used by the processing system to inspect the parison. In the described embodiments comprising a self-encoder network, the self-learning system is trained by a good (defect-free) plurality of parisons; thus, the self-learning system determines the reference data and image features used to correctly reconstruct the image of a good parison (i.e., with a high degree of similarity between the image captured by the detector and the corresponding reconstructed image); when these reference data and image features are applied to a defective parison, the reconstructed image is significantly different from the corresponding captured image and the processing system confirms the parison as defective when comparing them.

It has been observed that the comparison between the captured image and the reconstructed image is performed both during the network learning (or training) and during the examination; similar parameters used for training may be different from similar parameters to be used in the examination process. In particular, it is preferable to use the "PSNR", "haarpi" functions during the examination rather than during the training. For example, it is provided to use the "SSIM" function or norm I1 function or norm I2 function for training and the "PSNR" function or "haarpi" function for checking.

Preferably, the self-learning system (or processing system) comprises a first neural (sub-) network responsible for encoding the image to derive values of image features and a second neural (sub-) network responsible for decoding the values of image features to reconstruct the image; the first neural (sub) network and the second neural (sub) network are jointly subjected to training. The first neural (sub) network and the second neural (sub) network may be parts of a single neural network.

The present disclosure also provides a production line for manufacturing containers from thermoplastic material.

In one embodiment, a manufacturing line includes a molding machine configured to manufacture parisons. In one embodiment, the molding machine is an injection molding machine. In one embodiment, the molding machine is a compression molding machine (which may be a rotary machine).

In one embodiment, the production line comprises a thermal conditioning unit for heating and/or cooling the parisons. For ease of illustration, the thermal conditioning unit is hereinafter referred to as a "furnace," but does not limit the scope of the disclosure thereby. The oven is configured to receive the parisons sent out in the moulding machine and is equipped with heating means for heating the parisons.

In one embodiment, the manufacturing line includes a blow molding machine configured to receive the parisons and blow mold them in molds to form containers. Preferably, the blow molding machine is configured to receive the parisons heated in the oven. The furnace may be integrated in a blow molding machine.

In one embodiment, the blow molding machine and the parison moulding machine may be located in different production lines (even installed in separate plants) which work jointly to make containers of thermoplastic material: in fact, the parisons produced by the moulding machine are sent to a production line comprising a blow-moulding machine. The oven is preferably integrated upstream of the blow moulding machine in a production line comprising the latter to heat the parisons before they are blown.

In one embodiment, the production line comprises a storage unit (which may be automated) for receiving and storing parisons; the storage unit is configured to receive the parisons from the molding machine and to deliver them to the blow molding machine or oven.

In one embodiment, the production line comprises a device for optical inspection of parisons according to one or more aspects of the present disclosure. The optical inspection device is located downstream of the molding machine in the production line. The optical inspection device is located upstream of the blow molding machine in the production line. In one embodiment, the optical inspection device is located downstream of the molding machine and upstream of the furnace. In one embodiment, the optical inspection device is located downstream of the oven and upstream of the blow molding machine. Having the device upstream of the blow-moulding machine allows the defective parisons to be identified before they explode during the blow-moulding process due to their abnormal stress distribution. In one embodiment, the optical inspection device may be positioned in the storage unit or on a conveyor connecting the storage unit to other parts of the production line.

In one embodiment, the optical inspection device is positioned on the production line to inspect the parison at a temperature between 30 and 70 degrees celsius (preferably between 50 and 60 degrees celsius). For example, the parisons may be at such temperatures as they exit the molding machine.

In one embodiment, the optical inspection device is positioned on the production line to inspect the parison at ambient temperature (e.g., between 5 and 30 degrees celsius). The parisons can be stored or cooled to ambient temperature after they are formed.

Preferably, the optical inspection device is positioned on the production line to inspect the parison at a temperature below 60 degrees celsius (preferably below 50 degrees celsius); in fact, at higher temperatures, the parisons may deform, which may alter their stress distribution.

In one embodiment, the optical inspection device is integrated into a furnace. If the optical inspection device is integrated in the oven (preferably upstream of the heater, i.e. at the entrance of the oven), the positioning of the parisons in the oven can be exploited separately or sequentially.

In other embodiments, the optical inspection device according to the present disclosure is located off-line, integrated in a high speed viewer or low speed sampler.

The present disclosure also relates to a method for optical inspection of a parison. The optical inspection method comprises the step of emitting a light beam directed at the parison at the inspection position by means of a light emitter. The optical inspection method comprises the step of capturing an image of the parison at the inspection position by means of a probe. In the inspection position, the parison is operatively interposed between the light emitter and the detector. Thus, the captured image is a backlight image.

In one embodiment, the optical inspection method comprises a step of generating a polarized light beam by intercepting the light beam emitted by the light emitter on an emission polarizing filter interposed between the light emitter and the parison.

In one embodiment, the optical inspection method includes the step of receiving a polarized light beam on a receiving polarizing filter. In the inspection position, the parison is operatively interposed between the emission and reception polarizing filters.

In one embodiment, the parison is made of a material comprising polymer chains (preferably PET). In one embodiment, the image captured by the detector includes a color pattern representing the stress experienced by the polymer chains of the PET.

In one embodiment, the optical inspection method includes the step of processing the image (using a processing system). In one embodiment, in this processing step, the images captured by the detector are processed based on reference data contained in a memory. In one embodiment, the processing step comprises deriving a diagnostic indication relating to a defect in the parison.

In one embodiment, the processing step comprises the sub-step of encoding an image captured by the detector based on the reference data, thereby deriving values of a plurality of image features from the image.

In one embodiment, the processing step comprises the step of processing the plurality of image features to derive a diagnostic indication relating to a defect in the parison. More specifically, in one embodiment, the processing step includes the sub-step of generating an image reconstructed from values of a plurality of image features and based on the reference data. In one embodiment, the processing step comprises a sub-step of deriving a diagnostic indication relating to a defect of the parison by comparing the image captured by the detector with the reconstructed image.

In one embodiment, the method comprises a self-learning step (performed by a self-learning system connected to the processing system). In one embodiment, the self-learning step comprises the sub-step of capturing a plurality of images for a corresponding plurality of parisons. In one embodiment, the self-learning step comprises a sub-step of encoding each image of the plurality of images based on the reference data, comprising deriving from each image of the plurality of images a corresponding value of the plurality of image features based on a predetermined criterion (which may comprise a constraint on a maximum number of image features of the plurality of image features). In one embodiment, the self-learning step comprises the sub-step of generating, for each image of the plurality of images, a corresponding image by the corresponding values of the plurality of image features and reconstructing the corresponding image based on the reference data. In one embodiment, the self-learning step comprises the sub-step of comparing each image of the plurality of images with the corresponding reconstructed image and deriving a corresponding similarity parameter representing a similarity between the image captured by the detector and the corresponding reconstructed image. In one embodiment, the self-learning step comprises the sub-step of updating the reference data and/or the plurality of image features based on the similarity parameter and a predetermined threshold. More specifically, the reference data and the plurality of image features are updated such that the similarity parameter is below (or above) a predetermined threshold.

In one embodiment, the images of the plurality of images captured by the camera during the self-learning step represent a corresponding plurality of non-defective parisons. It is advantageous to have a self-learning system that does not need to receive an image of a defective parison as an input, because it is difficult to find a defective parison.

Preferably, the transmitting polarized filter and the receiving polarized filter are oriented with respect to each other according to the same predetermined orientation (e.g., parallel or perpendicular) as the self-learning step and the processing step.

In one embodiment, the method comprises the step of supplying one at a time (continuously) of the plurality of parisons to an inspection position. In one embodiment, the parison is supplied according to a predetermined orientation relative to the emission polarized filter and relative to the reception polarized filter. The orientation is the same as in the self-learning step and the processing step. An image of each of the plurality of parisons is captured while the parisons are in the inspection position.

In one embodiment, the processor is configured to process the randomly oriented images (e.g., rotate them to set them to a predetermined orientation); thus, in one embodiment, the orientation of the device with respect to the parison is constant.

In one embodiment, the processing system is configured to process the image of the parison in real time (immediately after the image is captured). In one embodiment, when the device provides a diagnostic indication that a parison is identified as defective, the production line is configured to stop or itself remove the defective parison in order to allow an operator to remove the defective parison.

In another embodiment, the processing system is configured to capture an image of the parison for processing in a post-processing mode. In this case, the system is configured to associate each image with a respective parison, so as to be able to identify parisons whose images represent a diagnostic indication of a defective parison.

The disclosure also relates to a method for processing an image of a parison. In one embodiment, the image processing method comprises the step of encoding an image, including deriving values for a plurality of image features. In one embodiment, the image processing method includes the step of generating a reconstructed image based on a plurality of image features. In one embodiment, the image processing method comprises a step of deriving a diagnostic indication relating to a defect of the parison by comparing the image captured by the camera with the reconstructed image.

In one embodiment, the image processing method comprises a self-learning step (according to one or more aspects of the present disclosure).

The present disclosure also relates to a computer program (software) comprising operating instructions configured (when run by a processor, in particular a processing unit of an apparatus according to one or more aspects of the present disclosure) to perform the steps of the processing method according to one or more aspects of the present disclosure.

Drawings

These and other features will become more apparent from the following detailed description of preferred embodiments, illustrated by way of non-limiting example in the accompanying drawings, wherein:

FIG. 1 illustrates an optical inspection apparatus according to the present disclosure;

FIG. 2 shows a process performed on an image by the optical inspection apparatus of FIG. 1;

FIG. 3 shows a self-learning procedure performed on an image by the optical inspection apparatus of FIG. 1;

4A, 4B and 4C show, respectively, an image captured by a camera, a reconstructed image and a comparison between the captured image and the reconstructed image for a parison that is defect-free;

5A, 5B and 5C show a camera-captured image, a reconstructed image and a comparison between the captured image and the reconstructed image, respectively, for a defective parison;

fig. 6 shows a production line for manufacturing containers of thermoplastic material comprising the apparatus of fig. 1.

Detailed Description

With reference to the figures, reference numeral 1 denotes an optical inspection device configured to inspect a parison 2.

The parison 2 comprises a body 200, which is substantially cylindrical in shape. The parison 2 (or the body 200) defines an axis of symmetry a. The body 200 is therefore cylindrically symmetric about the axis of symmetry a. The parison 2 comprises a closed bottom 201. The parison 2 includes a neck 202 defining an opening. The parison 2 comprises a ring 203.

The apparatus 1 is configured to receive parisons 2 at an inspection location 10. In one embodiment, the inspection position is defined by an inspection cavity. In one embodiment, the inspection chamber comprises a support element 11 configured to hold the parison 2 (preferably by supporting the ring 203).

The device 1 comprises a light emitter 3. The light emitter 3 comprises a light source 31. The light source 31 is configured to emit a light beam directed towards the parison 2 (i.e. towards the inspection position 10). The light emitter 3 comprises an emission polarizing filter 32. In one embodiment, an emission polarizing filter 32 is connected to the light source 31. The emission polarization filter 32 is configured to intercept and polarize the light beam emitted from the light source 31. Thus, the parison 2 receives the polarized beam from the emission polarizing filter 32 and refracts it.

The apparatus 1 comprises a detector 4. The detector 4 comprises a camera 41. The detector includes a receiving polarizing filter 42. In one embodiment, the receive polarizing filter 42 is connected to the camera 41. The reception polarization filter 42 is configured to receive and polarize the light beam refracted by the parison 2. Thus, the camera 41 receives the light beam polarized by the emission polarizing filter 32, refracted by the parison, and further polarized by the reception polarizing filter 42. The camera 41 is configured to capture (or acquire) an image 20 of the parison 2.

The light emitter 3 illuminates the parison 2 laterally on a first side 200A of the body 200. The probe 4 captures a transverse image of the parison 2 on a second side 200B of the body 200, opposite the first side 200A.

The apparatus 1 comprises a memory 5. The memory 5 contains reference data. More specifically, the memory 5 contains at least a first reference data set 51 and a second reference data set 52; in one embodiment, the first reference data set 51 and the second reference data set 52 are different from each other.

The apparatus 1 comprises a processor 6. The processor 6 is connected to the memory 5. The processor 6 is programmed to process the images 20 captured by the camera 41 on the basis of the reference data sets 51, 52, so as to derive a diagnostic indication 23 relating to the defects of the parisons 2. More specifically, the processor 6 is programmed to perform a step 61 of encoding the image 20 based on the first reference data set 51, thereby deriving values of the plurality of image features 21. The processor 6 is further configured to perform a step 62 of decoding the image feature 21 based on the second reference data set 52, thereby generating a reconstructed image 22.

The processor 6 is then configured to perform a step 63 of comparing the reconstructed image 20 with the captured image 22 to derive a diagnostic indication 23 relating to the defects of the parison 2.

In one embodiment, the diagnostic indication includes an error map 25 given the difference between (or conversely) the captured image 20 and the reconstructed image 22. In one embodiment shown in the figures, error map 25 shows uniform coloration when the parison is good or patchy coloration when the parison is defective.

In one embodiment, the diagnostic indication 23 includes a similarity parameter 24, the value of which is related to the similarity between the captured image 20 and the reconstructed image 22. In one embodiment, processor 6 is programmed to derive similar parameters 24 based on error map 25. In one embodiment, the diagnostic indication 23 comprises a binary parameter value that indicates whether the parison is good or defective (e.g., calculated by comparing the similar parameter 24 to a predetermined threshold).

In one embodiment, the apparatus 1 (or preferred processing system) includes a self-learning system 7. The self-learning system 7 is preferably integrated in the processor 6. The self-learning system 7 is connected to the memory 5.

The self-learning system 7 is configured to receive a plurality of images 20 captured for a plurality of corresponding parisons 2. The self-learning system 7 is preferably configured to perform the following steps for each image 20 it receives: a step 61 of encoding the image 20 based on the first reference data set 51 to derive a plurality of image features 21; step 62 of decoding the image feature 21 based on the second reference data set 51 to generate a reconstructed image 22; step 63, comparing the reconstructed image 22 with the captured image 20 to derive a similarity parameter 24 representing the similarity between the captured image 20 and the reconstructed image 22; step 70, evaluating the similarity parameter against a predetermined threshold 72 for the similarity parameter 24; the first reference data set 51, the second reference data set 52 and the image feature 21 are (iteratively) updated until the similarity parameter 24 is above (or below) the threshold parameter 72.

The self-learning system 7 thus solves the problem of optimizing the encoding operation 61 and the decoding operation 62, wherein variables are defined by the first reference data set 51 and the second reference data set 52 (and, if necessary, by the set of image features 21) such that the similarity parameter 24 is minimized, i.e. below a certain threshold (or maximized, i.e. above a certain threshold). Therefore, preferably, the first reference data set 51 and the second reference data set 52 are jointly updated.

Since the self-learning system 7 optimizes the encoding operation 61 and the decoding operation 62 by the image 20 of the good parison 2, the reference data set 51, 52 (and, if necessary, the set of image features 21) determined as a result of the optimization minimizes the difference between the captured image 20 and the reconstructed image 22 for the good parison 2. On the other hand, since these operations are not optimized for a defective parison 2, the reconstructed image 22 for the defective parison is significantly different from the captured image 20, and the processor 6 (knowing this difference) generates a diagnostic indication 23 indicating that the parison is defective.

Preferably, the steps of encoding 61, decoding 62, comparing 63, evaluating 70 and updating the reference data sets 51, 52 (and, if necessary, the image features 21) are performed iteratively by the self-learning system 7 for each successive image 20 (i.e. all iterations necessary to minimize or maximize the similarity parameter 24 are performed first for the first parison 2, then for the second parison 2 and so on). In one embodiment, the self-learning system may also perform a first iteration, in which it performs an encoding step 61, a decoding step 62, a comparison step 63 and an evaluation step 70 for all images 20; subsequently, starting from the similar parameters 24 obtained for all parisons 2, it updates the reference data sets 51, 52 (and, if necessary, the image features 21) and proceeds to perform a second iteration in which it again performs, for all images 20, an encoding step 61, a decoding step 62, a comparison step 63 and an evaluation step 70, and so on.

The present disclosure also relates to a production line 100 for manufacturing containers (e.g., bottles) of thermoplastic material.

The production line 100 includes a molding machine 101 configured for manufacturing (i.e., molding) parisons 2. In one embodiment, the molding machine 101 is a rotary machine. The line 100 also comprises a heating oven 102 configured to receive the formed parisons 2 and heat them. The production line 100 comprises a blow molding machine 103 configured to blow mold blanks 2 to make containers. In one embodiment, the blow molding machine 103 is a rotary machine.

Preferably, the line 100 comprises a first transfer carousel 106 configured to transfer the parisons 2 from the moulding machine 101 to the oven 102. Preferably, the line 100 comprises a second transfer carousel 107 configured to send the parisons 2 from the oven 102 to the blow-moulding machine 103. In one embodiment, the production line 100 comprises a storage unit 104 for storing the formed parisons 2 before they are blown. In one embodiment, the production line 100 includes a parison orientation device 105 configured to orient parisons 2 exiting and/or entering the storage unit 104. In one embodiment, the manufacturing line 100 includes a conveyor 108 configured to convey the parisons 2 into and/or out of the storage unit 104. The conveyor 108 conveys the parisons 2 from the storage unit 104 to the oven 102.