Orientation part feeding method

Abstract

A method for feeding objects such as tube tops or caps includes at least one orientation and quality inspection step incorporated into the feeding method that occurs continuously during production, the orientation and quality inspection step including a learning phase and a production phase.

Description

[Background technology]

[0001] [Corresponding application]

[0002] This application claims priority to an earlier European patent application number EP22191165.4, filed on August 19, 2022 in the name of AISAPACK Holding SA, the contents of which are incorporated herein by reference in their entirety. [Field of the Invention]

[0003] The present invention is in the field of mass-produced objects requiring high throughput feeding or dispensing systems such as vibratory or centrifugal bowls. More specifically, the present invention relates to a feeding method and device that utilizes visual inspection and artificial intelligence algorithms to feed oriented objects at high production throughput. [Prior art]

[0004] High throughput delivery systems for orienting objects are known in the prior art, see for example the following documents: US 5,311,977, US 4,608,646, DE 3,312,983, US 4,692,881 and US 5,853,078.

[0005] U.S. Patent No. 5,311,977 describes a system for feeding objects that allows for geometric inspection of the object and for determining its orientation by reorienting or rejecting it with the aid of a microprocessor output signal. In this publication, the geometric inspection is performed by an object detector having at least 1,000 linearly arranged pixels oriented to lighten or darken as a function of the object's geometric shape. The system described in this publication includes means for detecting points on the contour of an object placed within a scanning tranche and for comparing the positions of the contour points with a stored profile in real time. The system allows for the orientation or rejection of the object in response to a microprocessor output signal based on contour point signals from multiple scanning tranches.

[0006] U.S. Pat. No. 4,608,646 describes a microcontroller-based system for recognizing and identifying identical or different objects transported along a track in an object feeder, such as a bowl feeder, verifying the object's orientation, and sorting the objects oriented in a predetermined, repetitive sequence. Object recognition and ordering are programmable according to user requirements. Object recognition requires a device for recognizing the object's silhouette, which includes a set of optical sensors coupled to a perforated grid located within the feed track. An image of the silhouette of each object to be sorted is first digitized and stored in the microcontroller's memory at a location associated with the object's identification number. Similarly, the ordering of the different objects is stored in the microcontroller's memory. Subsequently, as the objects are fed onto the grid, each object is compared with the corresponding stored image of the correctly positioned sequence. Incorrect or incorrectly oriented objects are rejected by an air jet directed onto the feed track, while correct recognition of the object leads to the cessation of the air jet, allowing the object to be sent through the feed output station.

[0007] Publication DE 3312983 describes a vibratory bowl for sorting mechanical components using the component's position and contour as a decision criterion, the device comprising a conveying device for conveying the components essentially perpendicular to a line of electronic sensors, thereby enabling the contour of the component to be searched line by line, and using electronic comparators capable of transmitting output signals from the line of sensors, thereby enabling the component to be compared with previously stored setpoint values.

[0008] U.S. Patent No. 4,692,881 describes a device for feeding objects in a predetermined orientation. The device includes a detector consisting of a plurality of light-receiving elements arranged in one or more lines extending perpendicular to the object feeding direction, and at least one light-emitting element spaced apart from and facing the light-receiving elements. The device also includes a random access memory (RAM) for storing a reference signal model obtained by continuously detecting the shape of an object passing in front of the detector at a preselected required position. The device also includes a central processing unit (CPU) for comparing the reference signal model with a signal data model obtained when an object to be identified continuously passes any position in front of the detector. Improperly oriented objects are rejected into the bowl in response to each unfavorable comparison.

[0009] U.S. Patent No. 5,853,078 describes an apparatus for orienting and feeding objects that is particularly suitable for use in automated assembly systems. The apparatus includes a feed bowl with a spiral internal track that terminates at the level of the bowl's upper edge adjacent to an annular feed ring mounted for selective movement while rotating around the feed bowl. A control circuit including a fixed video camera positioned above the annular feed ring controls the rotational movement of the annular feed ring by an operatively connected motor to bring successive portions of the annular feed ring into a predetermined field of view of the video camera and distinguish correctly oriented objects from incorrectly oriented objects. A signal is then provided to a pick-and-place robot to remove the correctly oriented objects. A sweeper bar is positioned at selected locations to push incorrectly oriented objects off the annular feed ring and return them to the feed bowl for recycling. Another embodiment of a vibratory feed tank is also provided which utilizes a second selectively rotating disk in concentric and spaced relation to the feed ring to receive recovered objects removed from the annular feed ring into a container provided on the ring.

[0010] Objectives, constraints, and problems to be solved

[0011] The present invention aims to reduce the adjustment time of high-throughput systems, such as vibratory or centrifugal bowls, for delivering oriented objects. Despite improvements proposed in the prior art, particularly those described in U.S. Pat. Nos. 5,311,977, 4,608,646, 3,312,983, 4,692,881, and 5,853,078, these systems do not allow for rapid object changes, resulting in significant time wasted adjusting for each change of object. To overcome this difficulty, vibratory or gravity bowls are often used for unique object shapes, since the time required to replace a vibratory bowl on an assembly machine is shorter than the time required to adjust the bowl to deliver new objects at the required throughput. This situation has the disadvantage of requiring the investment and storage of multiple bowls for delivering objects individually tailored to a single object or a limited number of objects.

[0012] Another drawback of the devices described in the prior art relates to undetected defective objects, such as deformed or out-of-tolerance objects or products with poor aesthetics (e.g., appearance), which can cause untimely shutdowns of assembly machines or lead to defects in the assembled products.

[0013] The present invention makes it possible to remedy the aforementioned drawbacks thanks to a bowl equipped with a vision system associated with an artificial intelligence algorithm and possibly associated with an orientation means. The present invention also makes it possible to define rejection criteria for so-called defective parts. Rejection criteria can be related, for example, to the dimensions of the object, such as objects that are deformed or have dimensions outside the permitted range, or to aesthetic defects (for example scratches, stains, foreign bodies, incorrect color, etc.).

[0014] In accordance with the present invention, artificial intelligence algorithms associated with the vision system allow for rapid changeover of objects in the feed bowl with a high throughput of oriented objects. The present invention also allows for the rejection of defective objects, which avoids shutting down the assembly machine when objects are out of tolerance or deformed, and also avoids the use of objects with unsuitable aesthetics or appearance.

[0015] According to the invention, the learning phase allows the definition of a "norm" of what is acceptable for the supplied objects. This "norm" defines the range of object orientations and, where applicable, the acceptable dimensional and aesthetic ranges. In the context of the invention, the concepts of "acceptable or unacceptable defects", i.e., the concepts of objects that are considered "good" or "defective", are defined relative to a certain level of offset relative to a predetermined level established by learning.

[0016] The present invention makes it possible to guarantee a consistent object orientation and level of quality over time. Furthermore, templates, i.e., pre-established norms, can be reused to create the same object at a later time.

[0017] The orientation and quality level of an object can be adjusted over time by iterative learning as a function of observed differences. During production, the norms defined in the initial learning are fed into the normal production phase, but are refined by "supplemental" learning that takes into account objects with orientations or defects that are considered acceptable. Therefore, the norms must be adapted to integrate this information and prevent the process from rejecting these objects.

[0018] The present invention allows for the distribution of objects in very short time periods, and to achieve this performance it relies on a model of compression-decompression of images of objects, 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, inter alia:

[0020] - Inspection times are reduced, since visual inspection is performed while the objects are moving in the bowl, without the need to slow down production throughput and, at best, the inspection has a low impact on the latter. The acceptance range of the oriented object must be adjustable. - No known dimensional or visual defects (no defect library). -Aesthetic defects vary as a function of decor. - The defect acceptance level must be adjustable. An object that is considered misoriented is an object whose orientation is outside the range of orientation tolerances that are considered acceptable. -An object that is considered to be defective is an object whose defects are considered to be beyond the limits of tolerance.

[0021] The method proposed by the invention described below makes it possible to mitigate the above drawbacks and overcome the identified problems.

[0022] definition

[0023] -Object: the object to be dispensed (or dispensed) in the bowl, for example a cap or tube top. N: The number of objects forming the batch in the learning phase. N also corresponds to the number of secondary images that make up one batch. - Primary Image: A captured image of an object or part of an object. -K: Number of primary images per object. Ak: primary image with index k, 1 to K. - Secondary image: a part of the primary image. -Pk: Number of secondary images per primary image. -AkS k,p : Index k is primary image A k and the secondary image associated with index p is between 1 and Pk inclusive. -F k,p Model: Secondary Image S k,p The compression-decompression model associated with -Compression factor Qk,p :Model F k,p Compression factor of. -Reconstructed secondary image R k,p :Related Model F k,p Using the secondary image S k,p Reconstructed secondary image reconstructed from.

[0024] General description of the invention

[0025] The present invention relates to a method for dispensing oriented objects, e.g., packaging components such as tube tops or caps, which includes a visual inspection integrated into one or more steps of the method for dispensing the objects. The dispensing method according to the present invention has at least two phases for performing the visual inspection.

[0026] A learning phase in which a batch of objects considered to be "correctly oriented" and "good quality" is fed, after which criteria for the learning phase are defined based on images of said objects. -Generation forces that use images of the generated objects and criteria defined in the learning phase to quantify the orientation and quality of the objects being fed in real time and control the feeding process.

[0027] During the learning phase, the machine feeds N objects that are deemed to be of acceptable quality and orientation. One image (K=1) or multiple distinct images (K>1), called primary images, of each object are collected during the process of feeding the object. The collected KxN primary images undergo digital processing, which is described in more detail below and includes at least the following steps: -Repositioning of each primary image A k -Each primary image A k , S k,p Pk secondary images S denoted by k,p Here, k is 1 or more and K or less, and p is 1 or more and Pk or less. -Grouping secondary images into batches of N similar images. -Secondary Images k,p For each batch: Compression factor Q k,p The compressed representation F k,p Ask for. From each batch of secondary images, the compression factor Q k,p One particular example of the present invention is to derive the compression-decompression model Fk,p using k,p Each model F k,p Compression coefficient Q for k,p The adjustment of allows the adjustment of the level of defect detection and the optimization of the calculation time as a function of the observed area of the object.

[0028] Therefore, at the end of the learning phase, the model F k,p and the compression factor Q k,p are available for each observation area of the object, and each area is represented by a secondary image S k,p is defined by

[0029] As will be explained in more detail below, each secondary image of an object has its own dimensions. A special case of the present invention involves having all secondary images of the same size. In some cases, it is advantageous to be able to locally reduce the size of the secondary images in order to detect smaller defects. The size S of each secondary image k,p and the compression factor Q k,p By simultaneously adjusting the , the present invention allows for optimization of computation time while maintaining a high level of detection performance adjusted to match the level of requirements associated with the manufactured product. The present invention allows for local adaptation of the detection level to match the criticality level of the observed area.

[0030] During the generation phase K, we use so-called "primary" images of each object to monitor the orientation and quality of the objects being generated in real time. This allows us to: - Recycling of misoriented objects into the feeding system or correcting the misoriented objects. -Removing defective products from production as soon as possible.

[0031] To achieve real-time monitoring of the objects being produced, K primary images of the object are evaluated by the method described herein against a group of primary images acquired during a learning phase, from which a compression-decompression function and compression coefficients to be applied to the images of the object being produced are extracted. The 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 per object, whose values allow for classification of the object against thresholds corresponding to a level of orientation and a level of visual quality. Depending on the score values and predefined thresholds, incorrectly oriented objects are recycled or reoriented, and defective objects are discarded from the production process. Other thresholds can be used to detect batches of defective objects (too high a reject rate) and allow for the re-organization of the object batch without impairing the feed throughput of oriented objects.

[0032] Part of the invention resides in the calculation of a score that allows, thanks to several numerical values, to quantify the orientation and visual quality of the generated objects. Calculating the score for each generated object requires the following operations: - Acquisition of a primary image Ak of the object being generated - Realigning each primary image with respect to its respective reference image. - Using the same decomposition process used during the training phase, we decompose the K primary images into secondary images S k,p Divide into. - the model F defined during the learning phase k,p and coefficient Q k,p Using each secondary image S k,p Reconstructed image R k,p Calculate. -Secondary Images k,p and the reconstructed secondary image R k,p (All reconstruction errors are calculated for all secondary images of the object.) - Calculate the object score based on the reconstruction error.

[0033] Compression factor Q k,p A numerical model F withk,p The use of the method allows for a significant reduction in computation time, monitoring of the orientation and quality of the object during the orientation and delivery process, and control of the process. The method is particularly suited for methods of delivering oriented objects at high production throughput.

[0034] The present invention is advantageously used in the packaging field, for example, for dispensing packaging components such as tube tops or caps. The present invention is particularly advantageous for high-throughput dispensing of tube tops and caps in machines for producing tubes for so-called "oral care" or cosmetic products. The present invention is particularly advantageous for dispensing capping devices with caps.

[0035] The present invention can be used with many assembly methods, such as welding, gluing, clipping, or screwing. This is the case, for example, in the production of packaging tubes, where injected components (tube tops or shoulders and caps) are assembled at high speed by welding, clipping, or screwing to form the tube. It is highly advantageous to continuously control the orientation and appearance of the components fed to the assembly machine. This can improve efficiency and avoid rejects.

[0036] The present invention is primarily directed to an assembly method in an automated production line, and is particularly suitable for manufacturing objects with a high production throughput, such as objects produced in the packaging sector or any other sector with a high production throughput.

[0037] According to the present invention, allowable orientations are automatically defined based on a learning phase. No defect library is required, and the learning phase allows for the definition of objects that are acceptable in terms of their orientation, dimensions, and aesthetics. Improper orientations and defects are automatically detected during production once the learning procedure has been performed.

[0038] In some embodiments, the invention relates to a method for feeding oriented objects, e.g., packaging components such as tube tops or caps, by a feeder bowl, such as a vibratory or centrifugal bowl, said method comprising at least one orientation and quality inspection step integrated into the feeding method performed continuously during production, said inspection being based on images of the objects captured during feeding and using artificial intelligence algorithms, said inspection comprising a learning phase allowing for the definition of acceptable tolerances for object orientation and quality, and a production phase in which only objects whose orientation and quality fall within said acceptable tolerances are fed.

[0039] In some embodiments, the learning phase may include at least the following stages: -) generating N objects that are considered to have acceptable orientation and quality (i.e., within tolerances that are considered acceptable); -) capturing at least one reference primary image (Ak) of each of the N objects; -) Each reference primary image (A k ) rearrangement; -) Each reference primary image (A k ) to (P k ) secondary reference images (S k,p ) into the following steps: -) grouping corresponding reference secondary images into batches of N images; -) Batch compression factor (Q k,p ) using the compression-decompression model (F k,p ) is determined.

[0040] In some embodiments, the production phase may include at least the following stages: -) capturing at least one primary image of at least one object to be generated; -) each primary image into a secondary image (S k,p ) into the following steps: -) The compression-decompression model and compression coefficients defined in the training phase are applied to each secondary image (S k,p ) to obtain the reconstructed secondary image (R k,p) forming stage; -) Each reconstructed secondary image R k,p calculating the reconstruction error of -) assigning one or more scores to each object based on the reconstruction error; -) If possible, a stage of calculating the orientation index; -) Determine whether the object being fed has successfully passed its orientation and quality inspection based on the assigned score.

[0041] In some embodiments, if the orientation is not within tolerance, the object can be oriented to be within tolerance or recycled into the dispensing bowl for subsequent dispensing.

[0042] In some embodiments, if the quality of an object is not within tolerance, the object may be discarded from the production batch. If a defect is discovered, the object may be rejected in a manner that removes the defect (to within acceptable tolerances), or the defect may be corrected and reintroduced into the production batch.

[0043] In some embodiments, after the step of acquiring at least one primary image (during the training and / or generation phase), the or each primary image may be repositioned.

[0044] In some embodiments, each primary image may be processed, for example, numerically, which may rely, for example, on numerical filters (such as Gaussian blur filters) and / or edge detection and / or the application of masks to conceal certain areas of the image, such as, for example, the background or regions of no interest.

[0045] In another embodiment, multiple analyses can be performed on one or more primary images. Multiple analyses include applying multiple procedures to the same primary image simultaneously. Thus, a "mother" primary image can give rise to multiple "daughter" primary images as a function of the number of analyses performed. For example, a "mother" primary image can be subjected to a first processing by a Gaussian filter to generate a first "daughter" primary image, and a second processing by a Sobel filter to generate a second "daughter" primary image. The two "daughter" primary images undergo the same numerical processing defined by the present invention. Thus, one or more scores can be associated with each "daughter" primary image.

[0046] Multiple analyses are useful when very different properties are sought on an object. Thus, multiple analyses allow the analysis to be adapted to the properties sought. This method allows for more refined detection of each type of property. This property can be used to determine the orientation index of the object or to detect defects.

[0047] In some embodiments, the compression factor can be between 5 and 500,000, preferably between 100 and 10,000.

[0048] In some embodiments, the compression-decompression function may be determined based on principal component analysis (PCA).

[0049] In some embodiments, the compression-decompression function may be determined by an auto-encoder.

[0050] In some embodiments, the compression-decompression function may be determined by the so-called Orthogonal Matching Pursuit (OMP) algorithm.

[0051] In some embodiments, the reconstruction error may be calculated based on Euclidean and / or Minkovsky intervals and / or using the Tchebichev method.

[0052] In some embodiments, the score may correspond to a maximum reconstruction error and / or a mean reconstruction error and / or a weighted mean reconstruction error and / or a Euclidean interval and / or a p-interval and / or a Chebychev interval.

[0053] In some embodiments, N may be at least equal to 10.

[0054] In some embodiments, at least two primary images may be captured, the primary images being the same size or different sizes.

[0055] In some embodiments, each primary image may be divided into P secondary images of the same or different sizes.

[0056] In some embodiments, the secondary images S may be juxtaposed with or without overlap.

[0057] In some embodiments, some secondary images may be juxtaposed with overlap, and other secondary images may be juxtaposed without overlap.

[0058] In some embodiments, the secondary images may be the same size or different sizes.

[0059] In some embodiments, a comprehensive inspection of orientation and quality may be accomplished at least once in the feeding process.

[0060] In some embodiments, the learning phase may be iterated and repeated during production as objects are fed to account for differences that are not considered to be incorrect orientations or defects.

[0061] In some embodiments, the positioning may involve considering a predetermined number of interest points and descriptors distributed across the image and determining a relative movement between the reference image and the primary image that minimizes the registration error at the interest point level.

[0062] In some embodiments, the points of interest may be randomly distributed within the image or within a predetermined area of the image.

[0063] In some embodiments, the locations of the points of interest may be randomly or otherwise predefined.

[0064] In some embodiments, interest points may be detected by one of the following methods named "SIFT", "SURF", "FAST" or "ORB", and a descriptor defined by one of the methods named "SIFT", "SURF", "BRIEF" or "ORB".

[0065] In some embodiments, the image may be repositioned with respect to at least one axis, and / or may be repositioned by rotation about an axis perpendicular to the plane formed by the image, and / or may be repositioned by a combination of translational and rotational movements.

[0066] In some embodiments, the value of the score can be used to distinguish objects that are considered to be misoriented from objects that are considered to be defective.

[0067] In some embodiments, multiple scores can be used to distinguish objects that are considered to be misoriented from objects that are considered to be defective.

[0068] In some embodiments, the realignment of the image and at least one score can be used to distinguish objects that are considered to be misoriented from objects that are considered to be defective.

[0069] In some embodiments, the interest points and descriptors and at least one score can be used to distinguish objects that are considered to be misoriented from objects that are considered to be defective.

[0070] In some embodiments, objects that are deemed to be misoriented may be recycled into the feeder system.

[0071] In some embodiments, an object that is deemed to be misoriented may be correctly oriented before or after it leaves the feeder system, the orienting system being, for example, a robot or other equivalent means.

[0072] In some embodiments, an object deemed defective may be discarded from a production batch. For example, discarding it from a production batch can be accomplished by a jet of air diverting the object from the production stream and ejecting it into a reject bin. The object is either "fixed" by removing the defect, allowing it to be introduced into the production batch, or it is simply rejected. [Brief explanation of the drawings]

[0073] 1 to 7 are used to explain the present invention. [Figure 1] FIG. 1 illustrates an example of an object being fed into a bowl. [Figure 2] FIG. 2 illustrates the primary images acquired during the learning phase. [Figure 3] FIG. 3 illustrates the step of cutting a primary image into secondary images. [Figure 4] FIG. 4 illustrates the learning phase, in particular the formation of batches of secondary images to obtain a compression-decompression model for each batch. [Figure 5] FIG. 5 illustrates the use of the compression / decompression model in the generation phase. [Figure 6] FIG. 6 illustrates in block diagram form the main steps of the learning phase. [Figure 7] FIG. 7 illustrates in block diagram form the main steps of the generation phase.

[0074] Detailed Description of the Invention

[0075] FIG. 1 shows an object 1 being fed into a bowl at high throughput. To illustrate the invention and facilitate understanding, three decorative patterns are displayed on the object as a non-limiting example. The invention allows for monitoring the orientation of the object and the quality of these patterns on the fed object. The invention allows for the distribution and inspection of oriented objects at high production throughput. The invention allows for rapid changes in the fed object while reducing adjustment times. In the example shown in FIG. 1, the object may be considered a single part. The object may be made, for example, of plastic material, metal, wood, glass, or based on any other material or combination of these materials.

[0076] FIG. 2 illustrates an example of primary images of objects acquired during the learning phase. During the learning phase, N objects that are determined to be correctly oriented and of acceptable quality are fed by the bowl. To facilitate the explanation of the present invention, only four objects are shown in FIG. 2 as an example. To obtain a robust model, the number of objects required during the learning phase is greater than 10 (i.e., N>10), and preferably greater than 50 (i.e., N>50). Of course, these values are non-limiting examples, and N may be less than or equal to 10. FIG. 2 shows three primary images A1, A2, and A3, each representing a different pattern printed on the object. In the explanation of the present invention, A k denotes the primary image of the object, where the image index k varies between 1 and K, where K corresponds to the number of images per object.

[0077] As shown in Fig. 2, the primary images Ak are not necessarily of the same size. In Fig. 2, primary image A2 is smaller than primary image A1 and primary image A3. This allows, for example, to obtain a clearer (more pixelated) image A2. The primary images may constitute the entire surface of object 1 or, conversely, only partially cover its surface.

[0078] As shown in Figure 2, primary image A k targets specific areas of the object. This flexibility of the present invention allows for optimization of computation time while maintaining a highly accurate inspection of visual quality in the most critical areas, at the level of size, such as the location and number of primary images.

[0079] Figure 3 illustrates the division of a primary image into secondary images. Thus, as shown in Figure 3, a primary image A1 is divided into four secondary images S 1,1 , S 1,2 , S 1,3 , S 1,4 Therefore, each primary image A k The split index p is 1 and P k P changes between k Secondary image S k,p is decomposed into

[0080] As shown in Figure 3, the secondary images are not necessarily the same size. 1,2 and S 1,3 is the secondary image S 1、1 and S 1,4 This shows that the secondary image S 1,2 , S 1,3 This allows for more accurate detection of defects within the

[0081] As also illustrated in FIG. 3, the secondary image is not necessarily the same as the primary image A. k For example, secondary image S 2,p only partially covers the primary image A2. By reducing the size of the secondary image, the analysis is focused on a precise area of the object. Only the area of the object that is covered by the secondary image is analyzed.

[0082] FIG. 3 shows how the present invention can be applied to a secondary image S k,p This shows the fact that adjusting the number, size and position of the vertices allows for local adjustment of the observation area of the object.

[0083] FIG. 4 illustrates the learning phase, in particular the formation of batches of secondary images to obtain a compression-decompression model with compression coefficients for each batch.

[0084] Figure 4 shows how N similar secondary images S are used to form a batch. k,p Each batch is processed separately and the compression factor Q k,p Using the compression-decompression model F k,p As an example, N=4 secondary images S are generated as illustrated in FIG. 3,3 Using the compression factor Q 3、3 Model F with 3,3 Create a.

[0085] Figure 5 shows the use of the compression-decompression models obtained from the training phase in the production phase. In the production phase, each model F determined during the training phase is k,p Each secondary image S of the object being fed into the bowl is k,p Therefore, each secondary image of the object undergoes a compression-decompression operation using different compression coefficients and models from the training phase. The result of each compression-decompression operation is a reconstructed image that can be compared with the secondary image from which it was derived. k,p and its reconstructed image R k,p Comparing allows the calculation of the reconstruction error, which is used to define the score.

[0086] Figure 5 shows the model F 3,3 and its compression factor Q 3,3 Using the secondary image S 3,3 Reconstructed image from R 3,3 The particular case of obtaining is illustrated as an illustrative example. Figure 6 represents the main steps of the learning phase according to the invention. At the start of the learning phase, N objects that are determined to be correctly oriented and of acceptable quality are delivered by the bowl. A qualitative and / or quantitative determination of said objects can be carried out according to a visual inspection procedure or according to methods and means defined by the user's business. Thus, the number of objects delivered in the learning phase may be equal to or greater than N. The learning phase illustrated in Figure 6 comprises at least the following steps:

[0087] K × N so-called "primary" images are acquired of objects that are determined to be correctly oriented and of good quality during distribution. Each object can be associated with one primary image (K = 1) or several separate primary images (K > 1), depending on the dimensions of the area to be analyzed on the object and the size of the defects desired to be detected. Images can be acquired in a relatively constant light environment using lighting and magnification conditions suitable for industrial situations. Known lighting optimization techniques can be used to prevent environmental reflections or disturbances. Commonly used solutions include tunnels or black boxes that allow for the avoidance of interference with external lighting and / or light with specific wavelengths and / or lighting at grazing angles or indirect lighting. When multiple primary images are acquired of the same object (K > 1), the primary images can be spaced apart, side-by-side, or overlapping. Overlapping primary images is useful when it is desired to avoid cropping defects that may appear between the two images and / or to compensate for information loss on the image edges linked to the image realignment step. These approaches can be equally well combined as a function of the primary image and the information found therein. The image can also be pre-processed by optical or numerical filters, for example to improve the contrast.

[0088] The primary image is then repositioned relative to the reference image. As a general rule, the primary image of any object fed during the learning phase can serve as the reference image. The primary image of the first object fed during the learning phase is preferably used as the reference image. Methods for repositioning the primary image are explained in detail in the remainder of the description.

[0089] Each primary image Ak is then converted into a so-called "secondary" image P k The secondary images can be divided into smaller analysis areas than the primary image. Dividing the image can result in smaller analysis areas than the primary image. Reducing the size of the analysis area can be beneficial if the target area to search for possible defects is known in advance. The secondary images can be spaced apart to leave "unanalyzed" areas between them. This situation can be used, for example, when defects appear in the target area or when defects appear repeatedly and continuously. Reducing the size of the analysis area can reduce calculation time. Alternatively, the secondary images can be overlapped. Overlapping secondary images avoid cutting the defect into two parts if it appears at the junction between two secondary images. Overlapping secondary images are particularly useful when searching for small defects. Finally, the secondary images can be juxtaposed, without any gaps or overlaps. The primary image can be divided into secondary images of identical or various sizes, and the method of relative positioning of the secondary images (spaced, juxtaposed, or overlapping) can also be combined as a function of the defects being searched for.

[0090] In the next step, corresponding secondary images are grouped into a batch. The secondary images obtained from the KxN primary images generate a set of secondary images. Based on this set of secondary images, N corresponding secondary images are generated, i.e., the same secondary image S for each object. k,p This allows for the formation of a batch containing N secondary images S 1,1 are grouped together. N images S 1,2 Similarly for all images S k,p For N images S 1,3The same applies to the following:

[0091] The next step consists in determining a compressed representation for each batch of secondary images. This operation is a key step of the method according to the invention. It involves determining in particular the compression coefficient Q characterizing said batch. k,p A compression-decompression model F with k,p This includes obtaining the Model F k,p is used to monitor the quality of the object during the generation phase. In this way, the secondary image S 1,1 For a batch of 1,1 Model F with 1,1 Similarly, the image S 1,2 For a batch of Model F 1,2 is obtained, and then the image S 1,3 For a batch of Model F 1,3 and so on for image S k,p For each batch of model F k,p is obtained.

[0092] Secondary images k,p The compression factor Q for each batch of k,p The choice of depends on the available computation time and the size of the defects that it is desired to detect.

[0093] At the end of the training phase, a compression factor Q, related to the orientation and visual quality of the produced objects, is calculated. k,p A set of models F k,p is available.

[0094] According to the present invention, Model F k,p and the compression factor Q k,p The results of the learning phase, including the steps, are saved as "templates" that can then be reused during new production runs of the same object. Thus, objects of the same quality can be reproduced later by reusing predefined templates. This also avoids having to repeat the learning phase before starting each production run of the same object.

[0095] According to the present invention, iterative learning can be used during production. Thus, for example, additional (or complementary) learning can be achieved during production using new objects, and images of these objects can be added to the images of the objects initially considered during the learning phase. A new learning phase can be achieved based on a new set of images. Evolutionary learning is particularly suitable when differences in orientation or aesthetics between objects appear during production and are not considered defects. In other words, these objects are considered to be "good" as in the initial learning phase, and this is preferably taken into account. In this situation, iterative learning is necessary to avoid a high rejection rate, including objects with this difference. Iterative learning can be performed in a number of ways, for example, by pooling new images with previously captured images, by restarting learning with newly acquired images, or by retaining only a few initial images with the new images.

[0096] According to the invention, iterative learning is triggered by an indicator linked to the rejection of objects, such as the number of rejects per unit time or the number of rejects per quantity of objects fed, etc. When this indicator exceeds a fixed value, the operator is alerted and decides whether the increasing rejection rate requires further action. -Repetitive learning phase, -Adjustment of the feeding system, - Refusal to batch objects.

[0097] 7 illustrates the main steps of the object generation phase. The generation phase begins after the learning phase, i.e., when the characteristic criteria for "correctly" oriented and "acceptable" quality objects have been defined as described above. The present invention allows for recycling or orienting objects that are deemed to be incorrectly oriented, rejecting objects that are deemed to be defective from a production batch in real time, and avoiding the use of objects that are deemed to be defective if a drift in object quality is observed. The generation phase according to the present invention, as illustrated in FIG. 7, includes at least the following operations: Acquisition of K primary images of the object being fed into the bowl. The images of the object are captured in exactly the same way as the images captured in the training phase. The area, lighting, magnification, and calibration conditions are the same as those used in the training phase. The K images are realigned with respect to a reference image. The goal of the realignment operation is to avoid offsets between the images we want to compare. These offsets are related to variations in the position and orientation of the object during imaging. Next, each primary image A of the generated object k is P k The segmentation is achieved similarly to the segmentation of the images in the training phase. Thus, following this segmentation, a set of secondary images S is generated for each object. k,p is obtained. Next, for each secondary image S k,p is the compression factor Q predefined during the training phase k,p Model F with k,p This operation is performed for each secondary image S k,p For the reconstructed image R k,p In this way, a reconstructed image is obtained for the object being generated that can be compared with a secondary image of said object. From a numerical point of view, the term "reconstruction of a secondary image" does not necessarily mean the acquisition of a new image in the strict sense. The aim is to compare the image of the object being generated with the image obtained during the learning phase using the compression-decompression function and compression coefficients, and only the quantification of the differences between these images is strictly useful. For reasons of computation time, a choice may be made to limit the calculation to numerical objects that are representative of the reconstructed image and sufficient to quantify the differences between the secondary image and the reconstructed image. Model F k,p The use of is particularly advantageous as it allows the above comparison to be achieved in a very short time that is compatible with what is required and production throughput. The reconstruction error can be calculated based on a comparison of the secondary image and the reconstructed secondary image. The preferred method for quantifying this error is to calculate the mean squared error, although other equivalent methods are possible. Therefore, for each object, there are available secondary images and reconstructed images, and consequently, a reconstruction error. Based on this set of reconstruction errors, multiple scores can be defined for the generated object. Multiple calculation methods are possible for calculating an object's score, characterizing its similarity or dissimilarity from the training batch. Thus, according to the present invention, an object that is visually very different from the training batch due to a different orientation or defects will have one or more high scores. In contrast, an object that is visually very similar to the training batch will have one or more low scores and will be considered correctly oriented and of good quality (or acceptable quality). A third method for calculating an object's score involves taking the maximum value of the reconstruction errors. Other methods involve combining the reconstruction errors to calculate the value of the object's score. The next step is to recycle or correctly orient the "misoriented" objects and discard the defective objects from the production batch. If the value of one or more of the object's scores is below one or more predefined limits, the evaluated object complies with the orientation and visual quality criteria defined during the learning phase and the object remains in the production stream. Conversely, if the value or values of one or more of the object's scores are greater than said one or more limits, the object is either recycled (or oriented) into the feeder system because its orientation is outside the acceptable range, or discarded from the production stream because the object is defective and would not be useful to recycle.

[0098] Several methods can be used to distinguish misoriented objects from defective objects. The first method involves using the score values to distinguish misoriented objects from defective objects. Another method uses multiple scores to distinguish misoriented objects from defective objects. According to another method, the images are realigned and at least one score is used to distinguish misoriented objects from defective objects. According to another method, interest points and descriptors and at least one score are used to distinguish misoriented objects from defective objects.

[0099] The steps of the present invention are described in more detail below. Primary image repositioning

[0100] The method according to the present invention for rearranging an image comprises two steps. - Search for points of interest and descriptors in images. - Based on interest points and descriptors, the captured image is realigned with respect to a reference image to estimate the object orientation.

[0101] As described in this application, the reference image is typically defined for the first image or another image captured during the learning phase. The first step involves defining image interest points and associated descriptors. Interest points may be, for example, angular portions at the level of shapes present in the image. Furthermore, they may be areas of high contrast or color, or points of interest may be selected randomly. The identified interest points are then characterized by descriptors that define their characteristics.

[0102] Preferably, the points of interest are determined automatically using a suitable algorithm, although alternative methods involve arbitrarily predefining the locations of the points of interest.

[0103] The number of interest points used for reordering depends on the number of pixels per interest point. The total number of pixels used for reordering is generally between 100 and 10,000, preferably between 500 and 1,000.

[0104] A first method for defining interest points involves randomly selecting these points. This results in randomly defining a percentage of pixels called interest points, and the descriptors are properties of said pixels (position, color). This first method is particularly suitable for industrial production situations, especially for high-throughput production processes where the time available for computation is very short.

[0105] According to a first embodiment of the first method, the interest points are randomly distributed in the image.

[0106] According to a second embodiment of the first method, the points of interest are randomly distributed within a predetermined area of the image. This second embodiment is advantageous when it is known in advance where defects will appear.

[0107] A second method for defining interest points is based on the so-called "Scale Invariant Feature Transform (SIFT) method" (see U.S. Pat. No. 6,711,293), a method that allows preserving the same visual characteristics of an image regardless of scale. This method involves calculating image descriptors at interest points of the image. These descriptors correspond to numerical information derived from a local analysis of the image that characterizes the visual content of the image regardless of scale. The principle of this method involves detecting defined image areas around the interest points, preferably circular areas with a radius called the scale factor. In each of these areas, shapes and their contours are searched for, after which the local orientation of the contours is defined. Numerically, these local orientations result in vectors that constitute the SIFT descriptor of the interest point.

[0108] A third method for defining interest points is based on the "Speeded Up Robust Features ("SURF") method (see U.S. Patent Application Publication No. 2009 / 0238460), an accelerated method for defining interest points and descriptors. This method is similar to the SIFT method but has the advantage of being faster to execute. Like the SIFT method, this method involves extracting interest points and computing descriptors. The SURF method uses fast exact multiplication with a Hessian to detect interest points and an approximation of a Haar wavelet to compute the descriptors.

[0109] A fourth method for finding interest points based on features from the Feature Accelerated Segment Test ("FAST") method involves identifying potential interest points and then analyzing the intensity of pixels located around the interest points. This method allows for very rapid identification of interest points. Descriptors can be identified using the Binary Robust Independent Basic Features ("BRIEF") method.

[0110] The second step of the image realignment method involves comparing the primary image with the reference image using interest points and their descriptors. The best realignment is achieved by finding the best alignment between the descriptors of the two images.

[0111] In this example, the image may need to be repositioned about only one axis, or about two perpendicular axes, or rotated about an axis perpendicular to the plane formed by the image.

[0112] The image repositioning may be the result of a combination of translational and rotational movements. The optimal homographic transformation was searched for using a least squares method.

[0113] The interest points and descriptors are used in the image relocation operation. These descriptors may be, for example, pixel properties or SIFT, SURF, or BRIEF descriptors. The interest points and descriptors are used as marker points for relocating the image.

[0114] In the SIFT, SURF, and BRIEF methods, the realignment is performed by comparing descriptors. Inappropriate descriptors are discarded using a consensus method, for example, the Ransac algorithm. The optimal homographic transformation is then searched for using a least-squares method. Splitting a primary image into secondary images

[0115] The primary image can be divided into P secondary images in several ways.

[0116] An advantage of the present invention is that the visual analysis level can be adjusted to suit the observed area of the object. This adjustment is performed as a first step based on the number of primary images and the level of resolution of each primary image. Then, by decomposing into secondary images, the analysis level can be adjusted locally within each primary image. The first parameter that can be manipulated is the size of the secondary images. Smaller secondary images allow for local refinement of the analysis. Each secondary image S k,p size and compression factor Q k,p By simultaneously adjusting the detection levels, the present invention allows for optimization of computation time while maintaining a high performance detection level adjusted to meet the level requirements associated with the object being delivered. The present invention allows for the detection level to be locally adapted to the critical level of the observed area.

[0117] One particular example of the present invention involves all secondary images being the same size.

[0118] Thus, if all of the observed areas are equally important, a first method involves dividing the primary image into P secondary images of equal size that are juxtaposed without overlapping.

[0119] The second method involves dividing the primary image into P overlapping, juxtaposed secondary images of the same size, with the overlap adjusted as a function of the size of defects likely to appear on the object.

[0120] The smaller the defect, the smaller the overlap may be. Generally, the overlap is considered to be at least equal to half the characteristic length of the defect, the characteristic length being defined as the smallest diameter of a circle that can contain the entire defect.

[0121] Of course, it is also possible to combine these methods and use juxtaposed and / or overlapping and / or separate secondary images. Calculation of compression-decompression functions

[0122] According to a first method, which is also preferred, the compression-decompression functions and compression coefficients are determined based on principal component analysis (PCA). This method allows the definition of eigenvalues and vectors that characterize the batch resulting from the learning phase. In the new base, the eigenvectors are sorted in order of size. The compression coefficient comes from the number of dimensions retained in the new base. The higher the compression ratio, the fewer the number of dimensions in the new base. The invention allows the compression coefficient to be adjusted depending on the level of inspection required and the available computation time.

[0123] The first advantage of this method is related to the fact that the machine does not need instructions to define a new base: the eigenvectors are selected automatically by calculation.

[0124] The second advantage of this method is related to the reduced computation time for defect detection during the generation phase: the number of dimensions is reduced, and therefore the amount of data to be processed is reduced.

[0125] A third advantage of this method is that the generated image of the object can be assigned one or more scores in real time, which, when reconstructed using the model from the training phase, allows for quantification of the deviation / error level of the object being fed in the bowl relative to the object from the training phase.

[0126] The compression ratio is between 5 and 500,000, preferably between 100 and 10,000. The higher the compression ratio, the shorter the calculation time required to analyze the image during the generation phase. However, if the compression factor is too high, the model may be too coarse and unsuitable for detecting errors.

[0127] According to the second method, the model is an autoencoder. An autoencoder takes the form of a neural network that allows defining characteristics in an unsupervised way. An autoencoder consists of two parts: an encoder and a decoder. The encoder receives a secondary image S k,p and the decoder compresses the reconstructed image R k,p This allows you to obtain

[0128] According to the second method, an autoencoder is available for each batch of secondary images, each with its own compression factor.

[0129] According to the second method, the autoencoder is optimized during a training phase by comparing the reconstructed image with the initial image. This comparison allows the difference between the initial image and the reconstructed image to be quantified, thereby determining the encoder error. The training phase allows the optimization of the autoencoder by minimizing the image reconstruction error.

[0130] According to the third method, the model is based on the "Orthogonal Matching Pursuit (OMP) algorithm." This method involves searching for the best linear combination based on the orthogonal projections of several images selected in the library. The model is obtained by an iterative method. The reconstructed image improves with each image added from the library.

[0131] According to a third method, an image library is defined by a training phase, which is obtained by selecting a number of images from the training phase that are representative of the set of images. Computing the reconstructed image from the compression-decompression model

[0132] In the generation phase, each primary image Ak of the object to be inspected is repositioned using the method described above, and then P k S secondary images k,p Each secondary image S k,p undergoes a numerical reconstruction operation using the model defined in the training phase. Thus, at the end of the reconstruction operation, each secondary image S k,p Reconstructed image R available for k,p exists.

[0133] Compression factor Q k,p Model F with k,p Using each secondary image S k,p The reconstructing operation allows for very short computation times. k,p is 5 or more and 500,000 or less, preferably 10 or more and 10,000 or less.

[0134] Following the PCA method, which is also the preferred method, the secondary image S k,p is first converted to a vector. Then, this vector is applied to the function F defined during the training phase. k,p Then, the resulting vectors are transformed into an image, resulting in a reconstructed image R k,p is obtained.

[0135] According to the second method, the secondary image is reconstructed by an autoencoder with parameters defined in a training phase. k,p is the reconstructed image R k,p is processed by an autoencoder to obtain

[0136] According to a third method, the secondary images are reconstructed using the Orthogonal Matching Pursuit (OMP) algorithm, whose parameters are defined during a learning phase. Calculating the reconstruction error for each secondary image

[0137] The reconstruction error is the secondary image S k,p and the reconstructed image R k,pis obtained by comparing with

[0138] One method used to calculate the error is to use the secondary image S k,p and the reconstructed image R k,p The preferred method used to calculate the reconstruction error is the Euclidean distance or 2-norm method, which considers the square root of the sum of the squares of the errors.

[0139] Another method for calculating the error involves using the Minkowski distance, which is a generalization of Euclidean distance, p-distance. This method considers the pth root of the pth power of the sum of the absolute values of the errors. In this method, choosing a value of p greater than 2 assigns more weight to larger differences.

[0140] Another alternative is the Tchebichev or 3-norm method.

[0141] In this method, the maximum absolute value of the error is considered. Calculating the score

[0142] The object score value is obtained from the reconstruction error of each secondary image.

[0143] A preferred method involves assigning the maximum value of the reconstruction error to the score.

[0144] Another method involves calculating the score value by taking the average of the reconstruction errors.

[0145] Another alternative involves taking a weighted average of the reconstruction errors, which may be useful when the importance of defects is not the same in all areas of the object.

[0146] Alternative methods include using Euclidean distance or the 2-norm.

[0147] Another method involves using the p-distance.

[0148] Alternative methods include using Chebychev intervals or the 3-norm.

[0149] Other equivalent methods are of course possible in the context of the present invention.

[0150] Once the score or scores are calculated, the values are used to determine whether the object in question meets the required quality and orientation criteria. If so, it is retained in the feed stream. If the score does not meet the criteria because the object's orientation is outside of the acceptable range, the object is recycled or reoriented within the feed system. If the score does not meet the criteria because the object is defective, the object is discarded from the feed process.

[0151] Misoriented objects can be distinguished from defective objects based on the value of the score. So, for example, for a misoriented (upside down) cap, the score will vary from 7 to 10, while a defective cap will produce a score of 3 to 5. Therefore, an upside down cap can be easily distinguished from a defective cap.

[0152] In other cases, it is proposed to use multiple scores to distinguish defective objects from misoriented objects. In particular, the invention makes it possible to define scores for local regions of off-center objects. For example, consider an object containing an off-center orifice. A local image of the orifice allows for a score related to the object's orientation. Thus, by combining the orifice indentation with other indentations, it becomes possible to separate poorly oriented objects from defective objects.

[0153] According to another method, information about the repositioning of the objects and one or more scores are used to distinguish misoriented objects from defective objects.

[0154] According to another method, interest points and descriptors are used in conjunction with at least one score to distinguish misoriented objects from faulty objects.

[0155] Misoriented objects are preferably recycled within the bowl. One method involves blowing the components into the bowl with at least one air jet in the object's trajectory. Another method involves mechanically pushing the components into the bowl with a piston and cylinder. This system allows for recycling of objects by air jet or mechanical actuation.

[0156] In other embodiments, the orientation of a misoriented object is corrected before or after the object leaves the bowl. Many object-orientation systems are envisioned and can be associated with the present invention. These systems can include one or more axes as a function of the complexity of the orienting movement to be performed. The orienting system can, for example, be a robot.

[0157] It should be clearly understood that in this example, the method is implemented in a feeding system (such as a vibratory bowl or centrifugal bowl) capable of having a high throughput (e.g., at least 100 products per minute). In the examples, if the singular form is used to define the object being produced, it is for the sake of simplicity. In fact, the method is applied to successive objects in the production feeder. The method is therefore iterative and recursive for each successive object being fed, and the direction and quality are checked for all said successive objects.

[0158] The described embodiments are given as illustrative examples and should not be considered as limiting the invention. Other embodiments may, for example, rely on equivalent means to those described. Several embodiments may be equally combined with each other as a function of the circumstances or means and / or steps of the method used in one embodiment and may be used in another embodiment of the invention.

Claims

1. 1. A method for feeding oriented objects, e.g., packaging components such as tube tops or caps, through a feeder bowl, such as a vibratory bowl or centrifugal bowl, comprising: The method includes at least one orientation and quality inspection step integrated into the feeding method that is performed continuously during production, the inspection step being based on images of the object captured during feeding and using artificial intelligence algorithms; The inspection step comprises a learning phase allowing the definition of acceptable tolerances for the orientation and quality of the objects, and a production phase during which only objects whose orientation and quality are within the acceptable tolerances are delivered, the learning phase comprising at least: -) generating N objects deemed to have orientation and quality within acceptable tolerances; -) at least one reference primary image (A ｋ ) capturing the -) each reference primary image (A ｋ ) to (P ｋ ) reference secondary images (S ｋ，ｐ ) -) grouping corresponding reference secondary images into batches of N images; -) Compression factor per batch (Q ｋ，ｐ ) is used to calculate the compression-decompression model (F ｋ，ｐ ) determining Including, The generation phase includes at least: -) capturing at least one primary image of at least one object to be generated; -) each primary image into a secondary image (S ｋ，ｐ ) -) Apply the compression-decompression model and the compression coefficients defined in the learning phase to each secondary image (S ｋ，ｐ ) to obtain the reconstructed secondary image (R ｋ，ｐ ) forming a -) each reconstructed secondary image R ｋ，ｐ calculating the reconstruction error of -) assigning one or more scores to each object based on the reconstruction error; -) determining whether the object being fed has passed the inspection of its orientation and its quality based on the assigned score or scores; A method comprising:

2. 10. The method of claim 1, wherein if the object is deemed to be misoriented, the object is either oriented to be within an acceptable tolerance or reused in the feeder bowl.

3. The method of any one of claims 1 to 2, wherein if the object is deemed to be defective, the object is discarded from the production batch.

4. The method of any one of claims 1 to 3, wherein the value of the score is used to distinguish correctly oriented objects from faulty objects.

5. The method of any one of claims 1 to 4, wherein multiple scores are used to distinguish correctly oriented objects from defective objects.

6. a plurality of analyses are performed on at least one of the initially captured primary images; A method according to any one of claims 1 to 5, wherein the multiple analyses produce "daughter" primary images which are used in place of images originally captured at their source.

7. The method according to any one of claims 1 to 6, wherein after the step of acquiring at least one primary image, each primary image is repositioned.

8. A method according to any one of claims 1 to 7, wherein each primary image is processed using filters and / or contour detection and / or application of masks in order to hide certain areas of the image.

9. The score corresponds to the maximum value of the reconstruction error and / or the mean value of the reconstruction error and / or the weighted mean value of the reconstruction error and / or the Euclidean interval and / or the p-interval and / or the Tchebichev interval, the interval being a function of the distance between the secondary images S ｋ，ｐ and the reconstructed image R ｋ，ｐ The method according to any one of claims 1 to 8, wherein

10. The method of any one of claims 1 to 9, wherein at least two primary images are captured, said primary images being of the same size or of different sizes.

11. 11. The method according to any one of claims 1 to 10, wherein each primary image is divided into P secondary images S of the same or different size, the secondary images S being juxtaposed with overlapping and / or without overlapping.

12. 12. The method according to any one of claims 1 to 11, wherein the learning phase is iterative and is repeated during production with objects fed in to take into account any differences that are considered to be acceptable orientation or quality defects.

13. A rearrangement step is performed, the rearrangement step comprising: considering a predetermined number of interest points and descriptors distributed across the image; determining a relative motion between the reference image and the primary image that minimizes registration errors at the interest point and descriptor level; Including, 13. The method according to any one of claims 1 to 12, wherein the interest points are randomly distributed within the image or within a predetermined area of the image, the locations of the interest points being predetermined, arbitrarily or otherwise.

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

15. 15. The method of any one of claims 1 to 14, wherein the image reorientation and at least one score are used to distinguish misoriented objects from faulty objects, or the interest points and descriptors and at least one score are used to distinguish misoriented objects from faulty objects.

Images