Method for feeding an orienting component

Abstract

A method of feeding an object, such as a pipe top or a cap, includes at least one orientation and quality inspection step integrated into a feeding method performed continuously during production, the orientation and quality inspection including a learning phase and a production phase.

Description

[0001] Corresponding application

[0002] This application claims priority to an earlier European application No. EP22191165.4 filed on August 19, 2022 in the name of AISAPACK Holding AG, the contents of which are incorporated herein by reference in their entirety. Technical Field

[0003] The present invention is in the field of mass production of objects or dispensing systems requiring high throughput feeding, such as vibratory or centrifugal bowls. The present invention more particularly relates to a feeding method and apparatus for delivering oriented objects at high production throughput using visual inspection and artificial intelligence algorithms. Background Art

[0004] High throughput feeding systems for orienting objects are known in the prior art. Examples are given in the following publications: US5311977, US4608646, DE3312983, US4692881 and US5853078.

[0005] Publication US5311977 describes a system for feeding objects capable of determining the orientation of the objects by geometrical checking and reorienting or rejecting the objects with the aid of output signals of a microprocessor. In the publication, the geometrical checking is performed by an object detector having at least 1000 pixels arranged in a linear manner, which are oriented in such a way that they lighten or darken according to the geometry of the object. The system described in the publication comprises means capable of detecting points on the contour of the object located in the scanning section and comparing the positions of the contour points in real time with the stored contour. The system is capable of causing the objects to be located or rejected in response to the output signals of the microprocessor, based on the contour point signals from a plurality of scanning sections.

[0006] Publication US4608646 describes a microcontroller-based system for identifying and labeling identical or different objects conveyed along a track of an object feeder (e.g., a bowl feeder) to verify the orientation of the objects and to sort the oriented objects in a predetermined repetitive sequence. The identification and sorting of the objects can be programmed according to user requirements. The identification of the objects requires a device for identifying the silhouette of the object, which includes a set of light sensors connected to a perforated grid located in the feed track. The silhouette image of each object to be sorted is first digitized and stored in a location associated with the identification number of the object in the memory of the microcontroller. Similarly, the sequence of different objects is stored in the memory of the microcontroller. Thereafter, when the objects are fed onto the grid, each object is compared with the corresponding stored image of the sequence in the correct position, and the wrong or misoriented objects are rejected by the jet stream directed to the feed track, while the correct identification of the object causes the jet stream to stop, allowing the object to pass and be conveyed to the feed output station.

[0007] Publication DE 3312983 describes a vibrating bowl for sorting mechanical components using the position and contour of the component as decision criteria. The device comprises a conveying device for conveying the component substantially perpendicularly to an electronic sensor line, by which the component contour can be explored line by line, and an electronic comparator, to which an output signal from the sensor line can be transmitted and by which the output signal can be compared with a previously stored set point value.

[0008] Publication US4692881 describes a device for feeding objects in a predetermined orientation. The device includes a detector consisting of a plurality of light receiving elements and at least one light emitting element spaced apart from and facing the light receiving elements, the plurality of light receiving elements being arranged in a line or in a plurality of lines, the extension direction of the lines being perpendicular to the object feeding direction. The device also includes a random access memory (RAM) for storing a reference signal model obtained by continuously detecting the shape of the object when the object passes in front of the detector at a preselected desired position. The device also includes a central processing unit (CPU) for comparing the reference signal model with a signal data model obtained when the object to be distinguished passes continuously in front of the detector at any position. In response to each unfavorable comparison, the mispositioned object is rejected from entering the bowl.

[0009] Publication US5853078 describes a device for orienting and feeding objects, which is particularly suitable for use in automated assembly systems. The device includes a feed bowl including a spiral internal track that terminates at the level of the upper edge of the bowl, adjacent to an annular feed ring, which is mounted for selective rotational movement around the feed bowl. A control circuit including a fixed camera located above the annular feed ring operates by controlling the rotational movement of the annular feed ring through an operatively connected motor to bring a continuous portion of the annular feed ring into a predetermined field of view of the camera so as to distinguish between correctly oriented objects and misoriented objects, and then provide a signal to a pick and place robot to remove the correctly oriented objects. A cleaning bar is located in a selected position for pushing the misoriented objects out of the annular feed ring and returning them to the feed bowl for recovery. Another embodiment of a vibrating feed tank is also provided, which utilizes a second selective rotating disk in a concentric and spaced relationship with the feed ring to receive the recovered objects removed from the annular feed ring in a container provided on the ring. Summary of the invention

[0010] Goals, limitations, and problems to be solved

[0011] The object of the present invention is to reduce the settling time of high throughput systems for feeding oriented objects, such as vibratory bowls or centrifugal bowls. Although improvements have been proposed in the prior art and described in particular in publications US5311977, US4608646, DE3312983, US4692881 and US5853078, these systems do not allow for fast object changes, which results in a lot of time being wasted on settling each time an object is changed. To overcome this difficulty, vibratory or gravity bowls are usually dedicated to unique object geometries, since the time to change a vibratory bowl on an assembly machine is shorter than the time to adjust the bowl to dispense a new object at the desired throughput. The disadvantage of this case is the need to invest in and store a large number of bowls for dispensing the objects, which bowls are individually suitable for a single object or a limited number of objects.

[0012] Another disadvantage of the devices described in the prior art is related to undetected defective objects. Defective objects are, for example, deformed or out-of-tolerance objects or products that are aesthetically unsuitable (e.g., in their appearance). These objects cause the assembly machine to stop untimely or cause the assembled product to be defective.

[0013] The invention makes it possible to remedy the above-mentioned disadvantages by providing a bowl provided with a vision system associated with an artificial intelligence algorithm and possibly with an orientation device. The invention also makes it possible to define criteria for rejecting so-called defective parts. The rejection criteria can be related to the dimensions of the object (e.g. deformed or out-of-tolerance objects) or to aesthetic appearance defects (e.g. scratches, stains, foreign bodies, unsuitable colors, etc.).

[0014] According to the invention, an artificial intelligence algorithm associated with a vision system is able to quickly change objects in the feed bowl with a high throughput of oriented objects. The invention is also able to reject defective objects, thereby avoiding stopping the assembly machine when the object is out of tolerance or deformed, and also avoiding the use of objects that are not aesthetically or visually appropriate.

[0015] According to the invention, the learning phase enables the definition of an acceptable "norm" for the provided objects. This "norm" defines the range of orientations of the objects and, where applicable, the acceptable range of sizes and aesthetics. In the context of the invention, the concept of "acceptable or unacceptable defects", that is to say the concept of an object considered "good" or "defective", is defined relative to a certain level of deviation from the predefined norm established by learning.

[0016] The invention ensures that the orientation and quality level of an object remain constant over time. In addition, templates, that is to say previously established specifications, can be reused to produce identical objects later.

[0017] The orientation and quality level of the objects can be adjusted through iterative learning over time, based on the observed differences: during the production process, the specification defined by the initial learning is refined through a "supplementary" learning that takes into account objects that were provided during the normal production phase but had orientations or defects that were considered acceptable. It is therefore necessary to adjust the specification so that it incorporates this information and the process does not reject these objects.

[0018] The present invention is able to distribute objects in a very short period of time, and achieving this performance relies on a compression-decompression model of the object image as described in detail in this application.

[0019] In the context of the present invention, the constraints that arise and the problems to be solved are specifically as follows:

[0020] - Visual inspection is performed during the movement of the objects in the bowl, thus reducing inspection time, since there is no need to slow down the production throughput, or at most the inspection has a minimal impact on the latter.

[0021] - The acceptance range of oriented objects must be adjustable.

[0022] - Size and aesthetic defects are unknown (no defect library).

[0023] -Aesthetic defects vary with decoration.

[0024] -The level of acceptance of defects must be adjustable.

[0025] - An object considered to be misoriented is an object whose orientation is outside of a range of orientation tolerances that are considered acceptable.

[0026] - An object considered defective is one whose defects are outside the tolerance range considered acceptable.

[0027] The method proposed by the present invention described below is able to alleviate the above mentioned disadvantages and overcome the identified problems.

[0028] definition

[0029] - Object: The object that is fed (or dispensed) into the bowl, such as a lid or tube top

[0030] - N: the number of objects that form a batch during the learning phase; N also corresponds to the number of secondary images that form a batch

[0031] - Original image: A captured image of an object or a part of an object

[0032] -K: The number of original images for each object

[0033] -A k : The original image with index k, where 1≤k≤K

[0034] - Secondary image: a portion of the original image

[0035] -P k : Number of secondary images per original image

[0036] -A k S k,p : Compared with the original image A k The secondary image with index k associated with index p, where 1≤p≤P k

[0037] -F k,p Model: With secondary image S k,p Associative compression-decompression model

[0038] -Compression factor Q k,p : Model F k,p Compression factor

[0039] -Reconstructed secondary image R k,p :Use the association model F k,p From the secondary image S k,p Reconstructed secondary image

[0040] General Description of the Invention

[0041] The present invention relates to a method for feeding oriented objects (e.g. packaging components, such as tube tops or caps), comprising a visual inspection integrated into one or more steps of the method for dispensing said objects. The feeding method according to the invention comprises at least two stages for performing the visual inspection:

[0042] - A learning phase, during which a batch of objects considered to be "correctly oriented" and "good quality" is fed, and then a learning phase criterion is defined based on the images of said objects.

[0043] - A production phase, during which the images of the produced objects and the criteria defined during the learning phase are used to quantify in real time the orientation and quality of the fed objects and to control the feeding process.

[0044] During the learning phase, the machine is fed with N objects that are considered to have acceptable quality and orientation. During the process of feeding the objects, one image (K=1) or a plurality of different images (K>1) called raw images of each object are collected. The collected KxN raw images are digitally processed, which is described in more detail below and comprises at least the following steps:

[0045] - Reposition each original image A k

[0046] - For each original image A k Divide into P k secondary images, denoted as S k,p , where 1≤k≤K and 1≤p≤P k

[0047] - Group secondary images into batches of N similar images

[0048] -For each batch of images S k,p :

[0049] ○ Find the compression factor Q k,p The compressed representation F k,p

[0050] ○ The compression factor Q is derived from each batch of images k,p The compression-decompression model F k,p A specific embodiment of the present invention includes for all models F k,p have the same compression factor. Adjust each model F k,p The compression factor Q k,p It is possible to adjust the defect detection level and optimize the calculation time depending on the observed area of ​​the object.

[0051] Therefore, at the end of the learning phase, for each observed region of the object, a model F is obtained k,p and the compression factor Q k,p , each region consists of a secondary image S k,p definition.

[0052] As explained in more detail below, each secondary image of the object has its own size. A special case of the present invention is that all secondary images are of the same size. In some cases, it is advantageous to be able to locally reduce the size of the secondary images to detect smaller defects. By collectively adjusting each secondary image S k,p The size and compression factor Q k,p The invention enables optimization of computation time while maintaining a high level of detection performance adjusted to the level of requirements associated with the manufactured product. The invention enables local adjustment of the detection level to the level of criticality of the observed area.

[0053] During the production phase, K so-called “raw” images of each object are used to monitor the orientation and quality of the produced objects in real time, which enables:

[0054] - recovering the misoriented object into the feeding system or correcting the orientation of said misoriented object,

[0055] - Remove defective objects from production as quickly as possible.

[0056] In order to carry out real-time monitoring of the objects produced, K original images of said objects are evaluated relative to a set of original images acquired during a learning phase, from which a compression-decompression function and a compression factor applied to the images of said produced objects are extracted, by the method described in the present application. This comparison between the images acquired during the production phase and the images acquired during the learning phase leads to the determination of one or more scores for each object, the value of which enables the objects to be classified relative to thresholds corresponding to the orientation level and the visual quality level. Due to the value of the score and the predefined thresholds, the misoriented objects are recycled into the bowl or are redirected and the defective objects can be discarded from the production process. Other thresholds can be used to detect batches of defective objects (rejection rate too high) and to be able to change the batches of objects so as not to compromise the feeding throughput of the oriented objects.

[0057] Part of the invention is to calculate a score that, due to its multiple values, can quantify the orientation and visual quality of the produced object. Calculating the score for each produced object requires the following operations:

[0058] - Get the original image A of the produced object k

[0059] - Reposition each original image relative to the corresponding reference image

[0060] - Divide the K original images into secondary images S using the same decomposition process used during the learning phase k,p

[0061] - Using the model F defined during the learning phase k,p and factor Q k,p Calculate each secondary image S k,p The reconstructed image R k,p

[0062] - By comparing the secondary image S k,p and the reconstructed secondary image R k,p Calculate the reconstruction error for each secondary image (obtain all reconstruction errors from all secondary images of the object)

[0063] - Calculate the object score based on the reconstruction error

[0064] Using a compression factor Q k,p Numerical model F k,p It is possible to significantly reduce the computation time, monitor the orientation and quality of the object during the orientation and feeding process and control the process. The method is particularly suitable for methods of feeding oriented objects with high production throughput.

[0065] For example, the invention is advantageous in the packaging field for feeding packaging components, such as tube tops or caps. The invention is particularly advantageous for feeding tube tops and caps at high throughput on machines for producing tubes for so-called "oral care" or cosmetic products. The invention is particularly advantageous for feeding caps for capping equipment.

[0066] For example, the invention can be used for a variety of assembly methods, such as welding, gluing, clamping or screwing. This is the case, for example, with a method for producing packaging tubes, in which the injection components (tube top or shoulder and cover) are assembled at a high rate by welding, clamping or screwing to form the tube. Continuous control of the orientation and aesthetics of the components fed to the assembly machine is very advantageous. This makes it possible to increase efficiency and avoid defective products.

[0067] The invention is primarily directed to an assembly method on an automated production line. The invention is particularly suitable for manufacturing objects with a high production throughput, such as objects produced in the packaging industry or any other industry with a high production throughput.

[0068] According to the present invention, acceptable orientations are automatically defined based on a learning phase. No defect library is required, and the learning phase can define acceptable objects in terms of orientation, size and aesthetics. Once the learning process is performed, improper orientations and any defects are automatically detected during production.

[0069] In some embodiments, the present invention relates to a method for feeding an oriented object (for example, a packaging component such as a tube top or a lid) through a feed bowl (for example, a vibratory bowl or a centrifugal bowl), the method comprising at least one orientation and quality inspection step integrated into the feeding method performed continuously during production, the inspection being based on images of the object captured during feeding and using an artificial intelligence algorithm, and the inspection comprising a learning phase which makes it possible to define acceptable tolerances for the orientation and quality of the object; and a production phase during which only objects having an orientation and a quality within said acceptable tolerances are fed.

[0070] In some embodiments, the learning phase may include at least the following steps:

[0071] -) produce N objects that are considered to have acceptable orientation and quality (i.e., within tolerances that are considered acceptable);

[0072] -) capture at least one reference raw image (A) for each of the N objects k) ;

[0073] -) relocalize each reference original image (A k );

[0074] -) each reference original image (A k ) is divided into (P k ) secondary reference images (S k,p );

[0075] -) grouping the corresponding reference secondary images into batches of N images;

[0076] -) Determine the compression factor (Q) for each batch k,p ) compression-decompression model (F k,p ).

[0077] In some embodiments, the production phase may include at least the following steps:

[0078] -) capturing at least one original image of at least one produced object;

[0079] -) Divide each original image into secondary images (S k,p );

[0080] -) Apply the compression-decompression model and compression factor defined in the learning phase to each secondary image (S k,p ) to form a reconstructed secondary image (R k,p );

[0081] -) Calculate each reconstructed secondary image R k,p The reconstruction error of

[0082] -) assigning one or more scores to each object based on the reconstruction error;

[0083] -) It is possible to calculate directional indices;

[0084] -) Determine whether the fed object successfully passed its orientation and quality checks based on the assigned scores.

[0085] In some embodiments, if the orientation is not within acceptable tolerances, the object may be oriented to be within acceptable tolerances or recovered into the dispensing bowl for subsequent dispensing.

[0086] In some embodiments, if the quality of an object is not within acceptable tolerances, the object can be discarded from the production batch. The object can be rejected, or its defects can be corrected in a manner that eliminates the noted defects (so as to be within acceptable tolerances) and reintroduced into the production batch.

[0087] In some embodiments, after acquiring at least one original image (during a learning and / or production phase), the or each original image may be repositioned.

[0088] In some embodiments, for example, each raw image may be numerically processed. For example, the processing may rely on numerical filters (e.g., Gaussian blur filters) and / or edge detection and / or applying masks to hide certain areas of the image, such as background or areas of no interest.

[0089] In another embodiment, multiple analyses may be performed on one or more original images. Multiple analyses include applying multiple processes to the same original image at the same time. Thus, a "mother" original image may generate multiple "child" original images depending on the number of analyses performed. For example, a "mother" original image may be an object of a first process using a Gaussian filter, generating a first "child" original image, and an object of a second process using a Sobel filter, generating a second "child" original image. Both "child" original images undergo the same numerical processing as defined in the present invention. Thus, one or more scores may be associated with each "child" original image.

[0090] Multi-analysis is beneficial if very different features are being sought on an object. Multi-analysis thus enables the analysis to be adapted to the features being sought. This approach enables a finer detection of each type of feature. The feature can be used to determine the directional index of the object or to detect any defects.

[0091] In some embodiments, the compression factor may be between 5 and 500,000, inclusive, preferably between 100 and 10,000, inclusive.

[0092] In some embodiments, the compression-decompression function may be determined based on principal component analysis (PCA).

[0093] In some embodiments, the compression-decompression function may be determined by an autoencoder.

[0094] In some embodiments, the compression-decompression function may be determined by a so-called Orthogonal Matching Pursuit (OMP) algorithm.

[0095] In some embodiments, the reconstruction error may be calculated based on Euclidean distance and / or Minkovsky distance and / or using Tchebichev method.

[0096] In some embodiments, the score may correspond to a maximum value of the reconstruction error and / or an average value of the reconstruction error and / or a weighted average value of the reconstruction error and / or a Euclidean distance and / or a p-distance and / or a Chebychev distance.

[0097] In some embodiments, N may be at least equal to ten.

[0098] In some embodiments, at least two original images may be captured, the original images having the same size or different sizes.

[0099] In some embodiments, each original image may be divided into P secondary images of the same or different sizes.

[0100] In some embodiments, the secondary images S may be juxtaposed with or without overlapping.

[0101] In some embodiments, some secondary images may be juxtaposed overlappingly while other secondary images are juxtaposed non-overlappingly.

[0102] In some embodiments, the secondary images may be the same size or different sizes.

[0103] In some embodiments, a comprehensive orientation and quality check may be performed at least once during the feeding process.

[0104] In some embodiments, the learning phase may be iterative as objects are fed and repeated during production in order to account for differences that are not considered mis-orientations or defects.

[0105] In some embodiments, localization may include considering a predetermined number of interest points and descriptors distributed across the image, and determining a relative motion between the reference image and the original image that minimizes a superposition error at the interest point level.

[0106] In some embodiments, the points of interest may be distributed in a random manner in the image or in predefined areas of the image.

[0107] In some embodiments, the locations of points of interest may be predefined, random, or otherwise.

[0108] In some embodiments, interest points may be detected by one of the following methods named “SIFT”, “SURF”, “FAST”, or “ORB”, and descriptors may be defined by one of the following methods named “SIFT”, “SURF”, “BRIEF”, or “ORB”.

[0109] In some embodiments, the image may be repositioned relative to at least one axis and / or rotationally repositioned about an axis perpendicular to a plane formed by the image and / or repositioned by a combination of translational and rotational motions.

[0110] In some embodiments, the value of the score may be used to differentiate between objects that are considered misoriented and objects that are considered defective.

[0111] In some embodiments, multiple scores may be used to differentiate between objects that are considered misoriented and objects that are considered defective.

[0112] In some embodiments, the repositioning of the image and at least one score may be used to distinguish between objects that are considered misoriented and objects that are considered defective.

[0113] In some embodiments, the interest points and descriptors and at least one score may be used to distinguish between objects that are considered misoriented and objects that are considered defective.

[0114] In some embodiments, objects that are deemed to be misoriented may be retrieved into the feed system.

[0115] In some embodiments, an object that is considered misoriented may be correctly oriented before or after leaving the feeding system. The orientation system is, for example, a robot or other equivalent device.

[0116] In some embodiments, an object that is considered defective may be discarded from the production batch. For example, discarding it from the production batch may be accomplished by a jet of air that diverts the object from the production stream and ejects it into a reject bin. The object may be "corrected" by eliminating its defects, enabling it to be introduced into the production batch, or simply discarded. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figures 1 to 7 Used to illustrate the present invention.

[0118] · Figure 1 An example of an object being fed in a bowl is shown;

[0119] · Figure 2 shows the original image acquired during the learning phase;

[0120] · Figure 3 shows the cutting of the original image into secondary images;

[0121] · Figure 4 The learning phase is shown, in particular the formation of secondary image batches to obtain a compression-decompression model for each batch;

[0122] · Figure 5 shows the use of a compression-decompression model in the production phase;

[0123] · Figure 6 The main steps of the learning phase are shown in block diagram form;

[0124] · Figure 7 The main steps of the production phase are shown in block diagram form. DETAILED DESCRIPTION

[0125] Figure 1 The invention shows objects 1 being fed in a bowl at a high throughput. In order to illustrate the invention and facilitate its understanding, three decorative patterns are represented on the objects by way of non-limiting examples. The invention enables the orientation of the objects and the quality of these patterns on the objects being fed to be monitored. The invention enables the distribution and inspection of oriented objects at a high production throughput. The invention enables a fast change of the objects intended to be fed with reduced set-up times. Figure 1 In the example shown, the object can be considered as a unit part. The object can be made of plastic material, metal, wood, glass, or based on any other material or combination of these materials.

[0126] Figure 2 An example of a raw image of an object acquired during the learning phase is shown. During this learning phase, N objects judged to be correctly oriented and of acceptable quality are fed by the bowl. For ease of illustrating the present invention, Figure 2 Only four objects are shown in the example. In order to obtain a robust model, the number of objects required during the learning phase is greater than 10 (ie, N>10), preferably greater than 50 (ie, N>50). Of course, these values ​​are non-limiting examples, and N can be less than or equal to 10. Figure 2 Three original images A1, A2 and A3 are shown, each representing a different pattern printed on an object. k Representing the original images of the objects, the index k of the image varies between 1 and K, and K corresponds to the number of images of each object.

[0127] like Figure 2 As shown, the original image A k are not necessarily the same size. Figure 2 In the example, the original image A2 is smaller than the original images A1 and A3. This makes it possible, for example, to give the image A2 a better definition (greater number of pixels). The original images constitute the entire surface of the object 1 or, on the contrary, only partially cover its surface.

[0128] like Figure 2 As shown, the original image A k Targeting specific areas of an object. The flexibility of the invention in terms of size, location and number of raw images allows for optimization of computation time while maintaining a very accurate inspection of the visual quality of the most critical areas.

[0129] Figure 3 shows the division of the original image into secondary images. Figure 3 As shown, the original image A1 is divided into four secondary images S 1,1 , S 1,2 , S 1,3 and S 1,4 Therefore, each original image A k Decomposed into P k Secondary image S k,p , where the partition index p is between 1 and P k Changes between.

[0130] like Figure 3 As shown, the secondary images are not necessarily of the same size. Figure 3 For example, the secondary image S 1,2 and S 1,3 Smaller than the secondary image S 1,1 and S 1,4 This allows for a more precise search for the secondary image S 1,2 and S 1,3 defects in .

[0131] Figure 3 It also shows that the secondary image does not necessarily cover the entire original image A k For example, the secondary image S 2,p The original image A2 is only partially covered. By reducing the size of the secondary image, the analysis is focused on the exact area of ​​the object. Only the area of ​​the object covered by the secondary image is analyzed.

[0132] Figure 3 The present invention describes the method of adjusting the secondary image Sk,p The number, size and position of objects can be locally adjusted in the observation area.

[0133] Figure 4 The learning phase is described, in particular forming sub-batch images to obtain a compression-decompression model with a compression factor for each batch.

[0134] Figure 4 N similar secondary images S are displayed k,p The groups are grouped into batches. Each batch is processed individually and used to create a k,p The compression-decompression model F k,p For example and as Figure 3 As shown, N=4 secondary images S are used. 3,3 Create a compression factor Q 3,3 Model F 3,3 .

[0135] Figure 5 It is shown that the compression-decompression model obtained from the learning phase is used in the production phase. In the production phase, each model F determined in the learning phase k,p Each secondary image S of the object fed into the bowl is used to calculate k,p Thus, each secondary image of the object undergoes a compression-decompression operation, where the compression coefficients and the model are different from those in the learning phase. The result of each compression-decompression operation is a reconstructed image that can be compared with the secondary image from which it was obtained. Comparing the secondary images S k,p and its reconstructed image R k,p The reconstruction error can be calculated which will be used to define the score.

[0136] Figure 5 The use of Model F is illustrated by an illustrative example. 3,3 and its compression factor Q 3,3 From the secondary image S 3,3 Get the reconstructed image R 3,3 special circumstances.

[0137] Figure 6 = represents the main steps of the learning phase according to the invention. At the beginning of the learning phase, N objects judged to be correctly oriented and of acceptable quality are fed through the bowl. The qualitative and / or quantitative judgment of the objects can be performed according to a visual inspection procedure or according to methods and means defined by the user's business. Therefore, the number of objects fed during the learning phase can be equal to N or greater than N. Figure 6 The learning phase shown in includes at least the following steps:

[0138] · KxN so-called "raw" images of the objects judged to be correctly oriented and of good quality are acquired during the distribution of the objects. Each object can be associated with one raw image (K=1) or with a plurality of different raw images (K>1), depending on the size of the area to be analyzed on the object and the size of the defects that one wishes to detect. Lighting and magnification conditions suitable for an industrial environment are used to be able to acquire the images in a relatively constant lighting environment. Known lighting optimization techniques can be used to prevent reflection phenomena or interferences related to the environment. A variety of common solutions can be used, such as tunnels or black boxes, which are able to avoid interferences from lighting from the outside and / or light with a specific wavelength and / or lighting at a grazing angle or indirect lighting.

[0139] If multiple raw images (K>1) are acquired on the same object, the raw images may be spaced, juxtaposed or overlapped. Overlapping the raw images may be useful if one wishes to avoid cutting out defects that may appear between two images and / or to compensate for information loss at the edges of the images associated with an image repositioning step. These methods may also be combined based on the raw images and the information found therein. The images may also be preprocessed by optical or numerical filters, for example to increase contrast.

[0140] The original image is then repositioned relative to the reference image. As a general rule, the original image of any object fed in the learning phase can be used as the reference image. The original image of the first object fed in the learning phase is preferably used as the reference image. The method of repositioning the original image is described in detail in the rest of this application.

[0141] Then each original image A k Divide into P k A so-called "secondary" image. The division of the image may result in an analysis area that is smaller than the original image. If it is known a priori in which area of ​​the object a possible defect is to be looked for, it may be beneficial to reduce the size of the analysis area. The secondary images can be spaced apart from each other to leave a "non-analysis" area between them. This case can be used, for example, if a defect occurs in the target area, or if a defect occurs repeatedly and continuously. Reducing the size of the analysis area can reduce the calculation time. Alternatively, the secondary images can overlap. If a defect occurs at the junction between two secondary images, overlapping secondary images can avoid cutting the defect into two parts. Overlapping secondary images is particularly useful if small defects are to be looked for. Finally, the secondary images can be juxtaposed without spacing and without overlap. The original image can be divided into secondary images of equal or different sizes, and the way the secondary images are relatively positioned (spacing, juxtaposition or superposition) can also be combined depending on the defect to be looked for.

[0142] The next step consists in grouping the corresponding secondary images into batches. A set of secondary images is generated from the secondary images obtained from the KxN original images. Based on this set of secondary images, a batch containing N corresponding secondary images can be formed, i.e., the same secondary image S for each object k,p Therefore, the N secondary images S 1,1 are grouped into a batch. The same applies to N images S 1,2 , then applied to N images S 1,3 , and so on, for all images S k,p .

[0143] The next step consists in finding a compressed representation of each batch of secondary images. This operation is a key step of the method according to the invention. It consists in obtaining a compressed representation of the secondary images of the batch with a compression factor Q k,p The compression-decompression model F k,p Model F k,p It will be used to monitor the quality of the objects during the production phase. Therefore, for a batch S of secondary images 1,1 Get the compressed standard Q 1,1 Model F 1,1 Similarly, for an image batch S 1,2 Get Model F 1,2 , then for the image batch S 1,3 Get Model F 1,3 , and so on, so that for each batch of images S k,p Get Model F k,p .

[0144] Secondary images S for each batch k,p The compression factor Q k,p The choice depends on the available computing time and the size of the defects you wish to detect.

[0145] At the end of the learning phase, a compression factor Q k,p A set of models F k,p , which is associated with the orientation and visual quality of the produced object.

[0146] According to the present invention, the results of the learning phase (including the model F k,p and the compression factor Q k,p ) can be saved as a "template" and reused for subsequent new production runs of the same object. Thus, objects with the same quality can be reproduced by reusing the predefined template. This also makes it possible to avoid repeating the learning phase before starting each production run of the same object.

[0147] According to the invention, iterative learning can be used during the production process. Thus, additional (or supplementary) learning can be performed during the production process, for example with new objects, and the images of these objects are added to the images of the objects initially considered during the learning phase. A new learning phase can be performed based on a new set of images. Evolving learning is particularly suitable if orientation differences or aesthetic differences between objects arise during the production process and the differences are not considered as defects. In other words, these objects were considered "good" in the initial learning phase and this is preferably taken into account. In this case, iterative learning is necessary to avoid high rejection rates for objects containing such differences. Iterative learning can be performed in a variety of ways, for example, by merging new images with previously captured images, or by restarting the learning using newly acquired images, or by retaining only some of the initial images with new images.

[0148] According to the invention, iterative learning is triggered by an indicator related to the rejection of objects. This indicator is, for example, the number of rejections per unit time or the number of rejections per amount of fed objects. If this indicator exceeds a fixed value, the operator is alerted and decides whether an increase in the rejection rate is necessary:

[0149] - Iterative learning phase,

[0150] - Adjust the feeding system,

[0151] -Reject object batches.

[0152] Figure 7 represents the main steps of the object production phase. The production phase begins after the learning phase (that is, when the characteristic criteria of "correct" orientation and "acceptable" quality objects have been defined as described above). The present invention is able to recycle or orient objects that are considered to be misoriented, reject objects that are considered to be defective from the production batch in real time, and avoid the use of objects that are considered to be defective if deviations in the object quality are observed. Figure 7 The production phases according to the invention shown include at least the following operations:

[0153] Get K original images of the object being fed in the bowl. The object images are captured in exactly the same way as the images captured during the learning phase: the area, lighting, magnification, and adjustment conditions captured are the same as those used during the learning phase.

[0154] Reposition the K images relative to the reference image. The purpose of the repositioning operation is to avoid offsets between the images that one wishes to compare. These offsets are related to changes in the position and orientation of the object during imaging.

[0155] Then each original image A of the produced object k Divided into P kThe division is the same as the image division in the learning phase. After this division, a set of secondary images S is obtained for each object produced. k,p .

[0156] Then use model F k,p Compress-decompress each secondary image S k,p , Model F k,p With a compression factor Q predefined during the learning phase k,p This operation is performed for each secondary image S k,p Generate the reconstructed image R k,p . Thus, for the produced object, a reconstructed image is obtained which can be compared with a secondary image of said object. From a numerical point of view, the term "reconstruction of the secondary image" does not necessarily mean that a new image is obtained in the strict sense of the term. The goal is to compare the image of the produced object with the image obtained during the learning phase using the compression-decompression function and the compression factor, only the quantification of the difference between these images is strictly useful. For reasons of computational time, a choice may be made to limit the calculations to numerical objects representing the reconstructed image and sufficient to quantify the difference between the secondary image and the reconstructed image. Using the model F k,p This is particularly advantageous because it may enable such a comparison to be performed in a very short time compatible with requirements and production throughput.

[0157] • The reconstruction error can be calculated based on the comparison of the secondary image and the reconstructed secondary image. A preferred method to quantify this error is to calculate the mean square error, but other equivalent methods are possible.

[0158] ·Thus, for each object, a secondary image and a reconstructed image are obtained, and therefore a reconstruction error is obtained. A plurality of scores can be defined for the produced objects based on this set of reconstruction errors. A plurality of calculation methods can be used to calculate a score for an object, which score characterizes its similarity or difference to the learning batch. Thus, according to the present invention, an object that is visually very different from the learning batch due to having a different orientation or defects will have one or more high scores. Conversely, an object that is visually very similar to the learning batch will have one or more low scores and will be considered to be correctly oriented and of good quality (or of acceptable quality). A third method for calculating one or more scores for an object includes taking the maximum value of the reconstruction errors. Other methods include combining reconstruction errors to calculate the value of one or more scores for an object.

[0159] The next step consists in recycling or correctly orienting the "misoriented" objects and discarding the defective objects from the production batch. If the value of one or more scores of the object is below the predefined one or more limits, the evaluated object meets the orientation and visual quality criteria defined during the learning phase and the object remains in the production stream. On the contrary, if one or more values ​​of one or more scores of the object are greater than said one or more limits, the object is recycled into the feed system (or oriented) because its orientation is outside the acceptable range, or the object is discarded from the production stream because it is defective and it is useless to recycle it.

[0160] There are several methods that can be used to distinguish between misoriented objects and defective objects:

[0161] - The first method consists in using the value of the score to distinguish between misoriented objects and defective objects.

[0162] - According to another approach, multiple scores are used to distinguish between misoriented objects and defective objects.

[0163] - According to another method, the reoriented image and at least one score are used to distinguish between a misoriented object and a defective object.

[0164] According to another method, interest points and descriptors and at least one score are used to distinguish between misoriented objects and defective objects.

[0165] The steps of the present invention will be returned to below and described in more detail.

[0166] Reposition the original image

[0167] The method for repositioning an image according to the present invention comprises two steps:

[0168] - Search for interest points and descriptors in images

[0169] -Relocalize captured images relative to a reference image and infer the orientation of objects based on interest points and descriptors

[0170] One or more reference images are typically defined on the first image or another image captured during the learning phase, as described in the present application. The first step consists in defining interest points on the image and descriptors associated with the interest points. The interest points can be, for example, corner parts at the level of shapes present in the image; they can also be areas of high contrast or color, or interest points can be randomly selected. The identified interest points are then characterized by descriptors defining the characteristics of these interest points.

[0171] The points of interest are preferably determined automatically using a suitable algorithm, but alternatives include arbitrarily pre-defining the locations of the points of interest.

[0172] The number of interest points used for relocalization varies and depends on the number of pixels per interest point. The total number of pixels used for localization is typically between 100 and 10,000 (inclusive), and preferably between 500 and 1,000 (inclusive).

[0173] A first method for defining points of interest consists in randomly selecting these points. This amounts to randomly defining a percentage of pixels called points of interest, the descriptors being the characteristics of said pixels (position, color). The first method is particularly suitable for industrial production environments, especially in the case of high-throughput production processes, where the time available for calculations is very short.

[0174] According to a first embodiment of the first method, the interest points are randomly distributed in the image.

[0175] According to a second embodiment of the first method, the interest points are randomly distributed in predefined areas of the image. This second embodiment is advantageous when it is known a priori where any defects will appear.

[0176] A second method for defining points of interest is based on the so-called "Scale Invariant Feature Transform ("SIFT") method" (see publication US Pat. No. 6,711,293), i.e. a method that makes it possible to preserve the same visual features of an image independently of its scale. The method consists in calculating image descriptors at points of interest of said image. These descriptors correspond to numerical information derived from a local analysis of the image that characterizes the visual content of the image independently of its scale. The principle of the method consists in detecting in the image defined areas around the points of interest, said areas being preferably circular and having a radius known as the scale factor. In each of these areas a shape and its contours are sought and then the local orientation of the contours is defined. Numerically, these local orientations give rise to vectors that constitute the SIFT descriptors of the points of interest.

[0177] A third method for defining interest points is based on the "Speeded Up Robust Features ("SURF") method" (see publication US2009 / 0238460), an accelerated method for defining interest points and descriptors. This method is similar to the SIFT method, but has the advantage of execution speed. Like the SIFT method, this method includes the steps of extracting interest points and calculating descriptors. The SURF method uses Fast Exact Multiplication of the Hessian to detect interest points and uses an approximation of the Haar wavelet to calculate descriptors.

[0178] A fourth method for finding points of interest based on features from the Feature Accelerated Segment Test ("FAST") method involves identifying potential points of interest and then analyzing the intensity of pixels located around the points of interest. This method is able to identify points of interest very quickly. Descriptors can be identified using the Binary Robust Independent Elementary Features ("BRIEF") method.

[0179] The second step of the method for relocalizing the image consists in comparing the original image with the reference image using the interest points and their descriptors. The optimal relocalization is achieved by finding the best alignment between the descriptors of the two images.

[0180] In this case, the image may need to be repositioned relative to only one axis, or relative to two perpendicular axes, or rotationally repositioned about an axis perpendicular to the plane in which the image is formed.

[0181] The repositioning of the image can be the result of a combination of translation and rotation movements. The best homography transformation is found using the least squares method.

[0182] Interest points and descriptors are used for the operation of relocalizing the image. These descriptors can be features of pixels or SIFT, SURF, BRIEF descriptors for example. Interest points and descriptors are used as markers for relocalizing the image.

[0183] In SIFT, SURF and BRIEF methods, relocalization is achieved by comparing descriptors. Irrelevant descriptors are discarded using a consensus method (e.g., Ransac algorithm). The best homography is then found using the least squares method.

[0184] Divide the original image into secondary images

[0185] The original image can be divided into P secondary images in various ways.

[0186] One advantage of the present invention is the ability to adapt the visual analysis level to the observed area of ​​the object. This adjustment is performed as a first step based on the number of primary images and the resolution level of each primary image. Then, the decomposition into secondary images enables the analysis level to be adjusted locally in each primary image. The first parameter that can be manipulated is the size of the secondary images. Smaller secondary images enable local refinement of the analysis. By collectively adjusting each secondary image S k,p The size and compression factor Q k,p The invention makes it possible to optimize the computation time while maintaining a high performance detection level adjusted to the required level associated with the conveyed object. The invention makes it possible to locally adjust the detection level to the criticality level of the observation area.

[0187] One specific embodiment of the invention comprises that all secondary images have the same size.

[0188] Therefore, when all observation areas are of equal importance, a first method consists in dividing the original image into P secondary images of equal size juxtaposed without overlap.

[0189] The second method consists in dividing the original image into P secondary images of the same size which are juxtaposed in an overlapping manner. The overlap is adjusted according to the size of the defects that may appear on the object.

[0190] The smaller the defect, the smaller the overlap can be. It is usually assumed that the overlap is at least equal to the characteristic half-length of the defect, which is defined as the smallest diameter of a circle that can contain the entire defect.

[0191] Of course, it is possible to combine these approaches and use secondary images that are juxtaposed and / or have overlap and / or are at a distance from each other.

[0192] Calculation of compression-decompression function

[0193] According to a first method, which is also the preferred method, the compression-decompression function and the compression factor are determined based on principal component analysis (PCA). This method allows defining eigenvalues ​​and vectors that characterize the batches generated during the learning phase. In the new basis, the eigenvectors are sorted in order of size. The compression factor is derived from the number of dimensions preserved in the new basis. The higher the compression factor, the fewer the dimensions in the new basis. The invention allows adjusting the compression factor according to the desired level of inspection and the available computing time.

[0194] The first advantage of this method is related to the fact that the machine does not need instructions to define a new basis. The eigenvectors are chosen automatically by calculation.

[0195] The second advantage of the method is related to the reduction of computing time for detecting defects in the production phase. Due to the reduction in dimensionality, the amount of data to be processed is reduced.

[0196] A third advantage of the method is that one or more scores can be assigned in real time to the images of the produced objects. The one or more scores obtained are able to quantify the level of deviation / error of the object fed in the bowl with respect to the object from the learning phase, by reconstructing it using the model from the learning phase.

[0197] The compression factor is between 5 and 500000 (inclusive), and preferably between 100 and 10000 (inclusive). The higher the compression factor, the shorter the computation time for analyzing the image at the production stage. However, a compression factor that is too high may result in a model that is too coarse and unsuitable for detecting errors.

[0198] According to the second approach, the model is an autoencoder. The autoencoder takes the form of a neural network that is able to define features in an unsupervised manner. The autoencoder consists of two parts: an encoder and a decoder. The encoder can compress the secondary image S k,p , the decoder can obtain the reconstructed image R k,p .

[0199] According to the second approach, one autoencoder is available for each batch of images. Each autoencoder has its own compression factor.

[0200] According to the second approach, the autoencoder is optimized during the learning phase. The autoencoder is optimized by comparing the reconstructed image with the initial image. This comparison allows quantifying the difference between the initial and reconstructed image and thus determining the encoder error. The learning phase allows optimizing the autoencoder by minimizing the image reconstruction error.

[0201] According to a third approach, the model is based on the “Orthogonal Matching Pursuit (“OMP”) algorithm”. This method consists in finding the best linear combination based on orthogonal projections of several images selected from a library. The model is obtained through an iterative method. Each time an image is added from the library, the reconstructed image is improved.

[0202] According to a third method, a library of images is defined by a learning phase. The library is obtained by selecting a few images representing a set of images from the learning phase.

[0203] Computational reconstruction of the image from the compression-decompression model

[0204] In the production phase, each original image A of the inspected object is repositioned using the above method. k , and then divide it into P k Secondary image S k,p Each secondary image S k,p A numerical reconstruction operation is performed using its model defined in the learning phase. Thus, at the end of the reconstruction operation, each secondary image S k,p Both have reconstructed images R k,p Available.

[0205] Using a compression factor Q k,p Model F k,p Reconstruct each secondary image S k,p The operation can achieve very short computation time. The compression factor Q k,p Between 5 and 500,000 (inclusive), and preferably between 10 and 10,000 (inclusive).

[0206] According to the PCA method (also the preferred method), the secondary image S is first k,pThen use the function F defined in the learning phase k,p This vector is projected into the eigenvector basis. The reconstructed image R is then obtained by transforming the obtained vector into an image k,p .

[0207] According to the second method, the secondary image is reconstructed by an autoencoder whose parameters are defined during the learning phase. k,p To obtain the reconstructed image R k,p .

[0208] According to a third approach, the secondary image is reconstructed using an Orthogonal Matching Pursuit (OMP) algorithm, the parameters of which are defined during a learning phase.

[0209] Calculate the reconstruction error for each secondary image

[0210] By comparing the secondary image S k,p and the reconstructed image R k,p Get the reconstruction error.

[0211] One method of calculating the error involves measuring the secondary image S k,p and the reconstructed image R k,p The preferred method for calculating the reconstruction error is the Euclidean distance or the 2-norm method. This method considers the square root of the sum of the squares of the errors.

[0212] Another method for calculating the error is to use the Minkowski distance, or p-distance, which is a generalization of the Euclidean distance. This method considers the p-th root of the sum of the absolute values ​​of the errors raised to the pth power. This method can assign greater weight to larger differences by choosing a p-value greater than 2.

[0213] Another alternative is the Chebychev or 3-norm method. This method considers the maximum absolute value of the error.

[0214] Calculate one or more scores

[0215] The value of one or more scores of the object is obtained from the reconstruction error of each secondary image.

[0216] A preferred method comprises assigning to the score the maximum value of the reconstruction error.

[0217] An alternative approach consists in calculating the value of the score by taking the average of the reconstruction errors.

[0218] Another alternative approach consists in obtaining a weighted average of the reconstruction errors. A weighted average may be useful if the criticality of the defect is not the same in all areas of the object.

[0219] Another approach involves using the Euclidean distance or the 2-norm.

[0220] Another approach involves using p-distance.

[0221] Another approach involves using the Chebychev distance or the 3-norm.

[0222] Other equivalent methods are of course possible within the context of the present invention.

[0223] Once the score or scores are calculated, the value or values ​​thereof are used to determine whether the object under consideration meets the desired quality and orientation conditions. If so, it is retained in the feed stream. If the score does not meet the conditions because the orientation of the object is outside the acceptable range, the object is recycled into the feed system or reoriented. If the score does not meet the conditions because the object is defective, the object is discarded from the feed process.

[0224] Misoriented objects can be distinguished from defective objects based on the score values. Thus, for example, for a misoriented (inverted) cap, the score varies between 7 and 10, while a cap defect yields a score between 3 and 5. Thus, an inverted cap can be easily distinguished from a defective cap.

[0225] In other cases, it is suggested to use multiple scores to distinguish between defective objects and misoriented objects. In particular, the invention enables the definition of scores for local areas of eccentric objects. For example, consider an object comprising an eccentric aperture. A local image of the aperture enables the acquisition of a score related to the object orientation. Combining the score of the aperture with the other scores thus enables the distinction of misoriented objects from defective objects.

[0226] According to an alternative approach, information about the repositioned object and one or more scores are used to distinguish between a misoriented object and a defective object.

[0227] According to another approach, interest points and descriptors are used together with at least one score to distinguish between misoriented objects and defective objects.

[0228] The misoriented object is preferably recovered into the bowl. A first method consists in blowing the assembly into the bowl by at least one air jet in the trajectory of the object. Another method consists in mechanically displacing the assembly into the bowl by a piston and cylinder. The system is capable of recovering the object by air jet or mechanical actuation.

[0229] In other embodiments, the orientation of the misoriented object is corrected before or after the object leaves the bowl. A variety of object orientation systems can be envisioned and associated with the present invention. Depending on the complexity of the orientation movement to be performed, these systems can include one or more axes. The orientation system is, for example, a robot.

[0230] In the present example, it must be clearly understood that the method is implemented in a feeding system (e.g. a vibratory bowl or a centrifugal bowl) that can have a high throughput (e.g. at least 100 products per minute). If in the examples a singular number is used to define the produced object, it is for the sake of simplicity. In fact, the method is applicable to the production of a continuous stream of objects in a feeder: the method is therefore iterative and repeated for each continuous stream of objects fed, and the orientation and quality of all said continuous streams are checked.

[0231] The described embodiments are described by way of illustrative examples and should not be considered as limiting the present invention. For example, other embodiments may rely on devices equivalent to the described embodiments. The embodiments may also be combined with each other as appropriate, or the devices and / or method steps used in one embodiment may be used in another embodiment of the present invention.

Claims

1. A method for feeding oriented objects, such as vibratory or centrifugal bowls, through a feeding bowl, the oriented objects being, for example, packaging components such as tube tops or caps, comprising at least one orientation and quality inspection step integrated into the feeding method performed continuously during production, the inspection being based on an image of the object captured during feeding and using an artificial intelligence algorithm, The inspection comprises a learning phase which enables the definition of acceptable tolerances for the orientation and quality of the object; and a production phase during which only objects having orientation and quality within said acceptable tolerances are introduced, wherein: The learning phase includes at least the following steps: -) produce N objects that are deemed to have orientation and quality within acceptable tolerances; -) capturing at least one reference original image (Ak) of each of the N objects; -) Divide each reference original image (Ak) into (P k ) secondary reference images (S k，p ); -) grouping the corresponding reference secondary images into batches of N images; -) Determine the compression factor (Q) for each batch k，p ) compression-decompression model (F k，p ), And the production stage includes at least the following steps: -) capturing at least one original image of at least one object produced; -) Each original image is divided into secondary images (S k，p ); -) applying the compression-decompression model and the compression factor defined in the learning phase to each secondary image (S k，p ) to form a reconstructed secondary image (R k，p ); -) Calculate each reconstructed secondary image R k，p The reconstruction error of -) assigning one or more scores to each object based on the reconstruction error; -) Determine whether the fed object has successfully passed the check of its orientation and its quality based on the assigned score or scores.

2. The method according to claim 1, wherein: If an object is deemed to be misoriented, it is either oriented to within acceptable tolerances or retrieved into a feed bowl.

3. A method according to any one of the preceding claims, wherein: If an object is deemed defective, it is discarded from the production batch.

4. A method according to any one of the preceding claims, wherein: The value of the score is used to distinguish between correctly oriented objects and defective objects.

5. A method according to any one of the preceding claims, wherein: Multiple scores are used to distinguish correctly oriented objects from defective objects.

6. A method according to any one of the preceding claims, wherein: At least one raw image that was initially captured is subjected to multiple analyses that generate a "child" raw image that is used to replace the image that was initially captured at its source.

7. A method according to any one of the preceding claims, wherein: After the step of acquiring at least one raw image, each raw image is repositioned.

8. A method according to any one of the preceding claims, wherein: Each raw image is processed using filters and / or contour detection and / or applying masks to hide certain areas of the image.

9. A method according to any one of the preceding claims, wherein: The score corresponds to the maximum value of the reconstruction error and / or the average value of the reconstruction error and / or the weighted average value of the reconstruction error and / or the Euclidean distance and / or the p-distance and / or the Chebychev distance, which is the value of the distance between the secondary image S k，p and the reconstructed image R k，p between.

10. A method according to any one of the preceding claims, wherein: At least two original images are captured, the original images having the same size or different sizes.

11. A method according to any one of the preceding claims, wherein: Each original image is divided into P secondary images S of the same size or different sizes, which are juxtaposed overlappingly and / or non-overlappingly.

12. A method according to any one of the preceding claims, wherein: The learning phase is iterative as objects are fed and repeated during production in order to take into account any differences in what are considered acceptable orientation or quality defects.

13. A method according to any one of the preceding claims, wherein: A relocalization step is performed, wherein the relocalization step includes taking into account a predetermined number of interest points and descriptors distributed on the image, and determining a relative motion between the original image and the reference image that minimizes a superposition error at the level of interest points, and the interest points are randomly distributed in the image or in predefined areas of the image, and the positions of the interest points are predefined, arbitrary or in other ways.

14. The method according to claim 13, wherein: The image is repositioned in at least one axis and / or the image is rotationally repositioned about an axis perpendicular to a plane formed by the image and / or the image is repositioned by a combination of translational and rotational movements.

15. A method according to any one of the preceding claims, wherein: The repositioned image and the at least one score are used to distinguish between a misoriented object and a defective object, or the interest points and the descriptors and the at least one score are used to distinguish between a misoriented object and a defective object.

Images