Method and apparatus for inspecting glass containers according to at least two modalities to classify containers by glass defects thereof

Abstract

The invention relates to a method for inspecting glass containers (2), comprising the following steps: inspecting each container using an inspection system (10) in order to obtain at least one analysis image according to a first modality corresponding to an absorption image (Ia) and at least one analysis image according to a second modality corresponding to a birefringence image (Ib) or a refraction image (Ir), -defining a category list (D1, D2,... Dk,... Dp) comprising at least glass defects,-matching at least partial analysis images according to the first modality and according to the second modality,-from the at least one analysis image according to the first modality and the at least one analysis image according to the second modality which are matched together, determining a category list (D1, D2,... Dk,... Dp), classifying the analysis image by means of an image classifier (Cl) that determines which result category of the list of categories the analysis image belongs to, the image classifier having been trained by supervised learning,-classifying the container according to the result category.

Description

Technical Field

[0001] The present invention relates to the technical field of inspecting transparent or translucent containers, such as glass bottles, jars or flasks, for the purpose of quality control in order to detect and identify possible defects that may affect these containers.

[0002] The object of the present invention is to seek particularly advantageous applications for analyzing physical characteristics of containers to identify non-conforming physical characteristics corresponding to defects, such as surface defects (such as wrinkles or notches), as well as internal defects of the material (such as cracks, inclusions or bubbles). Background Art

[0003] For the manufacture of glass containers, it is known that the manufacturing process including melting glass and then transferring it to the molding unit is implemented by a manufacturing facility, which includes a melting furnace and a forehearth for supplying molten glass to a molding machine (usually a type specified by an IS machine). The containers just formed by the molding machine are placed on the output conveyor belt in turn to form a row of containers. The containers are transported by the conveyor belt in a row so as to be transferred to different processing stations in turn. Specifically, the molded container is placed in an annealing furnace, which increases the temperature of the container and then cools the container in a monitored manner so that the thermal stress generated by the molding process disappears. Other processes for molding glass containers are known for glass tableware, insulators, syringes and bulbs. For example, there are molding machines such as rotary and continuous presses, but there are no parallel aligned sections like the IS machine. There are also machines for processing preforms into tubes (especially borosilicate glass) to manufacture syringes and bulbs dedicated to pharmaceutical products.

[0004] It is known to systematically inspect all containers leaving the forming machine and travelling on an output conveyor using different inspection devices, in particular systems for inspecting the walls in transmission, for which purpose a light source is arranged on one side of the conveyor and at least one camera, generally at least two cameras, is arranged on the other side to acquire at least one image formed by light transmitted through the container wall.

[0005] It is also known to systematically inspect all containers leaving the annealing furnace using different inspection devices, in particular systems for inspecting the walls in transmission, for which a light source is provided on one side of the conveyor and at least one camera (usually 2 to 6, 12 or 24) is provided on the other side to acquire at least one image formed by light transmitted through the walls of the containers. The different devices for inspecting the walls in transmission are designed to acquire at least one image formed by light transmitted through vertical walls, while some other devices are suitable for acquiring at least one image formed by light transmitted vertically through the bottom of the container.

[0006] For example, patent EP2082216 describes a method for detecting low-contrast defects on the one hand (particularly defects that refract light, such as bubbles, surface wrinkles or local variations in the thickness of transparent materials), and high-contrast defects on the other hand (particularly defects that absorb transmitted light, such as inclusions or stains of opaque substances). The process aims to drive a light source so that it sequentially generates two types of illumination, the first type being uniform illumination and the second type being formed by alternating dark and light areas with discontinuous spatial variations. When a moving target is sequentially illuminated by the two types of illumination, an image of the target is captured. The image captured using illumination according to the first modality is analyzed in order to detect high-contrast defects, and the image captured using illumination according to the second modality is analyzed in order to detect low-contrast defects.

[0007] A method for analyzing an image of a container for cold inspection is also known from patent EP 1 109 008. A segmentation step detects features in the image and the area around the features. A discriminatory parameter for a region is calculated and combined with a fuzzy logic method to determine the type of feature that is most likely among a list of features corresponding to possible defects. The compliance of the region is then determined by applying different criteria depending on the type of feature retained. For example, a wrinkle-type defect is ejected for a certain surface, while an inclusion-type defect is rejected even if the surface is small. The patent teaches in particular that defects do not all have the same criticality, which justifies trying to determine their nature before deciding to reject the container. In this method, the illumination is a uniform extended source, which correctly reveals absorption defects, but where refractive defects are visible, but with low contrast and therefore insufficient sensitivity.

[0008] However, it seems difficult to identify certain defects with certainty, in particular “cosmetic defects”, i.e., all types of visual defects: inclusions (such as foreign bodies such as ceramics, metals, etc.), bubbles, wrinkles, river lines (surface grooves), cracks (cracks), burrs, trapezoidal defects, oil stains, extremely thin areas, unmelted particles. These cosmetic defects appear in the image as local optical changes, or pixels that deviate from the background. These cosmetic defects can be serious if they pose a risk to consumers, a risk of container breakage or loss of functionality. Since the identification of cosmetic defects from images can be ambiguous, a safety margin is taken into account during the inspection process. As a result, these containers are considered defective even if they are acceptable or meet the requirements.

[0009] Furthermore, it should be noted that acceptable artifacts such as engravings or decorations, lightly marked mold seams, etc. can be distinguished in the images. Furthermore, the nature of cosmetic defects needs to be accurately identified in order to identify serious defects by distinguishing them from other defects. Preventing these serious defects requires improving the reliability of cosmetic defect classification. In addition to the fact that improved defect identification will increase production efficiency, such identification of defects also makes it possible to determine their potential causes, so that manufacturing facilities can be driven according to the category of detected defects. Indeed, without reliable identification of detected defects, safe decisions cannot be made to correct the manufacturing process manually or automatically.

[0010] Patent application WO2021213864 proposes a process for more reliably inspecting containers conveyed by a conveyor belt on a production line, in particular a bottle making line. The container is conveyed to at least a first inspection unit and a second inspection unit, each of which includes a transmitter and a receiver. The inspection unit can inspect the container using white light, laser, high-frequency electromagnetic waves, gamma rays and / or X-rays as a transmitter. The container is conveyed between a transmitter and a receiver (such as a camera) so that first measurement data is acquired using a first inspection unit and second measurement data is acquired using a second inspection unit. According to the process, the first measurement data and the second measurement data are combined to form common input data for an artificial intelligence-based evaluation unit, and an inspection result such as a filling level is provided as an output.

[0011] As a single exemplary embodiment, the document describes the detection of the liquid level of a container by combining X-ray imaging and inspection using an infrared light source. The document indicates that the process can also be used to inspect the side walls, bottom, mouth, contents of the container, such as foreign contaminants or product residues. The results of the inspection may also be defects such as damage to the container (in particular cracks and / or fragments of the glass).

[0012] However, patent application WO2021213864 does not provide any teaching to accurately detect defects visible under transmission (especially appearance defects). Regardless of whether there is an obligation to detect these defects, it seems necessary to accurately identify the nature of appearance defects in order to identify serious appearance defects compared to other defects that may be considered non-serious. An incorrect identification of the type of defect on a container will result in the container being rejected, while a correct identification of the defect will make it possible to identify the container as qualified. Therefore, in order to improve sorting efficiency, it seems necessary to accurately identify the nature of defects visible under transmission.

[0013] A container inspection device for improving detection reliability, in particular in order to be able to reliably distinguish decorative elements from contaminants or stains, is also known from patent application EP 3 679 356. The device comprises a light source emitting radiation having different wavelengths and intensity ranges in spatially separated areas. The light source illuminates the container to be inspected, and a color camera is configured to detect the radiation emitted by the light source and having passed through the container.

[0014] The device further comprises an evaluation device designed to analyze the intensity image in order to determine pixels or areas therein having an intensity different from that of neighboring pixels or areas, from which the presence of light-absorbing defects (such as stains) is inferred. The evaluation device allows the color image to be analyzed to determine pixels or areas having a color different from that of neighboring pixels or areas, from which the presence of light-diffusing elements (such as decorative elements) is inferred. Thus, if a brightness contrast is observed locally and at the same time there is no color contrast in this area, the presence of a contaminant in this area is detected by the evaluation unit. If the local brightness contrast coincides with the local color contrast, the evaluation unit detects the presence of a decorative element.

[0015] The evaluation unit can also identify structures which lead to local color contrast but almost no local brightness contrast or only low local brightness contrast. For example, splinters or water droplets in the glass can lead to such local color contrast, while the light passing through can radiate through these areas substantially without brightness loss.

[0016] This inspection device allows to distinguish decorative elements from stains or contaminations. In other words, to distinguish light-refracting defects from light-absorbing defects. However, as pointed out in the application, this inspection device cannot distinguish various light-refracting defects from each other (in particular with regard to glass inclusions and water drops).

[0017] In the prior art, a process for optical inspection of containers is also known from WO2020 / 244815, wherein the container is conveyed to an inspection unit comprising an irradiation unit and a camera. The irradiation unit emits light from a light-emitting surface, and the light is locally encoded based on polarization characteristics, intensity characteristics and / or phase characteristics. It can be understood in the document WO2020 / 244815 that the polarization characteristics refer to light emitted from different emission points on the emission surface being emitted in different polarization directions. It can also be understood in the document WO2020 / 244815 that the polarization characteristics are linear, elliptical and / or circular polarization characteristics. For example, a polarization filter having a continuously changing polarization curve or several polarization filters with different orientations can be provided in the region of the light emission surface.

[0018] According to a variant of the embodiment, the light emitted by the light output surface is locally encoded based on a polarization feature (e.g., a polarization direction detected by a camera). It is possible to determine for a pixel in the camera image from which radiation position the corresponding light portion originates independently of the radiation feature of the light output surface. Because the image processing unit uses the image of (at least one) camera to obtain information about the position of the radiation point, it is possible to distinguish refractive defects, for example, based on local changes in the radiation position. In addition, the absorption of light by an absorption defect can be detected using information about the intensity. However, this inspection device has the same disadvantage as other known devices, namely, it is not possible to distinguish one glass defect from another glass defect (both of which have, for example, light refractive properties).

[0019] Also known from the prior art is document EP 0 957 355, which describes a container inspection device intended to detect opacity changes and stresses. In this device, a device for rotating the container about its axis, a light source with a diffuser and a polarizer, a first camera receiving the transmitted polarized light and a second camera receiving the light that has passed through the second polarizer are used. An image processing device is then used. Summary of the invention

[0020] The present invention aims to overcome the drawbacks of the prior art by proposing a method for monitoring the quality of glass containers designed to more effectively detect glass defects visible in transmission, by ensuring safe and definitive identification of the glass defects, in order to optimize the sorting of the containers.

[0021] Another object of the invention is to propose a method for monitoring the quality of glass containers which allows, after the determination and exact identification of the nature of the glass defect, to provide more complete information for the correction of the monitoring parameters of the glass container manufacturing process of the manufacturing facility.

[0022] In order to achieve these aims, the object of the present invention relates to a method for inspecting glass containers in order to classify them, the method comprising the following steps:

[0023] - inspecting each container using an inspection system comprising at least one light source for illuminating the container and at least one camera arranged to collect light that has passed through the container, so as to acquire an image of at least a portion of the container illuminated in transmission, thereby obtaining at least one analysis image according to a first modality corresponding to an absorption image and at least one analysis image according to a second modality corresponding to a birefringence image or a refraction image,

[0024] - defining a list of categories comprising at least glass defects, the list of categories comprising a number of categories, the number of categories being independent of the number of modes;

[0025] - ensuring a match between at least a portion of the analysis image according to the first modality and at least a portion of the analysis image according to the second modality,

[0026] - from the matched at least one analysis image according to the first modality and at least one analysis image according to the second modality, classifying these analysis images by means of an image classifier, which determines to which result class of the class list these analysis images belong,

[0027] - an image classifier has been trained by supervised learning on a learning set, the learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with a class from a list of classes, such that the trained image classifier classifies the containers according to classification features according to at least two modalities;

[0028] - Sort containers by result category.

[0029] The object of the present invention is based on a new method for detecting glass defects visible in transmission on glass containers. The method according to the invention takes into account different interactions between light and matter (in particular absorption, refraction and birefringence) to determine the physical nature of the defect. Taking these interactions into account in a combined manner makes the identification of defects visible in transmission more reliable.

[0030] One object of the present invention is to allow the classification of glass defects according to their physical characteristics by evaluating the changes of light transmitted through the container wall according to the absorption of light and at least the refraction of light and / or the change of the polarization state of light. These interactions each have a more specific emphasis according to the different observation modalities.

[0031] The invention is based on the idea that the defects modify the transmitted light not only by a single type of change, but by a variable combination of at least two types of change.

[0032] Thus, an opaque foreign body (such as a stone) will not only show some absorption of light (change in intensity), but also a change in polarization state through birefringence associated with the stresses in the glass surrounding it. Unmelted particles or devitrified glass will be transparent and therefore absorb little or no light, but will not only refract due to its different refractive index from ordinary glass, but will also cause a change in polarization state due to birefringence associated with the stresses it creates.

[0033] The present invention improves the quality of classification of defects (or observed features) due to the following facts:

[0034] - According to the first type of changes in the light transmitted through the wall, the defects may be similar,

[0035] - If two types of changes in the transmitted light are considered, the defects may be different.

[0036] The invention thus improves the quality of classification of glass defects, since it is not affected by the fact that the observation device cannot focus on a single type of variation of the light transmitted through the container wall. In fact, when observing a container illuminated by an extended luminous surface producing a uniform illumination, the image depends mainly on absorption, but also to a lesser extent on refraction. This is due to the finite size of the luminous surface and to the fact that strong refractive defects cause an attenuation of the light by deflecting it outside the pupil of the objective, an effect that is not easily distinguishable from absorption. Refractive defects therefore appear in the absorption image.

[0037] Similarly, when observing a container illuminated by an extended luminous surface producing an illumination that exhibits spatial variations in polarization characteristics (the case of a refracted image), the image depends primarily on refraction, but if the defect is birefringent, it also changes the polarization characteristics of the light. As a result, this observation modality is not sufficient to determine the nature of the defect.

[0038] In other words, the invention aims to improve the quality of classification of containers according to glass defects by the fact that it allows the combined consideration of the intensity according to at least two types of variations of the transmitted light and according to the morphological and photometric characteristics in different images of the defect, while overcoming the situation in which the observation modality does not give an independent estimate of one type of variation of the transmitted light.

[0039] According to a preferred variant implementing the three examination modalities, the method comprises the following steps:

[0040] - inspecting each container using an inspection system configured for acquiring images to obtain at least one analysis image according to a first modality, at least one analysis image according to a second modality, and at least one analysis image according to a third modality, the analysis image according to the second modality corresponding to a birefringence image, the analysis image according to the third modality corresponding to a refraction image,

[0041] - ensuring a matching of at least a portion of the analysis image according to the first modality, at least a portion of the analysis image according to the second modality and at least a portion of the analysis image according to the third modality,

[0042] - from the matched at least one analysis image according to the first modality, at least one analysis image according to the second modality and at least one analysis image according to the third modality, classifying these analysis images by means of an image classifier, which determines to which result class of the class list these analysis images belong,

[0043] - an image classifier is trained by supervised learning on a learning set, the learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with a class from a list of classes, so that the trained image classifier classifies the containers according to a classification criterion according to at least three modalities;

[0044] - Sort containers by result category.

[0045] According to a first exemplary embodiment of an inspection system, the method is intended to inspect a container using an inspection system configured for acquiring images and calculating, from several of these images, analysis images corresponding to absorption images, birefringence images and / or refraction images.

[0046] According to another exemplary embodiment of the inspection system, the method is directed to inspecting containers using an inspection system configured for obtaining polarization composite images and calculating absorption images and / or birefringence images and / or refraction images from these polarization composite images.

[0047] According to another exemplary embodiment of the inspection system, the method is directed to inspecting a container using an inspection system configured to acquire color composite images using a color camera and to calculate absorption images and refraction images from the color composite images.

[0048] According to another exemplary embodiment of an inspection system, the method is directed to inspecting a container using an inspection system configured for acquiring an image directly corresponding to an absorption image, a birefringence image and / or a refraction image.

[0049] Advantageously, in order to ensure a match of at least a portion of an analysis image according to the first modality and at least a portion of an analysis image according to the second modality and / or at least a portion of an analysis image according to the third modality, the method detects candidate regions in the analysis image of the first modality and in the analysis image of the second modality and / or in the analysis image of the third modality, the method ensuring, for each container:

[0050] - matching of a candidate region in an analysis image of the first, second or third modality with a corresponding region of an analysis image of at least one other modality,

[0051] -Or matching of candidate regions from two different modalities.

[0052] According to another advantageous example, the method ensures, as matching, a merging of at least one analysis image of the first modality with an analysis image of the second modality and / or an analysis image of the third modality to obtain a merged image, the method ensuring:

[0053] - Extract classification features from the merged image,

[0054] - and classifying the containers using a classification criterion applied to the classification features of the merged image.

[0055] According to another advantageous example, the method ensures, as matching, a merging of at least one analysis image of the first modality with an analysis image of the second modality and / or the third modality to obtain a merged image, the method ensuring:

[0056] - segment the merged image to detect merge candidate regions,

[0057] - Classify containers using a classification criterion applied to the features of the merged candidate regions.

[0058] According to another feature of the present invention:

[0059] - extracting classification features according to the first modality from the analysis image according to the first modality,

[0060] - extracting classification features according to the second modality and classification features according to the third modality from the analysis image according to the second modality and / or the analysis image according to the third modality, respectively,

[0061] - grading the containers using a classification criterion applied according to the characteristics of the first modality and the characteristics of the second modality and / or the characteristics of the third modality.

[0062] According to an advantageous variant, classification features according to a first modality, and classification features according to a second modality, and / or classification features according to a third modality, and / or combined features taking into account features that logically or mathematically combine an analysis image according to the first modality and an analysis image according to the second modality and / or an analysis image according to the third modality are selected, wherein these classification features according to the first, second and third modalities are features such as position, size, shape or numerical values ​​representing absorption and / or refraction and / or birefringence.

[0063] According to one embodiment, the containers are classified by a supervised learning image classifier, the input data of which is:

[0064] - according to the classification characteristics of the first modality and according to the classification characteristics of the second modality and / or according to the classification characteristics of the third modality,

[0065] - either according to an analysis image of the first modality and according to an analysis image of the second modality and / or according to an analysis image of the third modality,

[0066] - or a portion of the image is analyzed according to a first modality and a portion of the image is analyzed according to a second modality and / or according to a third modality.

[0067] According to another embodiment, the containers are graded by a supervised learning image classifier, whose input data is at least one merged image obtained by merging at least one analysis image according to a first modality and an analysis image according to a second modality and / or an analysis image according to a third modality, or by merging a region of at least one analysis image according to the first modality and a region of an analysis image according to the second modality and / or a region of an analysis image according to a third modality.

[0068] According to a preferred example, each container is graded according to at least one image category selected from a list of categories representing at least glass defects (such as, in particular, trapezoidal defects, inclusions, bubbles).

[0069] According to another object of the invention, at least one sorting feature is compared with a rejection criterion, the sorting feature and the rejection criterion determining whether the container is qualified or not depending on the category to which it belongs, the sorting feature being calculated on at least one image of the container according to a modality.

[0070] According to an advantageous example, the method implements a step of taking into account at least one detected glass defect in order to deduce therefrom adjustment information for at least one monitoring parameter of the container production facility.

[0071] According to another characteristic of the implementation of this process:

[0072] - an image classifier associates a confidence score with the classification of a container forming part of the inspected product;

[0073] - The classification of a container is taken into consideration only if the confidence score exceeds a confidence threshold, such that:

[0074] * Count defects by defect category;

[0075] * and / or decide on the rejection of containers;

[0076] * and / or trigger an alarm indicating the presence of at least one critical defect in the inspected product.

[0077] Another object of the invention is to propose a device for inspecting glass containers leaving a manufacturing facility in order to sort them according to glass defects, comprising:

[0078] - An inspection system, including at least one light source and at least one camera, the light source irradiating the container, and the camera being arranged to collect the light passing through the container to obtain an image of at least a part of the container irradiated by the light source in a transmission manner.

[0079] - An information processing unit, which is connected to the inspection system and is adapted to provide, for each container, at least one analysis image according to a first modality corresponding to an absorption image and at least one analysis image according to a second modality corresponding to a birefringence image or a refraction image. The information processing unit is configured to perform the following operations:

[0080] * Take into account a list of categories including at least glass defects, the list of categories including the number of categories, which is independent of the number of modalities;

[0081] * Match at least a part of the analysis image according to the first modality and at least a part of the analysis image according to the second modality.

[0082] * Classify the analysis images through an image classifier from the at least one analysis image according to the first modality and the at least one analysis image according to the second modality that are matched together, and the image classifier determines which result category in the list of categories the analysis images belong to.

[0083] * Take into account an image classifier that has been trained through supervised learning on a learning set, the learning set including a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or the plurality of image regions being matched together and associated with the image categories from the list of categories, such that the trained image classifier grades the containers according to classification features according to at least two modalities.

[0084] * Classify the containers according to the result categories.

[0085] According to an embodiment of the present invention, the inspection system is configured to obtain polarization composite images, and the information processing unit is configured to calculate absorption analysis images and / or birefringence analysis images and / or refraction analysis images according to these polarization composite images.

[0086] According to another embodiment of the present invention, the inspection system is configured to obtain images, and the information processing unit is configured to calculate absorption analysis images and / or birefringence analysis images and / or refraction analysis images according to some of these images.

[0087] According to another embodiment, the inspection system is configured to use a color camera to obtain color composite images, and the information processing unit is configured to calculate absorption images and refraction images according to these color composite images.

[0088] According to another embodiment, the inspection system is configured to acquire an image that directly corresponds to the absorption image, the birefringence image and / or the refraction image.

[0089] Various other characteristics emerge from the description given below with reference to the accompanying drawings which show, by way of non-limiting example, an embodiment of the object of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] [ Figure 1 ] Figure 1 is a simplified illustration of an apparatus according to the present invention for inspecting glass containers leaving a manufacturing facility.

[0091] [ Figure 2 ] Figure 2 An exemplary embodiment of a system according to the invention for acquiring images of containers leaving a manufacturing facility by means of a polarization camera and implemented in an inspection device is presented.

[0092] [ Figure 3 ] Figure 3 Shown by Figure 2 An example of a composite pixel obtained by the inspection system's polarization camera is shown.

[0093] [ Figure 4 ] Figure 4 Another exemplary embodiment of an image acquisition system according to the invention for a container leaving a manufacturing facility and implemented in an inspection device is presented.

[0094] [ Figure 5 ] Figure 5 is a simplified functional block diagram of an example of a first embodiment of an inspection apparatus according to the present invention, subjecting information contained in an analysis image according to a first modality and in an analysis image according to a second modality to processing referred to as conventional processing.

[0095] [ Figure 6 ] Figure 6 is a simplified functional block diagram of an example of a first embodiment of an inspection apparatus according to the present invention, performing a segmentation operation on a merged image obtained by matching an analysis image according to a first modality and an analysis image according to a second modality.

[0096] [ Figure 7 ] Figure 7 It is a simplified functional block diagram of an example of a first embodiment of an inspection device according to the present invention, implementing a convolutional neural network, which has matching candidate regions as input data, and the matching candidate regions are obtained by performing a segmentation operation on an analysis image according to a first modality and an analysis image according to a second modality.

[0097] [ Figure 8 ] Figure 8It is a simplified functional block diagram of an example of a first embodiment of an inspection device according to the present invention, implementing three convolutional neural networks, which have candidate regions as input data, and the candidate regions are obtained by performing segmentation operations on an analysis image according to a first modality, an analysis image according to a second modality, and an analysis image according to a third modality.

[0098] [ Fig. 9 ] Fig. 9 is a simplified functional block diagram of an example of a second embodiment of an inspection device according to the present invention, implementing a convolutional neural network having as input data an analysis image according to a first modality and a merged image of an analysis image according to a second modality.

[0099] [ Fig.10 ] Fig.10 is a simplified functional block diagram of an example of a second embodiment of an inspection device according to the present invention, implementing a convolutional neural network having as input data all or part of an analysis image according to a first modality and an analysis image according to a second modality without prior segmentation. DETAILED DESCRIPTION

[0100] Figure 1 A device 1 according to the invention is shown for inspecting glass containers 2 leaving all known types of manufacturing or forming facilities 3. The inspection device 1 is intended to detect, for each container, whether the container has a glass defect and, for containers that do have a glass defect, to identify the type of defect among a range of possible glass defects.

[0101] At the exit of the manufacturing facility 3, the containers 2, such as glass bottles or flasks in the example shown, have a high temperature, usually between 300° C. and 600° C. In a known manner, the containers 2 just formed by the facility 3 are supported by an output conveyor 5 so as to form a row of containers by being placed one after the other on the output conveyor in the example shown. The containers 2 are transported by the conveyor 5 in a row in the conveying direction in order to be transported one after the other to different treatment stations, in particular an annealing lehr 6, upstream of which a coating hood, not shown, is placed, which usually constitutes the first treatment station after forming. Advantageously, the inspection device 1 according to the invention inspects the containers downstream of the annealing lehr 6. It is conceivable that the inspection device 1 according to the invention is installed upstream of the coating hood or between the coating hood and the annealing lehr 6.

[0102] The manufacturing facility 3 is known per se and one example will be briefly described merely to allow an understanding of the interaction between the inspection device 1 according to the invention and the manufacturing facility 3 .

[0103] The manufacturing facility 3 comprises a production computer 7 for supervising the different functions of the manufacturing facility 3. The computer is generally a sequencer that controls pneumatic or electric actuators and valves that monitor the circulation of cooling air or blowing pressure. Traditionally, the manufacturing facility 3 comprises several different forming sections that operate in parallel and convey at least one glass container in sequence. In the example of an automatic glass bottle forming machine, the different individual forming sections each comprise at least one blank mold and at least one blow mold that receive the glass semi-finished product. In a known manner, it is possible to identify the forming section, blank mold and blow mold from which each container 2 comes, and for a given production, the order in which the containers travel is known until they enter the annealing furnace 6. When the inspection device 1 according to the invention is installed downstream of the annealing furnace 6, it is conceivable to equip it with a device for reading the information carried on the container and indicating the original section of the mold or container and / or their manufacturing timestamp. Alternatively, it can be proposed to connect and synchronize the inspection device 1 with an information reading device located near the inspection line.

[0104] The inspection device 1 according to the invention comprises an inspection system 10, which includes at least one light source 11, which illuminates the container 2, and at least one camera 12, 12a, ..., which is arranged to collect the light that passes through the container, so as to obtain an image of at least a part of the container illuminated by the light source 11 in a transmission manner. The camera 12, 12a generally comprises an objective lens with an optical center and an optical axis, and a substantially flat linear or matrix photoelectric located in the focal plane. The inspection system 10 is connected to an electronic information processing unit 13. The electronic information processing unit 13 is a computer system of all types, including a computer, peripherals (display unit, storage unit, keyboard, connection to different plant networks, ...), programs implementing in particular image processing algorithms, databases, etc.

[0105] The information processing unit 13 is connected to the production computer 7 in order to receive time information from the production computer when necessary, so that the container 2, its image and its detected defects can be associated with the mold number or molding cavity. Usually, the operation of the inspection system 10 is synchronized with the operation of the container molding cavity.

[0106] Furthermore, the information processing unit 13 transmits the identified glass defects and the measurements performed to the production computer 7, so that the production computer can automatically deduce adjustment information for at least one monitoring parameter of the manufacturing facility 3. This adjustment of the monitoring parameter is performed manually or automatically. Finally, the information processing unit 13 is connected to an ejector to control the ejection of containers identified as defective and / or to a display unit to present the identified glass defects and images of the containers to the operator.

[0107] The inspection system 10 is configured to collect light coming from the light source 11 and having passed through the container, in order to obtain images in a general sense of at least a portion of the container illuminated in transmission. According to the invention, these images are obtained according to at least two different inspection modalities corresponding to different interactions of light with the container wall and with the glass defects to be identified. According to a variant of the embodiment, the inspection system 10 is configured to obtain images according to two different inspection modalities. According to a preferred variant of the embodiment, the inspection system 10 is configured to obtain images according to three different inspection modalities.

[0108] The first inspection modality is called absorption modality. This first modality focuses mainly on the absorption of the penetrating light by the wall of the container. Some defects have the property of complete or partial absorption. Therefore, these defects appear opaque or dark when observed in transmission, that is, the light that passes through the defect-free glass wall undergoes an absorption called normal absorption, which corresponds to the assumed uniform color tone and thickness of the glass wall. However, absorption defects present local anomalies, sometimes with an absorption lower than the normal absorption (bubbles or thin spots) but usually with an absorption higher than the normal absorption. In the following, absorption will refer only to the abnormal absorption of absorption defects. Such defects include in particular inclusions in the glass (especially ceramic or metal inclusions) and / or stains on the glass (grease, etc.). However, such defects also include some cracks (cracks) that can present an orientation in the glass that blocks the inspection light, mainly because the inspection light is then reflected in a direction that is not visible to the camera.

[0109] The second inspection modality is called birefringence. The second modality implements the change of the polarization state of the light passing through the container wall mainly through a defect called stress defect, which gives the glass its birefringent characteristics. Some defects have a birefringent nature. Therefore, some defects are reflected as the presence of residual mechanical stresses in the material (sometimes called internal mechanical stresses). In the wall, birefringent or stress defects, such as the inclusion of foreign matter (ceramic, metal, "devitrified glass"), lead to a change of the polarization state, that is to say a polarization phase shift between the two components of the electric field or a change in the direction of linear complex polarized light.

[0110] The third inspection mode is called the refraction mode. Due to the angle between the intersecting refractive surfaces and due to the difference in refractive index, this third mode mainly implements the change of the propagation direction of the light passing through the container wall through refraction defects. Each surface is an air / glass or glass / air interface, and is therefore a diopter that refracts the light passing through it. In the absence of defects, the surfaces of the wall are essentially parallel, and the refraction does not cause any visible deviation of the light passing through the container. The defects called refraction defects are defects that locally cause abnormal refraction, mainly when the defects appear due to the slope deviation between the surface or diopters of the wall. Therefore, the mention of refraction defects is only to indicate the deviation of light caused by a specific refraction at the defect level called refraction defects. Refraction defects are defects that can be detected mainly by the refraction anomalies they produce (especially in penetrating light inspection). Typically, surface defects (wrinkles, river-like lines) or glass distribution defects (bubbles, thin spots, compression rings), trapezoidal defects and burrs are usually graded in refraction defects. It should be noted that trapezoidal defects and burrs usually cause strong refraction, so that these defects are usually also clearly visible in the absorption image.

[0111] When the present invention implements two modes, the first mode is an absorption mode and the second mode is a birefringence or refraction mode. When the present invention implements three modes, the first mode is an absorption mode, the second mode is a birefringence mode and the third mode is a refraction mode.

[0112] In a known manner, various configurations of the inspection system 10 can be used to perform inspections of containers 2 according to these different inspection modalities. It can be considered that the inspection system 10 is configured to acquire images in order to obtain images called analysis images according to any one or two of the three modalities mentioned above. In practice, the analysis images according to these modalities can be obtained directly from the acquired images or from calculation or processing operations, depending on the inspection system 10 used. Thus, the inspection system 10 associated with the processing unit 13 makes it possible to obtain:

[0113] - an analysis image according to the first modality, referred to in the rest of the description as absorption image Ia, which image consists of absorption pixels whose values ​​mainly represent the absorption of the penetrating light by the intersecting walls;

[0114] - an analysis image according to the second modality, referred to in the rest of the description as birefringence image Ib, which image consists of birefringence pixels whose values ​​mainly represent the changes in the polarization state of the light that has passed through the container wall;

[0115] - An analysis image according to the third modality, called refraction image Ir in the rest of the description, this image consisting of refraction pixels, the values ​​of which mainly represent the changes in the propagation direction of the light passing through the container wall.

[0116] The remainder of the specification describes various methods for obtaining absorption images Ia, birefringence images Ib and refraction images Ir by way of non-limiting examples. Note that in all the following methods, a light source illuminates the inspection area of ​​the container in transmission and is observed against the background of the container.

[0117] In order to obtain the absorption image Ia, a first simple method is to produce non-polarized uniform intensity illumination in the effective part of the light source. The second solution for obtaining the absorption image is to use a linear or circular polarized uniform intensity light source and generate the image with the aid of a camera and without any polarizing filter between the container and the camera. The intensity uniformity of the light source can be perfect, that is to say the intensity is constant over the plane emitting diffuse light and the entire effective area of ​​the extended light source. This uniformity (especially when the glass of the container is tinted) can also be local, and it can be stipulated that the light source area irradiating the neck with a thicker glass wall emits a stronger uniform intensity, while the area of ​​the container body with a thinner wall emits a weaker uniform intensity. As described in document FR 2794241, the intensity of a light source that produces a slower continuous spatial variation than that observed near a defect is also considered to be uniform or relatively uniform, so the image analysis algorithm that compares a pixel with its neighboring pixels to detect fast local intensity changes as defects cannot detect the slow intensity changes emitted by the light source.

[0118] Another solution for obtaining an absorption image is described in patent application EP 3 679 356, which proposes to produce the illumination using a light source with spatially varying color and to obtain a composite image (RGB) converted into the HSV (hue, saturation, value) color representation space. The absorption image is obtained by means of the V image, or a transformation of this image is obtained, for example, by means of a gradient calculation. It should be noted that the refraction image is obtained by means of the H image, or a transformation of this image is obtained, for example, by means of a hue gradient calculation.

[0119] Another solution for obtaining the absorption image is to use a linearly or circularly polarized monochromatic light source of uniform intensity and to form a composite polarization image with the aid of a polarization camera and then to calculate the absorption image from at least two local polarization images corresponding to observations through two linear filters with analysis directions at 90° to each other.

[0120] Another solution to obtain an absorption image is to use a light source with a varying polarization signature while maintaining relatively uniform intensity to produce the illumination, and use a polarization camera to acquire a composite image containing at least 2 to 4 partial images, and then calculate the absorption image based on the 2 to 4 partial images taken through a polarization analyzer at 90° to each other.

[0121] The following describes a method for obtaining an absorption image on the one hand and a refraction image on the other hand using a single inspection system and a single raw image acquisition. Figure 2 As shown, the inspection system includes a diffuse extended light source 11, which is composed of a plurality of elementary light sources 11a (e.g., LEDs) that can be driven independently of each other. The light source 11 also includes a matrix of linear polarization filters 11b and liquid crystal cells 11c for changing the polarization characteristics of the light emitted by each elementary light source. Therefore, the light source 11 is composed of elementary light sources whose intensities can be driven and driven according to polarization characteristics such as polarization direction.

[0122] By these means, the light source 11 can present a polarization characteristic variation function according to any desired spatial variation function over an effective area of ​​the area to be inspected of the container 2 illuminated with transmitted light. According to a variant of the embodiment, the polarization characteristic is the polarization direction. Preferably, the light source 11 has a variation of the polarization characteristic according to a piecewise continuous periodic law (preferably a symmetric trigonometric function).

[0123] The polarization camera 12 transmits a composite digital image In through its photoelectric sensor, which includes a set of composite pixels. Figure 3 As shown, each composite pixel is a group of 4 local pixels placed side by side. In front of each of the 4 local pixels of each composite pixel, a separate linear polarization filter is placed, each filter having four linear polarization directions, oriented at 0°, 45°, 90° and 135°, respectively. By combining the local pixels of each composite pixel with the same directional polarization filter, four local images filtered by 0°, 45°, 90° and 135° linear filters can be obtained, namely, images I0, I45, I90, I135, whose pixels are I0(x,y), I45(x,y), I90(x,y) and I135(x,y), respectively.

[0124] To calculate the absorption image, each pixel IAbs(x,y) is obtained by adding or averaging at least two local pixels corresponding to the two directions of the 90° polarizer.

[0125] IAbs(x,y)=I0(x,x)+I90(x,y) or

[0126] IAbs(x,y)=I45(x,x)+I135(x,y) or

[0127] IAbs(x,y)=1 / 2(I45(x,x)+I135(x,y)+I0(x,x)+I90(x,y))

[0128] In order to obtain the refraction image Iref(x,y), the polarization direction α(x,y) of the light received by the camera is calculated, for example, by using Stokes parameters.

[0129] The first formula for calculating the value of the received polarization direction α(x,y) from the values ​​of two local pixels can be written as:

[0130]

[0131] Or even according to 4 pixels:

[0132]

[0133] Since the polarization direction variation is trigonometrically periodic, it can be observed that, in the absence of refraction, for a defect-free wall with theoretically parallel faces, the polarization direction α(x,y) is half the phase of the trigonometric periodic signal, i.e. The measurement of the direction is equivalent to the measurement of the phase. Since the variation function is, for example, a symmetrical triangle and thus linear almost everywhere, the local phase deviation is proportional to the refraction.

[0134] In order to obtain the birefringence image Ib, a first simple method is to generate linearly polarized monochromatic uniform illumination in the effective part along the P1 direction on the light source. The image is acquired by a black and white camera or by a polarization camera, in front of which a linear polarization analyzer orthogonal to the P1 direction is placed, and the polarization camera considers the pixels corresponding to the analyzed local image through a linear filter orthogonal to the direction P1. The pixel value of the birefringence image is almost zero unless there is a stress defect. When light passes through a stress defect, the luminous intensity obtained depends on the direction and intensity of the stress.

[0135] The second method is to obtain a birefringence image Ib and generate circularly polarized monochromatic uniform illumination in the effective part along the path S1 on the light source. The image is acquired using a black and white camera, in front of which a 1 / 4 wave retardation plate and then a linear polarization analyzer are placed. The pixel value of the birefringence image is almost zero unless there is a stress defect. When light passes through a stress defect, the luminous intensity obtained depends on the intensity of the stress, but not on the direction of the stress.

[0136] A third method of obtaining a birefringence image Ib is to generate linearly polarized monochromatic uniform illumination along a single direction in the effective part on the light source or to generate circularly polarized monochromatic uniform illumination along a single path. The image is acquired by a polarization camera, in front of which a 1 / 4 wave retardation plate that transmits a composite image is optionally placed. The amount of birefringence is calculated based on 2 or 4 local images, which depends on the polarization phase shift between the electric field components Ex and Ey, and the amount of birefringence is a measure of stress. Those skilled in the art can find calculation formulas from the Mallus equation and the Stokes formalism. The polarization phase shift can be calculated to obtain a pixel value in the birefringence image, which depends on the intensity of the stress, but preferably does not depend on the direction of the stress, so the detection is isotropic and proportional to the stress. According to one of these variations, the polarization phase shift can be measured between 0° and 90° or even between 0° and 180°. This method also allows the absorption image Ia to be calculated using the composite image transmitted by the same polarization camera by calculating each pixel IAbs (x, y) as described above.

[0137] In order to obtain the refraction image Ir, a first method for obtaining the refraction image is shown in patent US4606634, which describes a refraction defect detection which consists in varying the "angular spectrum" of an extended light source. A light source of variable size (e.g. a luminous disk of variable diameter) is located at the focus of a converging projection lens. In the image obtained using a camera receiving the light passing through the container, an enhanced contrast is obtained on the refraction target, which can be enhanced by reducing the angular spectrum, which is achieved by reducing the diameter of the luminous disk.

[0138] A second group of methods for obtaining refraction images consists in spatially varying, along the emitting surface of an extended light source (for example a luminous panel), a characteristic of the light emitted by the light source, which the camera knows how to distinguish. As characteristic of the spatially varying light, initially used is its intensity. At least one image is acquired using a camera that is sensitive to intensity (and therefore a priori a monochrome camera that delivers monochrome images). The intensity of the light emitted by each luminous surface unit varies spatially according to a law of spatial variation in one or two dimensions. In other words, these inspection processes suitable for detecting refraction defects use an illumination device that provides light that is sometimes called "structured light", i.e. with an emitting surface, usually two-dimensional, that has intensity variations or intensity patterns.

[0139] These methods include a large number of variations with respect to their variation patterns and possible measurements of features by one or more cameras. The spatial variation pattern is unidirectional or bidirectional. In the bidirectional case, the features can be expected to be distributed according to a regular checkerboard pattern. More commonly, the spatial variation pattern is unidirectional in the vertical, oblique or horizontal direction. In addition, images with vertical spatial variation patterns can be generated and analyzed, and then images with horizontal spatial variation patterns can be generated and analyzed. These images can also be combined. The spatial variation pattern can be uniform and continuous over the entire inspection area. For example, documents US4487322, EP0344617, EP1006350, EP2082216, EP2558847, EP3552001 and EP2875339 describe various variations of these methods.

[0140] A third group of methods for obtaining a refraction image includes spatially varying polarization characteristics, color characteristics and / or phase characteristics along the emitting surface as characteristics of the light emitted by the light source. For example, documents WO2020 / 244815 and EP3679356 show variations of the third group of methods for obtaining a refraction image.

[0141] As can be seen from the examples given above, the inspection method using the inspection system 10 associated with the electronic information processing unit 13 makes it possible to obtain an absorption image Ia, a birefringence image Ib and a refraction image Ir, according to the different methods summarized below.

[0142] Therefore, the method according to the invention aims to inspect the container 2 using an inspection system 10, which is configured to acquire images and to calculate, from several of these images and with the aid of an electronic information processing unit 13, an analysis image corresponding to the absorption image Ia, the birefringence image Ib and / or the refraction image Ir.

[0143] According to another method, the method is intended to inspect the container 2 using an inspection system 10 configured to acquire polarization composite images and to calculate absorption images Ia and / or birefringence images Ib and / or refraction images Ir from these polarization composite images. The inspection system 10 can be configured to acquire color polarization composite images or monochrome polarization composite images.

[0144] According to another method, the method aims at inspecting the container 2 using an inspection system 10 configured to acquire color composite images using a color camera and to calculate the absorption image Ia and the refraction image Ir from these color composite images.

[0145] According to another method, the method aims to inspect the container 2 using an inspection system 10 configured to acquire images that directly correspond to the absorption image Ia, the birefringence image Ib and / or the refraction image Ir.

[0146] The purpose of the combination of these different illumination and image acquisition techniques is to acquire at least two and preferably at least three images from each inspected area of ​​each container, each image according to a different modality, for each modality the value of each pixel depending on the different modality of interaction of the light with the wall it intersects and the defects to be detected. Of course, the configuration of the illumination and the camera depends on the inspected area of ​​the container, which may correspond to, for example, the body, the bottom, the neck, the shoulder, the mouth thread, the mouth or an area where engraving is present.

[0147] Figure 4 The positioning of the cameras 12 is illustrated by way of example, with their viewing directions distributed around the vertical axis of the container 2 to ensure inspection of the entire periphery of the container. According to this example, the inspection system 10 comprises two inspection stations distributed along the path of the container 2. The first inspection station P1 comprises a light source 11, which is arranged along a first side of the trajectory, has a concave emission surface S1, and is composed of a plurality of elementary light sources, which are controlled to define three illumination areas with a first identical illumination configuration and a second identical illumination configuration. The first inspection station also comprises three cameras 12, which are arranged along a second side of the trajectory opposite to the first side and have optical axes D of different directions focused on the illumination areas.

[0148] The second inspection station P2 comprises a light source 11 arranged along the second side of the track, having a concave emitting surface S2, consisting of a plurality of elementary light sources controlled to define three illumination areas having a first identical illumination configuration and a second identical illumination configuration. The second inspection station comprises three cameras 12 arranged along the first side of the track and having optical axes of different directions focused on the illumination areas.

[0149] When the moving container 2 is sequentially substantially centered in the viewing direction of each camera 12, the light source and the camera are controlled to acquire an image of the container illuminated by the relevant illumination area, which is controlled to be turned on sequentially according to the first illumination configuration and the second illumination configuration during the acquisition. Therefore, for each container, six images according to the first modality (e.g., absorption) can be obtained in six viewing directions, and six images according to the second modality (e.g., birefringence modality) can be obtained in six viewing directions.

[0150] Regardless of the examination system used, the information processing unit 13 is suitable for providing an analysis image according to a first modality (absorption image) and an analysis image according to a second modality (birefringence image or refraction image), and according to a preferred variant of the embodiment, even an image according to a third modality. Furthermore, according to the invention, the information processing unit 13 is configured to perform the following operations:

[0151] - taking into account a list of categories including at least glass defects, said list of categories comprising a number of categories which is independent of the number of modes;

[0152] - matching at least a portion of the analysis image according to the first modality with at least a portion of the analysis image according to the second modality or even according to the third modality,

[0153] - from the matched together at least one analysis image according to the first modality and at least one analysis image according to the second modality or even the third modality, classifying the analysis image using an image classifier in order to determine to which result class of the class list the analysis image belongs,

[0154] - taking into account an image classifier that has been trained by supervised learning on a learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with classes from a list, such that the trained classifier classifies the containers according to classification features according to at least two modalities,

[0155] - Sort containers by result category.

[0156] The information processing unit 13 is thus suitable for implementing an inspection method for detecting defects on the containers and for classifying the containers according to previously defined categories.

[0157] According to one feature of the invention, the method aims to define a list of p categories D1, D2, ... Dk, ... Dp comprising at least glass defects (i.e. defects related to the container manufacturing process). The relevant glass defects are glass defects having optical properties that interact with the light passing through the container, such as at least a portion of absorption, and / or a portion of birefringence and / or a portion of refraction, so that they can be detected by the aforementioned device. For a portion of the list, each category corresponds to a glass defect. For example, these categories can correspond to bubbles, inclusions, wrinkles, particles, stones, burrs, large bubbles, trapezoidal defects or bird-wing-shaped defects. In addition, several categories can correspond to the same type of glass defects, such as trapezoidal glass defects. According to this example, one category can correspond to a large trapezoidal defect with thick glass filaments, while another category can correspond to a small trapezoidal defect with a small unconnected tip. These categories are for illustrative purposes only.

[0158] For another part of the list, some categories correspond to artifacts that do not correspond to defects. Thus, these categories can correspond to: protrusions with a technical function, such as positioning marks or stripes on a placement plane; protrusions with a decorative function, such as emblems; or protrusions with a technical or commercial indication function, such as brand, capacity, mold number. Other categories can correspond to distinguishable elements on the container, such as a molded seam, which can be circular at the bottom or linear on the vertical wall. Identifying the molded seam in the image and thus classifying the image element as a molded seam then allows for a specific analysis to determine whether the molded seam has a light imprint (which is not a defect) or a deep imprint (which requires ejecting the container with such a molded seam).

[0159] Advantageously, a severity can be associated with each category in the category list, that is, a higher value for a severe defect category (such as a trapezoidal defect), a lower value for a less severe defect category (such as a wrinkle), and an even lower value for a non-defect target category (such as a molded seam).

[0160] It should be noted that the number p of image categories is independent of the number of modalities. Generally, the category list includes a greater number of categories than the number of modalities implemented. In addition, the category list can include categories that do not correspond to glass defects. For example, the list can include categories that do not correspond to glass defects, categories corresponding to the container 2 without defects, categories corresponding to the container 2 with a molded seam, and categories corresponding to the container 2 with an emblem. According to an advantageous variant of the embodiment, the category list contains non-glass defect categories, at least one trapezoidal defect category, at least one inclusion category, and at least one bubble category. These categories are recorded and accessible by the information processing unit 13.

[0161] Thus, the number of categories is first determined by production requirements and quality control, and thus by the need to identify production defects in order to make correct decisions during sorting and to allow possible corrections to the process. In contrast, the number of modalities is determined only by the technical and economic limitations of the known means used to emphasize absorption, refraction, and birefringence characteristics. According to the prior art described in the patent application EP 3 679 356, based on the reasoning of a person skilled in the art based on the prior knowledge of the optical interaction of defects in two modalities A and B, it is concluded that the number of modalities is used to predict the determined number of categories, that is, in this example there are only four categories: A and B, A and not B, not A but B, A strong and B weak.

[0162] The number of categories also depends on the quality of sorting obtained with the help of a supervised image classifier. In fact, during the training of the image classifier, it is known to verify the good classification rate obtained on the test set. It has been observed that the classification is better when the list of categories contains several categories for the same defect, such as a trapezoidal defect. In other words, the number of categories can be increased to improve the quality of automatic classification.

[0163] from Figures 5 to 10 It can be seen that, for each container 2, the inspection method according to the invention comprises performing at least one acquisition operation Ac1 to obtain at least one analysis image according to a first modality, and performing one acquisition operation Ac2 to obtain at least one analysis image according to a second modality. Figure 8 In the example shown, for each container 2, the inspection method according to the present invention further comprises performing an acquisition operation Ac3 to obtain an analysis image according to a third modality. As described above, absorption, birefringence and refraction images can be directly obtained. Therefore, the absorption, refraction and birefringence images are the same as the acquired images.

[0164] Absorption, birefringence, and refraction images can also be obtained by performing computational operations on the obtained images. Figure 6 In the example shown, a polarization camera or a color camera allows acquisition of a polarization composite image or a color image, respectively. From this composite image, calculation operations C1, C2 are performed to obtain an analysis image according to the first modality and an analysis image according to the second modality. Thus, with a light source of uniform polarization, an absorption image and a birefringence image can be obtained, while with a light source of varying polarization direction, an absorption image and a refraction image can be obtained.

[0165] The method according to the invention consists in ensuring an analysis operation of the image, comprising an operation or step of matching MC at least a portion of an analysis image according to a first modality with at least a portion of an analysis image according to a second modality and possibly at least a portion of an analysis image according to a third modality. The method then consists in implementing a classification or classification step C1, using an image classifier, from the information contained at least in the analysis image according to the first modality and the analysis image according to the second modality and possibly the analysis image according to the third modality that are matched together, in order to classify the container in a resulting class Dk that forms part of a list of classes.

[0166] According to Figures 5 to 8 In a first embodiment implemented in an exemplary embodiment of the present invention, the analysis operation is intended to process the analysis image so as to extract therefrom a region corresponding to a visible feature or object, if such feature or object is present. Fig. 9 and Fig.10In the second embodiment described in detail in the exemplary embodiment of the present invention, the purpose of the image analysis operation is not to extract candidate regions therefrom, but to consider all or part of the analysis image according to the first modality and the analysis image according to the second modality and / or according to the third modality.

[0167] An object in a digital analytical image is typically a group of related pixels that have at least one common feature that is not shared by adjacent groups. Thus, an object is surrounded by a closed contour and is identified solely from features of the analytical image, such as luminosity, grayscale, intensity, texture, spatial frequency, contrast, color characteristics, or any measurement of optical features detected by the camera.

[0168] A target corresponds to a zone or region of the analyzed image where glass defects may exist. A target does not necessarily consist of a single image region surrounded by a single closed contour, because a target can be formed by several disconnected parts. Therefore, if different disconnected parts of the same target are close to each other, one region can contain the different disconnected parts. If they are far apart, it can be considered that two different regions form the same target. Regions with targets (also called candidate regions) represent targets that can be classified as belonging to one of the target categories in the list of possible categories defined above.

[0169] The operation of analyzing the analysis images according to the various modalities is performed by implementing one or more image processing operations by any type of digital processing known per se (such as thresholding, histogram correction, convolution operations or binary or grayscale mathematical morphological operations) in such a way as to extract all the areas with the target. Such image processing operations known to those skilled in the art can be applied during the operations C1, C2 of calculating the analysis images, or during the analysis prior to the matching MC, or, for example, in the segmentation step SR described below.

[0170] Therefore, the method implements an operation SR1 of segmenting and detecting candidate regions on an analysis image according to a first modality, an operation SR2 of segmenting and detecting candidate regions on an analysis image according to a second modality, and an operation SR3 of segmenting and detecting candidate regions on an analysis image according to a third modality ( Figure 5 , Figure 7 and Figure 8 ).according to Figure 6 The exemplary embodiment shown performs an operation SR of segmenting and detecting candidate regions on a merged analysis image IF of an analysis image according to the first modality and an analysis image according to the second modality, which will be explained in the remainder of the description.

[0171] The segmentation operation usually consists in cutting the image into regions or segments, that is to say, assigning pixels to regions. This segmentation operation aims to determine candidate regions in each analyzed image by filtering, thresholding, contour tracing, operations, etc., with the aim usually (but not necessarily) of measuring parameters characterizing these regions. This image segmentation operation is performed according to a filtering method suitable for the implemented modality or merged image IF.

[0172] These segmentation operations SR1, SR2, SR3, SR enable detection of candidate image regions defined by their contours confined to the target, and these candidate image regions RC1, RC2, RC3 and RCC are respectively according to an analysis image of the first modality, according to an analysis image of the second modality, according to an analysis image of the third modality, or according to a merged image of at least two modalities. It is also possible that these segmentation operations SR1, SR2, SR3, SR enable detection of candidate image regions RE1, RE2, RE3, REC defined by their rectangles framing the target, and these candidate image regions are respectively according to an analysis image of the first modality, according to an analysis image of the second modality, according to an analysis image of the third modality, or according to a merged image of at least two modalities. It is also possible that these segmentation operations SR1, SR2, SR3, SR enable detection of candidate image regions RL1, RL2, RL3, RLC defined by an enlarged rectangle framing the target so as to take into account the background of the target in the image during classification, these candidate image regions being respectively an analysis image according to a first modality, an analysis image according to a second modality, an analysis image according to a third modality or a merged image according to at least two modalities.

[0173] According to the method of the present invention, a matching operation MC is performed for each container to match the candidate region in the analysis image of the first modality, the second modality or the third modality with the corresponding region of the analysis image of at least one other modality. The method can also perform matching of candidate regions of two different modalities for each container.

[0174] This matching operation MC aims to ensure the matching of the regions in the analysis images of at least two different modalities by comparing their positions on the container. In the most common case, a geometric transformation from one analysis image to another is determined, which allows, starting from a region or a pixel of an analysis image of the container, to locate a region or a pixel of another analysis image corresponding to the same region or essential part of the container. The geometric transformation can be of any necessary type and includes, for example, translation / rotation, deformation, scaling, etc. When two different analysis images 1b, 1r or 1a of the same container 2 are transmitted by two different cameras with two different objectives, by taking into account the three-dimensional geometry of the container 2, its position relative to the cameras 12, 12a when the images are acquired, and the geometric and optical parameters of the inspection system 10 (for example the direction of the optical axis, the position of the optical center and the focal length of the objective of the camera 12, 12a, which determines the optical projection of the container on the flat image sensor of the camera's objective), it is possible to determine a geometric transformation that associates the regions or pixels of the two analysis images 1b, 1r or 1a corresponding to the same region of the container wall.

[0175] According to a variant of the embodiment, a pixel-to-pixel matching of two analysis images of different modalities of a container or of two analysis image regions of different modalities of a container is performed. To this end, a geometric transformation is determined for all pixels. A transformed image that can be superimposed on the other image or a portion of an image can also be calculated for one of the two analysis images or a portion of an analysis image. The geometric transformation and interpolation of the pixel values ​​(e.g. bilinear interpolation) are then applied to all pixels of the relevant region.

[0176] In case the analysis images of different modalities are matched pixel-to-pixel due to the inspection system 10, the matching is straightforward. This is particularly the case during the acquisition of monochrome or color polarization composite images or color composite images, as described above.

[0177] According to another variation of the embodiment, the area whose center or center of gravity is close to the container, that is, the area matched or adjacent by geometric transformation is matched, or the area where the rectangle that frames the object or the enlarged rectangle that frames the object intersects or overlaps with a certain ratio of a given surface area on the container is matched.

[0178] Furthermore, the matching MC may be performed from candidate region to candidate region, or from a candidate region to a corresponding region determined during matching, or as in the illustrated exemplary embodiment, from pixel to pixel.

[0179] according to Figure 6 and Fig. 9The exemplary embodiment shown forms a merged image IF by matching an image or a portion of an analysis image according to a first modality with an image or a portion of an analysis image according to a second modality and / or an image or a portion of an analysis image according to a third modality. Thus, as a matching, the method ensures a merging of the analysis images according to the first modality and according to the second modality and / or the third modality to obtain a merged image IF. Then, the method ensures a segmentation of the merged image to detect a merge candidate region.

[0180] The matching operation MC can involve all analysis images or only a part of these analysis images. As mentioned before, this matching MC is done pixel by pixel. A value depending on the modality is assigned to each pixel pc(x,y) of the coordinates x, y of the composite image. This value is, for example, a 16-bit scalar pc(x,y) with 8 absorption bits and 8 refraction or birefringence bits, or a vector vc(x,y) whose components {vt(x,y), vr(x,y)} are respectively the absorption scalar and the refraction or birefringence scalar. The simplest involves directly obtaining the pixel value of the analysis image in absorption and the pixel value of the image in the refraction or birefringence image that matches the pixel of the analysis image in absorption. However, it is also possible to construct each merged pixel from a combination of several adjacent pixels in the image or from interpolated values.

[0181] It should be noted that in the described example, the merged image IF corresponds to a merge by matching of the analysis images according to the first modality and according to the second modality and / or the third modality. According to a variant of the embodiment not shown in the drawings, the merged image IF can be directly equivalent to a composite image transmitted by a monochrome polarization composite image sensor, a color polarization composite image sensor, or a color composite image sensor. In other words, the monochrome polarization composite image, the color polarization composite image, or the color composite image can be directly analyzed as the merged image IF.

[0182] according to Figure 5 , Figure 7 and Figure 8 In the exemplary embodiment shown, matching is performed from candidate region to candidate region. Registration of one image on another image can be performed so that the candidate regions coincide on two images of different modalities. It is also possible to directly search for candidate regions located in the same area of ​​the container.

[0183] The method according to the invention also aims to select classification features according to the first modality and classification features according to the second modality and / or classification features according to the third modality, and / or combined features, which take into account the features of the analysis image according to the first modality and the analysis image according to the second modality and / or the analysis image according to the third modality combined in a logical or mathematical way. These classification features according to the first, second and third modalities are features of position, size, shape (concavity, perimeter, surface, etc.) or values ​​representative of absorption and / or refraction and / or birefringence (photometric values, such as average level, contrast, variance, texture, etc.).

[0184] according to Figure 5 and Figure 6 In the exemplary embodiment shown, operations EC1, EC2, EC are performed according to the method of the present invention to extract classification features. Figure 5 In the exemplary embodiment shown, the method implements an operation EC1 of extracting classification features for candidate regions RC1, RE1, RL1 from an analysis image according to a first modality, so that an n-dimensional vector Ct representing n features m1i obtained from an analysis image according to the first modality can be defined for each candidate region. Similarly, the method according to the present invention implements an operation EC2 of extracting classification criteria for candidate regions RC2, RE2, RL2 from an analysis image according to a second modality, so that an m-dimensional vector Cr representing m features m2i obtained from an analysis image according to the second modality can be defined for each candidate region.

[0185] It should be noted that according to Figure 5 In the exemplary embodiment shown, the operation MC of matching the candidate regions RC1, RE1, RL1 of the analysis image according to the first modality with the candidate regions RC2, RE2, RL2 of the analysis image having the second modality makes it possible to obtain an n+m-dimensional vector Cc, which represents the n+m features m1i, m2i obtained for each candidate region matched between the analysis images according to the first and second modalities.

[0186] exist Figure 6 In the exemplary embodiment shown, the method according to the invention implements an operation EC of extracting classification features for the merged candidate regions RCC, REC or RLC according to the first modality and according to the second modality obtained after the segmentation operation SR. This extraction operation makes it possible to obtain a vector Cc of n+m dimensions representing n+m features m1i, m2i for each candidate region matched between the analysis image according to the first modality and the analysis image according to the second modality, or for a merged candidate region or a composite candidate region from a monochromatic polarization sensor, a color sensor or a color polarization sensor.

[0187] exist Figure 5 and Figure 6 In an exemplary embodiment of , the classification features are determined by preliminary analysis (ie, expert knowledge or statistical research). The image analysis algorithm determines the selected features m1i, m2i of position, size, shape and luminosity. Figures 7 to 10 In the exemplary embodiment shown, the classification features are determined by supervised learning embedded in the trained neural networks CNN, CNN1, CNN2, as described in the rest of the description. In fact, the role of the layers of the convolutional neural network, called convolutional layers, is to be able to determine by learning the parameters that allow the extraction of important image features for classification, which are then automatically extracted by the trained neural network during classification. For example, the neural network used can be a model of the known type according to the abbreviations RESNET or VGG or any other known network, which is generally available in open source.

[0188] As an indication, the classifier of the present invention can be implemented by a neural network as described above, or by a model of the Transformer type.

[0189] Using previously determined classification criteria applied to the classification features, the method classifies defects and therefore containers with these defects. The classification operation makes it possible to decide on the target class Dk of a candidate region or container from a list of p possible classes D1, D2, ... Dp. If a candidate region is found in only one of the two images according to the first modality, the analysis of the defect is performed according to the features associated with the type of image utilized, but also taking into account the features associated with the other modality: an analysis based on the merging of the features associated with both types of images is performed. The principle of the invention is based on taking into account at least two inspection modalities in order to provide additional and reliable information to make a decision on the classification of an object or container and therefore to identify defects. It should be noted that according to the prior art methods, when the information in the first modality enables the detection of a candidate region, but when the information in the corresponding region of the image according to the second modality is very weak, the information in the image according to the second modality is ignored and therefore does not contribute to the classification.

[0190] According to the present invention, the classification operation is performed by an image classifier trained by supervised learning. Figure 5 and Figure 6 In the exemplary embodiment shown, the image classifier CI can be, for example, a support vector machine (SVM), a Bayesian classifier or a neural network NN. Figure 7 and Fig.10 In the exemplary embodiment shown, a convolutional neural network CNN is implemented as image classifier C1.

[0191] It should be understood that the method according to the invention aims at implementing a learning step of an image classifier for classifying the analysis image. This learning step is of course performed before the step of implementing the image classifier for classifying the containers. This image classifier is trained by supervised learning, that is to say by applying to it the operation of classification to be performed.

[0192] The image classifier C1 is trained by a supervised learning method, which involves determining the parameters of the image classifier from a set of objects or images of known categories (called a learning set). According to this supervised learning method, the system is provided with sorted, labeled data assigned according to a predefined number of categories. Each data is associated with one of the categories by sorting and labeling, which will allow the algorithm to calculate a more general model, which then allows any unknown, unlabeled data to be associated with one of the previously defined categories. This supervised learning method is different from the unsupervised learning method (no prior knowledge about the categories). According to this method, the data is provided to the system in batches without any sorting or labeling. The system itself is responsible for defining the number of categories it considers most relevant and associating one of the categories with each data (for example: X-means clustering algorithm). The supervised learning method is also different from unsupervised learning (with prior knowledge about the number of categories): the data is provided to the system in batches without any sorting or labeling. On the other hand, the number of expected categories N is indicated to the system. Then, the system automatically associates each data with one of the N expected classes (for example: K-means clustering algorithm).

[0193] The image classifier for container inspection implemented within the framework of the invention has been trained on a learning set comprising a plurality of records, each consisting of an image or an image region of the same container according to each modality associated with a class in a list of classes, so that the trained image classifier classifies the containers according to classification features according to at least two modalities. Of course, as mentioned above, the list of classes considered includes at least the representation of glass defects.

[0194] For an example or reference container, each record in the learning set includes:

[0195] - at least one analysis image according to the first modality and at least one analysis image according to the second modality and / or one analysis image according to the third modality matched together, and at least one label assigned to an exemplary container, at least one target class from a list of possible classes, or,

[0196] - at least one analysis image region of an exemplary container according to a first modality and at least one image analysis region of a second modality and / or a third modality of the exemplary container matched together, and at least one label assigned to the corresponding region of the exemplary container, at least one target category from a list of possible categories.

[0197] Therefore, the learning set includes pairs or triplets of matching regions, preferably pairs or triplets of feature vectors associated with a glass defect.

[0198] This image classifier, which has been trained during the learning phase, is used during the inspection of the container to ensure that the container is classified according to the input data applied to the image classifier and representing the inspected container. Therefore, the information processing unit 13 is configured to ensure the execution or implementation of the image classifier previously trained by supervised learning so that the trained image classifier classifies the containers.

[0199] according to Figure 5 and Figure 6 In the example shown, the input data of the image classifier is information of a vector Cc of dimension n+m, which represents the features m1i, m2i obtained for each candidate region matched between the analysis images of the first and second modalities. Typically, the classification features according to the first modality and the classification features according to the second modality and / or the classification features according to the third modality are the input data of the image classifier.

[0200] according to Figure 7 The exemplary embodiment shown, as an image classifier of a convolutional neural network CNN, has as input data a pair of candidate regions (RL1, RL2), which are obtained after a matching operation MC of these candidate regions.

[0201] according to Figure 8 In the exemplary embodiment shown, the image classifier includes a first convolutional neural network CNN1 having candidate regions (RC1, RE1, RL1) according to a first modality as input data, the input data being obtained after an operation SR1 of segmenting and detecting the candidate regions on an analysis image according to the first modality. The image classifier also includes a second convolutional neural network CNN2 having candidate regions (RC2, RE2, RL2) according to a second modality as input data, the input data being obtained after an operation SR2 of segmenting and detecting the candidate regions on an analysis image according to the second modality. The image classifier also includes a third convolutional neural network CNN3 having candidate regions (RC3, RE3, RL3) according to a third modality as input data, the input data being obtained after an operation SR3 of segmenting and detecting the candidate regions on an analysis image according to the third modality.

[0202] The first convolutional neural network CNN1, the second convolutional neural network CNN2 and the third convolutional neural network CNN3 work in parallel on the candidate area of ​​absorption, the candidate area of ​​birefringence and the candidate area of ​​refraction respectively, and these three areas are associated according to three modes through the matching operation MC according to the above-mentioned technology.

[0203] The outputs of the first convolutional neural network CNN1, the second convolutional neural network CNN2 and the third convolutional neural network CNN3 are input data for an image classifier, which is for example of SVM, random forest, Bayesian type, and preferably a neural network NN, allowing classification according to three modalities. The outputs of the first convolutional neural network CNN1, the second convolutional neural network CNN2 and the third convolutional neural network CNN3 are, for example, hypotheses of the categories to which they belong, but they can be more complex data with vectors whose dimensions are greater than the number of categories p. It should be noted that the learning set of the neural network contains candidate regions in groups of three according to the three inspection modalities, with the target category in the list of possible categories as a label.

[0204] According to Fig. 9 and Fig.10 In a second embodiment of the exemplary embodiment of the present invention, the purpose of the image analysis operation is not to extract candidate regions therefrom, but to take into account all or part of the analysis image according to the first modality and all or part of the analysis image according to the second modality and / or according to the third modality without performing a previous segmentation operation. If only parts of the image are analyzed, these parts preferably correspond to one or more areas of interest of the container, such as the mouth, the collar, the shoulder, the body, the mouth thread, or the left and right halves, or areas where there are engravings. Fig. 9 and Fig.10 In the two exemplary embodiments shown, the analysis operation is based on implementing a neural network as an image classifier trained by supervised learning.

[0205] exist Fig. 9In the exemplary embodiment shown, an operation MC of matching an analysis image according to a first modality and an analysis image according to a second modality is performed to obtain a merged image IF. The fused image IF is obtained by merging at least one analysis image of the container according to the first modality with at least one analysis image according to the second modality, or by merging a region of at least one image according to the first modality with a corresponding region of at least one image according to the second modality. In general, as already explained, the fused image IC is obtained by merging at least one analysis image of the container according to the first modality with at least one analysis image according to the second modality and / or one analysis image according to the third modality, or by merging a region of at least one analysis image according to the first modality with a corresponding region of at least one analysis image according to the second modality and / or one analysis image according to the third modality. Similarly, the merged image IF can be directly equivalent to the composite image transmitted by the image sensor.

[0206] The analysis images according to the various modalities are taken during acquisition operations Ac1, Ac2 performed by the inspection system 10, as explained in the above description. This matching MC of the pixel-to-pixel images is performed, as Figure 6 As explained in the exemplary embodiment of . The merged image is used as input data for a convolutional neural network CNN that is able to take into account, by supervised learning, location, size, shape and photometric features that are important for the intended classification. It should be noted that the learning set contains the merged image IF or a region of the merged image whose label is the target class from a list of possible classes. The classification features according to the various modalities are taken into account in the weights resulting from learning and defining the convolutional neural network CNN. It should be noted that, unlike Figure 6 and Figure 7 Unlike the example in , in this variant the segmentation operation is not necessary, because the convolutional neural network step is able to classify images based on their content by learning, without prior segmentation, and is able to locally determine the position, size, shape, and photometric features that are important for classification. However, the segmentation operation is possible, for example by replacing the image with an image classifier based on a convolutional neural network CNN. Figure 6 Feature extraction EC and image classifier CL in.

[0207] exist Fig.10 In the exemplary embodiment shown, the analysis image (partially or completely) according to the first modality is used as input data for a first convolutional neural network CNN1, while the analysis image (partially or completely) according to the second modality is used as input data for a second convolutional neural network CNN2. As mentioned above, these convolutional neural networks have previously been trained by supervised learning in order to be able to determine morphological features important for the intended classification, such as position, size, shape and luminosity.

[0208] The first convolutional neural network CNN1 and the second convolutional neural network CNN2 work in parallel on two candidate images in each modality, respectively. The output of the first convolutional neural network CNN1 and the output of the second convolutional neural network CNN2 are input data for a classifier, for example of the SVM, random forest, Bayesian type, and preferably a neural network NN, allowing the container to be classified according to the two modalities. It should be noted that the two candidate images in each modality on which the first convolutional neural network CNN1 and the second convolutional neural network CNN2 work are associated by a matching operation.

[0209] The output of the first convolutional neural network CNN1 and the output of the second convolutional neural network CNN2 are, for example, hypotheses of the categories to which they belong, but they can also be more complex data whose vector dimension is greater than the number of categories p. Figure 8 Unlike the example in , in this variant, a segmentation operation is not necessary because the convolutional neural network step is able to classify images based on their content by learning without prior segmentation and is able to locally determine morphological features such as position, size or shape, and photometric features that are important for classification.

[0210] As can be seen from the above description, the method according to the present invention is not only capable of identifying glass defects of containers in a transmission manner, but also capable of classifying these defects to allow a shift from inspection to optimization of the manufacturing process. One of the features of the present invention is the definition of a category list including defect categories, which makes it possible to associate glass defects with characteristics of the manufacturing process to be adjusted. The improvement of glass defect classification enables better tracing of the causes of glass defects.

[0211] The object of the invention is advantageously exploited within the framework of a manufacturing facility to allow better detection and classification of defects present in containers.Thanks to the combination of two or three modalities, some defects can be more easily found, detected and classified.

[0212] Preferably, the inspection method according to the invention is designed so that the image classifier associates a confidence score with the classification of the container resulting from the production. The confidence score is generally the probability that the container belongs to the class. The score can be expressed as a % (percentage) or a value between 0 and 1.

[0213] It should be noted that one container can contain several defects. There are several methods for classifying such containers. In a variant of the method comprising a segmentation step, analysis images of the same container can be extracted, several image regions and, for example, several segments SR being identified as belonging to a defect class. In this case, according to a first variant, the class to which the container belongs will be the class of the segments classified in the defect class with the highest severity. It is also possible to take into account a confidence score for the classification, that is, the class assigned to the container will be the class of the segments with a confidence score above a confidence threshold. According to a second variant, it is possible to count all the defects carried by a single container (in particular when a container carries several severe defects). Thus, the production statistical analysis to be described later can take into account the distribution of defects independently of the number of containers rejected.

[0214] The object of the invention is to sort the production of containers in the following manner. After the containers have been classified, at least one container sorting feature is compared with a rejection criterion and, when the sorting feature exceeds the rejection criterion for the container, the container is considered non-compliant and is rejected. In practice, the facility comprises a rejector for removing defective containers from production. The sorting feature and the rejection criterion determine whether the container is qualified or not depending on the category to which it belongs, the sorting feature being calculated on at least one image of the container according to one of two or three modalities. The sorting feature and the rejection criterion are, for example, defect dimensions, such as its surface or its length measured in at least one analysis image. Confidence levels can be taken into account for sorting in the following way, containers belonging to the category of low severity defects are rejected only when the confidence level is high and, conversely, containers belonging to the category of severe defects are rejected even when the confidence level is low.

[0215] The object of the invention is to carry out a statistical analysis of the container production, that is to say to analyse the frequency or distribution of different types of defects and their severity, the defect types and their severity being included in a list of categories which are predetermined for the purpose of monitoring the process.

[0216] Some defects are caused in the step of molding the container in the mold and are therefore related to different molding parameters between different sections or between different cavities. Therefore, it is preferred to count the defect categories according to the original section or cavity of the container. When the facility is installed at the exit of the molding machine (thus upstream of the annealing furnace 6), the inspection is carried out immediately after manufacturing, so the timestamp of the container manufacturing is known, and by synchronization, the original cavity or section of the container is also known, because the order in which the containers leave the molding machine is known. When the inspection device 1 according to the invention is installed downstream of the annealing furnace 6, it is preferably equipped with a device for reading the information carried on the container and indicating the original section of the mold or container and / or their manufacturing timestamp, or is connected to such a reading device. Therefore, it is feasible and preferred to perform a statistical analysis of the production starting from the classification of the containers through the inspection process, according to the distribution of defects directly related to the production parameters when manufacturing each container and / or according to the different cavities and sections of the manufacturing machine.

[0217] Therefore, statistical analysis of container production makes it possible to link defects to their causes in order to obtain two results:

[0218] - On the one hand, correlations between manufacturing parameters and resulting defects can be determined, allowing more effective process adjustment methods to be defined;

[0219] On the other hand, knowledge of the causal relationships makes it possible to provide, in real time, the production computer 7 with the probability of forming a feedback loop in order to regulate the method by correcting the defects and thus the difference between the expected quality and the estimated quality of the container.

[0220] In summary, the purpose of the present invention is advantageously used in the following various production operations:

[0221] - Sorting of production;

[0222] - Conduct statistical analysis of production;

[0223] - Determine the cause-effect relationship between production parameters and defects;

[0224] - monitor the method by reducing observed defects;

[0225] -Alarms are used to warn operators of serious defects.

[0226] According to a preferred variant of the invention, the classification of a container is considered only if the confidence score of the assigned class exceeds a confidence threshold, so that:

[0227] - count defects by defect category in defect frequency statistics,

[0228] - and / or decide to reject the container,

[0229] - and / or triggering an alarm regarding the presence of one or more serious defects in the inspected product.

[0230] It should be noted that the confidence threshold corresponds to a predetermined or adjustable minimum value of the confidence score.

[0231] Taking these different modalities into account through a supervised learning classification algorithm allows not only the automatic management of different aspects of light / matter interactions (absorption, birefringence, refraction), but also the union of these interactions resulting from the analytical image acquisition method. Through supervised learning, the image classifier takes into account photometric and morphological features in both modalities or even all three modalities. Thus, the image classifier is able to determine, by learning, the geometric and photometric features that are meaningful for the intended classification. In addition, such an image classifier is able to associate unknown and unlabeled image data with previously defined categories during inspection.

[0232] The method according to the invention is therefore different from the classification methods of the prior art, the classification process of which is based solely on physical or logical considerations, applying predefined business rules. In contrast, the image classifier implemented according to the invention makes it possible to construct a model of association between input data and categories. In view of supervised learning, the input data are directly associated by the image classifier with one of the categories selected from the list of categories.

Claims

1. A method for inspecting glass containers (2) to classify the containers, the method comprising: The following steps are involved: - inspecting each container using an inspection system (10) comprising at least one light source (11) illuminating the container and at least one camera (12, 12a) arranged to collect light that has passed through the container to acquire an image of at least a portion of the container illuminated in transmission, thereby obtaining at least one analysis image according to a first modality corresponding to an absorption image (Ia) and at least one analysis image according to a second modality corresponding to a birefringence image (Ib) or a refraction image (Ir), - defining a list of categories (D1, D2, .. Dk, .. Dp), said list of categories comprising at least glass defects, said list of categories comprising a number of categories, said number of categories being independent of the number of modes; - ensuring a match between at least a portion of an analysis image according to the first modality and at least a portion of an analysis image according to the second modality, - from the matched together at least one analysis image according to the first modality and at least one analysis image according to the second modality, classifying the analysis image by means of an image classifier (C1), the image classifier determining to which result class of the class list the analysis image belongs, - the image classifier (C1) has been trained by supervised learning on a learning set, the learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with a class from the class list, so that the trained image classifier classifies the container according to classification features according to at least two modalities; - sorting said containers according to said result categories.

2. The method according to claim 1, in, The method comprises the following steps: - inspecting each container using the inspection system (10), the inspection system being configured to acquire images to obtain at least one analysis image according to the first modality, at least one analysis image according to the second modality and at least one analysis image according to a third modality, the analysis image according to the second modality corresponding to the birefringence image (Ib), the analysis image according to the third modality corresponding to the refraction image (Ir), - ensuring a matching of at least a portion of an analysis image according to the first modality, at least a portion of an analysis image according to the second modality and at least a portion of an analysis image according to the third modality, - classifying the analysis image from the matched at least one analysis image according to the first modality, at least one analysis image according to the second modality and at least one analysis image according to the third modality by means of an image classifier (C1), the image classifier determining to which result class of the class list the analysis image belongs, - the image classifier has been trained by supervised learning on a learning set, the learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with a class from a list of classes, such that the trained image classifier classifies the container according to a classification criterion according to at least three modalities; - sorting said containers according to said result categories.

3. The method according to any one of the preceding claims, in, The method is intended to inspect the container using the inspection system (10) configured to acquire images and to calculate, from several of these images, an analysis image corresponding to an absorption image (Ia), a birefringence image (Ib) and / or a refraction image (Ir).

4. The method according to any one of claims 1 or 2, in, The method is intended to inspect the container using the inspection system (10), which is configured to acquire polarization composite images and to calculate absorption images and / or birefringence images and / or refraction images from these polarization composite images.

5. The method according to any one of claims 1 or 2, in, The method is directed to inspecting the container using the inspection system (10), the inspection system being configured to acquire color composite images using a color camera and to calculate absorption images and refraction images from these color composite images.

6. The method according to any one of claims 1 or 2, in, The method is directed to inspecting a container using the inspection system (10), the inspection system being configured to acquire an image that directly corresponds to an absorption image, a birefringence image and / or a refraction image.

7. The method according to any one of the preceding claims, in, In order to ensure a match of at least a portion of an analysis image according to the first modality and at least a portion of an analysis image according to the second modality and / or at least a portion of an analysis image according to the third modality, the method detects candidate regions (RC1, RC2, RC3, RCC) in the analysis image of the first modality and in the analysis image of the second modality and / or in the analysis image of the third modality, and for each container the method ensures that: - matching of a candidate region in an analysis image of the first modality, the second modality or the third modality with a corresponding region of an analysis image of at least one other modality, -Or matching of candidate regions from two different modalities.

8. The method according to any one of the preceding claims, in, As a match, the method ensures a merging of at least one analysis image of the first modality with an analysis image of the second modality and / or an analysis image of the third modality to obtain a merged image (IF), the method ensuring: - extracting classification features (m1i, m2i) from the merged image, - and classifying the container using a classification criterion applied to the classification features of the merged image.

9. The method according to any one of the preceding claims, in, As a match, the method ensures a merging of at least one analysis image of the first modality with an analysis image of the second modality and / or the third modality to obtain a merged image, the method ensuring: - segmenting the merged image to detect merge candidate regions, - classifying the container using a classification criterion applied to the features of the merge candidate region.

10. The method according to any one of the preceding claims, in: - extracting classification features according to the first modality from the analysis image according to the first modality, - extracting classification features according to the second modality and classification features according to the third modality from the analysis image according to the second modality and / or the analysis image according to the third modality, respectively, - grading the containers using classification criteria applied according to features of the first modality and the second modality and / or the third modality.

11. The method according to the preceding claim, in, Classification features according to the first modality and classification features according to the second modality and / or classification features according to the third modality, and / or combined features that logically or mathematically combine features of the analysis image according to the first modality and the analysis image according to the second modality and / or the analysis image according to the third modality are selected, and these classification features according to the first, second and third modalities are features of position, size, shape or features representing values ​​of absorption and / or refraction and / or birefringence.

12. The method according to any one of the preceding claims, in, The containers are classified by a supervised learning image classifier, the input data of which is: - according to the classification characteristics of the first modality and according to the classification characteristics of the second modality and / or according to the classification characteristics of the third modality, - either according to an analysis image of the first modality and according to an analysis image of the second modality and / or according to an analysis image of the third modality, - or a portion of an analysis image according to the first modality and a portion of an analysis image according to the second modality and / or according to the third modality.

13. The method according to any one of claims 1 to 10, in, The containers are graded by a supervised learning image classifier, whose input data is at least one merged image (IF) obtained by merging at least one analysis image according to the first modality and an analysis image according to the second modality and / or an analysis image according to the third modality, or by merging a region of at least one analysis image according to the first modality and a region of an analysis image according to the second modality and / or a region of an analysis image according to the third modality.

14. The method according to any one of the preceding claims, in, Each container (2) is classified according to at least one category selected from a list of categories representing at least glass defects, such as, in particular, trapezoidal defects, inclusions, bubbles.

15. The method according to any one of the preceding claims, in, At least one sorting feature is compared with a rejection criterion, the sorting feature and the rejection criterion determining whether the container is acceptable or not depending on the category, the sorting feature being calculated on at least one image of the container according to a modality.

16. Inspection method according to any one of the preceding claims, in, A step of taking into account at least one detected glass defect is carried out in order to deduce therefrom adjustment information for at least one monitoring parameter of the container production facility (3).

17. Inspection method according to any one of the preceding claims, in: - the image classifier associates a confidence score with the classification of a container forming part of the inspected product; - the classification of the container is considered only if the confidence score exceeds a confidence threshold; - Count defects by defect category; - and / or decide to reject said container; - and / or triggering an alarm indicating the presence of at least one serious defect in the inspected product.

18. A device for inspecting glass containers (2) leaving a manufacturing facility (3) to sort said containers according to glass defects, said device include: an inspection system (10) comprising at least one light source (11) illuminating the container and at least one camera (12, 12a) arranged to collect light passing through the container to acquire an image of at least a portion of the container illuminated by the light source in transmission, - an information processing unit (13) connected to the inspection system (10) and adapted to provide, for each container, at least one analysis image according to a first modality corresponding to an absorption image (Ia) and at least one analysis image according to a second modality corresponding to a birefringence image (Ib) or a refraction image (Ir), the information processing unit (13) being configured to perform the following operations: * taking into account a list of categories (D1, D2, ... Dp) representing at least a glass defect, said list of categories comprising a number of categories which is independent of the number of modes; * matching at least a portion of an analysis image according to the first modality with at least a portion of an analysis image according to the second modality, * from the matched together at least one analysis image according to the first modality and at least one analysis image according to the second modality, classifying the analysis image by means of an image classifier (C1), the image classifier determining to which result class of the class list the analysis image belongs, * taking into account an image classifier (CI) that has been trained by supervised learning on a learning set, the learning set comprising a plurality of records, each record consisting of a plurality of images or a plurality of image regions of the same container according to each modality, the plurality of images or a plurality of image regions being matched together and associated with a class from a list of classes, such that the trained image classifier classifies the container according to classification features according to at least two modalities; * Classify the container according to the result category.

19. The device according to claim 18, in, The inspection system (10) is configured to acquire polarization composite images, and the information processing unit is configured to calculate absorption analysis images and / or birefringence analysis images and / or refraction analysis images based on the polarization composite images.

20. The device according to claim 18, in, The inspection system (10) is configured to acquire images, and the information processing unit (13) is configured to calculate an absorption analysis image and / or a birefringence analysis image and / or a refraction analysis image from several of these images.

21. The device according to claim 18, in, The inspection system (10) is configured to acquire color composite images using a color camera, and the information processing unit (13) is configured to calculate absorption images and refraction images based on the color composite images.

22. The device according to claim 18, in, The inspection system (10) is configured to acquire an image that directly corresponds to an absorption image, a birefringence image and / or a refraction image.

Images