Device and method for inspecting an object

Abstract

A device (1) for inspecting an object (O) constituting a package which is rigid and has an outside surface comprises: a conveyor (T) for conveying a succession of objects (O) along a feed path (P); an illuminator (I) for illuminating the outside surface of the object (O) at an inspection position 5 (PI) of the feed path (P); a camera (R) for capturing image data (DI) of the outside surface of the object (O) at the inspection position (PI); a processing unit (E) having access to a memory (M) containing a machine- learned model (ML). The processing unit (ML) is programmed to generate a plurality of artificial images (IA), representing an artificially simulated 10 outside surface of the object (O), to feed them to the machine-learned model (ML), when the model is in a learning configuration (A), to train it to recognize a defectiveness condition of the object (O), and to feed the image data (DI) to the machine-learned model (ML), when the model is in a working configuration (L) so as to derive diagnostic information regarding 15 the defectiveness condition of the object (O).

Description

[0001] DEVICE AND METHOD FOR INSPECTING AN OBJECT

[0002] Technical field

[0003] This invention relates to a device and a method for inspecting an object. In particular, the technical field of this invention regards the inspection of an object constituting a package which is rigid and has an outside surface.

[0004] Background art

[0005] In the manufacturing sector, ensuring defect free quality control for objects produced is crucial for maximum efficiency and speed in the context of continuous cycle production lines.

[0006] Patent documents WO2020 / 121239 and IT102022000011345, in the name of the present Applicant, describe a device for the continuous cycle inspection of objects. The device involves use of a machine-learned model, which is trained through training data to identify and classify different categories of defects in the objects and which forms part of an inspection system. The training data used, however, may be inadequate to allow a sufficiently accurate quality control.

[0007] In this context, it is of crucial importance to ensure that the training data are diverse enough to allow the machine-learned model to identify and classify the defects of the objects in the most complete and accurate manner possible. In effect, diversifying the training data allows the model to distinguish the defects more efficiently, thus contributing to providing a robust and reliable quality control system. When the training data are inadequate, the quality control may be slow and ineffective.

[0008] Disclosure of the invention

[0009] The aim of this disclosure is to provide a device and a method for inspecting an object to overcome the above mentioned disadvantages of the prior art.

[0010] In particular, this disclosure has for an aim to provide a device and a method for inspecting an object capable of deriving information regarding the defectiveness of the object quickly and effectively.

[0011] This aim is fully achieved by the device and the method of this disclosure as characterized in the appended claims.

[0012] It should be noted that this invention can be applied to all sectors which require quality control of objects such as, for example, rigid packaging. In this sector, the objects that are inspected for defects may be plastic objects (caps, parisons, containers ...) or objects made from other materials (glass, aluminium, jars, tins ...). In particular, the object to be inspected has an outside surface. Moreover, the object to be inspected has an inside surface. Specifically, the object extends around a longitudinal axis; therefore, it has an inside surface facing the longitudinal axis and an outside surface opposite to the internal surface.

[0013] The device comprises a conveyor for conveying a succession of objects along a feed path. Preferably, the objects are conveyed in an orderly flow. In particular, the objects are conveyed in such a way that they adopt an inspection position along the feed path.

[0014] The device comprises an illuminator, configured to illuminate the outside surface of the object positioned at the inspection position.

[0015] The device comprises a camera, configured to view the outside surface of the object positioned at the inspection position and to capture image data of the outside surface.

[0016] In an example, the device comprises a plurality of cameras, configured to view the same object and to capture the image data. Thus, the image data are defined by different contributions of the different cameras of the plurality of cameras. The different contributions may be processed by the cameras to derive the image data. The image data may be transmitted to the processor. Alternatively or in addition, the different contributions may be transmitted to the processor and the processor may be programmed to process the different contributions to derive the image data.

[0017] More generally, the camera (and / or the plurality of cameras) can be configured to view the object positioned in the inspection position and capture image data representing the external appearance of the object. The external appearance of the object may include the outside surface and / or the inside surface. In fact, when the object is transparent, both the outside and inside surfaces can be observed. The captured image data may represent either the outside surface, the inside surface, or both.

[0018] In an example, the inspection position is a single inspection position. Each camera may be configured to view the outside surface of the object positioned at the inspection position and to capture respective image data relating to a corresponding portion of the outside surface (and / or an inside surface of the object, or more generally, an external appearance of the object). Thus, the cameras of the plurality of cameras may be disposed around the inspection position. In an example, the inspection position comprises a plurality of inspection positions. Each camera may be configured to view the inspection surface of the object positioned at a corresponding inspection position of the plurality of inspection positions. In such a case, the cameras of the plurality of cameras may be disposed around a respective inspection position of the plurality of inspection positions.

[0019] The device comprises a processing unit (or a processing system). It should be noted that the processing unit (or system) may consist of a single piece of hardware or two more distinct pieces of hardware. The processing unit has access to a memory containing a machine-learned model.

[0020] The processing unit is programmed to generate a plurality of artificial images representing an artificially simulated outside surface (and / or an inside surface of the object, or more generally, an external appearance of the object) of the object.

[0021] The processing unit is programmed to feed the plurality of artificial images to the machine-learned model, when the model is in a learning configuration, to train it to recognize a condition of defectiveness of the object.

[0022] The processing unit is programmed to feed the image data captured by the camera to the machine-learned model, when the model is in a working configuration, to derive diagnostic information about the condition of defectiveness of the object.

[0023] Artificially generating a plurality of images in order to train the model has the advantage of being able to generate a considerable quantity of examples, potentially diversified from one another, so as to ensure that training is completer and more accurate.

[0024] It should be noted that the artificial images may represent an outside surface (and / or an inside surface of the object, or more generally, an external appearance of the object) which is defective or one which is free of defects.

[0025] The expression condition of defectiveness of the object is used to indicate the presence or absence of a defect (that is to say, a good object or a defective object) and / or a type of defect. Thus, the machine-learned model is configured to distinguish between good objects and defective objects and / or to identify the type of defect present in the objects. In particular, the term condition of defectiveness of the object may also include the anomaly detection method for object identification, in addition to methods for identifying specific types of defects.

[0026] The processing unit may be programmed to generate the artificial images according to a variability parameter, where the variability parameter represents an aesthetic variability of real images representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object). For example, the variability parameter may be characteristic of a predetermined defect or of a plurality of predetermined defects. The variability parameter may regard a range of dimensions typically revealing the predetermined defect (or the plurality of predetermined defects) on the object. The purpose of this aspect is to provide the machine-learned model with a set of sufficiently diverse training data to make the machine-learned model more efficient in recognizing defective objects. The aesthetic variability, however, may also be characteristic of an exterior aspect of the object, where such an aspect is free of defects. For example, the aesthetic variability may regard a plurality of types of outside surfaces of the object (for example, the outside surface of the object may be smooth, textured and so on). Thus, providing the machine-learned model with artificial images according to a variability parameter has the advantage of training the model to analyse captured images even of different objects.

[0027] According to an aspect, the processing unit may be programmed to process a plurality of real images representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object), so as to derive the variability parameter to characterize the aesthetic variability of the surface of the real images. Thus, the variability parameter is preferably derived from a plurality of real images. The variability parameter might, however, also be transmitted to the processing unit by a user. In particular, the processing unit may apply the variability parameter representing the variability of the real images to the surface of the virtual images.

[0028] In an example, the plurality of real images comprises a group of real, three-dimensional images, created, for example, by tomography (i.e., using a tomography apparatus). These real images (acquired through an optical sensor / camera or, more specifically, a tomograph) are used in a preliminary phase before the inspection. The purpose is to derive, from these real images, the variability parameter that will later be used to generate artificial images. Therefore, the real images are obtained outside the inspection device and at a different time from when the inspection device operates.

[0029] Alternatively or in addition, the plurality of real images may comprise a group of real, two-dimensional images. Thus, the variability parameter may be derived from complex images (such as the three-dimensional images) or from simpler images (such as the two-dimensional ones).

[0030] The memory may include a plurality of real images, representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object). Associated with each of the real images there may be a target representing the defectiveness of the outside surface (and / or an inside surface of the object, or more generally, an external appearance of the object). The processing unit may be programmed to feed the plurality of real images to the machine-learned model when the model is in the learning configuration. That way, the machine-learned model is trained by feeding it real images with which a target is associated, according to the supervised learning method. The target may be indicative of the presence and / or the absence of defects and / or the type of defect.

[0031] In an example, the artificial images comprise a group of defective images, each defective image being characterized by a predetermined defect (or a plurality of predetermined defects). The processing unit may be programmed to label the images of the group of defective images by means of labels representing the predetermined defect, so as to train the machine-learned model to recognize the predetermined defects. For this purpose, the processing unit may be programmed to feed the labelled defective images to the machine-learned model when the model is in the learning configuration.

[0032] The processing unit may be programmed to provide the machine-learned model with artificial images labelled with a target (in particular, the target represents the condition of defectiveness of the object) when the model is in the learning configuration.

[0033] Thus, supervised learning of the machine-learned model may occur by feeding real images with which a target is associated and / or artificial images labelled with a target.

[0034] The processing unit may be configured to receive an alteration parameter from a user through a user interface. The alteration parameter may represent a predetermined defect to be applied to an artificial image. The processing unit may be configured to apply the alteration parameter to an artificial image (in particular, modifying the artificial image). Thus, the alteration parameter may indicate a defect (for example, a type of defect) and / or a zone of the object on which to apply the defect, and / or an extent of the defect (for example, a dimension or visibility of the defect). That way, the user can, at will, generate a plurality of artificial images to be fed to the machine-learned model in order to train it.

[0035] The machine-learned model may be programmed to process a captured image in order to derive a compressed image. In particular, the model may derive the compressed image by reducing the size of the captured image and / or by extracting the predetermined characteristics from the captured image. The machine-learned model may be programmed to reconstitute the compressed image to derive the captured image anew. In particular, the model may reconstitute the compressed image to derive the captured image anew so as to obtain the original image or an image similar to the original one. To compress and reconstitute the captured image, the machine-learned model may extract the predetermined characteristics and use them to reconstruct the original image. Compressing and reconstituting the captured image allow the machine-learned model to concentrate on the more important details of the captured image, making the model more robust against variations and minor defects in the original captured image. The machine-learned model may be programmed to compare the reconstituted image with the captured image and to derive an efficiency parameter. The efficiency parameter may represent how much the reconstituted image corresponds to the original image. The machine- learned model may be programmed to distinguish between images without defects and images with defects, based on the efficiency parameter. That way, the machine-learned model may be programmed to perform anomaly detection on the captured images.

[0036] The machine-learned model may be programmed to process a captured image to derive a compressed image and to reconstitute the compressed image to derive the captured image anew, according to an autoencoding method.

[0037] The processing unit may be programmed to feed the machine-learned model with a plurality of artificial images with defects or without defects. Preferably (in particular according to the anomaly detection method), the processing unit feeds the defect free artificial images to the machine- learned model in the learning mode. This aspect has the purpose of training the machine-learned model and to make it more efficient in using the anomaly detection method; in effect, since there are usually more defect free objects than defective objects, the machine-learned model, in the working configuration, receives more captured images of defect free objects than it does captured images of defective objects.

[0038] Thus, the machine learning model generates a compressed image from an initially captured image and then reconstructs the compressed image to derive the original captured image. Following this compression and reconstruction process, the machine learning model computes an efficiency parameter. This efficiency parameter represents how closely the reconstructed image matches the original real image.

[0039] For example, an encoder compresses and reconstructs images. If this encoder has been trained using normal (defect-free) images, it will struggle more when compressing and reconstructing anomalous images compared to normal ones. Consequently, the efficiency parameter will be low. In this case, a low efficiency parameter indicates that the reconstructed image significantly differs from the original, allowing the machine learning model to detect the anomaly.

[0040] In an example, the processing unit may be programmed to receive a plurality of setup parameters representing a setting of the camera and / or of the illuminator. The processing unit may be programmed to generate the plurality of artificial images according to the plurality of setup parameters. The setup parameters may comprise a type of illuminator light source, an intensity of the illuminator light source, an orientation of the illuminator light source, and others.

[0041] The setup parameters may be used in the learning configuration and / or in the working configuration of the machine-learned model.

[0042] The inspection device might also be connected, via the processing unit, for example, to a production apparatus which makes the objects. For example, the processing unit may be programmed to provide one or more hints, together with the diagnostic information regarding the condition of defectiveness of the object, where the hints represent a setting of one or more object production control parameters. Thus, the processing unit may be programmed to communicate with the production apparatus so as to provide it with one or more hints regarding a setting of one or more object production control parameters. The hints may be used by a user of the production apparatus. In addition or alternatively, the processing unit may be programmed to set one or more object production control parameters based on the hints.

[0043] In an embodiment, the processing unit comprises a first processor and a second processor which is physically distinct from the first processor. The machine-learned model may be run on the first processor and the artificial images generated on the second processor. In other words, the first processor may preside over the execution of (that is to say, is responsible for executing) the machine-learned model and the second processor may preside over (that is to say, is responsible for) the generating of the artificial images. For example, the first processor may be a processor at the client end and the second processor may be a processor at the server end.

[0044] The conveyor, the illuminator, the camera (or the plurality of cameras, if present), and / or the object to be inspected can define a physical laboratory. In the physical laboratory, the conveyor transports the object to the inspection position, and the camera captures images representing the illuminated (real) object. In one embodiment, the plurality of artificial images is generated within a virtual laboratory. Preferably, the virtual laboratory includes: a virtual conveyor to transport a virtual object, a virtual illuminator to illuminate the virtual object and a virtual camera (or a plurality of virtual cameras) to acquire image data related to the virtual object (from which the plurality of artificial images is generated).

[0045] Thus, the virtual laboratory constitutes a digital twin of the physical laboratory.

[0046] The virtual laboratory can be parameterized; in other words, the virtual environment is designed to allow modifications and / or adjustments to characteristics related to its virtual components (i.e., the virtual conveyor, illuminator, camera, and / or object). Specifically, one or more of these virtual components can be parameterized; one or more of these virtual components can be configured to allow modifications and / or adjustments to its properties.

[0047] For example, the virtual illuminator can be parameterized to allow a user to modify its characteristics, such as illumination intensity, type of illumination (grazing, collimated, or diffused), illumination direction, illumination wavelength, and other properties. Similarly, the virtual object can be parameterized to allow modifications to its attributes, such as the presence (or absence) of one or more defects, defect position, defect size, number of defects (identical or different), object dimensions, object material, and more.

[0048] The variation of one or more object characteristics can be based on digital images representing the object and its defects and / or real images (e.g., tomographic scans of the object) showing the object and its defects in 2D or 3D. For example, the virtual object can be modified by applying a virtual texture (or skin) to alter its external appearance.

[0049] Thus, the term "artificially simulated external surface of the object" refers to the virtual object's external appearance obtained by applying a virtual texture (or skin) according to one or more of the methods described above.

[0050] The virtual (or digital) laboratory is a software tool that digitally replicates a real vision system, with the goal of simulating the images that the real vision system would produce.

[0051] The virtual laboratory can be implemented using known image rendering techniques, according to one or more of the methods described, for example, in patent document CN117078853A, the content of which is incorporated herein by reference.

[0052] Specifically, the virtual laboratory is generated using simulation software. More specifically, TracePro® is used to simulate the virtual illuminator, and VRED® is used to render the images acquired by the virtual camera.

[0053] The virtual laboratory provides a tool to support the development and improvement of existing artificial vision systems.

[0054] The virtual laboratory is designed to receive a plurality of input parameters from an operator. These parameters may relate to the virtual object to be inspected (e.g., object type, with or without defects, defect types, and more), the illuminator (e.g., type of illumination, illumination intensity, and more), and the camera (e.g., number of cameras, camera positions, and more).

[0055] In particular, the input parameters can be used to modify one or more characteristics of the virtual illuminator, camera, and / or object. The virtual laboratory is configured to output one or more artificial images.

[0056] Preferably, the virtual illuminator is simulated by a processing unit. Specifically, the virtual illuminator is modeled by simulating a light flow through a surface. This approach takes into account both passive elements (e.g., diffusers) and active elements (e.g., light sources) in a single object known as a RayFile.

[0057] Regarding the types of models (objects) that can be inspected, different rendering approaches can be used depending on the information to be extracted from the simulated images: ideal geometries (generated via a CAD) can be rendered, primarily for geometric verification (e.g., inspection field dimensions), or measured geometries can be rendered to directly simulate the appearance of defects caused by the manufacturing process. Artificial images can be generated under different lighting conditions, such as diffused light, collimated light, and / or grazing light.

[0058] Thus, the digital laboratory functions as a real laboratory and can be used for various purposes, including geometric verification of the inspected field of view, feasibility assessment of defect inspectability, optimization of image acquisition, prototyping of new lighting systems, and automatic generation of datasets for training artificial intelligence networks.

[0059] This disclosure also provides a method for inspecting an object constituting a package which is rigid and has an outside surface.

[0060] Moreover, the object to be inspected has an inside surface. Specifically, the object develops around a longitudinal axis; therefore, the object has an inside surface facing the longitudinal axis and an outside surface opposite to the inside surface

[0061] The method comprises a step of conveying a succession of objects along a feed path via a conveyor. Preferably, the objects are conveyed in an orderly flow. In particular, the objects moving along the feed path adopt an inspection position.

[0062] The method comprises a step, via an illuminator, of illuminating an outside surface of the object positioned at an inspection position along the path.

[0063] The method comprises a step, via a camera, of capturing image data of the outside surface of the object positioned at the inspection position with the camera directed towards the outside surface of the object (and / or an inside surface of the object, or more generally, an external appearance of the object).

[0064] The method comprises a step, via a processing unit having access to a memory containing a machine-learned model, of generating a plurality of artificial images representing an artificially simulated outside surface of the object. The method comprises a step, via the processing unit, of feeding the plurality of artificial images to the machine-learned model, when the model is in a learning configuration, to train it to recognize a condition of defectiveness of the object.

[0065] The method comprises a step, via the processing unit, of feeding the image data captured by the camera to the machine-learned model, when the model is in a learning configuration, to derive diagnostic information about the condition of defectiveness of the object.

[0066] The method may comprise a step of generating the artificial images according to a variability parameter representing an aesthetic variability of real images representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object). The method may comprise a step, via the processing unit, of processing a plurality of real images representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object), in order to derive the variability parameter.

[0067] In an example, the method comprises a step, via the processing unit, of feeding a group of example images of the plurality of artificial images to the machine-learned model when the model is in a validating configuration, so as to validate the machine-learned model.

[0068] The memory may include a plurality of real images, representing an outside surface of the real object (and / or an inside surface of the object, or more generally, an external appearance of the object) and the method may comprise a step of feeding the plurality of real images to the machine- learned model when the model is in the learning configuration.

[0069] According to an aspect, the method comprises a step, via the machine- learned model, of processing a captured image to derive a compressed image, of reconstituting the compressed image and deriving the captured image anew, of comparing the reconstituted image with the captured image and deriving an efficiency parameter, and of distinguishing between artificial images without defects and artificial images with defects, based on the efficiency parameter. In an example, the method comprises a step, via the processing unit, of receiving a plurality of setup parameters representing a setting of the camera and / or of the illuminator, and of generating the plurality of artificial images according to the plurality of setup parameters.

[0070] Brief description of drawings

[0071] This and other features of the invention will become more apparent from the following description of a preferred embodiment of it, illustrated purely by way of example in the accompanying drawings, in which:

[0072] - Figure 1 illustrates a device 1 according to one or more aspects of this disclosure;

[0073] - Figures 2A and 2B respectively illustrate a step of learning A and a step of working L, according to one or more aspects of this disclosure.

[0074] Detailed description of preferred embodiments of the invention

[0075] With reference to the accompanying drawings, the numeral 1 denotes a device for inspecting an object O. The object O constitutes a rigid package and has an outside surface and an inside surface. Specifically, the object O develops around a longitudinal axis, so that the inside surface faces the longitudinal axis and the outside surface is opposite to the inside surface.

[0076] The device 1 comprises a conveyor T for conveying a succession of objects O along a feed path P. Preferably, the objects O are conveyed in an orderly flow of objects O. The objects O along the feed path P adopt an inspection position PI.

[0077] The device 1 comprises an illuminator I, configured to illuminate the outside surface of the object O. The illuminator I is positioned at the inspection position PI so as to illuminate the object O when it is at the inspection position PI. The illuminator I can illuminate the object O from above (that is, the light of the illuminator I may be directed substantially parallel to the inspection position PI, that is, perpendicularly to the feed path P), and / or from the side (that is, the light of the illuminator I may be directed substantially perpendicularly or transversely to the inspection position PI, that is, substantially parallel or transversely to the feed path).

[0078] The device 1 comprises a camera R, configured to view the object O positioned at the inspection position PI. The camera R may be configured to view the object O from above (that is, the optical path of the camera R may be substantially parallel to the inspection position PI) and / or from the side (that is, the optical path of the camera R may be perpendicular or transverse to the inspection position PI, that is, substantially parallel or transverse to the feed path P). In an example, the device 1 comprises a plurality of cameras R so as to view the entire outside surface of the object O. In addition or alternatively, the object O may be configured to rotate at the inspection position PI so as to be illuminated by the illuminator I and the whole of its outside surface viewed by the camera R.

[0079] The camera R is configured for capturing image data DI of said outside surface (and / or inside surface).

[0080] The device 1 comprises a processing unit E having access to a memory M containing a machine-learned model ML.

[0081] The processing unit E is programmed to generate a plurality of artificial images IA, representing an artificially simulated outside surface of the object O (and / or inside surface). The processing unit is programmed to feed the plurality of artificial images IA to the machine-learned model ML, when the machine-learned model ML is in a learning configuration A, to train it to recognize a condition of defectiveness of the object O.

[0082] The processing unit E is programmed to feed the image data DI captured by the camera R to the machine-learned model ML, when the machine- learned model ML is in a working configuration L, to derive diagnostic information about the condition of defectiveness of the object O.

[0083] In the learning configuration A, the processing unit E may associate a target TA with each artificial image of the plurality of artificial images IA and feed the plurality of artificial images IA, together with the target TA, to the machine-learned model ML. The target TA represents the defectiveness of the object O represented in the artificial image IA.

[0084] In the working configuration L, the processing unit E feeds the image data DI and the machine-learned model ML derives the condition of defectiveness of the object O, that is, of the target TA.

[0085] Thus, operatively, the objects O are conveyed by the conveyor T along the feed path P in an orderly sequence. One after the other along the feed path P, the objects O adopt an inspection position PI. The object O at the position PI is illuminated by the illuminator I and framed by the camera R. The camera R captures the image data DI representing the outside surface of the object O (and / or inside surface). The image data DI are transmitted to the processing unit E.

[0086] The processing unit E generates a plurality of artificial images IA, representing an artificially simulated outside surface of the object O (and / or inside surface), and feeds them to the machine-learned model ML during a step of learning A so as to train it to recognize a condition of defectiveness of the object O. In a step of working L of the machine- learned model ML, the processing unit E feeds the image data DI captured by the camera R to derive diagnostic information about the object O.

[0087] After being used to train the machine-learned model ML, the artificial images IA may be stored in the memory or they be deleted from the processing unit E.

[0088] The processing unit E is programmed to receive a plurality of real images IR, representing an outside surface of the real object (and / or inside surface), where the real image IR is a three-dimensional image, obtained preferably by tomography. The processing unit E is programmed to process the real images IR to extract a variability parameter representing an aesthetic variability of real images IR representing an outside surface of the real object (and / or inside surface). The processing unit E is programmed to generate the artificial images IA according to the variability parameter.

[0089] The processing unit E may be programmed to associate with each real image IR a target representing the defectiveness of the outside surface (and / or inside surface). The target may represent the presence or absence of a defect. Thus, during the step of learning A, the processing unit E feeds the real images to the machine-learned model ML together with the target. A group of real images IR of the plurality of real images IR associated with a target representing the defectiveness of the outside surface (and / or inside surface) may define images which are defective, that is, characterized by the presence of at least one defect. In such a case, the processing unit E associates a label representing the type of defect with each defective image.

[0090] In an example, the processing unit E receives from a user an alteration parameter, representing a predetermined defect, and applies it to one or more artificial images of the plurality of artificial images IA. For this purpose, the device 1 may comprise a user interface, through which the user communicates with the processing unit E.

[0091] The processing unit E may be programmed to receive a plurality of setup parameters representing a setting of the camera R and / or of the illuminator I and to generate the plurality of artificial images IA according to the setup parameters..

[0092] In an embodiment, the processing unit E comprises a first processor E', which presides over the execution of the machine-learned model ML, and a second processor E", physically distinct from the first processor E', which presides over the generating of the artificial images IA.

Claims

CLAIMS1. A device (1 ) for inspecting an object (O) constituting a package which is rigid and has an outside surface, the device (1 ) comprising:- a conveyor (T) configured to convey a succession of objects (O) along a feed path (P) in an orderly flow, so that the objects (O) adopt an inspection position (PI) along the feed path (P);- an illuminator (I), configured to illuminate the outside surface of the object (O) positioned at the inspection position (PI);- a camera (R), configured to view the outside surface of the object (O) positioned at the inspection position (PI) and to capture image data (DI) of said outside surface;- a processing unit (E) having access to a memory (M) containing a machine-learned model (ML), and programmed to generate a plurality of artificial images (IA), representing an artificially simulated outside surface of the object (O), feed the plurality of artificial images (IA) to the machine-learned model (ML), when the model is in a learning configuration (A), to train it to recognize a condition of defectiveness of the object (O), feed the image data (DI) captured by the camera (R) to the machine-learned model (ML), when the model is in a working configuration (L), to derive diagnostic information about the condition of defectiveness of the object (O).

2. The device (1 ) according to claim 1 , wherein the processing unit (E) is programmed to generate the artificial images (IA) according to a variability parameter representing an aesthetic variability of real images (IR) representing an outside surface of the real object (O).

3. The device (1 ) according to claim 2, wherein the processing unit (E) is programmed to process a plurality of real images (IR) representing an outside surface of the real object (O), so as to derive the variability parameter to characterize the aesthetic variability of the surface of the real images (IR).

4. The device (1 ) according to claim 3, wherein the plurality of real images (IR) comprises a group of real, three-dimensional images made by tomography.

5. The device (1 ) according to any one of the preceding claims, wherein the memory (M) includes a plurality of real images (IR), representing an outside surface of the real object (O), associated with each of the real images (IR) is a target representing the defectiveness of the outside surface, and the processing unit (E) is programmed to feed the plurality of real images (IR) to the machine-learned model (ML) when the model is in the learning configuration (A).

6. The device (1 ) according to any one of the preceding claims, wherein the artificial images (IA) comprise a group of defective images, each defective image being characterized by a predetermined defect, the processing unit (E) being programmed to label the images of the group of defective images by means of labels representing the predetermined defect, so as to train the machine-learned model (ML) to recognize the predetermined defects.

7. The device (1 ) according to claim 6, wherein the processing unit (E) is configured to receive from a user through a user interface, an alteration parameter representing a predetermined defect to be applied to an artificial image (IA).

8. The device (1 ) according to any one of the preceding claims, wherein the machine-learned model (ML) is programmed to process a captured image to derive a compressed image; reconstitute the compressed image to derive the captured image anew; compare the reconstituted image with the captured image and to derive an efficiency parameter, and distinguish between artificial images (IA) without defects and artificial images (IA) with defects, based on the efficiency parameter.

9. The device (1 ) according to any one of the preceding claims, wherein the processing unit (E) is programmed to receive a plurality of setup parameters representing a setting of the camera (R) and / or of the illuminator (I) and the processing unit (E) is programmed to generate the plurality of artificial images (IA) according to the plurality of setup parameters.

10. The device (1 ) according to any one of the preceding claims, wherein the processing unit (E) comprises a first processor (E') and a second processor (E") which is physically distinct from the first processor, the machine-learned model (ML) being run on the first processor (E') and the artificial images (IA) generated on the second processor (E").

11. The device (1 ) according to any one of the preceding claims, comprising a plurality of cameras (R), configured to view the same object (O) and to capture the image data (DI).

12. A method for inspecting an object (O) constituting a package which is rigid and has an outside surface, the method comprising the following steps:- via a conveyor (T) conveying a succession of objects (O) along a feed path (P) in an orderly flow, so that the objects (O) adopt an inspection position (PI) along the feed path (P);- via an illuminator (I), illuminating the outside surface of the object (O) positioned at an inspection position (PI) along the path (P);- via a camera (R), capturing image data (DI) of the outside surface of the object (O) positioned at the inspection position (PI) with the camera (R) directed towards the outside surface of the object (O);- via a processing unit (E) having access to a memory (M) containing a machine-learned model (ML): generating a plurality of artificial images (IA), representing an artificially simulated outside surface of the object (O), feeding the plurality of artificial images (IA) to the machine-learned model (ML), when the model is in a learning configuration (A), to train it torecognize a condition of defectiveness of the object (O), feeding the image data (DI) captured by the camera (R) to the machine-learned model (ML), when the model is in a working configuration (L), to derive diagnostic information about the condition of defectiveness of the object (O).

13. The method according to claim 12, comprising a step, via the processing unit (E), of processing a plurality of real images (IR) representing an outside surface of the real object (O), to derive a variability parameter representing an aesthetic variability of the real images (IR), wherein the step of generating the artificial images (IA) is carried out according to the variability parameter.

14. The method according to claim 12 or 13, comprising a step, via the processing unit (E), of feeding a group of example images of the plurality of artificial images (IA) to the machine-learned model (ML) when the model is in a validating configuration, so as to validate the machine-learned model (ML).

15. The method according to any one of claims 12 to 14, comprising the following steps, carried out by the machine-learned model (ML): processing a captured image to derive a compressed image; reconstituting the compressed image and deriving the captured image anew; comparing the reconstituted image with the captured image and deriving an efficiency parameter, and distinguishing between artificial images (IA) without defects and artificial images (IA) with defects, based on the efficiency parameter.

16. The method according to any one of claims 12 to 15, comprising the following steps, carried out by the processing unit (E): receiving a plurality of setup parameters representing a setting of the camera (R) and / or of the illuminator (I) and generating the plurality of artificial images (IA) according to the plurality of setup parameters.

Images