Method and device for inspecting glass containers according to at least two modalities in order to classify the containers according to glass defects

Abstract

Method and device for inspecting glass containers according to at least two modalities in order to classify the containers according to glass defects. 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 list of classes (D1, D2,...Dk,…Dp) including at least glass defects, - ensure a matching of at least a portion of the analysis images according to the first modality and according to the second modality, - from at least one analysis image according to the first modality and at least one analysis image according to the second modality, matched, classify the analysis images using an image classifier (Cl) that determines membership in a result class from the list of classes, the image classifier having been trained by supervised learning, - classify the container according to the result class. Figure for the abstract: Fig. 1.

Description

Title of the invention: Method and device for inspecting glass containers in at least two ways in order to classify the containers according to glass defects. Technical field

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

[0002] The object of the invention finds particularly advantageous applications for analyzing the physical characteristics of containers in order to identify non-conforming physical characteristics corresponding to defects, such as surface defects, such as folds or crevices, internal defects in the material, such as cracks, inclusions, or bubbles. Previous technique

[0003] For the manufacture of glass containers, it is known that the manufacturing process, comprising melting the glass and then conveying it to forming units, is implemented using a manufacturing installation comprising a melting furnace, a feed core for the molten glass to a forming machine, generally of the type designated as an IS machine. The containers that have just been formed by the forming machine are placed successively on an outfeed conveyor to form a line of containers. The containers are transported in a line by a conveyor to be sent successively to different processing stations. In particular, the formed containers are brought into an annealing furnace, which raises their temperature and then cools them in a controlled manner to eliminate the thermal stresses created by the forming process.Other glass container forming processes are known for tableware, insulators, syringes, and ampoules. For example, there are forming machines such as rotary and sequential presses, rather than parallel-section forming machines like IS machines. There are also machines that transform preforms into tubes, particularly borosilicate glass, to produce syringes and ampoules for pharmaceutical products.

[0004] It is known to systematically inspect all containers exiting the forming machine, moving along the outfeed conveyor, using various inspection equipment, including transmitting wall inspection systems for which a light source is arranged on one side of the conveyor and at least one A camera (typically at least two) is positioned on the other side to acquire at least one image formed by the light transmitted through the walls of the container.

[0005] It is also known to systematically inspect all vessels exiting the annealing furnace using various inspection equipment, including transmitting wall inspection systems in which a light source is positioned on one side of the conveyor and at least one camera (typically 2 to 6, 12, or 24) is positioned on the other side to acquire at least one image formed by the light transmitted through the vessel walls. Different transmitting wall inspection systems are designed to acquire at least one image formed by the light transmitted through the vertical walls, while other systems are adapted to acquire at least one image formed by the light transmitted vertically through the bottom of the vessel.

[0006] For example, patent EP2082216 describes a method for detecting, on the one hand, low-contrast defects, particularly light-refracting defects such as air bubbles, surface creases, or local variations in the thickness of transparent material, and on the other hand, high-contrast defects, particularly defects that absorb passing light, such as non-transparent inclusions or soiling. This method aims to control a light source so that it successively produces two types of illumination: the first type is homogeneous illumination, while the second type consists of alternating dark and light areas with discontinuous spatial variation. Images of the moving object are taken when it is successively illuminated by the two types of illumination.Images taken with lighting according to the first modality are analyzed to detect high-contrast defects, and images taken with lighting according to the second modality are analyzed to detect low-contrast defects.

[0007] It is also known from patent EP 1 109 008, a method for analyzing container images for cold inspection. A segmentation step detects features in the images and regions around these features. Discriminatory parameters for a region are calculated and combined with a fuzzy logic method to determine the most probable type of feature from a list of features corresponding to possible defects. The conformity of the region is then determined by applying different criteria depending on the type of feature selected. For example, a defect of the pleat type will be rejected for a certain area, whereas a defect of the inclusion type will be rejected even if it has a small area. This patent notably teaches that not all defects have the same criticality, which justifies seeking to determine their nature before deciding to reject a container.In this method, the lighting is a homogeneous extended source, which reveals. correctly absorbs defects, but in which refracting defects are visible but with low contrast and therefore with insufficient sensitivity.

[0008] However, a difficulty arises in reliably identifying certain defects, particularly "appearance defects," namely visual defects of any type: inclusions (of foreign bodies such as ceramics or metal), bubbles, folds, rivers (surface grooves), glazes (cracks), fins, trapezoids, grease spots, very thin areas, and unmelted areas. These appearance defects manifest themselves in images as local optical variations or pixels that deviate from the background. These appearance defects can be critical if they lead to a risk to the consumer, a risk of breakage, or a loss of functionality of the container. Since recognizing an appearance defect from an image can be ambiguous, safety margins are taken during detection. Consequently, some containers are considered defective even though they are acceptable or compliant.

[0009] Furthermore, it should be noted that acceptable artifacts, such as engravings or decorations, and faint mold seams, can be distinguished in the images. Therefore, it is necessary to precisely identify the nature of surface defects in order to identify critical defects and distinguish them from other defects. Preventing critical defects requires improving the reliability of surface defect classification. Besides improving defect identification, this identification allows for the determination of their potential causes so that the manufacturing installation can be controlled according to the category of detected defects. Indeed, without reliable identification of detected defects, no sound decision regarding correction of the manufacturing process can be made, whether manually or automatically.

[0010] Patent application WO2021213864 proposes a method for more reliably inspecting containers transported by a conveyor on a line, particularly a bottling line. The containers are transported to at least one first inspection unit and one second inspection unit, each comprising a transmitter and a receiver. The inspection units can inspect the containers using white light, laser light, high-frequency electromagnetic waves, gamma rays, and / or X-rays as transmitters. The containers are transported between the transmitter and a receiver, such as a camera, which allows the first inspection unit to acquire initial measurement data and the second inspection unit to acquire further measurement data. According to this method, the initial and second measurement data are combined to form common input data for an intelligence-based evaluation unit. artificial and providing as output an inspection result, such as a fill level.

[0011] As a single example of an embodiment, this document describes the detection of the fill level of containers by combining X-ray imaging and inspection with an infrared light source. This document specifies that such a method can also be used to check the side wall, the bottom, the mouth, and the contents of the container, such as for contamination by foreign bodies or product residues. The inspection result may also reveal defects, such as damage to the containers, including cracks and / or glass chips.

[0012] However, patent application WO2021213864 provides no guidance on reliably detecting transmission-visible defects, particularly surface defects. Regardless of the requirement to detect these defects, it is necessary to precisely identify the nature of surface defects in order to distinguish critical defects from other defects that may be considered non-critical. Incorrectly identifying a defect type on a container can lead to its rejection, whereas correctly identifying the defect would have allowed the container to be marked as acceptable. Therefore, to improve sorting efficiency, it appears necessary to precisely identify the nature of transmission-visible defects.

[0013] Also known from patent application EP 3,679,356 is a container inspection device for increasing detection reliability, particularly for reliably distinguishing decorative elements from contamination or soiling. The device comprises a light source emitting radiation with different wavelength ranges and intensities in spatially separated areas. This light source illuminates the container to be examined, and a color camera is configured to detect the radiation emitted by the source and having passed through the container.

[0014] The device also includes an evaluation device designed to analyze the intensity image to identify pixels or regions with a different intensity than their surroundings, thereby inferring the presence of a light-absorbing defect such as dirt. The evaluation device also analyzes color images to identify pixels or regions with a different color than their surroundings, thereby inferring the presence of light-diffusing elements such as decorative features. Thus, if a local brightness contrast is observed and there is no color contrast in that area, the evaluation unit detects contamination in that area. If a local brightness contrast coincides with a local color contrast, the evaluation unit detects the presence of a decorative feature.

[0015] The evaluation unit can also identify structures that cause local color contrast but virtually no local brightness contrast or only a weak local brightness contrast. For example, chips in glass or water droplets can cause such local color contrast, while light shining through them can radiate through these areas substantially without loss of brightness.

[0016] Such an inspection device makes it possible to distinguish decorative elements from dirt or contamination. In other words, light-refracting defects are distinguished from light-absorbing defects. However, this inspection device does not make it possible to distinguish between different types of light-refracting defects, as indicated by this application, particularly with regard to glass inclusions and water droplets.

[0017] In the prior art, a method for the optical inspection of containers is also known from document WO2020 / 244815, in which the containers are transported to an inspection unit comprising a lighting unit and a camera. The lighting unit emits light from a light-emitting surface that is locally coded based on a polarization property, an intensity property, and / or a phase property. In document WO2020 / 244815, the term "polarization property" means that the light emitted from different emission points of the light-emitting surface is emitted with different polarization directions. In document WO2020 / 244815, the term "polarization property" also means linear, elliptical, and / or circular polarization.For example, a polarization filter with a continuously changing polarization curve or several polarization filters with different orientations can be arranged in the area of ​​the light-emitting surface.

[0018] According to one embodiment, the light emitted by the light-emitting surface is locally encoded based on its polarization property, such as the polarization direction detected by the camera. It is possible to determine, independently of the radiation characteristic of the light-emitting surface for the pixels of the camera image, the radiation location from which the corresponding portion of light originates. Since the image processing unit uses the camera image, of which there is at least one, to obtain information on the location of radiation points, it is possible, for example, to distinguish a refracting defect based on a local change in the radiation location. Similarly, the intensity information can be used to detect the absorption of light by absorbing defects.However, this inspection device has the same drawbacks as other devices. known with regard to the impossibility of distinguishing one glass defect from another glass defect, both exhibiting, for example, a light-refracting characteristic. Description of the invention

[0019] The present invention aims to remedy the drawbacks of the prior art by proposing a method for quality control of glass containers, designed to achieve more efficient detection of glass defects visible in transmission by ensuring their safe and certain identification in order to optimize the sorting of containers.

[0020] Another object of the invention is to propose a method of quality control of glass containers, allowing, after the safe and certain identification of the nature of the glass defects, to give more complete information on the corrections to be made to the control parameters of a manufacturing process of glass containers of a manufacturing installation.

[0021] To achieve such objectives, the object of the invention relates to a method for inspecting glass containers in order to classify a container, the method comprising the following steps: - inspect each container using an inspection system comprising at least one light source illuminating the container and at least one camera arranged to recover the light that has passed through the container, so as to acquire images of at least a portion of the container illuminated in transmission, with a view to 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, - define a list of classes including at least glass defects, the list of classes having a number of classes independent of the number of modalities; - ensure a correspondence between at least part of an analysis image according to the first modality and at least part of an analysis image according to the second modality, -using at least one image analyzed according to the first modality and at least one image analyzed according to the second modality, and having matched them, classify the analysis images using an image classifier that determines membership in a resulting class from the list of classes, - the image classifier having been trained by supervised learning, on a training set comprising recordings composed each of the images or image regions of the same container according to each modality, matched and associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification characteristics according to at least two modalities; - classify the container according to the result class.

[0022] The object of the invention is based on a new approach for detecting transmission-visible glass defects in glass containers. The method according to the invention takes into account the various interactions between light and matter, and in particular, absorption, refraction, and birefringence, in order to determine the physical nature of the defects. Considering these interactions in combination makes it possible to reliably identify transmission-visible defects.

[0023] One objective of the invention is to allow the classification of glass defects according to their physical characteristics by evaluating the modifications of light transmitted through the wall of the containers, based on light absorption and at least light refraction and / or changes in the polarization state of the light. Each of these interactions is highlighted more specifically using different observation methods.

[0024] The invention is based on the idea that defects modify the transmitted light not according to a single type of modification, but by a variable combination of at least two types of modification.

[0025] Thus, an opaque foreign body such as a stone manifests itself through a certain absorption of light (modification of intensity) but also through a modification of the polarization state due to birefringence linked to the stresses in the glass surrounding it. Unmelted or devitrified glass will be transparent and therefore absorb little or nothing, but will produce both refraction because it has a different refractive index than normal glass, and a modification of the polarization state due to birefringence linked to the stresses it creates.

[0026] The invention improves the quality of classification of defects (or observed features) by the fact that defects can be: - similar according to a first type of modification of the light transmitted through the wall, - different if we consider two types of modifications to the transmitted light.

[0027] The invention thus improves the quality of glass defect classification also because it is insensitive to the fact that observation devices do not allow for the detection of only one type of modification of the light transmitted through the wall. Indeed, when observing a container illuminated by a large luminous surface producing uniform illumination, the image depends primarily on absorption but also, to a lesser extent, on refraction. This is due to the finite size of the luminous surface and the fact that highly refracting defects can cause light extinction by deflecting the light outside the The pupil of the objective lens makes this effect indistinguishable from absorption. Thus, in an absorption image, refracting defects appear.

[0028] Similarly, when observing a container illuminated by an extended luminous surface producing illumination that exhibits a spatial variation in a polarization property (a case of a refracted image), the image depends primarily on refraction, but if a defect is birefringent, it also modifies the polarization property of the light. It follows that this method of observation is not sufficient to determine the nature of a defect.

[0029] In other words, the object of the invention aims to increase the quality of classification of containers according to glass defects by the fact that it allows to take into account together, in combination, according to the intensity of at least two types of modifications of the transmitted light and according to morphological and photometric characteristics in different images of the defects, while freeing oneself from the fact that no mode of observation gives an independent estimate of a type of modification of the transmitted light.

[0030] According to a preferred embodiment implementing three inspection methods, the process comprises the following steps: - inspect each container using the inspection system configured to acquire images in order 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 a birefringence image while the analysis image according to the third modality corresponds to a refraction image, - ensure a matching of at least part of an analysis image according to the first modality, at least part of an analysis image according to the second modality, and at least part of an analysis image according to the third modality, - using 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, classify the analysis images using an image classifier that determines membership in a result class from the list of classes. - the image classifier having been trained by supervised learning, on a training set comprising recordings each composed of images or image regions of the same container according to each modality, matched and associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification criteria according to at least three modalities, - classify the container according to the result class.

[0031] According to a first embodiment of the inspection system, the method aims to inspect the containers using the inspection system configured to acquire images and to calculate from several of these images, analysis images corresponding to absorption images, birefringence images and / or refraction images.

[0032] According to another embodiment of the inspection system, the method aims to inspect the containers using the inspection system configured to acquire polarimetric composite images and to calculate from these polarimetric composite images, absorption images, and / or birefringence images and / or refraction images.

[0033] According to another embodiment of the inspection system, the method aims to inspect the containers using the inspection system configured to acquire, using a color camera, color composite images and to calculate from these color composite images, absorption images and refraction images.

[0034] According to another embodiment of the inspection system, the method aims to inspect the containers using the inspection system configured to acquire images corresponding directly to absorption images, birefringence images and / or refraction images.

[0035] Advantageously, to ensure the matching 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 images of the first modality and in the analysis images of the second modality and / or in the analysis images of the third modality, the method ensuring, for each container: - a matching of candidate regions in the analysis images of the first, second, or third modality, with the corresponding regions in the analysis images of at least one other modality, - or a matching of candidate regions of two different modalities.

[0036] According to another advantageous example, the method ensures, as a matching, a fusion of at least one analysis image of the first modality and an analysis image of the second modality and / or an analysis image of the third modality to obtain a fused image, the method ensuring: - an extraction of classification features from the fused image, - and a classification of the container using classification criteria applied to the classification features of the fused image.

[0037] According to another advantageous example, the method ensures, as a matching operation, a fusion of at least one analysis image of the first modality and an analysis image of the second modality and / or the third modality to obtain a fused image, the method ensuring: - segmentation of the merged images to detect merged candidate regions, - a classification of the container using classification criteria applied to the characteristics of the merged candidate regions.

[0038] According to another feature of the invention: - We extract analysis images according to the first modality, classification characteristics according to the first modality, - We extract analysis images according to the second modality and / or analysis images according to the third modality, classification features respectively according to the second modality and according to the third modality, - we classify the container using classification criteria applied to features according to the first modality and the second modality and / or the third modality.

[0039] According to an advantageous variant, classification features according to the first modality and classification features according to the second modality and / or classification features according to the third modality are chosen, and / or fused features which take into account features combining in a logical or mathematical way analysis images according to the first modality and analysis images according to the second modality and / or analysis images according to the third modality, these classification features according to the first, second and third modality being position, size, shape or value features expressing absorption and / or refraction and / or birefringence.

[0040] According to one embodiment, the container is classified by a supervised learning image classifier whose input data are: - the classification characteristics according to the first modality and the classification characteristics according to the second modality and / or the classification characteristics according to the third modality, - or the analysis images according to the first modality and the analysis images according to the second modality and / or analysis images according to the third modality, - or parts of the analysis images according to the first modality and parts of the analysis images according to the second modality and / or according to the third modality.

[0041] According to another embodiment, the container is classified 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 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 regions of 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.

[0042] According to a preferred example, each container is classified according to at least one class of images taken from a list of classes representing at least glass defects such as, in particular, trapezoid, inclusion, bubble.

[0043] According to another object of the invention, at least one sorting characteristic is compared to a rejection criterion, the sorting characteristic and the rejection criterion being dependent on the class of membership to decide whether or not the container conforms, the sorting characteristic being calculated on at least one image of the container according to a modality.

[0044] According to an advantageous example of the process, a step is implemented to take into account at least one identified glass defect in order to deduce adjustment information for at least one control parameter of a container manufacturing installation.

[0045] According to another implementation feature of the process: - the image classifier associates a confidence score with the classification of containers that are part of an inspected production; - the ranking of containers is taken into account only when the confidence score exceeds a confidence threshold for; * count the defects by defect class; * and / or decide to reject the container * and / or trigger an alarm indicating the presence of at least one critical defect in the inspected production

[0046] Another object of the invention is to provide a device for inspecting glass containers leaving a manufacturing facility in order to classify the containers according to glass defects, the device comprising: - an inspection system comprising at least one light source illuminating the container and at least one camera arranged to recover the light that has passed through the container, so as to acquire images of at least part of the container illuminated in transmission by the light source, - an information processing unit connected to the inspection system and 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, this information processing unit being configured to perform operations: * taking into account a list of classes including representations of at least the glass defects, the list of classes having a number of classes independent of the number of modalities; * matching at least part of an image analyzed according to the first modality and at least part of an image analyzed according to the second modality, * from at least one analysis image according to the first modality and at least one analysis image according to the second modality, matched, classification of the analysis images by means of an image classifier that determines membership in a result class from the list of classes, * taking into account the image classifier having been trained by supervised learning, on a training set comprising recordings composed each of the images or image regions of the same container according to each modality, matched and associated with an image class from the list of classes, so that the trained image classifier classifies the containers according to classification characteristics according to at least two modalities; * classification of the container according to the resulting class.

[0047] According to one embodiment of the invention, the inspection system is configured to acquire polarimetric composite images while the information processing unit is configured to calculate from these polarimetric composite images, absorption analysis images, and / or birefringence analysis images and / or refraction analysis images.

[0048] According to another embodiment of the invention, the inspection system is configured to acquire images while the information processing unit is configured to calculate, from several of these images, absorption analysis images, and / or birefringence analysis images and / or refraction analysis images.

[0049] According to another embodiment, the inspection system is configured to acquire color composite images using a color camera while the information processing unit is configured to calculate absorption and refraction images from these color composite images.

[0050] According to another embodiment, the inspection system is configured to acquire images corresponding directly to absorption images, birefringence images and / or refraction images.

[0051] Various other features emerge from the description given below with reference to the attached drawings which show, by way of non-limiting examples, forms of embodiment of the object of the invention. Brief description of the drawings

[0052] [Fig.1] Fig.1 is a simplified view of a device according to the invention for inspecting glass containers coming out of a manufacturing facility.

[0053] [Fig.2] Fig.2 represents an example of an embodiment of a polarimetric camera image acquisition system for containers coming out of a manufacturing facility and implemented in the inspection device according to the invention.

[0054] [Fig.3] [Fig.3] illustrates an example of a composite pixel obtained by a polarimetric camera of the inspection system illustrated in [Fig.2].

[0055] [Fig.4] Fig.4 represents another example of the implementation of an image acquisition system for containers leaving a manufacturing facility and implemented in the inspection device according to the invention.

[0056] [Fig.5] The [Fig.5] is a simplified functional block diagram of an example of a first embodiment of the inspection device according to the invention, implementing a so-called conventional processing of the information contained in the analysis images according to a first modality and in the analysis images according to a second modality.

[0057] [Fig.6] The [Fig.6] is a simplified functional block diagram of an example of a first embodiment of the inspection device according to the invention, implementing a segmentation operation on a fused image obtained by matching an analysis image according to a first modality and an analysis image according to a second modality.

[0058] [Fig.7] The [Fig.7] is a simplified functional block diagram of an example of a first embodiment of the inspection device according to the invention, implementing a convolutional neural network having as input data, candidate regions matched and resulting from segmentation operations of an analysis image according to a first modality and of an analysis image according to a second modality.

[0059] [Fig.8] The [Fig.8] is a simplified functional block diagram of an example of a first embodiment of the inspection device according to the invention, implementing three convolutional neural networks having as input data, a candidate region resulting from a segmentation operation of 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.

[0060] [Fig.9] Fig.9 is a simplified functional block diagram of an example of a second embodiment of the inspection device according to the invention, implementing a convolutional neural network having as input data, a fused image of an analysis image according to a first modality and an analysis image according to a second modality.

[0061] [Fig. 10] The [Fig. 10] is a simplified functional block diagram of an example of a second embodiment of the inspection device according to the invention, implementing convolutional neural networks 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. Description of the implementation methods

[0062] Figure 1 illustrates a device 1 according to the invention for inspecting glass containers 2 exiting a manufacturing or forming installation 3 of all types known per se. The inspection device 1 is designed to detect, for each container, whether the container has a glass defect and, for a container with a glass defect, to identify a type of defect from among a family of possible glass defects.

[0063] Upon exiting the manufacturing installation 3, the containers 2, such as glass bottles or flasks in the illustrated example, are at a high temperature typically between 300°C and 600°C. As is known, the containers 2, freshly formed by the installation 3, are taken up by an outfeed conveyor 5 to form a line of containers, in the illustrated example, placed successively on the outfeed conveyor. The containers 2 are transported in a line by the conveyor 5 in a transfer direction to be conveyed successively to different processing stations, and in particular to an annealing arch 6, upstream of which is placed a surface treatment hood (not shown), which generally constitutes the first of the processing stations after forming. Advantageously, the inspection device 1 according to the invention inspects the containers downstream of the annealing arch 6.It may be provided that the inspection device 1 according to the invention is installed upstream of the surface treatment hood or between the surface treatment hood and the annealing arch 6.

[0064] The manufacturing facility 3 is known in itself and an example will be described succinctly only to allow an understanding of the interaction between the inspection device 1 according to the invention and the manufacturing facility 3.

[0065] The manufacturing installation 3 includes a production computer 7 for monitoring the various functionalities of the manufacturing installation 3. The computer is typically a sequencer that controls pneumatic or motorized actuators as well as valves controlling the circulation of cooling air or the blowing pressure. Typically, the manufacturing installation 3 comprises several separate forming sections operating in parallel and successively delivering at least one glass container. In the example In the IS machine, the various distinct forming sections each comprise at least one roughing mold receiving a glass parison and at least one finishing mold. It is possible to identify, in a known manner, the forming section, the roughing mold, and the finishing mold from which each container 2 originates, the order in which the containers pass through being known for a given production run until the containers enter the annealing arch 6. When the inspection device 1 according to the invention is installed downstream of the annealing arch 6, it can be equipped with a device for reading information on the containers, indicating the mold or section of origin of the containers and / or a timestamp of their manufacture. Alternatively, it can be proposed that the inspection device 1 be connected and synchronized with the information reading device located near the inspection line.

[0066] The inspection device 1 according to the invention comprises an inspection system 10 including at least one light source 11 illuminating the container 2 and at least one camera 12, 12a, ... arranged to recover the light that has passed through the container, so as to acquire images of at least a portion of the container illuminated in transmission by the light source 11. The camera 12, 12a generally comprises a lens having an optical center and an optical axis, and a photoelectric, linear or matrix, generally planar, positioned in the focal plane. The inspection system 10 is connected to an electronic information processing unit 13. This electronic information processing unit 13 is a computer system of any type comprising computers, external peripherals (display unit, storage unit, keyboards, connection to various factory networks, ...programs implementing, in particular, image processing algorithms, databases, etc.

[0067] This information processing unit 13 is connected to the production computer 7 in order to receive, if necessary, time-based information from the production computer, enabling the association of the containers 2, their images, and their detected defects with the mold number or the forming cavity. Typically, the operation of the inspection system 10 is synchronized with the operation of the container forming cavities.

[0068] Furthermore, this information processing unit 13 transmits the identified glass defects and measurements taken to the production computer 7, so that the production computer can automatically deduce adjustment information for at least one control parameter of the manufacturing installation 3. Such adjustment of the control parameters is carried out 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 unit display to present to an operator the identified glass defects and images of the containers.

[0069] The inspection system 10 is configured to recover light from the light source 11 that has passed through the container in order to acquire images, in a general sense, of at least a portion of the container illuminated by transmission. According to the invention, these images are obtained using at least two different inspection methods corresponding to different interactions of light with the container wall and with the glass defects to be identified. In one embodiment, the inspection system 10 is configured to obtain images using two different inspection methods. In a preferred embodiment, the inspection system 10 is configured to obtain images using three different inspection methods.

[0070] The first inspection method is called absorption. This method primarily highlights the absorption of light passing through the wall of the container. Some defects have an absorbing character, either totally or partially. These defects thus appear opaque or dark when viewed in transmission, that is to say, when light passing through a flawless glass wall undergoes what is called normal absorption, corresponding to the color and thickness, assumed to be homogeneous, of the glass wall. However, absorbing defects exhibit a local anomaly with absorption that is sometimes lower (bubble or thin) but generally higher than normal absorption. In the following, the term "absorption" will refer only to the abnormal absorption of absorbing defects. Such defects include, in particular, inclusions in the glass, especially of ceramics or metals, and / or soiling (grease, etc.) on the glass.But such defects also include certain glazes (cracks) that would be oriented in the glass in such a way as to block the inspection light, mainly because the inspection light is then reflected in a direction that is not seen by the camera.

[0071] The second inspection method is called birefringence. This method primarily involves modifying the polarization state of light passing through the container wall by means of a so-called stress defect, which gives the glass a birefringent property. Some defects have a birefringent character. Thus, some defects result in the presence of residual mechanical stresses in the material (sometimes called internal mechanical stresses or, particularly in English, "stress"). In the wall, a birefringent or stress defect, such as an inclusion of foreign bodies (ceramic, metal, "devitrified glass"), causes a modification of the polarization state, that is to say, a polarization phase shift between two components of the electric field or a change in the direction of linearly repolarized light.

[0072] The third inspection method is called refraction. This third method primarily involves modifying the direction of light propagation through the container wall by a refracting defect, due to an angle between the dioptric surfaces traversed and also a difference in refractive index. Each surface is an air / glass or glass / air interface, therefore a diopter that refracts the light passing through it. In the absence of a defect, the wall surfaces are substantially parallel, and refraction does not cause any visible deviation of the light rays passing through the container. A refracting defect is a defect that locally causes abnormal refraction, mainly when the defect manifests as differences in slope between surfaces or diopters of the wall(s).We will therefore use the term "refractive error" only to refer to deviations of light caused by refraction at specific locations known as refracting defects. Refracting defects are those primarily detectable by the refractive anomalies they generate, particularly in through-light inspections. Typically, refracting defects include surface defects (folds, rivers) or glass distribution defects (bubbles, thins, compression rings), trapezoids, and fins. It should be noted that trapezoids and fins generally cause such strong refractions that these defects are usually clearly visible even in absorption images.

[0073] When the invention implements two modes, the first mode is absorption while the second mode is birefringence or refraction. When the invention implements three modes, the first mode is absorption, the second mode is birefringence, and the third mode is refraction.

[0074] As is known, the inspection of containers 2 according to these different inspection methods can be carried out using various configurations of the inspection system 10. It is assumed that the inspection system 10 is configured to acquire images in order to obtain so-called analysis images according to one or more of the three methods described above. Indeed, depending on the inspection system 10 used, the analysis images according to these methods can be obtained directly from the acquired images or from calculations or processing.Thus, the inspection system 10 associated with the processing unit 13 makes it possible to obtain: - an analysis image according to the first modality, called in the following description, absorption image la, composed of absorption pixels whose value mainly represents the absorption of the light passing through by the wall traversed; - an analysis image according to the second modality, called in the following description, birefringence image Ib, composed of birefringence pixels of which . the value mainly represents a change in the polarization state of the light passing through the wall of the container; - an analysis image according to the third modality, called in the rest of the description, Ir refraction image, composed of refraction pixels whose value mainly represents a modification of the direction of propagation of light passing through the wall of the container.

[0075] The following description describes, by way of non-limiting example, various methods of obtaining Ia absorption images, Ib birefringence images, and Ir refraction images. It is recalled that in all the methods that follow, a light source illuminates transparently the inspected region of the container, with observation of the light source in the background of the container.

[0076] To obtain absorption images, a first simple method is to produce uniform intensity, unpolarized illumination on the light source in an active portion. A second solution for obtaining the absorption image is to use a linearly or circularly polarized light source with uniform intensity and to acquire the image using a camera without any polarizing filter between the container and the camera. The uniformity of intensity of the light source can be perfect, that is to say, the intensity is constant over the entire active area of ​​a flat and extended light source emitting diffuse light.Uniformity can also be local, particularly when the glass of the containers is tinted. It can be expected that a region of the light source illuminating the neck of the container, where the glass wall is thicker, will emit a stronger, more uniform intensity, while a region illuminating the body of the container, where the wall is thinner, will emit a weaker, more uniform intensity. The intensity of a light source that generates, as described in document FR 2794241, a continuous spatial variation in intensity that is slower than that observed near a defect is also considered uniform or relatively uniform, such that an image analysis algorithm that compares pixels to their neighbors to detect rapid local variations in intensity as defects will not detect the slow variations in intensity emitted by the source.

[0077] Another solution for obtaining the absorption image is described in patent application EP 3 679 356, which proposes producing illumination with a source whose color varies spatially and acquiring a composite (RGB) image that is transformed in the HSV (Hue, Saturation, Value) color representation space. The absorption image is obtained from the V image or a transformation of this image, for example, by means of a gradient calculation. It should be noted that a refraction image is obtained from the H image, or a transformation of this image, for example, by means of a hue gradient calculation.

[0078] Another solution for obtaining the absorption image is to use a uniformly polarized monochrome light source with linear or circular polarization and to produce a composite polarimetric image using a polarimetric camera, and to calculate the absorption image from at least two partial polarimetric images corresponding to observations through two linear filters with analysis directions at 90° to each other.

[0079] Another solution for obtaining an absorption image is to produce the illumination using a source exhibiting variations in a polarization characteristic with the intensity remaining relatively uniform and to acquire a composite image containing at least 2 to 4 partial images using a polarimetric camera and to calculate an absorption image from 2 to 4 partial images taken through polarization analyzers at 90° to each other.

[0080] The following description describes a method for obtaining, with a single inspection system and a single raw image acquisition, both an absorption image and a refraction image. The inspection system, as illustrated in [Fig. 2], comprises an extended diffuse light source 11 composed of independently controllable elementary sources 1a, such as LEDs. The light source 11 also includes a linear polarizing filter 11b and a liquid crystal cell array 11c, allowing the polarization properties of the emitted light to be modified for each elementary source. The light source 11 is therefore composed of elementary sources whose intensity and polarization properties, such as the polarization direction, can be controlled.

[0081] By these means, the light source 11 is capable of exhibiting, over an active area illuminating a region of the container 2 to be inspected in transmitted light, a variation function of a polarization property according to any desired spatial variation function. The polarization property is, in one embodiment, the polarization direction. Preferably, the light source 11 exhibits a variation of the polarization property according to a piecewise continuous periodic law, preferably a symmetric triangular function.

[0082] A polarimetric camera 12, through its photoelectric sensor, delivers a composite digital image In, which comprises a set of composite pixels. Each composite pixel is a group of four juxtaposed partial pixels as illustrated in [Fig. 3]. In front of each of the four partial pixels of each composite pixel is placed an individual linear polarizing filter, each presenting the four linear polarization directions oriented respectively at 0°, 45°, 90°, and 135°. By grouping together the partial pixels of each composite pixel with a polarizing filter of the same orientation, four filtered partial images can be obtained. linear 0° 45° 90° 135° that is to say the images 10,145 190 1135, with pixels I0(x,y) I45(x,y) I90(x,y) and I135(x,y).

[0083] To calculate the absorption image, each pixel IAbs(x,y) is obtained by adding or averaging 2 by 2 at least two partial pixels corresponding to two directions of the 90° polarizer.

[0084] lAbs (x,y) = I0(x,x) + I90(x,y) or

[0085] lAbs (x,y) = I45(x,x) + I135(x,y) or

[0086] lAbs (x,y) = Vi (I45(x,x) + I135(x,y) + I0(x,x) + I90(x,y))

[0087] To obtain the refraction image Iref(x,y), we calculate for example the polarization direction of the light received by the camera a(x,y) using Stokes parameters.

[0088] A first formula for calculating a value of the received polarization orientation a(x,y) from the values ​​of two partial pixels can be written:

[0089] TT_ / 190 (x,y) IRef (x,y) =a(x,y) =arctand-ïoyy-

[0090] Or, alternatively, starting from the 4 pixels: r00911 and \ / x 1 / 145 (x,y)-I135 (x,y) \ L J IRef (x,y) = a(x,y)=2 atan ()

[0092] Since the variation in polarization direction is triangular periodic, it can be observed that in the absence of refraction, i.e., for a flawless wall with theoretically parallel faces, the polarization orientation a(x,y) is half the phase of the triangular periodic signal q ≥ 2a. Measuring the direction is equivalent to measuring the phase. A local phase shift is directly proportional to the refraction since the variation function is, for example, triangular symmetric and therefore linear almost everywhere.

[0093] To obtain Ib birefringence images, a simple first method is to produce uniform monochrome illumination linearly polarized along a PI direction on the light source, in an active portion. The image is acquired either by a black and white camera in front of which is placed a linear polarization analyzer whose direction is orthogonal to the PI direction, or by a polarimetric camera in which the pixels of the partial image corresponding to the analysis are taken into account through a linear filter orthogonal to the PL direction. The value of the pixels of the birefringence image is then almost zero except in the presence of a stress defect. When the light passes through a stress defect, the light intensity obtained depends on the direction and intensity of the stresses.

[0094] A second method of obtaining a birefringence image Ib is to produce on the light source a uniform monochrome illumination circularly polarized in a direction SI in an active portion. The image is acquired with a black camera and A white background is placed in front of which is a wave retarder plate followed by a linear polarization analyzer. The pixel values ​​of the birefringence image are then almost zero except in the presence of a stress defect. When light passes through a stress defect, the resulting light intensity depends on the stress intensity but not on the stress direction.

[0095] A third method for obtaining a birefringence image Ib is to produce on the light source a uniform monochrome illumination that is linearly polarized in a single direction or circularly polarized in a single direction in an active portion. The image is acquired by a polarimetric camera, in front of which a wave retarder plate is optionally placed, producing a composite image. From 2 or 4 partial images, a birefringence quantity is calculated that depends on the polarization phase shift between the Ex and Ey components of the electric field and is a measure of the stress. Those skilled in the art will be able to derive the calculation formulas from the Mallus equation and Stokes' formalism.The polarization phase shift can be calculated to obtain pixel values ​​in the birefringence image. This value depends on the intensity of the stresses, but preferably not on their direction; therefore, the detection is isotropic and proportional to the stresses. Depending on one of these variants, the polarization phase shift can be measured between 0 and 90° or even between 0 and 180°. This method also allows the calculation of an absorption image using the composite image delivered by the same polarimetric camera, by calculating each pixel IAbs(x,y) as explained previously.

[0096] To obtain infrared refraction images, a first method for obtaining a refraction image is illustrated in US patent 4606634, which describes a type of refracting defect detection that consists of modifying the "angular spectrum" of an extended light source. A light source of variable dimensions, for example, a luminous disk of variable diameter, is at the focus of a converging projection lens. In the image obtained using a camera that receives the light having passed through the container, there is enhanced contrast on the refracting objects, this contrast being able to be increased by reducing the angular spectrum, which is produced by reducing the diameter of the luminous disk.

[0097] A second set of methods for obtaining a refraction image consists of spatially varying, along the emitting surface of an extended light source, such as a light panel, a property of the light emitted by the source that the camera can distinguish. Initially, the intensity of the light was used as the spatially varying property. At least one image is acquired with an intensity-sensitive camera, which is therefore a priori monochrome, producing monochrome images. The intensity of the light emitted by each unit of the light-emitting surface varies spatially according to a one- or two-fold spatial variation law. dimensions. In other words, these inspection methods adapted for the detection of refracting defects implement illumination devices that provide light which is sometimes called "structured", that is to say having an emitting surface, generally two-dimensional, which presents variations or patterns of intensity.

[0098] These methods have a large number of variations with respect to their variation law and the possible measurement of the property using one or more cameras. The spatial variation law is unidirectional or bidirectional. In the bidirectional case, the property can be expected to be distributed according to a regular checkerboard pattern. It is more common for the spatial variation law to be unidirectional in a vertical direction, alternatively oblique or horizontal. It is also possible to acquire and analyze images with a vertical spatial variation law, and then acquire and analyze images with a horizontal spatial variation law. These images can also be combined. The spatial variation law can be uniform and continuous over the entire inspected region. By way of example, documents US4487322, EP0344617, EP1006350, EP2082216, EP2558847, EP3552001, and EP2875339 describe various variations of these methods.

[0099] A third set of methods for obtaining a refraction image consists of spatially varying, along the emitting surface, a polarization property, a color property, and / or a phase property of the light emitted by the source. By way of example, documents WO2020 / 244815 and EP3679356 illustrate variants of this third set of methods for obtaining a refraction image.

[0100] It is clear from the examples given above that the inspection process using the inspection system 10 associated with the electronic information processing unit 13 allows the obtaining of absorption images la, birefringence images Ib and refraction images Ir, according to different methods summarized below.

[0101] Thus, the method according to the invention aims to inspect the containers 2 using the inspection system 10 configured to acquire images and to calculate from several of these images and thanks to the electronic information processing unit 13, analysis images corresponding to absorption images Ia, birefringence images Ib and / or refraction images Ir.

[0102] According to another method, the process aims to inspect the containers 2 using the inspection system 10 configured to acquire polarimetric composite images and to calculate, from these polarimetric composite images, Ia absorption images, and / or Ib birefringence images, and / or Ir refraction images. The inspection system 10 can be configured to acquire images color polarimetric composites or monochrome polarimetric composite images.

[0103] According to another method, the process aims to inspect the containers 2 using the inspection system 10 configured to acquire, using a color camera, color composite images and to calculate from these color composite images, absorption images and Ir refraction images.

[0104] According to another method, the process aims to inspect the containers 2 using the inspection system 10 configured to acquire images corresponding directly to absorption images Ia, birefringence images Ib and / or refraction images Ir.

[0105] The objective of combining these various lighting and image acquisition techniques is to obtain, for each inspected region of each container, at least two and preferably at least three images, each under a different modality. For each modality, the value of each pixel depends on a different mode of interaction of the light with the wall through which it passes and with the defects to be detected. Naturally, the lighting and camera configuration depends on the inspected region of the container, which may correspond to the body, the base, the neck, the shoulder, the snout, the ring, or an area containing engravings, for example.

[0106] Figure 4 illustrates, by way of example, the positioning of cameras 12 with observation 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 containers 2. The first inspection station PI comprises a light source 11 arranged along one side of the path and having a concave emitting surface SI composed of a plurality of elementary light sources arranged to define three illumination zones, each having an identical first illumination configuration and an identical second illumination configuration. This first inspection station also comprises three cameras 12 arranged along a second side of the path opposite the first side and with optical axes D in different directions centered on the illumination zones.

[0107] The second inspection station P2 comprises a light source 11 arranged along the second side of the trajectory and having a concave emitting surface S2 composed of a plurality of elementary light sources arranged to define three lighting zones, each having an identical first lighting configuration and an identical second lighting configuration. This second inspection station comprises three cameras 12 arranged along the first side of the trajectory and with optical axes in different directions centered on the lighting zones.

[0108] The light sources and cameras are controlled so that, when a moving container 2 is successively substantially centered on an observation direction, each of the cameras 12 acquires images of the container illuminated by the associated lighting area, which is successively switched on during the acquisition, according to the first lighting configuration and the second lighting configuration. For each container, six images with six observation directions can thus be obtained according to a first modality (absorption, for example) and six images with six observation directions according to a second modality (birefringence, for example).

[0109] Regardless of the inspection system implemented, the information processing unit 13 is adapted to provide analysis images according to the first modality (absorption image) and analysis images according to the second modality (birefringence or refraction image), or even images according to the third modality according to the preferred embodiment. Furthermore, in accordance with the invention, this information processing unit 13 is configured to perform the following operations: - taking into account a list of classes including at least glass defects, the list of classes having a number of classes independent of the number of modalities; - matching at least part of an image analyzed according to the first modality with at least part of an image analyzed according to the second modality, or even according to the third modality, - from 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, matched, implementation of an image classifier to classify the analysis images in order to determine membership in a result class from the list of classes, - taking into account the image classifier having been trained by supervised learning, on a training set comprising recordings each composed of images or image regions of the same container according to each modality, matched and associated with a class from the list, so that the trained classifier classifies the containers according to classification characteristics according to at least two modalities, - classification of the container according to the result class.

[0110] The information processing unit 13 is thus adapted to implement an inspection process to detect defects on containers and to classify the containers according to predefined classes.

[0111] According to one feature of the invention, the method aims to define a list of p classes D1, D2,...Dk,...Dp including at least some glass defects, that is, defects related to the manufacturing process of the containers. The glass defects concerned are glass defects having optical properties of interaction with light passing through the container, such as at least some absorption, and / or some birefringence, and / or some refraction, such that they are detectable by means of the aforementioned devices. For a portion of this list, each class corresponds to a glass defect. For example, these classes may correspond to a bubble, an inclusion, a fold, a grain, a stone, a fin, a large bubble, a trapezoid, or a seesaw. Furthermore, several classes may correspond to the same type of glass defect, such as, for example, the trapezoidal glass defect.According to this example, one class could correspond to large trapezoids with thick glass strands and another class to small trapezoids with small, unconnected points. These classes are given for illustrative purposes only.

[0112] For another part of this list, certain classes correspond to image artifacts that are not defects. Thus, some classes may correspond to reliefs with a technical function, such as positioning notches or the striations of the mounting surface, with a decorative function, such as coats of arms, or with a function of technical or commercial indications, such as brand, capacity, or mold number. Other classes may correspond to distinguishable elements on the container, such as mold seams, which may be circular at the base or linear on the vertical wall. Recognizing a mold seam in the image, and therefore classifying an image element as a mold seam, then allows for specific analysis to determine whether the mold seam is faint, which is not a defect, or deeply marked, which requires the ejection of the container bearing such a mold seam.

[0113] Advantageously, it is possible to associate with each class in the list of classes, a criticality, that is to say a high value for a class of critical defects such as a trapezoid, lower for a class of less critical defects such as a fold, even lower for a class of non-defect objects such as a mold joint.

[0114] It should be noted that the number p of image classes is independent of the number of modalities. Generally, the class list contains more classes than the number of modalities implemented. Furthermore, the class list may include classes that do not correspond to glass defects. For example, the list may include, as classes not corresponding to glass defects, a class corresponding to a defect-free container 2, a class corresponding to a container 2 with a mold seam, and a class corresponding to a container 2 with a coat of arms. According to an advantageous embodiment, the class list contains a non-glass defect class, at least one trapezoidal class, at least one inclusion class, and at least one bubble class. These classes are recorded and accessible to the information processing unit 13.

[0115] The number of classes is therefore determined first by the needs of production and quality control, and thus by the necessity of identifying production defects in order to make the right decisions during sorting and to allow for the possible correction of the process. Conversely, the number of modalities is determined solely by the technical and economic limitations of the known means of demonstrating the absorption, refraction, and birefringence properties. According to the prior art described in patent application EP 3 679 356, reasoning based on a priori knowledge by a person skilled in the art of the optical interactions of defects according to two modalities A and B leads to predicting a number of classes determined by the number of modalities, namely, in this example, only four classes: A and B, A and not B, not A and B, and strong A and weak B.

[0116] The number of classes also depends on the quality of the sorting obtained using a supervised image classifier. Indeed, during image classifier training, it is known to verify the correct classification rate on test sets. It has been observed that the classification is better when the class list contains several classes for the same defect, such as trapezoids. In other words, the number of classes can be increased to improve the quality of the automatic classification.

[0117] As shown in Figures 5 to 10, the inspection method according to the invention consists, for each container 2, of performing at least one acquisition operation A1 to obtain at least one analysis image according to the first modality and one acquisition operation Ac2 to obtain at least one analysis image according to the second modality. In the example illustrated in [Fig. 8], the inspection method according to the invention also consists, for each container 2, of performing an acquisition operation Ac3 to obtain an analysis image according to the third modality. As explained previously, the absorption, birefringence, and refraction images can be obtained directly. Thus, the absorption, refraction, and birefringence images are identical to the acquired images.

[0118] Absorption, birefringence, and refraction images can also be obtained by computational operations performed on the acquired images. In the example illustrated in [Fig. 6], a polarimetric camera or a color camera allows the acquisition ACf of a polarimetric composite image or a color image, respectively. From this composite image, computational 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 uniformly polarized light source, it is possible to obtain an absorption image and an image of birefringence, while with a light source with a variation in polarization direction, an absorption image and a refraction image can be obtained.

[0119] The method according to the invention then consists of performing an image analysis operation comprising a matching operation or step MC of at least a portion of the analysis image according to the first modality and at least a portion of an analysis image according to the second modality, and optionally at least a portion of an analysis image according to the third modality. The method then consists of implementing a classification or ranking step Cl, using an image classifier, based on the information contained in at least the analysis image according to the first modality and the analysis image according to the second modality, and optionally the analysis image according to the third modality, matched together, in order to classify the container into a result class Dk from the list of classes.

[0120] According to a first embodiment implemented in the examples of Figures 5 to 8, the analysis operations aim to process the analysis images in such a way as to extract, in the event of the presence of a visible feature or object, a region corresponding to that object. According to a second embodiment, which will be described in detail in the examples of Figures 9 and 10, the image analysis operations do not aim to extract a candidate region but to take into account all or part of the analysis images according to the first modality and the analysis images according to the second modality and / or according to the third modality.

[0121] An object in a digital analysis image is generally a set of connected pixels possessing at least one common property not possessed by neighboring sets. An object is therefore surrounded by a closed contour and is recognized as such solely on the basis of properties of the analysis image: properties for example photometric, gray level, intensity, texture, spatial frequency, contrast, color or any measurements of an optical property detected by the camera, etc.

[0122] An object corresponds to an area or region of an analysis image that potentially presents a glass defect. It is not necessary for an object to be composed of a single image region closed by a single contour, as an object can be formed by several disjoint parts. Thus, a region can encompass the different disjoint parts of the same object if they are close together. If they are far apart, two distinct regions can be considered to form a single object. A region with an object, also called a candidate region, presents an object that can be classified as belonging to a class of objects from the list of possible classes as defined above.

[0123] The image analysis operations according to the various modalities implement one or more image processing operations using any type of digital processing known per se, such as thresholding, histogram correction, convolution, or binary or grayscale mathematical morphology operations, performed in such a way as to extract all regions containing an object. Such image processing techniques known to those skilled in the art can be applied during the calculation operations C1, C2 of the analysis images, either during the analysis before the MC matching, or, for example, in the SR segmentation step described below.

[0124] Thus, the method implements an SRI operation for segmentation and candidate region detection on the analysis images according to the first modality, an SR2 operation for segmentation and candidate region detection on the analysis images according to the second modality, and an SR3 operation for segmentation and candidate region detection on the analysis images according to the third modality (Figures 5, 7, and 8). According to the embodiment illustrated in [Fig. 6], an SR operation for segmentation and candidate region detection is performed on an IF fused analysis image of an analysis image according to the first modality and an analysis image according to the second modality, as will be explained later in the description.

[0125] A segmentation operation classically consists of dividing the image into regions or segments, that is, assigning each pixel a region membership. This segmentation operation aims to determine candidate regions in each analysis image, through filtering, thresholding, edge tracking, etc., generally but not necessarily with a view to measuring parameters that characterize these regions. This image segmentation operation is performed using a filtering method adapted to the modality implemented or to the fused image.

[0126] These SRI, SR2, SR3, SR segmentation operations allow the detection of candidate image regions defined by their contour limited to the object, these candidate image regions RC1, RC2, RC3, and RCC being respectively analysis images according to the first modality, analysis images according to the second modality, analysis images according to the third modality, or images merged according to at least two modalities. It is also possible that these SRI, SR2, SR3, SR segmentation operations allow the detection of candidate image regions defined by their rectangle enclosing the object RE1, RE2, RE3, REC, these candidate image regions being respectively analysis images according to the first modality, analysis images according to the second modality, analysis images according to the third modality, or images merged according to at least two modalities. It is also possible that these SRI, SR2, SR3, SR segmentation operations allow the detection of candidate image regions defined by their expanded rectangle framing the object RL1, RL2, RL3, RLC, so as to take into account during the classification, the context of the object in the image, these candidate image regions being respectively analysis images according to the first modality, analysis images according to the second modality, analysis images according to the third modality or images merged according to at least two modalities.

[0127] The method according to the invention implements, for each container, a MC operation of matching candidate regions in the analysis images of the first, second, or third modality with the corresponding regions in the analysis images of at least one other modality. The method can also implement, for each container, a matching of candidate regions from two different modalities.

[0128] This MC matching operation aims to ensure that regions in the analysis images of at least two different modalities are matched by comparing their respective positions on the container. In the most general case, a geometric transformation is determined from one analysis image to the other, which allows, starting from a region or a pixel of one container analysis image, the location of a region or a pixel of the other analysis image corresponding to the same region or elementary part of the container. The geometric transformation can be of any necessary type and includes, for example, a translation / rotation, an anamorphosis, a scaling change, etc.When the two different analysis images Ib, Ir or la of the same container 2 are delivered by two different cameras with two different lenses, the geometric transformation linking two by two the regions or pixels of the two analysis images Ib, Ir or la corresponding to the same region of the wall of the container can be determined, taking into account the three-dimensional geometry of the container 2, its position relative to the cameras 12, 12a at the time of image acquisition and the geometric and optical parameters of the inspection system 10, such as the direction of the optical axis, the position of the optical center and the focal length of the lenses of the cameras 12, 12a, these parameters determining the optical projection of the container by the lenses of the cameras onto the flat image sensors.

[0129] According to one embodiment, a pixel-by-pixel matching is performed between analysis images of two different modalities of a container or between analysis image regions of two different modalities of a container. To do this, the geometric transformation is determined for all pixels. It is also possible to calculate, for one of the two analysis images, or portion of an analysis image, a transformed image that is superimposable on the other image or portion of an image. Then, the transformation is applied. to all pixels in the region concerned the geometric transformation as well as interpolations, for example bilinear, of the pixel values.

[0130] In the case where the analysis images of the different modalities correspond pixel by pixel due to the inspection system 10, the matching is direct. This is the case in particular, as explained above, when acquiring a monochrome or color polarimetric composite image or a color composite image.

[0131] According to another embodiment, regions whose midpoint or center of gravity are close on the container are matched, that is, they correspond or are adjacent under the geometric transformation. Or, regions whose rectangles framing the object, or enlarged rectangles framing the object, intersect or overlap on the container within a certain proportion of a given surface area are matched.

[0132] Also, this MC matching can be carried out either from candidate regions to candidate regions, or from candidate regions to corresponding regions determined during the matching or from pixel to pixel as in the illustrated embodiment example.

[0133] According to the embodiments illustrated in Figures 6 and 9, an IF merged image is created by matching an image or portion of an image from the first analysis method with an image or portion of an image from the second analysis method and / or an image or portion of an image from the third analysis method. The method thus ensures, by way of matching, a fusion of the analysis images from the first method and from the second and / or third methods to obtain the IF merged image. The method then segments the merged images to detect merged candidate regions.

[0134] The MC matching operation can apply to all the analysis images or only parts of them. This MC matching is performed pixel by pixel as explained previously. Each pixel pc(x,y) with coordinates x, y in the composite image is assigned a value that depends on the modalities. This value is, for example, either a 16-bit scalar pc(x,y) with 8 bits for absorption and 8 bits for refraction or birefringence, or a vector vc(x,y) whose components {vt(x,y), vr(x,y)} are each an absorption scalar and a refraction or birefringence scalar. The simplest method is to directly retrieve the value of a pixel from the analysis image in absorption and a pixel from the image in refraction or birefringence, and match it with the pixel from the analysis image in absorption.However, it is possible to construct each merged pixel from a combination of several neighboring pixels in the images, or from interpolated values.

[0135] It should be noted that in the examples described, an IF merged image corresponds to a merging, with matching, of the analysis images according to the first modality and according to the second modality and / or the third modality. According to an alternative embodiment not illustrated by the drawings, the IF merged image can be directly identical to a composite image delivered by a monochrome polarimetric composite image sensor, a color polarimetric composite image sensor, or a color composite image sensor. In other words, it is possible to directly analyze a monochrome polarimetric composite image, a color polarimetric composite image, or a color composite image as an IF merged image.

[0136] According to the embodiment examples illustrated in Figures 5, 7, and 8, the matching is performed from candidate region to candidate region. It is possible to register one image to the other in order to make the candidate regions coincide in the two images of different modalities. It is also possible to directly search for candidate regions located in the same area of ​​the container.

[0137] 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 fused features that take into account features logically or mathematically combining analysis images according to the first modality and analysis images according to the second modality and / or analysis images according to the third modality. These classification features according to the first, second, and third modality are features of position, size, shape (concavity, perimeter, surface area, etc.) or values ​​expressing absorption and / or refraction and / or birefringence (photometric values ​​such as average level, contrast, variance, textures, etc.).

[0138] According to the embodiments illustrated in Figures 5 and 6, the method according to the invention implements an EC1, EC2, EC operation for extracting classification features. According to the embodiment illustrated in [Fig. 5], the method implements an EC1 operation for extracting classification features for the candidate region RC1, RE1, RL1 of the analysis images according to the first modality, enabling the definition of a vector Ct of dimension n representing n m1 features obtained from the analysis image according to the first modality for each candidate region. Similarly, the method according to the invention implements an EC2 operation for extracting classification criteria for the candidate region RC2, RE2, RL2 of the analysis images according to the second modality, enabling the definition of a vector Cr of dimension m representing m2i features obtained from the analysis image according to the second modality for each candidate region.

[0139] It should be noted that according to the embodiment example illustrated in [Fig.5], the MC matching operation of candidate regions RC1, RE1, RL1 of the analysis images according to the first modality with candidate regions RC2, RE2, RL2 of the analysis images with the second modality allows obtaining a vector Ce of dimension n+m representing n+m characteristics mli, m2i obtained for each candidate region matched between the analysis images according to the first and the second modality.

[0140] In the embodiment illustrated in [Fig. 6], the method according to the invention implements an EC operation for extracting classification features for the candidate region merged according to the first modality and according to the second modality RCC, REC or RLC, obtained after the SR segmentation operation. This extraction operation makes it possible to obtain a vector Ce of dimension n+m representing n+m mli, m2i features obtained for each candidate region matched between the analysis images according to the first modality and the analysis images according to the second modality or for the merged candidate regions or the composite candidate regions from a monochrome polarimetric sensor, a color sensor or a color polarimetric sensor.

[0141] In the embodiment examples in Figures 5 and 6, the classification features are determined by a preliminary analysis, namely domain knowledge or statistical studies. Image analysis algorithms determine the selected mli and m2i features of position, size, shape, and photometry. In the embodiment examples illustrated in Figures 7 to 10, the classification features are determined by supervised learning by being embedded in trained neural networks CNN, CNN1, and CNN2, as described below. Indeed, the role of the so-called convolutional layers of the convolutional neural networks is to determine, through learning, parameters that allow them to extract image features significant for classification, and then to automatically extract these features during classification by the trained neural network.For example, the neural network used may be based on a model of a type known by the acronym RESNET or VGG or any other known network, these networks being often available as open source.

[0142] Using predetermined classification criteria applied to the classification characteristics, the method classifies defects and, consequently, the containers bearing these defects. The classification operation determines the object class Dk of the candidate region or container from among the list of p possible classes D1, D2, ...Dp. If a candidate region is found in only one of the two images according to a first method, an analysis of the defect is performed according to the characteristics associated with the type of image used, and also takes into account Taking into account the characteristics associated with the other modality: an analysis is performed based on the fusion of the characteristics associated with the two types of images. The principle of the invention relies on considering at least two inspection modalities to provide additional and reliable information for classifying objects or containers, and therefore for identifying defects. It should be noted that, according to prior art methods, when the information in a first modality allows the detection of a candidate region, but the information in a corresponding region of an image from a second modality is very weak, then the information in the image from the second modality is ignored and thus does not contribute to the classification.

[0143] According to the invention, the classification operation is performed by an image classifier trained by supervised learning. In the embodiments illustrated in Figures 5 and 6, the image classifier Cl can be, for example, a support vector machine (SVM), a Bayesian classifier, or a neural network (NN). In the embodiments illustrated in Figures 7 and 10, convolutional neural networks (CNNs) are implemented as image classifiers Cl.

[0144] It should be understood that the method according to the invention aims to implement a learning step for the image classifier used for classifying the analysis images. This learning step is, of course, carried out prior to the implementation step of the image classifier for classifying the containers. This image classifier is trained by supervised learning, that is to say, by operations that impose the classification to be performed on it.

[0145] The image classifier Cl is trained using supervised learning methods that determine the image classifier's parameters from a set of objects or images whose class is known, called the training set. According to this supervised learning method, the system is provided with sorted, labeled data, distributed according to a predefined number of classes. Each piece of data is associated, through sorting and labeling, with one of these classes, which allows the algorithm to calculate a more general model that can subsequently associate any unknown, unlabeled data with one of the predefined classes. This supervised learning method differs from the unsupervised learning method (which has no prior knowledge of the classes). According to the latter method, data is provided to the system in bulk, without any sorting or labeling.The system itself determines the most relevant number of classes and associates each piece of data with one of these classes (example: X-Means clustering algorithm). Supervised learning also differs from unsupervised learning (which has a priori assumption about the number of classes): on. We provide the system with raw data, without any sorting or labeling. Instead, we tell the system the expected number N of classes. The system then automatically associates each piece of data with one of the N expected classes (example: K-Means clustering algorithm).

[0146] The image classifier implemented within the framework of the invention for inspecting containers was trained on a training set comprising recordings, each composed of images or image regions of the same container according to each modality associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification characteristics in at least two modalities. Of course, as explained above, the list of classes considered includes representations of at least the glass defects.

[0147] Each record in this training set includes, for an example or reference container: - 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 and at least one label assigning to the example container, at least one class of objects from the list of possible classes or, - at least one region of analysis image according to the first modality of the example container, at least one region of analysis image of the second modality and / or of the third modality of the example container matched and at least one label assigning to the corresponding region of the example container at least the class of objects from a list of possible classes.

[0148] The training set thus comprises pairs or triplets of corresponding regions, preferably pairs or triplets of characteristic vectors associated with a type of glass defect.

[0149] This image classifier, which has been trained during a learning phase, is used during container inspection to classify the containers according to the input data applied to the image classifier and representative of the inspected containers. The information processing unit 13 is thus configured to execute or implement the image classifier, which has been previously trained by supervised learning, so that this trained image classifier classifies the containers.

[0150] According to the examples illustrated in Figures 5 and 6, the input data of the image classifier are the information in the n+m dimension vector Ce representing the m1, m2i features obtained for each candidate region matched between the analysis images according to the first and second modality. In general, the classification features according to the first The modality and classification characteristics according to the second modality and / or classification characteristics according to the third modality, are the input data of the image classifier.

[0151] According to the embodiment illustrated in [Fig.7], the image classifier, which is a convolutional neural network CNN, has as input data a pair of candidate regions (RL1, RL2), obtained after the MC matching operation of these candidate regions.

[0152] According to the embodiment illustrated in [Fig. 8], the image classifier comprises a first convolutional neural network CNN 1 (Convolutional Neural Network) whose input data is a candidate region according to the first modality (RC1, RE1, RL1) obtained after the SRI operation of segmentation and candidate region detection on the images analyzed according to the first modality. The image classifier also comprises a second convolutional neural network CNN2 (Convolutional Neural Network) whose input data is a candidate region according to the second modality (RC2, RE2, RL2) obtained after the SR2 operation of segmentation and candidate region detection on the images analyzed according to the second modality.The image classifier also includes a third convolutional neural network CNN3 (Convolutional Neural Network) with as input data a candidate region according to the third modality (RC3, RE3, RL3) obtained after the SR3 operation of segmentation and detection of candidate regions on the analysis images according to the third modality.

[0153] The first convolutional neural network CNN1, the second convolutional neural network CNN2 and the third convolutional neural network CNN3 each work in parallel respectively on a candidate region in absorption, on a candidate region in birefringence and on a candidate region in refraction, these three regions according to the three modalities, being associated by the MC matching operation according to the techniques explained above.

[0154] The outputs of the first convolutional neural network CNN1, the second convolutional neural network CNN2, and the third convolutional neural network CNN3 are the input data for an image classifier, for example, of the SVM, Random Forest, Bayesian, and preferably NN type, allowing classification according to the 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, class membership hypotheses, but they can be more complex data with vectors of dimensions greater than the number p of classes. It is recalled that the training set of the neural networks contains triplets of candidate regions according to the three modalities inspection, with a label indicating an object class from the list of possible classes.

[0155] According to a second embodiment implemented in the examples of Figures 9 and 10, the image analysis operations do not aim to extract a candidate region but rather to consider all or part of the images analyzed according to the first modality and the images analyzed according to the second modality and / or according to the third modality, without implementing a prior segmentation operation. If only parts of images are analyzed, these parts preferably correspond to one or more regions of interest of the container, such as the ring, the neck, the shoulder, the body, the mouth, or a right or left half-side, or an area containing engravings. According to these two embodiments illustrated in Figures 9 and 10, the analysis operations rely on the implementation of neural networks as image classifiers trained by supervised learning.

[0156] In the embodiment illustrated in [Fig. 9], a MC operation is performed to match an analysis image according to the first modality with an analysis image according to the second modality in order to obtain a fused IF image. The fused IF image is obtained by merging at least one analysis image according to the first modality with at least one analysis image according to the second modality of a container or by merging regions of at least one image according to the first modality with corresponding regions of at least one image according to the second modality.Generally, as already explained, the IC fused image is obtained by merging at least one analysis image 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 of a container, or by merging regions of at least one analysis image according to the first modality with corresponding regions of at least one analysis image according to the second modality and / or one analysis image according to the third modality. Similarly, the IF fused image can be directly identical to a composite image delivered by an image sensor.

[0157] The analysis images according to the various modalities are taken during the Acl, Ac2 acquisition operations performed by the inspection system 10, as explained in the description above. This pixel-by-pixel MC matching of the images is performed as explained in the embodiment example in [Fig. 6]. This fused image is used as input data for a convolutional neural network (CNN) which, through supervised learning, is capable of taking into account the position, size, shape, and photometry features that are significant for the intended classification. It should be noted that the training set contains IF fused images or regions of fused images labeled with a class of objects. from the list of possible classes. The classification characteristics according to the various modalities are taken into account in the weights resulting from the training and defining the convolutional neural network (CNN). It should be noted that, unlike the examples in Figures 6 and 7, the segmentation operation is not necessary in this variant because the stages of the convolutional neural network are capable, through training, of classifying images according to their content without prior segmentation, and of locally determining the position, size, shape, and photometry features that are significant for classification. However, a segmentation operation is possible, for example, by replacing the feature extraction (EC) and image classifier (CL) in [Fig. 6] with an image classifier based on a convolutional neural network (CNN).

[0158] In the embodiment illustrated in [Fig. 10], the image analyzed according to the first modality (in part or in whole) is used as input data for a first convolutional neural network CNN1, while the image analyzed according to the second modality (in part or in whole) is used as input data for a second convolutional neural network CNN2. As explained above, these convolutional neural networks have been previously trained by supervised learning in order to determine the morphological features such as position, size, shape, and photometry that are significant for the intended classification.

[0159] The first convolutional neural network CNN1 and the second convolutional neural network CNN2 each work in parallel on two candidate images in each modality. The outputs of the first convolutional neural network CNN1 and the second convolutional neural network CNN2 are the input data for a classifier, for example, of the SVM, Random Forest, Bayesian, and preferably an NN neural network type, allowing the containers 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.

[0160] The outputs of the first convolutional neural network CNN1 and the second convolutional neural network CNN2 are, for example, class membership hypotheses, but they can be more complex data with vectors of dimensions greater than the number p of classes. It should be noted that, unlike the example in [Fig. 8], the segmentation operation is not necessary in this variant because the stages of the convolutional neural network are capable, through learning, of classifying images according to their content without prior segmentation, and of locally determining the morphological characteristics, for example, position, size or shape, and photometric characteristics that are significant for the classification.

[0161] It is clear from the preceding description that the method according to the invention not only allows for the identification of glass defects in containers during transmission but also for the classification of these defects, thus enabling a transition from inspection to optimization of the manufacturing process. One of the features of the invention is the definition of a class list comprising classes of defects, which makes it possible to link glass defects with characteristics of the manufacturing process to be regulated. Improving the classification of glass defects allows for a better understanding of their causes.

[0162] The object of the invention is advantageously exploited in manufacturing facilities to allow for better detection and categorization of defects present within containers. Certain defects can be seen, detected, and categorized more easily through the combination of two or three methods.

[0163] Preferably, the inspection method according to the invention is designed such that the image classifier associates a confidence score with the classification of containers from a production run. The confidence score is typically the probability of the container belonging to the assigned class. The score can be expressed as a percentage or a value between 0 and 1.

[0164] It should be noted that a container may have several defects. There are several ways to classify such containers. In variants of the process that include a segmentation step, analysis images of the same container can be extracted. Several image regions, and for example, several SR segments, are recognized as belonging to defect classes. In this case, according to a first variant, the container's class will be that of the segment classified in the defect class with the highest criticality. It is also possible to take into account the confidence score of the classification; that is, the class assigned to the container will be that of the segment classified with a confidence score higher than the confidence threshold. According to a second variant, all the defects present in a single container can be counted, particularly when a container has several critical defects.Thus, the statistical analysis of production, which will be presented later, can account for the distribution of defects independently of the number of rejected containers.

[0165] The object of the invention is used for sorting container production in the following manner. After a container is classified, at least one container sorting characteristic is compared to a rejection criterion, and when the sorting characteristic exceeds the rejection criterion for a container, the container is considered non-compliant and rejected. The installation includes an ejector to remove defective containers from production. The sorting characteristic and the rejection criterion depend on the class to which the container belongs to determine whether or not it conforms. The sorting characteristic is calculated on at least one image of the container according to one of two or three modalities. The sorting characteristic and the rejection criterion are, for example, a defect dimension such as its area or length, measured in at least one analysis image. The confidence score can optionally be taken into account for the sorting, rejecting containers belonging to a less critical defect class only if the confidence score is high, and conversely, rejecting containers belonging to a critical defect class even if the confidence score is low.

[0166] The object of the invention is exploited to carry out a statistical analysis of a container production, that is to say an analysis of the frequency or distribution of the different types of defects and their criticality, the types of defects included in the list of classes and their criticality being determined in advance for the purpose of process control.

[0167] Certain defects are caused during the container forming stages in the molds, and are therefore related to forming parameters that differ from one section to another or from one cavity to another. It is therefore preferable to classify the defect classes according to the original section or cavity of the containers. When the installation is located at the exit of the forming machines, i.e., upstream of the annealing arch 6, the inspection is immediate after manufacturing, so the timestamp of the container's manufacture is known, and the original cavities or sections of the containers are also known through synchronization, since the order in which the containers exit the forming machine is known.When the inspection device 1 according to the invention is installed downstream of the annealing arch 6, it is preferably equipped with, or connected to, a device for reading information on the containers indicating the original mold or section of the containers and / or a timestamp of their manufacture. It is therefore possible and preferable to perform statistical analysis of production, based on the classification of the containers by the inspection process, according to the distribution of defects directly related to the production parameters at the time of manufacture of each container and / or according to the different cavities and sections of the manufacturing machine.

[0168] Statistical analysis of container production thus makes it possible to link defects with their causes, to obtain two results: - on the one hand, correlations can be determined between manufacturing parameters and the resulting defects, thus allowing for the definition of more efficient process regulation methods; - on the other hand, knowing the cause-and-effect relationships, deliver in real time to the production computer 7, the possibility of creating a feedback loop for regulate the process by correcting defects, therefore discrepancies between the desired quality and the estimated quality of the containers.

[0169] To summarize the above, the object of the invention is advantageously exploited for various production operations described below: - sorting of production; - statistical analysis of production; - determination of cause-and-effect relationships between production parameters and defects; - control of the process by reducing the defects observed; - warning operators of the occurrence of critical faults by means of an alarm.

[0170] According to a preferred embodiment of the invention, the classification of containers is only taken into account when the confidence score of the assigned class exceeds a confidence threshold for: - to count the defects by defect class in a defect frequency statistic, - and / or decide to reject the container - and / or trigger an alarm for the presence of one or more critical defects in the inspected production.

[0171] It should be noted that the confidence threshold corresponds to a predetermined or adjustable minimum value of the confidence score.

[0172] Taking these various modalities into account by a supervised learning classification algorithm makes it possible to automatically manage both the different aspects of light / matter interactions (absorption, birefringence, refraction) and the coupling of these interactions resulting from the methods used to obtain the analysis images. Through supervised learning, the image classifier considers photometric and morphological characteristics in both or even all three modalities. The image classifier is thus able, through learning, to determine the geometric and photometric characteristics significant for the intended classification. Furthermore, such an image classifier is able, during inspection, to associate unknown and unlabeled image data with a predefined class.

[0173] The method according to the invention thus differs from prior art classification methods, whose classification process relies solely on physical or logical considerations, applying predefined business rules. Conversely, the image classifier implemented according to the invention allows it to construct, on its own, the association model between the input data and the class. Given the supervised learning, it is the image classifier that directly associates the input data with a class selected from the list of classes.

Claims

1.

2. Demands Method for inspecting glass containers (2) for the purpose of classifying a container, the method comprising the following steps: - inspect 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 recover the light that has passed through the container, so as to acquire images of at least a part of the container illuminated in transmission, with a view to obtaining at least one analysis image according to a first modality corresponding to an absorption image (la) and at least one analysis image according to a second modality corresponding to a refraction image (Ir), - define a list of classes (Dl, D2,..Dk,...Dp) including at least glass defects, the list of classes having a number of classes independent of the number of modalities; - ensure a correspondence between at least part of an analysis image according to the first modality and at least part of an analysis image according to the second modality, -using at least one image analyzed according to the first modality and at least one image analyzed according to the second modality, and having matched them, classify the analysis images using an image classifier (Cl) that determines membership in a result class from the list of classes, - the image classifier (Cl) having been trained by supervised learning, on a training set comprising recordings composed each of the images or image regions of the same container according to each modality, matched and associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification characteristics according to at least two modalities, ; - classify the container according to the result class. A method according to claim 1, wherein the method comprises the following steps: - inspect each container using the inspection system (10) configured to acquire images in order 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 image of analysis according to a third modality, the image analyzed according to the second modality corresponding to a refraction image (Ir) while the image analyzed according to the third modality corresponds to a birefringence image (Ib), - ensure a matching of at least a part of an image analyzed according to the first modality, at least a part of an image analyzed according to the second modality and at least a part of an image analyzed according to the third modality, - from at least one image analyzed according to the first modality, at least one image analyzed according to the second modality and at least one image analyzed according to the third modality, classify the analysis images by means of an image classifier (Cl) which determines membership in a result class from the list of classes, - the image classifier having been trained by supervised learning,on a training set comprising recordings, each composed of images or image regions of the same container according to each modality, matched and associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification criteria according to at least three modalities; - classify the container according to the resulting class.

3. A method according to any one of the preceding claims, wherein the method aims to inspect the containers using the inspection system (10) configured to acquire images and to calculate from several of these images, analysis images corresponding to absorption images (la) and refraction images (Ir), and where appropriate to birefringence images (Ib).

4. A method according to any one of claims 1 or 2 wherein the method aims to inspect the containers using the inspection system (10) configured to acquire polarimetric composite images and to calculate from these polarimetric composite images, absorption images, and / or birefringence images and / or refraction images.

5. A method according to claim 1 or 2, wherein the method is for inspecting containers using the inspection system (10) configured to acquire images using a color camera color composites and to calculate from these color composite images, absorption images and refraction images.

6. A method according to any one of claims 1 or 2 wherein the method aims to inspect containers using the inspection system (10) configured to acquire images corresponding directly to absorption images, birefringence images and / or refraction images.

7. A method according to any one of the preceding claims, in order to ensure the matching of at least a part of an analysis image according to the first modality and at least a part of an analysis image according to the second modality, and where applicable at least a part of an analysis image according to the third modality, the method detects candidate regions (RC1, RC2, RC3, RCC) in the analysis images of the first modality and in the analysis images of the second modality, and where applicable in the analysis images of the third modality, the method ensuring, for each container: - a matching of the candidate regions in the analysis images of the first modality, of the second modality, or where applicable of the third modality, with the corresponding regions of the analysis images of at least one other modality, - or a matching of candidate regions of two different modalities.

8. A method according to any one of the preceding claims, wherein the method ensures, as a matching step, a fusion of at least one analysis image of the first modality and an analysis image of the second modality and optionally an analysis image of the third modality, to obtain a fused image (FI), the method ensuring: - an extraction of classification features (mli, m2i) from the fused image, - and a classification of the container using classification criteria applied to the classification features of the fused image.

9. A method according to any one of the preceding claims, wherein the method ensures, as a matching process, a fusion of at least one analysis image of the first modality and an image analysis of the second modality, and where applicable the third modality, to obtain a fused image, the process ensuring: - a segmentation of the fused images to detect fused candidate regions, - a classification of the container using classification criteria applied to the characteristics of the fused candidate regions.

10. A method according to any one of the preceding claims wherein: - analysis images according to the first modality are extracted, classification features according to the first modality are extracted, - analysis images according to the second modality and where applicable analysis images according to the third modality are extracted, respectively according to the second modality and where applicable according to the third modality, - the container is classified using classification criteria applied to the features according to the first modality and the second modality and where applicable the third modality.

11. A method according to the preceding claim wherein classification features according to the first modality and classification features according to the second modality and, where applicable, classification features according to the third modality are selected, and / or fused features which take into account features combining in a logical or mathematical manner analysis images according to the first modality and analysis images according to the second modality and, where applicable, analysis images according to the third modality, these classification features according to the first, second and, where applicable, third modality being position, size, shape or value features expressing absorption and / or refraction and / or, where applicable, birefringence.

12. A method according to any one of the preceding claims, wherein the container is classified by a supervised learning image classifier whose input data are: - the classification features according to the first modality and the classification features according to the second modality and, where applicable, the classification features according to the third modality, - or the analysis images according to the first modality and the analysis images according to the second modality and where applicable analysis images according to the third modality, - or parts of the analysis images according to the first modality and parts of the analysis images according to the second modality and where applicable according to the third modality.

13. A method according to any one of claims 1 to 10 wherein the container is classified by a supervised learning image classifier whose input data is at least one fused image (FI) obtained by merging at least one analysis image according to the first modality and one analysis image according to the second modality and optionally one analysis image according to the third modality or by merging regions of at least one analysis image according to the first modality and one analysis image according to the second modality and optionally one analysis image according to the third modality.

14. A method according to any one of the preceding claims in which each container (2) is classified according to at least one class taken from a list of classes representing at least glass defects such as, in particular, trapezoid, inclusion, bubble.

15. A method according to any one of the preceding claims in which at least one sorting characteristic is compared to a rejection criterion, the sorting characteristic and the rejection criterion being dependent on the class of membership to decide whether or not the container conforms, the sorting characteristic being calculated on at least one image of the container according to one modality.

16. An inspection method according to any one of the preceding claims, wherein a step is implemented for taking into account at least one identified glass defect in order to deduce adjustment information for at least one control parameter of a manufacturing installation (3) for containers.

17. An inspection method according to any one of the preceding claims, wherein: - the image classifier associates a confidence score with the classification of containers that are part of an inspected production; - the classification of the containers is taken into account only when the confidence score exceeds a confidence threshold for: - counting defects by defect class; - and / or deciding to reject the container;

18. -and / or trigger an alarm for the presence of at least one critical defect in the inspected production. A device for inspecting glass containers (2) leaving a manufacturing facility (3) for the purpose of classifying the containers in relation to glass defects, the device comprising: - an inspection system (10) comprising at least one light source (11) illuminating the container and at least one camera (12, 12a) arranged to recover the light that has passed through the container, so as to acquire images of at least part of the container illuminated in transmission by the light source, - 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 (la) and at least one analysis image according to a second modality corresponding to a refraction image (Ir), this information processing unit (13) being configured to perform operations: * taking into account a list of classes (D1, D2,... .Dp) including representations of at least the glass defects, the list of classes having a number of classes independent of the number of modalities; * matching at least part of an image analyzed according to the first modality and at least part of an image analyzed according to the second modality, * from at least one analysis image according to the first modality and at least one analysis image according to the second modality, matched, classification of the analysis images by means of an image classifier (Cl) which determines membership in a result class from the list of classes, * taking into account the image classifier (Cl) having been trained by supervised learning, on a training set comprising recordings composed each of the images or image regions of the same container according to each modality, matched and associated with a class from the list of classes, so that the trained image classifier classifies the containers according to classification characteristics according to at least two modalities; * classification of the container according to the resulting class.

19. Device according to claim 18 wherein the inspection system (10) is configured to acquire polarimetric composite images while the information processing unit is configured to calculate from these polarimetric composite images, absorption analysis images, and / or birefringence analysis images and / or refraction analysis images.

20. Device according to claim 18 wherein the inspection system (10) is configured to acquire images while the information processing unit (13) is configured to calculate from several of these images, absorption analysis images, and / or refraction images, and where appropriate birefringence images.

21. Device according to claim 18 wherein the inspection system (10) is configured to acquire color composite images using a color camera while the information processing unit (13) is configured to calculate absorption and refraction images from these color composite images.

22. Device according to claim 18 wherein the inspection system (10) is configured to acquire images corresponding directly to absorption images, birefringence images and / or refraction images.

Images