CONTAINER HANDLING MACHINE AND METHOD FOR ALIGNING A CONTAINER IN A CONTAINER RECEPTION OF A CONTAINER HANDLING MACHINE

Description

[0001] The present invention relates to a container treatment machine according to claim 1 and to a method for aligning a container in a container holder of a container treatment machine according to claim 7. Stand the Technology

[0002] It is known from the prior art that containers, for example after they leave a cleaning machine or a blow molding machine and are fed to a machine in which they must arrive with a defined orientation, are aligned before they can be fed to the actual treatment in the downstream container treatment machine.

[0003] The desired orientation, also called target position, may depend, for example, on certain physical features of the surface of the container, such as mold seams, notches or bulges or similar.

[0004] EP 2 251 269 and DE 10 2011 007 520 disclose methods for aligning a container by capturing a full rotation of the container with a camera, or at least during a rotation long enough to detect a feature. Based on a detected feature, the container is rotated so that it can be moved into the target position. However, this requires precise knowledge of the container's surface and the detection of such a specific feature.

[0005] Furthermore, the specific target position must be specified by the user. This requires considerable user experience, as even a slightly incorrect specification of the actual target position relative to, for example, a print head can significantly impair the print result.

[0006] Furthermore, the method specified in the documents described is not suitable for drawing conclusions for future containers from a series of containers already examined and aligned, or even for taking into account alignment features not yet processed, so that a complete examination of the container is always necessary.

[0007] This ensures that the efficiency of the process in the known state of the art can only be increased up to a limit essentially determined by the skills of the operator.

[0008] DE 10 2008 059 229 A1 discloses a method for aligning containers using a neural network. Task

[0009] Based on the known state of the art, the problem to be solved is to specify a method and a container treatment machine for aligning a container before carrying out a treatment step on the container, which can realize the alignment of the container with high reliability and low effort for the operator, wherein the method also reacts as reliably as possible to modifications of alignment features. Solution

[0010] This object is achieved according to the invention by the container treatment machine according to claim 1 and the method for aligning a container in a container holder of a container treatment machine according to claim 7. Advantageous developments of the invention are covered in the subclaims.

[0011] The container treatment machine according to the invention for treating containers such as bottles, cans or the like is characterized in that the alignment module comprises a neural network which, by processing the image of a container transported in a container holder upstream of the treatment unit, can determine a necessary rotation of the container from its current position to the target position and the alignment module can control the rotation of the container holder depending on the determined rotation.

[0012] According to the invention, the target position is the orientation of the container under which the container is to be treated. The treatment of the container can, for example, be the application of a decorative element or an inspection process or similar, whereby the treatment essentially depends on the correct orientation of the container. The target position can be defined in particular by an alignment feature on the surface of the container in the form of a physical characteristic. For example, it can be provided that a label is wound around the container starting at a forming seam of the container. For this purpose, however, the container with the forming seam must be moved into a position suitable for the labeling unit, so that the suitable position of the forming seam defines the target position.

[0013] The current position of the container is the position that the container occupies before its alignment takes place and before it is subjected to treatment in the treatment unit, so that it is possible to align the container from its current position to the target position before the start of treatment.

[0014] This results in the necessary rotation as the difference angle between the current orientation and the target position.

[0015] According to the invention, the neural network is preferably able to compare the current position with the target position based on pattern recognition, for example by detecting certain surface characteristics from the neural network and translating them into a detection of a current position compared with the target position.

[0016] The difference angle determined by the neural network and / or the alignment module can then be used to control the rotation of the container holder by the alignment module, whereby the detection of the necessary rotation by the neural network is robust even with small deviations in the shape / color / size of the containers and requires little additional input from the operator.

[0017] The container processing machine, and in particular the use of a neural network, can also be advantageously used to determine a necessary rotation when the containers to be rotated have embossed elements in the material, such as coats of arms, symbols, logos, or the like embossed into glass or plastic, especially those with a "handcrafted" appearance, i.e., those that look as if they were manually embossed. Due to the transparency of the container material and the inconsistent shape of such embossing, these are difficult and unreliable to identify using conventional image processing programs.

[0018] By using a neural network with the properties according to the invention, the necessary rotation of the container can be determined, for example, with regard to such an embossing or based on such an embossing. Regardless of the rotation, the neural network can detect the embossing on the container in order to obtain further information, for example, about the quality of the embossing, regardless of the rotation to be performed.

[0019] Furthermore, the container treatment machine according to the invention, through the use of a learning neural network, can take into account slightly changing container characteristics over time, such as the changing shape of the mold seams due to wear on the blow molds. A particular advantage of neural networks is that they can learn over time and thus flexibly and reliably use changing container properties (e.g., due to wear within a certain tolerance limit) to determine the necessary rotation.

[0020] The neural network can be a deep neural network (DNN) or a convolutional neural network. Deep neural networks, and especially convolutional neural networks, are particularly well suited for pattern recognition of images and / or surfaces, so they can be used particularly well to detect the current positions, in particular one or more alignment features on the surface of the container, to determine the necessary rotation, which further improves the alignment of the container to its target position.

[0021] Furthermore, the neural network is designed to learn current positions of containers in relation to a target position from images obtained during operation of the container handling machine.

[0022] For example, if the rotation performed did not result in the correct orientation of the container, the neural network can learn from an initial attempt to determine the orientation of the container and the necessary rotation. For example, if a second image of the container is taken after the rotation, the neural network examines whether the target position has been reached. If this is not the case and the target position needs to be corrected, appropriate design of the neural network can lead to a learning process that continuously improves the detection of the current position and the determination of the necessary rotation as the machine operates.

[0023] In one embodiment, the target position is determined based on an alignment feature of the container.

[0024] As already described above, the alignment feature may in particular be a physical characteristic, such as a molded seam or a material bulge or indentation, which may be decisive for the further treatment of the container, for example the application of a label or a printed image.

[0025] In this way, the objective can be determined in a simple geometric way.

[0026] In a further development of this embodiment, the container treatment machine comprises an input unit with which an operator can enter a container type and / or a type of alignment feature, based on which the neural network can determine the target position.

[0027] This embodiment allows an operator to easily determine the desired rotation result, such as the target position, without requiring precise input from the operator or even specifying a precise angular position relative to a preferred zero position of the container. For example, the input can be configured so that the operator can select between different types of features, such as a forming seam, material bulge, material indentation, notches, or points. Simply entering this term as an alignment feature is sufficient to provide the container handling machine with the necessary information, which is then essentially automatically converted by the alignment module and the neural network into correct container alignment.

[0028] Furthermore, the container handling machine can comprise at least one of a labeling machine, a printing machine, a direct printing machine, an inspection machine, or a packer. These machines typically require the alignment of the container in a specific target position, so the implementation of the invention in these machines is particularly advantageous.

[0029] It can also be provided that the container holder comprises a turntable and a (rotatable) centering bell, whereby a container can be clamped and rotated between the turntable and the (rotatable) centering bell. This design of the container holder enables effective and control-friendly rotation of the containers.

[0030] It can also be provided that the container is transported by means of a centering bell (e.g., a centering bell that can be rotated by a servo motor) that rotates around an axis parallel to the container's longitudinal axis and is rotated in front of the camera(s). The containers can thus be transported suspended, which also leaves the floor free for camera viewing without obstruction.

[0031] The method according to the invention for aligning a container in a container holder of a container treatment machine, wherein the container is aligned in a target position, is characterized in that the alignment module comprises a neural network which, by processing the image of the container transported in a container holder upstream of the treatment unit, determines a necessary rotation of the container from its current position to the target position and the alignment module controls the rotation of the container holder depending on the determined rotation.

[0032] According to the method according to the invention, the container is rotated into the target position before the treatment takes place in the treatment unit. This does not necessarily mean that the container is rotated into the target position before reaching the treatment unit. The container can also be rotated while the container is already positioned within the effective range of the treatment unit, but the treatment unit has not yet begun the treatment.

[0033] This method provides a simple yet reliable alignment of the container.

[0034] In one embodiment, the neural network is a pre-trained neural network.

[0035] By pre-learning the neural network, i.e. training the neural network on one or more alignment features and target positions of containers or container types, it is ensured that even when the container handling machine is put into operation, an essentially completely correct alignment of the containers into the target position is possible starting from any position.

[0036] Furthermore, the neural network can be designed to learn current positions of containers in relation to a target position from images of containers taken during operation of the container treatment machine.

[0037] Even with a pre-trained neural network, this embodiment can lead to further improvement in pattern recognition of the container's surface and thus in further processing, such as determining the necessary rotation, over the operating period. This can further minimize errors.

[0038] In one embodiment, the container handling machine comprises an input unit with which an operator enters a container type and / or a type of alignment feature, and wherein the neural network determines the target position based on the entered container type and / or the entered alignment feature.

[0039] This embodiment provides a way to equip the container handling machine, and in particular the alignment module and the neural network, with the necessary information to determine the target position and the necessary rotation in a manner that is less prone to errors.

[0040] In one embodiment, it is provided that the neural network for determining the rotation in exactly one image of the container in its current position, in a first step, searches for an alignment feature on the basis of which the target position of the container is defined, and, if the alignment feature is at least partially found in the image of the container, determines the rotation from the position of the alignment feature in the image and the target position of the alignment feature;and wherein in a second step, if the alignment feature is not at least partially found in exactly one image, the neural network determines a possible position of the alignment feature based on the information present in the image and the alignment module causes a rotation of the container in the container receptacle based on the possible position of the alignment feature, wherein in a third step, a second image of a container is recorded in the rotated position and the neural network searches for the alignment feature in the second image and, if the alignment feature is at least partially found in the image of the container, determines the rotation from the position of the alignment feature in the image and the target position of the alignment feature;

[0041] This embodiment enables "self-correction" of the neural network, as the neural network can check the result it determined in the second and third steps and, for example, correct the first result in a subsequent first step. In combination with a learning process of the neural network, this can advantageously contribute to improving the container alignment process.

[0042] In a further development of this embodiment, the neural network learns from the results of the second and third steps. This realizes the advantage discussed above. Short description of the characters

[0043] Fig. 1 shows a container handling machine according to an embodiment. Fig. 2 shows the processing of a captured image of a container to determine the angle of rotation according to an embodiment. Fig. 3 shows an embodiment of a training method for a neural network. Detailed description

[0044] Fig. 1 shows a container treatment machine 100 according to one embodiment of the invention. This container treatment machine can be designed, for example, as a labeling machine or direct printing machine. However, the invention is not limited with regard to the specific choice of the container treatment machine 100 or its design. It is only intended that the container treatment machine be a container treatment machine that requires a specific alignment of a container before it is treated with a treatment unit 104.

[0045] For example, for applying labels using a treatment unit 104 designed as a labeling unit, it may be envisaged that one side of the label is applied to a formed seam of a plastic container, such as PET, and the remainder of the label is then at least partially wrapped around the container. However, the containers are usually fed to such a labeling machine unaligned, so subsequent alignment is required before the label is applied.

[0046] This also applies to other implementations of container handling machines, including inspection machines.

[0047] In principle, the container treatment machines are provided with a container holder 102 in addition to a treatment unit. This container holder allows containers to be removed from a feed device 105, for example, and transferred, in particular, into the active area of the treatment unit 104, where the treatment of the container (for example, applying a label or printing a print image on the container) is carried out.

[0048] According to one embodiment, it is provided that the container is rotated into the target position (i.e. into the desired position of the container) at least before the start of the actual treatment step with the treatment unit 104, wherein it is preferably rotated about its longitudinal axis.

[0049] The longitudinal axis is the axis along the longest dimension of the container and extends, in particular, from a base of the container, on which the container usually stands, toward the opening of the container. The container receptacle can, for example, comprise a turntable and a centering bell associated with it, between which the container is clamped. This is particularly known for plastic bottles or cans. However, the invention is not limited in this regard, and other variants for transporting and / or rotating the container, for example, with neck-handling devices that can turn the container over on its support ring, are also conceivable.

[0050] In order to move the container to the desired target position, the current orientation of the container must be known so that the distance the container needs to be rotated around its axis can be determined. For this purpose, the container handling machine comprises a recording device, in particular a camera 103, with which at least one image of the container can be recorded while it is in its current position or orientation.

[0051] This captured image is then sent by the camera to an alignment module 130. This alignment module can be understood as a control unit or at least comprise a control unit and is designed to process the captured image of the container in order to determine the necessary rotation of the container from its current position to the target position. Furthermore, the alignment module is preferably designed to control the rotation of the container holder depending on the thus determined necessary rotation, for example, by controlling a servo drive of the container holder and causing the container to rotate in the container holder by a specific angle.

[0052] According to the invention, the alignment module comprises a neural network that processes the image captured by the camera 103 (or a somehow pre-processed image, as described below) and, through this processing, in particular pattern recognition, detects the current position of the container and, on the basis thereof, determines the necessary rotation.

[0053] According to one embodiment, the necessary rotation is ultimately determined by the position of a specific alignment feature on the surface of the container in its current position compared to the position of this alignment feature in the target position.

[0054] According to the invention, the neural network is trained in such a way that it recognizes the current position of the alignment feature or another structure of the container that allows conclusions to be drawn about the current position of the alignment feature by means of image recognition or pattern recognition and derives from this the necessary rotation of the container in order to bring the alignment feature into the desired position.

[0055] The alignment feature can, as already mentioned above, be a molded seam. However, other physical characteristics of the container, such as the position of material bulges or indentations, embossings, certain markings (which, for example, were already applied in a previous step using digital printing techniques), or the like, are also conceivable. The invention is not limited in this regard. However, it must fundamentally be possible to determine the current position of the container based on physical properties of the container by taking an image of the container and comparing it with a target position.

[0056] Once the required rotation or angle of rotation has been determined to rotate the container from its current position to the target position, in which the alignment feature has the desired orientation, the alignment module, as already mentioned above, controls the relevant container holder or a suitable device for rotating the container such that the container is moved from its current position to the target position. Treatment by treatment unit 104 can then take place at this point.

[0057] In the Fig. 1An operating terminal 107 is also shown. This operating terminal allows the operator to interact with the container handling machine and can, for example, serve as a control unit for the entire container handling machine. While the operating terminal 107 is shown here separately from the alignment module 130, it can also be provided that the alignment module is integrated, in particular together with the neural network, in the control unit or the operating terminal 107. Via the operating terminal 107, an operator can preferably communicate to the container handling machine and in particular to the alignment module information regarding the container type and / or a type of alignment feature or a physical characteristic, depending on which the neural network should determine the current position and the target position.This can be done by selecting alignment features and / or container types from a list, making it easy for the operator to do so, thus reducing the risk of errors. It may also be possible for the operator to select new containers as the container type that have not previously been processed by the neural network or the alignment module. In this case, a training process can be initiated in which, for example, new alignment features or the general container type are learned by the neural network, as described with reference to [Figure 1]. Fig. 3 described.

[0058] In the Fig. 1In the embodiment shown, only one camera 103 is provided. The container treatment machine shown here is designed as a carousel 101 and equipped with container receptacles arranged along its periphery. In order to realize active learning of the neural network even during operation of the container treatment machine, it can be provided that another camera is arranged along the direction of rotation of the carousel, downstream of the camera 103 but upstream of the treatment unit 104. This camera preferably takes another image of the container in the "new" current position after the container has been rotated from its current position to the presumed target position, as determined by the neural network.To give the neural network the opportunity to learn during operation of the container handling machine, one embodiment can provide for this new image to be processed again by the neural network of the alignment module 130 and to check whether, for example, a specific alignment feature that should actually be positioned in the target position after the first rotation is actually located in the target position. If this is the case, the neural network can learn from this processing of the second image that the previously determined rotation was correct. If this is not the case, the neural network also learns from this and attempts to rotate the container again to achieve the target position.

[0059] With a pre-trained neural network, it is expected that any deviations from the actual target position that may occur will be so minimal that a slight second rotation will very likely position the container correctly in its target position so that it can be treated accordingly by the treatment unit. Since the learning process usually makes the neural network more and more reliable as time progresses during the operation of the container treatment machine, the additional camera (not shown here) can also be provided only during an initial period, for example, the first week, of operation of the container treatment machine. Alternatively or additionally, such an additional camera can also be provided if a new type of container that has not previously been processed by the neural network is to be treated by the container treatment machine.With a sufficiently pre-trained neural network, it is possible for the neural network to already rotate the new container type essentially correctly when the orientation feature used for identification is specified. To improve the quality here as well, the second camera can be used to enable the neural network to learn.

[0060] The neural network is preferably a deep neural network and particularly preferably a convolutional neural network. These networks are particularly well suited for pattern recognition of images and can therefore be used advantageously for the invention.

[0061] The Fig. 2 shows a merely schematic representation of image processing in a convolutional neural network for determining the current position of a container and deriving a corresponding angle of rotation.

[0062] Fig. 2This process is essentially presented in the form of a flow chart with a schematic representation of the container and the camera, but is fundamentally to be understood as a process flow within the neural network.

[0063] First, the camera 103, as already mentioned with reference to the Fig. 1described, one or more images of the container 231 are taken. Preferably, the container is rotated in front of the camera over an angle of 360° (i.e., a full rotation), and images of the container are taken. In particular, it can be provided that the container is rotated in the recording range of the camera (also called the camera's field of view), in which an image of the container is taken, while several images 232 of the container are taken by the camera. For example, the images can be taken at specific angles of rotation relative to the initial position, for example starting at 0° and then each time after a rotation of 90° (i.e., 90°, 180°, and 270°). Other angles of rotation, such as taking a picture every 45° during the rotation, are also conceivable. Preferably, the container is rotated in front of the camera at least far enough until the identifying feature to be identified can be imaged at least once.

[0064] Preferably, the images are rectified and / or edited using conventional image processing means (for example, the sharpness or contrast is changed) and / or, in one embodiment, can be combined to form a panoramic image.

[0065] These images 232 are then made available in the form of a suitable file, such as an image file 233, to the alignment module and here in particular to the neural network 240, which in the embodiment shown here is designed as a convolutional neural network.

[0066] A convolutional neural network processes images by successively multiplying a matrix representing the image by a smaller matrix, forming the dot product of each. The smaller matrix is often referred to as the "kernel" and is abbreviated K below.

[0067] This can be understood as follows. A matrix M of size S x T serves as the starting point and is multiplied by the matrix K of size P x Q (PCS,QCT). Starting with the first entry of the matrix M, the inner product of a submatrix U (of size P x Q) of the image matrix M with the matrix K is formed. The indices of the initial entries for forming the inner product are then incremented by 1 (for example, only the columns and / or only the rows), and the inner product is determined again using the resulting submatrix. These inner products each result in exactly one number. If this number is represented together with the remaining inner products as a new matrix by using the corresponding indices that were used to determine the inner product, a new, reduced matrix R is obtained, which, compared to the initial size of the original matrix M (S x T), now has the size (S - P+1) x (T - Q+1).Let this be the matrix R, then its entries R ij each have the value of the corresponding inner product resulting from the original matrix or image matrix M and the matrix K.

[0068] The values of the entries in matrix K, as well as the quantities P and Q of matrix K, are ultimately parameters of the neural network and are typically trained using training methods, for example, using previously known images. These parameters can also be trained, as in the previous embodiment, when measuring the already rotated container again.

[0069] A convolutional neural network, as used in Fig. 2 is shown, usually comprises several layers 241-243 (and further layers not shown separately here), each of which performs a corresponding transformation of the matrix input to it with a corresponding matrix K.

[0070] Accordingly, the originally input image 233 is further processed by the layers 241-243 of the neural network, resulting in "intermediate images" 234-236 and, at the end of the process in the convolutional neural network, a final reduced image 237.

[0071] This final image 237 ultimately allows for inference regarding the presence or absence and the exact position of a corresponding alignment feature, the recognition of which the neural network was trained. Within the meaning of the invention, this is an alignment feature or another physical characteristic of the container. The reduced image 237 now comprises a reduced size corresponding to its passage through the layers of the neural network. However, it can also be expanded back to its original size, for example, in order to determine the actual position of the alignment feature or the physical characteristic of the container in the image based on the pattern recognition in the neural network.

[0072] This final image 237 can then be used by the neural network or, more generally, the alignment module to determine the current position of the physical characteristic or alignment feature, from which the necessary rotation of the container can then be derived by comparing it with the target position.

[0073] For this purpose, a virtual shift of the final image 237 (i.e., a rotation of the container in the image) can be performed until the alignment feature's position matches the position of the alignment feature in the target position. The virtual rotation performed is then the required rotation of the container in the container holder to move it from its current position to the target position.

[0074] After the Fig. 2Once the final image 237 has been determined, the current position of the container can be determined, as already described. Once this has been determined, the necessary angle of rotation for the container can also be determined in order to bring it into the target position (or to position the alignment feature in the target position). In a next step, the alignment module then controls the container holder, so that in step 250 the container is rotated from its current position to the target position. As already described, it can also be provided here that a second camera subsequently checks whether the container is actually positioned in the target position, after which a subsequent correction is enabled and, at the same time, further learning of the neural network can be carried out.

[0075] The accuracy with which the necessary rotation can be determined depends largely on the state of the neural network, and in particular, its training. An insufficiently trained neural network will typically not determine the position of the alignment feature with high accuracy and therefore will not correctly determine the required angle of rotation when subsequently determining it. Likewise, an insufficiently trained neural network may mistakenly interpret features on the surface of a container as the alignment feature to be sought, which will also lead to errors in determining the angle.

[0076] For this reason, it is necessary to train the neural network before the container handling machine begins operating. Continued learning during the operation of the container handling machine is also advantageous in order to achieve a continuous improvement in the quality of the neural network's results.

[0077] Learning carried out before the operation of the container handling machine, but ultimately also learning during operation of the container handling machine, basically proceeds in such a way that images of containers are made available to the neural network for processing during a learning phase. These can be images from ongoing operation, for example, but also images taken from a large database that are used to train the neural network (e.g., before the container handling machine is put into operation). In addition to these images, the neural network is informed which alignment feature or physical characteristic on the surface of the container it should search for. In addition, a target position of this alignment feature or physical characteristic can be passed to the neural network, so that the task of the neural network is to determine the necessary angle of rotation.

[0078] This is done in step 301 according to the Fig. 3 .

[0079] The neural network now processes the images provided to it, usually one after the other, and determines the presumed position of the alignment feature for each image and, if necessary, derives the necessary rotation angle. This is done in steps 302 and 303 according to the Fig. 3 Instead of determining the angle of rotation, the neural network's task can also be simply to precisely determine the position of the alignment feature. Instead of determining the angle of rotation in step 303, the result of the neural network's processing will then be the determination of the position of the alignment feature.

[0080] For training a neural network, the desired results (within the scope of the implementation of the Fig. 3The angle of rotation (i.e., the angle of rotation to be determined) must be known in order to compare the results of the neural network with the correct results. Therefore, in a next step 305, the angle of rotation determined by the neural network and the actual angle of rotation 304, which is available as additional input, are compared. This comparison can, for example, involve calculating the difference or a weighted comparison of the determined and actual angles of rotation.

[0081] Based on this comparison, the neural network can now modify its parameters (the size of the matrices K and / or the values of the parameters contained therein) in step 306 as part of a learning process. Using these new parameters, the last processed image or all recently processed images are reprocessed, and the necessary rotation angle is determined. A further comparison is then made in step 305 with the actual rotation angle 304, and the parameters are modified again, if necessary, in step 306.

[0082] This process is typically performed until the deviation between the determined angle of rotation and the actual angle of rotation for all training data falls below a certain threshold. The parameters then obtained are used in step 307 as the final parameters of the neural network after this training, until the next training cycle of the neural network is performed, for example, during a break in operation of the container handling machine.

Claims

1. Container treatment machine for treating containers, such as bottles, cans or the like, the container treatment machine comprising a treatment unit for treating containers and container receptacles in which containers can be received so as to be rotatable about an axis, the container treatment machine comprising a camera for capturing an image of a container transported in a container receptacle upstream of the treatment unit and an alignment module, wherein the alignment module is configured to rotate a container to a target position by controlling the container receptacle, characterized in that the alignment module comprises a neural network which is configured to determine a necessary rotation of the container from its current position to the target position by processing the image of a container transported in a container receptacle upstream of the treatment unit, the alignment module is configured to control the rotation of the container receptacle depending on the determined rotation and the neural network is adapted to learn current positions of containers in relation to a target position from images acquired during operation of the container treatment machine.

2. Container treatment machine of claim 1, wherein the neural network is a Deep Neural Network (DNN) or a Convolutional Neural Network.

3. Container treatment machine according to any one of claims 1 to 2, wherein the target position is determined based on an alignment feature of the container.

4. Container treatment machine according to claim 3, wherein the container treatment machine comprises an input unit for an operator to input a type of container and / or a type of an alignment feature, based on which the neural network enables determination of the target position.

5. Container treatment machine according to any one of claims 1 to 4, wherein the container treatment machine comprises at least one of a labeling machine, a printing machine, a direct printing machine, an inspection machine, a packer.

6. Container treatment machine according to any one of claims 1 to 5, wherein the container receptacle comprises a turntable and a rotatable centering bell, enabling a container to be clamped and rotated between the turntable and the rotatable centering bell.

7. Method for aligning a container in a container receptacle of a container treatment machine, wherein the container is aligned into a target position before a treatment step is performed on the container by means of a treatment unit of the container treatment machine, wherein the container is rotated about an axis from a current position to the target position in a container receptacle, wherein the container treatment machine comprises a camera, which captures an image of the container transported in the container receptacle upstream of the treatment unit, and an alignment module that rotates a container to a target position by controlling the container receptacle, characterized in that the alignment module comprises a neural network that determines a necessary rotation of the container from its current position to the target position by processing the image of the container transported in a container receptacle upstream of the treatment unit, the alignment module controls the rotation of the container receptacle depending on the determined rotation and the neural network is adapted to learn current positions of containers relative to a target position from images of containers captured during operation of the container treatment machine.

8. Method of claim 7, wherein the neural network is a pre-learned neural network.

9. Method of any one of claims 7 to 8, wherein the container treatment machine comprises an input unit with which an operator inputs a container type and / or a type of an alignment feature, and wherein the neural network determines the target position based on the input container type and / or the input alignment feature.

10. Method according to any one of claims 7 to 9, wherein the neural network, for determining the rotation in exactly one image of the container in its current position, in a first step searches for an alignment feature by means of which the target position of the container is defined and, if the alignment feature is found at least partially in the image of the container, determines the rotation from the position of the alignment feature in the image and the target position of the alignment feature; and wherein in a second step, if the alignment feature is not found at least partially in the exactly one image, the neural network determines a possible position of the alignment feature based on the information present in the image and the alignment module causes a rotation of the container in the container receptacle based on the possible position of the alignment feature, wherein in a third step a second image of a container is taken in the rotated position and the neural network searches for the alignment feature in the second image and, if the alignment feature is found at least partially in the image of the container, determines the rotation from the position of the alignment feature in the image and from the target position of the alignment feature.

11. Method of claim 10, wherein the neural network learns from a result of the second step and the third step.