Method and system for performing quality control of objects in plant for producing objects in continuous cycle

Abstract

A method for performing quality control of objects in a plant for producing objects in a continuous cycle, comprising the steps of: capturing, for each object (O), an image (I); applying, for each image, a first processing step (102A) to classify the image and the corresponding object to one of the following two categories: a defective object category and a non-defective object category; if the image is attributed to a defective object category, a second processing step (102B) is applied to image data relating to the image, and the image and the corresponding object are further classified according to a plurality of defect categories.

Description

Technical Field

[0001] The invention relates to a method and a system for performing a quality control of objects in a plant for producing objects in a continuous cycle. Background Art

[0002] Quality control is critical in production lines, especially in high-volume production lines.

[0003] In these lines, some objects may be defective and quality checks must be performed on such objects before they leave the factory so that the defective objects can be removed.

[0004] Quality control can include manual visual inspection. However, this method is not precise enough and is often replaced by automated visual inspection.

[0005] Methods for automatically detecting defects in objects are known in the prior art. In these methods, one or more images of the object to be inspected are captured by optical means and any defects are identified as far as possible based on the image data. Furthermore, after the initial process of identifying defects, these methods usually also involve further data processing and analysis to classify the identified defects.

[0006] In this context, patent documents US20090324057A1, CN110349150, CN110838107, US2013129185 and WO2004111618 describe automated inspection methods for detecting defects. Patent document US2021 / 010953A1 discloses a system for high-speed inspection and inspection of objects using X-rays; the system focuses on inspecting integrated circuits by analyzing individual parts of the integrated circuits. However, the system is quite complex and does not allow real-time (online) control quality of objects manufactured at high production rates (which is a typical case in the field of rigid packaging).

[0007] In fact, in this field, there is a growing demand for methods capable of performing quality control of objects in a shorter time space and with higher precision.

[0008] It must be said that the invention can be applied in all fields where quality control of objects is required, such as the field of rigid packaging, where the products to be inspected for defects can be made of plastic (caps, parisons, containers...) or other materials (glass, aluminum, cans, tins...). Summary of the invention

[0009] One object of the present disclosure is to overcome the above-mentioned drawbacks of the prior art by providing a method and a system for performing quality control of objects in a plant producing objects in a continuous cycle.

[0010] This object is achieved entirely by the method and system for performing quality control of objects in a plant for producing objects in a continuous cycle according to the present disclosure, as characterized in the appended claims.

[0011] According to one aspect of the present disclosure, the present disclosure provides a method for performing quality control of objects in an apparatus for producing objects in a continuous cycle. The method includes the step of feeding the objects individually to an inspection station. The method includes the step of capturing an image of each object located in the inspection station.

[0012] In one example, the image is taken by an optical device. The optical device may include a camera. The optical device may include an illuminator for illuminating an object in the inspection station. The optical device observes the object located in the inspection station. Thus, when the object located in the inspection station is illuminated by the illuminator, an image of the object is taken by the optical device (camera). Preferably, the illuminator illuminates the object with light in the visible light or IR or UV spectrum.

[0013] The method further comprises the step of applying a first processing step to each image. The first processing step is performed to classify the image and the corresponding object into one of two categories: a defective object category and a non-defective object category.

[0014] If the image is classified as a defective object, the method includes applying a second processing step to image data associated with the image, and further classifying the image and the corresponding object according to a plurality of defect classes. The step of classifying the image and the corresponding object is performed based on a plurality of identification features. In one example, the plurality of identification features are extracted from the image data. The plurality of identification features are extracted from the image data in real time or alternatively in post-processing. Thus, the image data is processed to extract the plurality of identification features.

[0015] Such a concept allows defective objects to be separated from non-defective objects and furthermore allows the defects to be classified in a particularly efficient manner.

[0016] It should be noted that, since different types of defects in an object can be identified, the present disclosure also relates to taking actions in response to detected defects to adjust production equipment. The actions can be automatic or manual. In this way, the production equipment can be provided with a feedback control system. For example, criteria based on defect identification (e.g., criteria related to avoiding specific types of defects) can be used to update or adjust one or more control parameters (which control corresponding steps of continuous cycle production and / or update the settings of one or more components of the equipment).

[0017] In one example, the optical device includes a camera. In one example, the image of each object captured represents the visible appearance of the object. The image is taken by a camera.

[0018] In one example, the method comprises a step of storing in a database images belonging to the class of defective objects. Such a solution allows, for example, to refer to the database during the classification step.

[0019] As a result of the process of extracting multiple identification features from each image data, an array of each image data is generated, wherein the array includes the value of the identification feature of the image data. Such an array constitutes a unique mark of the image data and therefore constitutes a unique mark of the corresponding object. Multiple identification features define a working space, wherein each identification feature constitutes a dimension of the working space. Therefore, the working space has multiple dimensions. Each dimension of the multiple working space dimensions corresponds to a feature of the multiple identification features extracted from the image data of each image. In particular, the value of the identification characteristic of each image extracted defines the position of the image data of each image in the working space.

[0020] In one example, unsupervised clustering is used in the second processing step. During unsupervised clustering, each image data (image data associated with each image of the object captured) can be represented as a (data) point in the workspace (in fact, the array of image data provides a plurality of coordinates in the workspace). In one example, in unsupervised clustering, defect classes are generated by grouping data points with similar positions in the workspace. This approach allows the identification of different defect classes, including classes that were not considered before the quality control started. Furthermore, the step of generating defect classes allows the determination of the class with the highest number of defects.

[0021] "Unsupervised clustering" refers to a grouping system used to subdivide data points in a workspace into groups in an unsupervised manner.

[0022] It should be noted that classifying (or identifying) defects according to the present disclosure facilitates making the classification (or identification) results available to the user in a particularly simple and easy-to-read manner. For example, the output of the unsupervised classification may be a report (or graph) on different types of defects on the identified objects (e.g., considering a population of objects). Thus, the output does not need to be checked by a professional technician, and even a non-professional operator can read the output to see what the different types of defects are and the number of defects in each defect category.

[0023] According to another aspect of the present disclosure, unsupervised classification (ie, the clustering step) can be started at any time. In this way, a "continuous classification" system can also be created.

[0024] In practice, by means of unsupervised clustering, the classification of objects may be repeated each time an object is identified as defective, or at predetermined time intervals, or after a certain number of objects have been identified as defective, or according to other predetermined criteria.

[0025] According to another aspect, if an object is classified as defective in a first processing step, but is not identified as belonging to one of the identified defect categories in a second processing step, the system (due to unsupervised clustering) can create new defect categories (clusters) in the workspace. Thus, new defect categories can be added to continuously (i.e., the entire time the device is in operation) update the existing categories.

[0026] In one example, images classified as non-defective objects are excluded from the storing step.

[0027] This approach allows using less database space. In addition, by not storing data related to non-defective objects, the quality control process is faster.

[0028] In one example, the first processing step provides position information. The position information is related to the position of the defect in each defective object.

[0029] In one example, the position information is fed to a second processing step. This information can be used to classify the defect.

[0030] Additionally, in one example, the plurality of identification features includes at least one feature representing location information.

[0031] In one example, the first processing step is performed by a machine learning model. The machine learning model is trained to classify each image as either a defective object or a non-defective object. The machine learning model is trained based on training data. The training data may only include images of non-defective objects. This approach allows the machine learning model to be trained using images of non-defective objects. Defects can therefore be identified without requiring a complete defect database.

[0032] Furthermore, the first processing step may include, for each image, extracting diagnostic signatures from the image data and applying predetermined diagnostic rules (ie algorithms).

[0033] In one example, the first processing step includes, at a first stage, a machine learning model. The machine learning model is trained to classify each image as either a defective object class or a non-defective object class. The machine learning model is trained based on training data. In one example, the training data may include only images of non-defective objects. The first processing step may also include, at a second stage, for each image, extracting diagnostic markers from the image data and applying predetermined diagnostic rules (i.e., algorithms). In addition, both the results of the first stage and the results of the second stage of the first processing step are considered to classify the images and corresponding objects as either defective object classes or non-defective object classes.

[0034] In one example, in the second processing step, output data of the first stage and the second stage of the first processing step are received and processed in combination with each other.

[0035] In one example, in a first processing step, both the first stage and the second stage are applied to the image data acquired from each object.

[0036] In another example, in the first processing step, the image data of each object may be divided into a first subset and a second subset according to a predetermined criterion. In this scheme, for each object, the first stage is applied to the first subset and the second stage is applied to the second subset.

[0037] In one example, the first processing step may also include a plurality of tasks that provide a corresponding plurality of conditions related to the object to be inspected according to a predetermined order. In this scheme, the first set of tasks is performed by a machine learning model, and the second set of tasks is performed by extracting diagnostic markers from the image data and applying predetermined diagnostic rules.

[0038] According to one aspect of the present disclosure, the present disclosure also provides a system for performing quality control of objects in a device for producing objects in a continuous cycle. The system for performing quality control of objects in a device for producing objects in a continuous cycle (hereinafter, the system) includes an optical device. The optical device is configured to capture an image of each object located in an inspection station. The system may include a conveyor. The conveyor is configured to feed the objects to the inspection station individually. The system also includes a processing unit. The processing unit is programmed to process each image in a first processing step. The processing unit is programmed to classify the image and the corresponding object into one of the following two categories: a defective object category and a non-defective object category.

[0039] The processing unit is further configured to, in a second processing step, process the image data associated with each image that is classified as a defective object in response to the result of the first processing step, so as to classify the image and the corresponding object according to a plurality of defect classes. The second processing step is performed based on a plurality of identification features. In one example, the plurality of identification features are extracted from the image data.

[0040] In one example, the system includes a storage unit configured to store images belonging to a defective object category in a database.

[0041] In one example, the processing unit is configured to perform unsupervised clustering in the second processing step. The unsupervised clustering is programmed to define a workspace. The workspace has a plurality of dimensions. Each dimension corresponds to one of a plurality of identification features extracted from the image data of each image. The value of the identification feature extracted for each image defines the position of the image data of each image in the workspace, so that each image data is shown as a data point in the workspace. Therefore, the unsupervised clustering is programmed to show each image data as a data point in the workspace. The unsupervised clustering is programmed to generate defect categories by grouping data points having similar positions in the workspace.

[0042] In one example, the processing unit is configured to obtain position information related to the position of the defect in each defective object in the first processing step.

[0043] In a first processing step, the processing unit may include a machine learning model. The machine learning model is trained to classify each image as a defective object or a non-defective object. In one example, the machine learning model is trained based on training data. In one example, the training data may include only images of non-defective objects.

[0044] Furthermore, in the first processing step, the processing unit may comprise extracting diagnostic signatures from the image data and applying predetermined diagnostic rules (ie algorithms) for each image.

[0045] In a first processing step, the processing unit may include a first stage machine learning model. The machine learning model is trained to classify each image as a defective object or a non-defective object. The machine learning model is trained based on training data. In one example, the training data may include only images of non-defective objects.

[0046] Furthermore, in the first processing step, the processing unit may include extracting diagnostic markers from the image data and applying predetermined diagnostic rules for each image in a second stage.

[0047] In one example, both the results of the first stage and the results of the second stage of the first processing step are considered to classify the image and the corresponding object into a defective object class or a non-defective object class.

[0048] In one example, in the second processing step, output data of the first stage and the second stage of the first processing step are received and processed in combination with each other.

[0049] According to one aspect of the present disclosure, the present disclosure provides an apparatus for producing objects in a continuous cycle. The apparatus includes one or more machines for producing objects. The apparatus also includes a system for performing quality control of the objects. A system for performing quality control of objects is made according to the present disclosure.

[0050] According to one aspect of the present disclosure, the present disclosure provides a computer program, which includes instructions configured to perform quality control of objects in a device for producing objects in a continuous cycle according to the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] These and other characteristics will become more apparent from the following description of a preferred embodiment illustrated by way of non-limiting example in the accompanying drawings, in which:

[0052] - Figure 1 A system for performing quality control of objects in a plant producing objects in a continuous cycle according to the present disclosure is shown;

[0053] - Figure 2 The steps of processing the captured image of the object to perform quality control are shown;

[0054] - Figure 3 A first stage and a second stage of the processing steps are shown. DETAILED DESCRIPTION

[0055] With reference to the drawings, numeral 1 denotes a system for performing quality control of objects in a device producing objects O in a continuous cycle. The system 1 comprises an optical device 101. The optical device 101 is configured to capture an image I of each object O located in an inspection station IP. In another example, the optical device 101 is configured to capture multiple images of each object. In one example, the optical device 101 comprises a camera.

[0056] The optical device 101 comprises an illuminator for illuminating the object in the inspection station. Preferably, the illuminator is configured for illuminating the object in the inspection station IP with light in the visible light or IR or UV spectrum. Furthermore, the optical device comprises a camera configured for observing the object in the inspection station.

[0057] In one example, the image I acquired of each object O located in the inspection station represents the visible appearance of the object. In other words, in one example, each object is illuminated with light in the visible spectrum. In one example, the object may be exposed to infrared radiation. According to one example, in the inspection station, the object O is exposed to light in the visible or IR or UV spectrum (more generally, light in a spectrum other than X-rays). In particular, the image I acquired of each object shows the visible aspects of the object. The image I acquired of each object is an image of the entire object (represents the entire object).

[0058] The system 1 may also include a conveyor C. The conveyor C is configured to feed objects O individually to the inspection station IP. In other words, in a preferred embodiment, the objects are fed to the inspection station one at a time. In this scheme, each object is conveyed to the inspection station in a predetermined orientation. In addition, each object may be illuminated according to a predetermined orientation in the inspection station. In another example, the conveyor may be configured to feed objects in a disordered flow so that more than one object is present in the inspection station at any one time. Therefore, a captured image of an object in the inspection station may include more than one object. The conveyor is configured to feed objects O along a feed direction F. The system also includes a processing unit 102. The processing unit 102 is programmed to process each image in a first processing step. The processing unit 102 is configured to classify the image I and the corresponding object O into one of the following two categories: a defective object category and a non-defective object category.

[0059] The processing unit 102 is further configured to process the image data associated with each image I belonging to the defective object category in a second processing step. The processing unit 102 is configured to process the image data associated with each image I belonging to the defective object category in response to the result of the first processing step. In one example, the processing unit 102 is configured to process only the image data associated with the image belonging to the defective object category. The processing unit 102 is configured to perform the second processing step to classify the images and the corresponding objects according to a plurality of defect categories. The processing unit 102 is configured to classify the images and the corresponding objects based on a plurality of identification features. In one example, a plurality of identification features are extracted from the image data.

[0060] In one example, the processing unit 102 includes a storage unit. The storage unit 1021 is configured to store images belonging to the defective object category in the diagnosis database. In addition, in one example, images belonging to the non-defective object category are eliminated. Preferably, the storage unit includes a non-volatile memory.

[0061] In one example, the processing unit is configured to obtain position information related to the position of the defect in each defective object O in a first processing step.

[0062] In one example, in the first processing step 102A, the processing unit 102 includes a machine learning model. The machine learning model is trained to classify each image as a defective object or a non-defective object. The machine learning model is trained based on training data. In one example, the training data only includes images of non-defective objects.

[0063] Furthermore, in a first processing step, the processing unit comprises for each image extracting a diagnostic signature from the image data. In a first processing step, the processing unit comprises for each image applying a predetermined diagnostic rule.

[0064] More specifically, each image taken of the object to be inspected provides a set of pixels, and defects can take the form of inconsistent areas, for example, different values ​​of luminous intensity or color in contrast to the area containing it (defect-free area). The value of the contrast depends on the variability of the intensity compared to the defect-free object and is usually different at each point of the image to be inspected. This definition applies to defects in the form of uniform color areas.

[0065] For textured areas, defects can take the form of pixel areas containing variations (i.e., contrast) when compared to shading or color variations (including position variations) that are considered normal on a non-defective object. Thus, texture variations can be considered defects.

[0066] In one example, in the first processing step 102A, the processing unit 102 includes a machine learning model in the first stage. The machine learning model is trained to classify each image as a defective object or a non-defective object. The machine learning model is trained based on training data. The training data only includes images of non-defective objects.

[0067] Furthermore, in one example, in the second processing step 102B, the processing unit 102 includes extracting diagnostic markers from the image and applying predetermined diagnostic rules for each image in the second stage.

[0068] More specifically, both the results of the first stage and the results of the second stage of the first processing step are considered to classify the image and the corresponding object into a defective object class or a non-defective object class.

[0069] In one example, the processing unit 102 is configured to perform unsupervised clustering in a second processing step 102B. The unsupervised clustering is programmed to show each image data as a data point in a workspace. In particular, the workspace has multiple dimensions. Each dimension corresponds to a feature of multiple identification features. The value of the identification feature extracted for each image (providing an array of multiple coordinates in the workspace constituting the image to identify the points in the workspace) defines the position of the image data of each image in the workspace, so that each image data can be represented as a (data) point in the workspace. The unsupervised clustering is programmed to generate defect categories by grouping data points with similar positions in the workspace. Therefore, according to an example, for each image of each object acquired, multiple identification features are extracted from the image data of each image, and the value of each feature extracted from the image data of each image determines the position of the image data of the image in the workspace.

[0070] More specifically, "unsupervised clustering" refers to a grouping system for subdividing data points in a workspace into groups in an unsupervised manner. In other words, an unsupervised grouping system (i.e., unsupervised clustering) divides the data points in the workspace based on the location similarity of the data points in the workspace, and thus creates different groups of data points (or defect classes) in the workspace. Unsupervised clustering may also label each group of data points in the workspace.

[0071] According to one aspect of the present disclosure, the present disclosure provides a method for performing quality control of an object O in an apparatus for producing objects in a continuous cycle. The method includes the step of feeding the objects O individually to an inspection station IP. In other words, in a preferred embodiment, the objects are delivered to the inspection station one at a time. In this scheme, each object is delivered to the inspection station in a predetermined orientation. The method may also include the step of irradiating each object according to a predetermined orientation in the inspection station. In another example, the method may include the step of feeding the objects in a disordered flow so that more than one object is present in the inspection station at any one time. Therefore, the image of the object taken in the inspection station may include more than one object. The method includes the step of capturing an image I of each object O located in the inspection station IP. The method also includes the step of applying a first processing step 102A to each image. The first processing step 102A is performed to classify the image and the corresponding object into one of the following two categories: a defective object category and a non-defective object category. The first processing step may include extracting features from the input data (i.e., the image). The extracted data is processed in the first processing step to detect defects. In addition, if the image is classified as a defective object, the method includes the step of applying a second processing step 102B to the image data. The image data is from each image. The image data may be data from an original image of the object. The image data may be semi-processed data derived from the first processing step 102A. The second processing step 102B is performed to classify the image I and the corresponding object O according to a plurality of defect categories. The classification of defects is performed based on a plurality of identification features. In one example, identification features are extracted from the image I. For example, the second processing step 102B is applied only to objects that are classified as defective objects.

[0072] The method further comprises the step of storing images that are classified as defective objects in a (diagnostic) database. In one example, images that are classified as non-defective objects are excluded from the storing step. Preferably, images that are classified as non-defective objects are eliminated.

[0073] In one example, the first processing step 102A is performed by a machine learning model. The machine learning model is trained to classify each image as either a defective object or a non-defective object. The machine learning model is trained based on training data. In one example, the training data may include only images of non-defective objects.

[0074] In addition, the first processing step 102A may include extracting diagnostic markers from the data of the image I for each image I. The first processing step 102A may include applying predetermined diagnostic rules. In one example, for each image, a map of the diagnostic markers extracted from each image is obtained. Predetermined rules (or algorithms) are applied to the map to identify defects (if any) in the image and in the corresponding object.

[0075] For example, a defect may be in the form of a brightness variation of a portion of the image (resulting in a contrast in that area). Thus, detecting such a difference on the map means that a defect has been detected.

[0076] In one example, the first processing step 102A includes a machine learning model in the first stage 1021A. In addition, the step of extracting diagnostic markers is performed in the second stage 1022A of the first processing step 102A. In one example, the processing steps of the first stage and the second stage are performed simultaneously.

[0077] Furthermore, during the first processing step 102A, position information related to the position of the defect in each defective object is obtained. The position information is fed to the second processing step 102B. For example, the plurality of identification features includes at least one feature representing the position information. In one example, the position information is obtained in the second stage of the first processing step 102A.

[0078] In one example, unsupervised clustering is used in the second processing step 102B. During unsupervised clustering, each image data is represented as a data point in a workspace, and a defect category is generated by grouping data points with similar positions in the workspace. In one example, if at least two different data points in the workspace cannot be identified, the system attempts to show a second set of data points to distinguish the groups of data points. Through the interface, the user can also add new defect categories or modify (highlight, separate or mark) identified defect categories. In one example, both the results of the first stage and the results of the second stage of the first processing step 102A are considered to classify the image and the corresponding object into a defective object category or a non-defective object category.

[0079] In one example, in the second processing step, the output data of both the first stage and the second stage of the first processing step 102A are sent as input to the second processing step. More specifically, in the second processing step 102B, the output data of the first stage and the second stage of the first processing step 102A are received and processed in combination with each other.

[0080] In one example, in a first processing step, both the first stage and the second stage are applied to each image data obtained from each object. Thus, each image obtained from each object can be checked and classified as a defective object class or a non-defective object class by a machine learning model or by extracting diagnostic markers from the image data and applying predetermined diagnostic rules.

[0081] In addition, in one example, in the first processing step, the image data of each object can be divided into a first subset and a second subset according to a predetermined criterion. In this scheme, for each object, the first stage is applied to the first subset and the second stage is applied to the second subset. For example, a predetermined portion of the object is analyzed using a machine learning model, and another predetermined portion is analyzed by extracting diagnostic markers from the image data and applying predetermined diagnostic rules (e.g., artificial neural networks). Therefore, according to the predetermined criterion, the image data of each object can be subdivided into a first subset and a second subset, and a combination of the first stage and the second stage can be applied to each object.

[0082] The first processing step may also include a plurality of tasks. The plurality of tasks may provide a corresponding plurality of conditions to be satisfied according to a predetermined order. The plurality of conditions to be satisfied may be related to the object to be inspected. In such an approach, the first set of tasks may be performed by a machine learning model, and the second set of tasks may be performed by extracting diagnostic markers from the image data and applying predetermined diagnostic rules.

[0083] According to one aspect of the present disclosure, the present disclosure provides an apparatus for producing objects in a continuous cycle. The apparatus comprises one or more machines for producing objects. The apparatus also comprises a system 1 for performing quality control of the objects, wherein the system 1 is a system according to the present disclosure.

[0084] According to another aspect of the present disclosure, the present disclosure provides a computer program, wherein the computer program includes instructions configured to execute the steps of the method according to the present disclosure.

Claims

1. A method for performing quality control of objects in a device for producing objects in a continuous cycle, the method comprising the following steps: feeding said objects (O) individually to an inspection station (IP); For each object (O) located in said inspection station (IP), capturing an image (I); For each image, a first processing step (102A) is applied to classify the image and the corresponding object into one of two categories: a defective object category and a non-defective object category; If the image is classified as a defective object, a second processing step (102B) is applied to the image data associated with the image, and the image and the corresponding object are further classified according to a plurality of defect categories based on a plurality of identification features extracted from the image data.

2. The method according to claim 1, wherein: The optical device (101) illuminates the object in the inspection station (IP) with light in the visible or IR or UV spectrum and comprises a camera, wherein the camera observes the object and takes the image (I) of the object.

3. The method according to claim 1 or 2, comprising the step of storing said images belonging to said class of defective objects in a database.

4. The method according to claim 3, wherein: Unsupervised clustering is used in the second processing step (102A), wherein during the unsupervised clustering, a workspace having multiple dimensions is defined, wherein each dimension corresponds to a feature of the multiple identification features extracted from the image data of each image, wherein the values ​​of the identification features extracted for each image define the position of the image data of each image in the workspace, so that each image data is represented as a data point in the workspace, and the defect category is generated by grouping data points having similar positions in the workspace.

5. The method according to claim 3 or 4, wherein: The images classified as non-defective objects are excluded from the storing step.

6. A method according to any one of the preceding claims, wherein: During said first processing step (102A), position information relating to the location of a defect in each defective object is obtained.

7. The method according to claim 6, wherein: The position information is fed to the second processing step (102B).

8. The method according to claim 7, wherein: The plurality of identification features include at least one feature representing the location information.

9. A method according to any one of the preceding claims, wherein: The first processing step (102A) is performed by a machine learning model, which is trained to attribute each image to the defective object category or the non-defective object category, wherein the machine learning model is trained based on training data that only includes images of non-defective objects.

10. A method according to any one of the preceding claims, wherein: The first processing step (102A) comprises, for each image, extracting diagnostic markers from the image data and applying predetermined diagnostic rules (algorithms).

11. A method according to any one of the preceding claims, wherein: The first processing step (102A) comprises: a machine learning model trained to classify each image as belonging to the defective object class or the non-defective object class, and For each image, in a second stage (1022A) diagnostic markers are extracted from the image data and predetermined diagnostic rules are applied, Therein both the result of the first stage and the result of the second stage of the first processing step are taken into account in order to classify the image and the corresponding object into the defective object class or the non-defective object class.

12. The method according to claim 11, wherein: During the first processing step, one of the following conditions occurs: i) applying both the first stage and the second stage to image data acquired from each object; ii) dividing the image data of each object into a first subset and a second subset according to a predetermined criterion, wherein, for each object, the first stage is applied to the first subset and the second stage is applied to the second subset; iii) The first processing step includes a plurality of predetermined tasks, wherein the plurality of predetermined tasks provide a corresponding plurality of conditions to be satisfied according to a predetermined order, wherein a first group of tasks among the plurality of tasks are performed by the machine learning model, and a second group of tasks among the plurality of tasks are performed by extracting diagnostic markers from the image data and applying predetermined diagnostic rules.

13. A system (1) for performing quality control of objects in a device for producing said objects in a continuous cycle, the system comprising: an optical device (101) configured to capture an image (I) of each object (O) located in the inspection station (IP); a conveyor (C) for feeding the objects (O) individually to the inspection station (IP); A processing unit (102), the processing unit (102) being programmed to: In a first processing step (102A) each image (I) is processed to classify said image (I) and the corresponding object (O) into one of two classes: a class of defective objects and a class of non-defective objects, In a second processing step (102B), responsive to the results of the first processing step (102A), image data associated with each image belonging to the defective object category is processed so as to classify the images and corresponding objects according to multiple defect categories based on multiple identification features extracted from the image data.

14. The system (1) according to claim 13, further comprising a storage unit configured to store the images belonging to the defective object class in a database.

15. The system (1) according to claim 14, wherein: The processing unit (102) is configured to perform unsupervised clustering in the second processing step, the unsupervised clustering being programmed to define a multi-dimensional workspace, wherein each dimension corresponds to one of the plurality of identification features extracted from the image data of each image, wherein the value of the identification feature extracted for each image defines the position of the image data of each image in the workspace, so that each image data is shown as a data point in the workspace, and the unsupervised clustering is programmed to generate the defect category by grouping data points having similar positions in the workspace.

16. System (1) according to any one of the preceding claims, wherein: The processing unit (102) is configured to obtain position information related to the position of a defect in each defective object in the first processing step (102A).

17. System (1) according to any one of the preceding claims, wherein: In the first processing step (102A), the processing unit (102) includes a machine learning model, which is trained to attribute each image to the defective object category or the non-defective object category, wherein the machine learning model is trained based on training data that only includes images of non-defective objects.

18. System (1) according to any one of the preceding claims, wherein: In the first processing step (102A), the processing unit (102) comprises, for each image, extracting diagnostic markers from the image data and applying predetermined diagnostic rules.

19. System (1) according to any one of the preceding claims, wherein: In the first processing step (102A), the processing unit (102) comprises: a machine learning model trained to classify each image as belonging to the defective object class or the non-defective object class, and For each image, in a second stage (1022A) diagnostic markers are extracted from the image data and predetermined diagnostic rules are applied, Therein both the result of the first stage and the result of the second stage of the first processing step are taken into account in order to classify the image and the corresponding object into the defective object class or the non-defective object class.

20. The system (1) according to any one of the preceding claims, wherein: The optical device (101) comprises: an illuminator for illuminating an object in said inspection station (IP) with light in the visible or IR or UV spectrum; A camera is configured to observe an object in the inspection station (IP) and capture an image (I) of the object.

21. An apparatus for producing objects in a continuous cycle, the apparatus comprising: one or more machines for producing said object; A system (1) for performing quality control of the object, wherein the system (1) is a system according to any one of claims 13 to 20.

22. A computer program comprising instructions configured to, when executed on a processor, perform the steps of the method according to any one of claims 1 to 12.

Images