Method and device for inspecting three-dimensional object

Abstract

The invention relates to a method and a device for inspecting three-dimensional objects, such as pouch cells, each object comprising a substantially pouch-or cuboid-shaped housing having a top side and a bottom side, the top side of the housing being composed of at least one upper surface section and a plurality of side surface sections, the side surface sections extend obliquely, parallel or vertically with respect to the at least one upper surface section or represent corner sections, for each object, image data captured in a matrix manner by the matrix camera is generated and transmitted to the data processing unit in a stationary state of the object in accordance with light reflected from the top side by the area illumination unit, the image data captured in the matrix form comprises light reflected from the side surface portion, the image data captured in the matrix form is processed by the data processing unit into a first overall matrix, and the following steps are performed by the data processing unit: segmenting the first overall matrix, a defect type and / or severity of the detected defect is identified and / or a mass score is determined that allows the mass of the object to be evaluated.

Description

[0001] Description

[0002] The present invention relates to a method for inspecting three-dimensional objects, in particular so-called pouch battery cells (in the following referred to as pouch cells), and a corresponding apparatus.

[0003] Pouch cells are a type of battery, in particular for lithium-ion batteries. Pouch cells typically consist of a pouch-like housing or package formed from a plastic-coated metal foil, such as aluminum foil. This type of battery cell is also referred to as a polymer battery. The housing is designed as a flexible, flat and lightweight pouch or cushion that is relatively sealed from the outside. Inside the housing, there is typically a stack of superimposed electrode layers, active layers and separator layers. The terminals are formed as two tabs that protrude from the pouch-like housing adjacent on one side, on adjacent sides or on opposite sides. Pouch cells are known for their high energy density, compact design and flexibility, making them suitable for a variety of applications, including electric vehicles. Pouch cells can be easily sized to meet the specific requirements of different electric vehicle models. Their flat and flexible design also allows for easier integration into different vehicle spaces, enabling more efficient packaging and improved space utilization. A disadvantage of the pouch cell design is that due to their construction, they are typically sensitive to mechanical damage. This can easily lead to the release of gases or electrolytes or can cause the battery cell to swell significantly or cause internal short circuits.

[0004] Therefore, it is desirable to thoroughly inspect such and other three-dimensional objects during quality control to detect damaged objects at an early stage.

[0005] Various options for quality control of flat objects such as battery cells have been disclosed. For example, from the document US 2022 / 0 390 387 A1 a method is known in which optical coherence tomography (OCT) is used to inspect the gap between the lead foil and the tabs of a pouch cell. This can provide information about the quality of the sealing of the pouch cell, but this has very limited significance for the quality of the pouch cell. The document EP 4 117 081 A1 describes a very complex inspection system which comprises a thickness measuring unit, a unit for measuring electrical properties, a printing unit, a tab cutting unit, a weighing unit, a tab testing unit and a defect selection unit. The thickness measuring unit measures the thickness of the pouch cell and the printing unit is used to print information about the pouch cell on the surface of the pouch cell. The tab inspection unit uses visual inspection to determine the length and shape of the tabs. Defective pouch cells are separated by the defect selection unit into a hopper provided for this purpose. The document DE 10 2019 109 703 A1 shows and describes an arrangement for checking the quality of a battery cell, the transparent outer skin of which encloses an inner space. Within the inner space, i.e. under the outer skin, an (additional) glass pin or lithium metal wafer is arranged, which changes its optical appearance in the presence of a predetermined concentration of hydrogen fluoride. This glass pin or lithium metal wafer is therefore analyzed in detail by optoelectronic measurement in order to determine the hydrogen fluoride concentration and thus the quality of the battery cell. Finally, the document EP 3869 603 A1 describes a method for checking the quality of a laminated electrode-separator composite and a battery having the electrode-separator composite, which is suitable for large-scale production and ensures that the individual layers are firmly and reliably connected to one another. The inspection comprises detecting at least a portion of a surface of the electrode-separator composite by a detection device to generate a measurement result and evaluating the measurement result. The detection device is particularly suitable for determining the surface topography, the surface temperature and / or the surface color. This can be done with the aid of optical sensors, photographic equipment and / or cameras. In this case, the detection device can comprise at least one illumination device which can emit light onto the surface of the electrode-separator composite to be inspected. The evaluation can comprise image processing and / or image analysis.

[0006] Further methods for inspecting objects or for defect detection or tab detection are known from WANG Xu, CHENG Pan, “Deep learning-based visual defect inspection system for pouch battery packs”, in: Proceedings of the 6th International Conference on Cognitive Computation, ICCC 2022, as part of the Service Conference Federation, SCF 2022, held 10-14 December 2022 in Honolulu, Hawaii, USA, Cham, Switzerland: Springer, 2022 (Lecture Notes in Computer Science), WO 2023 / 284 712 Al, EP 4 166 935 Al and DE 10 2021 002 262 B3.

[0007] The known methods mentioned above are either relatively complex or only allow a very limited assessment of the quality of three-dimensional objects, such as, for example, pouch batteries. It is therefore an object of the present application to provide a fast and reliable method for inspecting objects which allows a comprehensive assessment of the quality of the objects. Similarly, it is an object of the present application to provide a corresponding inspection device.

[0008] The above-mentioned objects are achieved by a method for inspecting three-dimensional objects, in particular battery cells, for example in the form of pouch batteries, having the features described below and by a corresponding inspection device having the features described below.

[0009] In particular, the above-mentioned objects are solved by a method for inspecting three-dimensional objects, for example pouch batteries, wherein each object comprises a substantially pouch-like or cuboid-like housing having a top side and a bottom side,

[0010] wherein the top side of the housing consists of at least one upper surface section and a plurality of side surface sections which extend parallel, obliquely or perpendicularly or form corner sections with respect to the at least one upper surface section, optionally not at the same height as the upper surface section,

[0011] wherein the bottom side of the housing consists of at least one lower surface section and a plurality of side surface sections which extend parallel, obliquely or perpendicularly or form corner sections with respect to the at least one lower surface section, optionally not at the same height as the lower surface section,

[0012] wherein, for each object, the image data captured in matrix fashion by the matrix camera is generated from the light reflected from the top side according to the area lighting unit in a stationary state of the object to be inspected and is transmitted to the data processing unit (optionally including the reflected light at the upper tab surface of the first tab and / or the second tab), wherein the image data captured in matrix fashion includes the light reflected from the side surface segments, wherein the image data captured in matrix fashion is further processed by the data processing unit as a first overall matrix,

[0013] wherein the following steps are further performed by the data processing unit:

[0014] segmenting the first overall matrix into:

[0015] a first image data portion comprising image data of the upper surface segment, and

[0016] at least one second image data portion, wherein each second image data portion comprises image data of at least one predetermined portion of the side surface segments and / or at least one predetermined corner segment,

[0017] subdividing the first image data portion into a plurality of individual patches,

[0018] identifying a defect type of the detected defect and / or a severity of the detected defect and / or determining a quality score allowing to assess a quality of the object based on:

[0019] specifically determining for each patch of the plurality of patches by a correspondingly trained first NN algorithm whether the respective patch of the first image data portion comprises one or more anomalies, wherein a defect is identified if there are anomalies, and

[0020] identifying by a correspondingly trained second NN algorithm different from the first NN algorithm whether there are defects in the at least one second image data portion and corresponding classifying the respective second image data portion.

[0021] The data processing unit can output the identified or determined results at a predetermined interface so as to make them accessible to a user. A display device can be connected to the interface to display the output results. In addition, the bottom side of the object can be inspected in a similar manner to the method steps explained above and below, i.e. a defect type of the detected defect and / or a severity of the detected defect and / or a quality score of the object can be determined / identified. The bottom side of the housing consists of at least one lower surface segment and a plurality of side surface segments extending obliquely, parallel or perpendicularly with respect to the at least one lower surface segment or representing corner segments.

[0022] The method is used for inspecting three-dimensional objects, for example flat objects in the form of pouches or cuboids, for example for battery cells, for example pouch cells. In one embodiment, the present application can be used for flat objects, wherein a three-dimensional object is referred to as a flat object if it exhibits a spatial dimension in one spatial direction (for example, height) that is significantly smaller than in the other two spatial directions and thus essentially has the shape of a flat cuboid or pouch or a shape similar to these shapes. Alternatively, the dimension in one spatial direction can also be larger, so that the object is described as essentially a cuboid. In this context, "essentially" means that the shape of the object approximates the shape of a pouch or cuboid. For example, a cuboid can have steeply sloping edges. In many cases, such objects also comprise a first connection tab (for short: tab, for example anode) and, if applicable, at least a second connection tab (for short: tab, for example cathode), each of which protrudes laterally. Each object comprises a housing having a top side and a bottom side opposite the top side, wherein any tabs protruding belong to the housing. The method according to the present application can be used both for inspecting three-dimensional objects comprising one or more such tabs and for inspecting three-dimensional objects without such tabs. In particular, the method is suitable for objects comprising stepped or terraced sections, in particular comprising these sections on their edges, or the aforementioned connection tabs. Thus, the object is observed in such a way that one of the two largest sides forms a section of the top side, while the likewise large opposite side forms a section of the bottom side. When the top side is on top and the bottom side is on the bottom, the top side of the housing has at least one upper surface section which extends essentially horizontally and is the surface section of the top side having the largest dimension. Further horizontally extending surface sections can be provided which extend parallel to the upper surface section of the top side, for example terraced surface sections of the top side. The top side also comprises a plurality of side surface sections which extend obliquely, parallel or perpendicularly with respect to the at least one upper surface section (for example, edges or side surfaces) or form corner sections. The side surface sections also comprise sections which extend parallel to the upper surface section or a surface of a protruding tab (tab surface). Thus, the bottom side of the housing comprises at least one essentially horizontally extending bottom surface section which is the surface section having the largest dimension. Further horizontally extending surface sections can be provided which extend parallel to the "lower" surface section of the bottom side, for example terraced surface sections of the bottom side. The bottom side also comprises a plurality of side surface sections which extend obliquely, parallel or perpendicularly with respect to the at least one lower surface section (for example, edges or side surfaces) or constitute corner sections. The first tab and the at least one second tab, if present, can for example protrude from a short side and / or a long side and each comprise an upper tab surface and a lower tab surface. For example, the first tab and the second tab protrude from a single short side or long side. In this case, the tabs are arranged adjacent to one another.Alternatively, the first and second tabs can protrude from opposite short sides or long sides. When viewed from above on the top side or the bottom side of the housing, the housing can have a substantially rectangular shape (not taking into account any tabs that can be present). The short sides are the short sides of this rectangle, and the long sides represent the long sides of this rectangle.

[0023] For obtaining image data from the three-dimensional object to be inspected, the three-dimensional object is moved in the device for performing the method, wherein temporary stationary states of the three-dimensional object in the device, in which the three-dimensional object does not move, are part of this movement. The movement of the three-dimensional object is performed by means of a drive unit, which causes the object to be inspected to move (movement state) at a predetermined speed relative to the line illumination unit, for example substantially parallel to the upper surface section, for example in the direction of the maximum dimension (length) of the upper surface section or transversely thereto. The predetermined speed is for example at least 500 mm / s, for example at least 800 mm / s. Furthermore, the drive unit is configured to arrange the object at a predetermined position relative to the area illumination unit and for a predetermined period of time during further movement of the respective object in the stationary (stationary state). In this case, the predetermined period of time for the arrangement of the object in the stationary state can precede the movement state, but also follow the movement state. The predetermined period of time in the stationary state can for example be at least 300 ms, for example at least 400 ms. In one embodiment, the drive unit is realized by a slide that can move in a predetermined manner over a linear unit. The slide comprises for example a suction cup, by means of which the housing of the object can be attached to the slide at its bottom side. The movement data for moving the three-dimensional object to be inspected (i.e. its arrangement in the movement state and the stationary state, the position of the object and / or its speed, etc.) are captured by a movement detection unit and transmitted to the data processing unit. In the data processing unit, the captured movement data (movement information) are used together with the captured image data from the matrix camera for determining / identifying the presence of defects and / or for determining the quality score.

[0024] The area illumination unit illuminates the top side of the housing (e.g. the entire top side), optionally including the upper tab surface of the first tab and / or the second tab of the object to be inspected, from above. In one embodiment, the entire top side of the housing (optionally including the upper tab surface of the first tab and / or the second tab) or at least a larger section of the top side of the housing (optionally including the upper tab surface of the first tab and / or the second tab), e.g. at least 70%, e.g. at least 80%, of the entire top side of the housing, is illuminated by the area illumination unit. For example, the light from the area illumination unit is incident perpendicularly or obliquely onto the top side of the housing (optionally including the top tab surface of the first tab and / or the second tab), e.g. with an angle of incidence in the range of 10° to 60° with respect to the horizontal direction. By means of the oblique illumination by the area illumination unit, defects such as dents, protrusions, scratches, folding defects, edge cracks, defects at the seal and similar topological defects can be easily detected. Defects in the form of absorption defects (e.g. contamination, foreign bodies on the surface) can also be detected. The area illumination unit is realized by means of LED spots or other quasi-spotlights. In one embodiment, at least one second deflection mirror is arranged above the position of the object to be inspected in the rest state, which extends perpendicular to the horizontal direction and deflects the light from the area illumination unit such that it falls obliquely from above onto the top side of the bag-shaped housing (optionally with tabs). This can reduce the overall external dimensions of the inspection device.

[0025] The matrix camera captures the reflected light of the top side of the housing (optionally including the upper tab surface of the object to be inspected arranged in the rest state) illuminated obliquely from above, in the form of image data captured in a matrix manner (intensity, and in one embodiment, additionally color values), and transmits this captured image data to the data processing unit. In one embodiment, the matrix camera can capture the entire top side of the housing. For example, the matrix camera is arranged above the object when the object is in its rest state at a specified position, i.e. in this embodiment, the matrix camera is positioned above the rest state position of the object to be inspected.

[0026] The matrix camera can be designed, for example, as a CCD or CMOS camera. The matrix camera captures the light intensity of a large number of pixels in the field of view, which are arranged in rows and columns, i.e. in a matrix. To this end, the matrix camera comprises a light-sensitive element (e.g. a CCD or CMOS sensor) for each pixel. The size of the area captured by each light-sensitive element determines the resolution of the matrix camera. The matrix camera can for example comprise a field of view of 9344 x 7000 pixels or 8192 x 8192 pixels and thus capture image data with dimensions of 805 x 603 mm. The line-by-line capture can for example comprise a region of 16 to 128 x 1000 to 8192 pixels accordingly. The matrix camera is also arranged in such a way that it looks vertically from above at the object to be examined in the resting state, so that it clearly sees this part of the field of view. The matrix camera is focused in such a way that it comprises a sharpness that is as uniform as possible over the entire field of view. In particular, this is achieved for the line of sight in which the image data from the object (e.g. from the side surface section) reaches the matrix camera via the mirror. This is achieved by a corresponding aperture setting, which achieves the necessary depth of field.

[0027] The examination method according to the application is characterized by the fact that different elements of the image data with different examination analysis requirements are separated and analyzed by different methods by means of a segmentation of the image data captured in matrix fashion ("first overall matrix of image data"). In an advantageous manner, the image data of the relatively flat region of the upper surface section, which is prominent in terms of its extension ("first image data portion"), is initially processed and analyzed separately from the image data of the side surface section comprising the corner section ("second image data portion"), which has a smaller size and is expected to have a greater degree of concave-convexity. In the case of the first image data portion, it has proven advantageous in terms of computing speed to subdivide this portion into a large number of blocks and then, as described in more detail below, to use a first NN algorithm to determine whether one or more anomalies are present in the respective blocks. If at least one anomaly is present, a defect is detected / identified. In the second image data portion, it can be determined on the basis of a second NN algorithm whether a defect is present, which will be described in more detail below. In addition, a classification of this defect can also be carried out.

[0028] For the overall quality assessment, the individual analyses are combined for consideration. In particular, a defect type of the detected defects and / or their severity and / or a quality score of the object is identified, which allows assessing the quality of the object. For example, the identified defect type and / or their severity are also used for this purpose. As described above, it is advantageous to use different algorithms suitable for the possible defect types for analyzing the respective image data portions obtained by means of the segmentation. This can also create a time advantage and an accuracy advantage for the evaluation of the image data. In other words, by analyzing the different portions individually, meaningful results can be obtained quickly during the inspection, since the image data processing is adapted to the specific properties of the object.

[0029] In particular, the segmentation is performed by means of so-called layout recipes. Each layout recipe predefines a given section of the object with respect to the field of view of the matrix camera / cameras. Since the object is not always exactly in the predefined ideal position when the image is captured by the matrix camera, but can be shifted / rotated by a few pixels, a position correction is performed, i.e. a registration to the intended position, using for example a designated fixed point of the object, so that the first overall matrix (or correspondingly n first overall matrices determined from the row-wise observation or the second overall matrix) is adapted accordingly to the ideal position of the object. The above-described matrix with the image data is rotated and / or shifted accordingly. Once this adaptation has been performed, the desired image data portions can be reliably identified and extracted using the designated layout recipe.

[0030] As described in more detail below, defect types such as inclusions, pits (dents), protrusions (bumps), contaminations (dirt, electrolyte residues), pseudo-edges, orange skin, porosity, cracks, grinding grooves, spots, surface defects, blisters, scratches and wet prints are determined from the image data (image information) transmitted by the matrix camera to the data processing unit by processing the data in the data processing unit.

[0031] In one embodiment of the method, the classification of the at least one second image data portion is performed by means of a classifier with two states or a classifier with at least 3 states, wherein the classifier with at least 3 states allows for example to assign different types of defects, while by means of the classifier with two states it can be assessed whether a defect is present or not.

[0032] In one embodiment of the method, for each object, a plurality of line-by-line captured image data of reflected light from a line lighting unit at a linear region of the top side (optionally including the upper tab surface of the first tab and / or the second tab) in a moving state of the object to be inspected is generated by means of a camera, for example, a matrix camera, and transmitted to a data processing unit, wherein the following steps are further performed by means of the data processing unit:

[0033] merging the image data captured row by row into a second overall matrix comprising image data of the top side of the object,

[0034] Additionally, based on the image data of the second overall matrix, the defect type of the detected defect and / or the severity of the detected defect is identified and / or a quality score is determined, which allows the quality of the object to be assessed.

[0035] In an embodiment, the identification of the defect type of the detected defect and / or the severity of the detected defect based on the image data of the composite second overall matrix and / or the determination of a quality score allowing the quality of the object to be assessed may also be performed in place of the aforementioned determination and in combination with the aforementioned identification and corresponding classification, wherein the aforementioned determination is based on a separate determination for each of the plurality of blocks, by means of a correspondingly trained first NN algorithm, as to whether the corresponding block of the first image data portion includes one or more anomalies, and the aforementioned identification and corresponding classification is based on the identification of the presence or absence of a defect in at least one second image data portion and the corresponding classification of the corresponding second image data portion by means of a correspondingly trained second NN algorithm. Alternatively, in an embodiment, the identification of the defect type of the detected defect and / or the severity of the detected defect based on the image data of the composite second overall matrix and / or the determination of a quality score allowing the quality of the object to be assessed may also be performed in place of the aforementioned identification of the presence or absence of a defect in at least one second image data portion, by means of a correspondingly trained second NN algorithm, and in combination with the aforementioned determination, wherein the aforementioned determination is based on a separate determination for each of the plurality of blocks, by means of a correspondingly trained first NN algorithm, as to whether the corresponding block of the first image data portion includes one or more anomalies.

[0036] When the image data of the object is captured line by line, the captured "image lines" are combined / merged to form a matrix image (second overall matrix). The merging comprises a position alignment of the line by line captured image data of the object, so that a matrix of image data (second overall matrix) is created, wherein, if necessary, a correction of differences / overlaps is made. In other words, the second overall matrix contains the image data determined line by line for the entire upper side of the object or for a predetermined section of the upper side corresponding to the position on the top side of the object at which the incident light of the line illumination unit is reflected, respectively, and thus also contains an image of the top side of the object. In this respect, the image of the second overall matrix can be corrected so that the resulting image is equivalent to the first overall matrix or one of the n first overall matrices, for example, in terms of the size of the matrix and / or the position of the captured portion of the upper side of the object. The second overall matrix is used for further analysis of the image data as described below.

[0037] The line illumination unit illuminates a linear area of the top side of the housing (optionally including the upper tab surface of the first tab and / or the second tab). For example, the line illumination unit is formed by a lamp having a plurality of LEDs arranged to illuminate the desired linear area. In this case, one LED line can be provided, or for a wider linear area, several LED lines arranged adjacent to each other (for example, 2 to 10 LED lines) can be provided. In one embodiment, the line illumination unit can be switched in such a way that it illuminates each point of the linear area with light of two different intensities, i.e. high intensity A and low intensity B. Thus, the line-by-line detection of the light reflected from the linear area of the top side (optionally the upper tab surface of the first tab and / or the second tab) is performed with an adapted switching rhythm in the form of ABABAB... (i.e. alternating switching of the two different intensities A, B). For this purpose, the line-by-line capturing of the image data (capturing frequency and capturing time) and the feed rate of the drive unit are synchronized. This illumination is also referred to as HDR reflection bright field illumination and has advantages in the identification of certain types of defects, for example, the object is contaminated with a transparent substance, due to the special illumination technique. The line-by-line capturing of the image data can be performed by a camera with an appropriate field of view and resolution, for example, by a matrix camera, which also performs the matrix-by-matrix capturing of the image data. After the capturing by the matrix camera, the data (image data) is transferred to the data processing unit and is additionally used for the inspection of the three-dimensional object, wherein, as described above, before the analysis, the line-by-line captured image data is combined / merged to form an overall image of the respective top side of the object (second overall matrix).

[0038] It should be emphasized that in one embodiment the reflected light of the linearly illuminated area of the top side of the housing of the object under inspection in motion is recorded in the form of image data (image information, e.g. intensity, and in one embodiment additionally color values) by means of a single matrix camera, and the reflected light of the top side of the housing (optionally tilted illumination, if applicable including the top tab surface) illuminated from above in the form of image data (image information, e.g. intensity, and in one embodiment additionally color values) of the object under inspection arranged in a stationary state, and both captured image data are transmitted to the data processing unit. The line-by-line capture represents a sub-area of the field of view of the matrix camera and results in one pixel line or several adjacently positioned pixel lines (e.g. pixel lines with 16 to 128 pixels) with image data, whereas the matrix capture results in a pixel matrix with image data, wherein the pixel matrix also represents a sub-area of the field of view. In one embodiment, the image data can be determined in a predetermined wavelength range. The field of view of the matrix camera is designed in such a way that the matrix and line-by-line image data are captured by a single fixed (i.e. not moved during image data capture) matrix camera, which is subsequently assigned to the respective object by the data processing unit. In this context, the entire top side of the housing (optionally including the upper tab surface of the first and / or second tab) or at least a segment of the top side of the housing illuminated by the area illumination unit, i.e. at least a larger segment of the top side of the housing (optionally including the upper tab surface of the first and / or second tab), e.g. at least 70%, e.g. at least 80% of the entire top side of the housing of the object under inspection, can be captured during the matrix capture. The advantage of using a matrix camera for the matrix and line-by-line capture of the image data is that the obtained image data do not have to be coordinated with respect to the capturing instrument. They contain the same camera properties. The inclusion of line-by-line image data in the inspection further improves the quality assessment.

[0039] In one embodiment of the method, the following further steps are performed by means of the data processing unit:

[0040] The second overall matrix is divided into:

[0041] a third image data portion comprising image data of the upper surface segment, and / or

[0042] at least one fourth image data portion, wherein each fourth image data portion comprises image data of at least one predetermined segment of the side surface segment (wherein corner portions can be excluded in this portion) and / or at least one predetermined corner segment,

[0043] the third image data portion is subdivided into a plurality of individual blocks,

[0044] wherein the defect type of the detected defect and / or the severity of the detected defect and / or the quality score allowing to assess the quality of the object is identified based on the following:

[0045] by means of the first NN algorithm, it is determined for each of the plurality of blocks in particular whether the respective block of the third image data portion comprises one or more anomalies, wherein, if anomalies are present, a defect is identified, and / or

[0046] by means of the second NN algorithm, it is identified whether a defect is present in the at least one fourth image data portion and the respective fourth image data portion is classified accordingly.

[0047] The line-by-line merged image data and the previously line-by-line captured image data can be analyzed in the third image data portion defined above after decomposition into a plurality of individual blocks in a matrix-like manner by means of the first NN algorithm and / or in the fourth image data portion with regard to the presence of defects by means of the second NN algorithm. This method also applies to the line-by-line captured image data and has the advantages described above.

[0048] In one embodiment of the method, for each object, at least n (n > 2) recordings of the top side (e.g. the top side (optionally the entire top side, optionally including the upper tab surface) of the matrix-captured image data are generated by temporally successive capturing of reflected light (capture of the matrix camera) and are transmitted to the data processing unit, wherein the matrix-captured image data are further processed by the data processing unit into n first overall matrices, wherein the reflected light is generated by temporally separated illumination from n different directions obliquely from above the entire top side of the shell of the object in a stationary state of the object to be checked. The n recordings are transmitted to the data processing unit. Thus, in this embodiment, the area illumination unit is configured to illuminate the top side of the shell of the object to be checked in a stationary state obliquely from above, e.g. the entire top side of the shell (optionally including the upper tab surface of the first tab and the second tab of the object to be checked), from at least n different directions in time succession, and the matrix camera is correspondingly configured to capture the image data of the light reflected from the upper side of the shell in a matrix manner in time succession during the at least n directional illumination of the area illumination unit. The data processing unit is correspondingly configured to receive and process the n image data captured in a matrix manner during the n directional illumination of the area illumination unit, wherein the image data are assigned to the respective object. The image data captured in a matrix manner in the form of n recordings are further processed by the data processing unit into n first overall matrices.

[0049] In one embodiment of the inventive method, the following steps are further performed by the data processing unit:

[0050] Split the n first overall matrices into:

[0051] n fifth image data portions, each comprising image data of an upper surface segment of each of the n first overall matrices and / or

[0052] n sixth image data portions and optionally further image data portions of each of the n first overall matrices, wherein each sixth image data portion and optionally further image data portion comprises image data of at least one predetermined portion of the side surface segment and / or at least one predetermined corner segment,

[0053] in each case determining a maximum image and / or an absorption image and / or a topological image from the image data of the fifth image data portion and / or the sixth image data portion and / or of the possible further image data portion,

[0054] determining defects in the maximum image and / or absorption image and / or topology image of the fifth image data portion and / or the sixth image data portion and / or the further image data portion (if applicable) and analyzing and characterizing the defects,

[0055] Therein, the identification of the defect type and / or the severity of the detected defect and / or the determination of a quality score allowing an assessment of the quality of the object is based on the results of the analysis and / or characterization of the respective detected defect.

[0056] As explained in more detail below, the inspection can be further accelerated by additional defect analysis and determination using methods that do not use an NN algorithm, since, for example, the threshold analysis of the pixels of the maximum image and / or absorption image and / or topology image and the creation of the maximum image and / or absorption image and / or topology image can be performed very quickly. Both matrix-generated and row-by-row image data can be used for this purpose.

[0057] In this embodiment of the method, n records of image data captured in a matrix format, in particular image data of the fifth image data portion, the sixth image data portion, or another image data portion, a maximum image and / or a topological image and / or an absorption image, are used as image data for further analysis. A matrix image, a topological image, and / or an absorption image are each generated from n matrix-like captures of the corresponding image data portion captured consecutively in time. The maximum image represents the image data of the area that is best accessible for the respective lighting conditions and is therefore identified as the brightest. The advantage of a topological image is that it emphasizes topological changes in the image, while an absorption image highlights defects caused by light absorption (e.g., contamination on a surface).

[0058] For example, the n captured image data are generated in a pixel-identically manner, i.e. at least two matrix-wise captured image data of the top side (optionally the entire top side, optionally including the upper tab surface of the first and / or second tab) are each generated from the same point on the surface. Each of these matrix-wise captures is referred to as an image data matrix M, wherein at least two image data matrices Mk(k > 2, k = 2... n) are captured for each object. Thus, a pixel Pi of the first captured image data matrix M1 corresponds to the same pixel Pi on the surface of the top side (optionally including the upper tab surface) as a second (third, fourth, etc.) captured image data matrix Mk(M2, M3, M4,... Mn). The captured light intensity in a pixel Pi is denoted as i(Pi). The captured light intensity of the first image data matrix M1 at a pixel Pi is referred to as i1(Pi). Each image data matrix comprises image data of, for example, the first image data portion defined above.

[0059] The maximum image can be determined by forming the maximum of the light intensities of all image data matrices Mk in the respective pixel Pi, i.e. Max(i1(Pi), i2(Pi)) for two determined image data matrices M1, M2 of two illuminations of two different directions, or Max(i1(Pi), i2(Pi),... in if n illuminations of n different directions are used. In one embodiment, n = 4. The maximum is calculated for each pixel Pi and represented in the entire matrix (maximum matrix), resulting in a maximum image. n (Pi))· In one embodiment, n = 4. The maximum is calculated for each pixel Pi and represented in the entire matrix (maximum matrix), resulting in a maximum image.

[0060] The topological image and the absorption image can be determined by first applying two different parameterized low-pass filters (e.g. box filters) to each image data matrix Mk of each illumination situation independently of each other and subtracting them from each other:

[0061] Fk = low-pass1(Mk) - low-pass2(Mk)

[0062] In this context, the parameters of the two low-pass filters low-pass 1 and low-pass 2 differ, for example, in that a first parameter of the first low-pass filter low-pass 1 is smaller than a second parameter of the second low-pass filter low-pass 2. By this operation the light intensity assigned to each pixel Pi of the matrix Fk is called fk(Pi) (k = 2... n). Subsequently, from the resulting matrices Fk, similar to the maximum image above, the minimum or maximum is determined over all matrices, pixel by pixel, so that a minimum matrix MinM and a maximum matrix MaxM are determined, wherein each point Pi of the minimum matrix MinM is calculated as Min(f1(Pi), f2(Pi),... fn(Pi)) and each point Pi of the maximum matrix MaxM is calculated as Max(f1(Pi), f2(Pi),... fn(Pi)). Subsequently, a matrix H with values h(Pi) is determined, which is determined (again pixel by pixel) from the product of the minimum and maximum calculated at the respective point Pi with a scaling factor a (for example a = 64). This means that for each point Pi the value is

[0063] h(Pi) = Min(f1(Pi), f2(Pi),... fn(Pi)) * Max(f1(Pi), f2(Pi),... fn(Pi)) * a

[0064] Finally, from this a matrix Q with values q(Pi) is determined, wherein

[0065] q(Pi) = sqrt(abs(h(Pi))),

[0066] where abs(q(Pi)) is the absolute value of the value q(Pi) and sqrt() is the root function. This results in a value of the topology matrix T, wherein the value t(Pi) is as follows:

[0067] If h(Pi) < 0, then t(Pi) = q(Pi); or if h(Pi) > 0, then t(Pi) = 0.

[0068] Thus, the values of the absorption matrix A with values a(Pi) are obtained as follows:

[0069] If h(Pi) > 0, then a(Pi) = q(Pi); or if h(Pi) < 0, then a(Pi) = 0.

[0070] The topology matrix T with values t(Pi) calculated in this way is also referred to as a topology image, and the absorption matrix A with values a(Pi) is also referred to as an absorption image.

[0071] If the matrix-like capturing of the light reflected upwards by the area lighting unit from the top side, optionally from the entire top side and / or optionally including the upper tab surface, is carried out four times in the case of oblique illumination from above from four different directions, these directions are chosen, for example, such that illumination from two opposite long sides and two opposite short sides of the housing takes place. Alternatively, the lighting can illuminate the top side from the direction of each of the four corners of the housing. In one embodiment, it is advantageous if the capturing is produced with illumination, in which all illumination directions together cover an angle of 360°, optionally in terms of their component extending in the plane of the upper surface section (i.e. when illuminated from four different directions, directions offset by 90° from each case are provided with illumination, or when illuminated from six different directions, directions offset by 60° from each case are provided with illumination, etc.).

[0072] In this embodiment, the maximum image and / or the absorption image and / or the topological image generated for the fifth image data portion (e.g. one upper surface section) and / or the sixth image data portion (e.g. four side surface portions) and possibly further image data portions (e.g. four corner portions) and / or the third image data portion and / or the at least one fourth image data portion are then further used for the inspection. For example, by one or more threshold values specified for the respective image (maximum image, absorption image, topological image) and / or for the third image data portion and / or for the at least one fourth image data portion, the positions of the respective pixels in the respective image / portion whose respective values exceed or fall below the respective threshold value are analyzed pixel by pixel. If the respective threshold value is exceeded or not reached, it is assumed that a defect is present. Subsequently, further properties of these defects are determined, for example their extension in pixels, a histogram of the gray value in the defect area. This can be used to identify the respective defect type and / or the severity of the defect. The assignment of the analysis values and the defect types can be carried out, for example, using a corresponding table stored in the data processing unit.

[0073] In one embodiment, four corner image data matrices are extracted by the data processing unit from the segmented image data or by the segmentation itself, which at least contain the second image data portion or the fourth image data portion or the sixth image data portion or the further image data portion, wherein each matrix contains one corner of the housing and its position is precisely known, for example after a position correction described below. Alternatively, other image data portions of the object can be extracted. If necessary, several recorded maximum matrices are generated in advance.

[0074] The corner image information matrix of each corner (i.e., each corner segment) can be analyzed separately, for example, using a convolutional neural network (CNN) algorithm as the second NN algorithm, which includes a binary classifier (a classifier with two states (i.e., “complete corner” and “defective corner”). The corner image data matrix or alternatively the image data matrix of other image data portions can be defined to be relatively small (e.g., 128 x 128 pixels) and the underlying image data is taken only from the area of the respective corner of the housing. This CNN model is specifically developed for this low class classification task and low pixel matrix. The CNN model includes a compact structure, for example, which is more compact than regular deep architectures. For example, it consists of three strands with different convolution sizes, which are then merged. This structure reduces the number of parameters to be trained, so that the model has to learn fewer features. As a result, it is ideal for binary classification or other low-dimensional classification. For the training and evaluation of the CNN model, a large and extensive data set is used, which contains matrices of corner structures and corresponding expected defects. This data set is compiled with respect to the respective objects to be analyzed and annotated by engineers. The data set contains corner images of the respective objects (i.e., the soft pack batteries to be checked), which are divided into two classes: “defective” and “complete”. These images are carefully selected and annotated to ensure that they cover a variety of defects and variations of the corners. The training of the model is performed on the data set, wherein the images in the data set can be divided into a training set and a validation set. For example, a ratio of 80% can be used for the training data and a ratio of 20% can be used for the validation data. In addition, five-fold cross-validation can be performed to ensure that all images are present in the training data and the validation data. The model consists of several convolutional layers, pooling layers and fully connected layers, which enable the model to extract important features in the images and identify subtle differences between defective and complete corners. The convolutional layers are used to train the trainable weights of the convolution operation, which are then used to identify image features. These features are then aggregated with the pooling layers. Subsequently, the weights of the fully connected layers are iteratively trained to determine the probability of the associated class from the features. In addition, all images / segments can be cut to the same size as the corner image data matrix (128 x 128 pixels) to be analyzed before the analysis with the CNN algorithm, and the images / segments are oriented in such a way that they have the same orientation in order to achieve a consistent view.

[0075] In another embodiment, at least one predetermined portion of the side surface section (e.g. the platform section of the pouch cell) is extracted from the segmented image data by the data processing unit, which segmented image data comprises at least one second image data portion and / or at least one fourth image data portion and / or a sixth image data portion or a further image data portion, or is extracted by the segmentation itself. Each predetermined portion is analyzed using an object detection network based on a Mask-RCNN model as a further second NN algorithm. This algorithm identifies defects of various different defect types (e.g. 6 different defect types such as protrusion / nose / protrusion, indentation, fold, scratch, contamination, particle) and adds corresponding bounding boxes to the data of the corresponding second image data portion in the defect area. In order to train the Mask-RCNN model as an algorithm, data representing the corresponding second image data portion is used, wherein the corresponding defects are annotated in these image data matrices and provided with bounding boxes. The architecture of the Mask RCNN model is carefully chosen. This model is a further development of the faster R-CNN model and is capable of generating bounding boxes and masks for defects in a specified area. Rarely occurring defect types are artificially inserted into the corresponding image data portion of a given area to train the model. The performance of each model is evaluated using a separate validation data set. For example, a validation data set with a ratio of 80% for training data and 20% for validation data can be used. This ensures that the model effectively and accurately identifies defects and correctly classifies these defects.

[0076] In a further embodiment, at least one predetermined portion of the upper surface section is extracted by the data processing unit from the segmented image data comprising at least the first image data portion, or by the segmentation itself, in one embodiment the segmentation process comprises the entire upper surface section. In addition, the resolution in the predetermined (sub) regions can be reduced to a predetermined value (e.g. from 5120x2216 to 841x265 pixels) to speed up the process. The image is divided into several small patches (sub regions). Subsequently, each patch is examined using a pre-trained CNN algorithm “Wide ResNet-50” as the first NN algorithm to determine whether one or more predetermined features (defects / anomalies) are present in the respective patch. In “Wide ResNet-50”, the layers of the network are made “wider” by increasing the number of channels in the convolutional layers. This kind of CNN is able to recognize complex patterns and textures. It was also observed that this wider CNNs are generally able to generalize better, which means that they can process new, unknown data more effectively. This method is also known as PaDiM (Patch Distribution Modeling Framework for Anomaly Detection and Localization) and is an algorithm for anomaly detection and localization tasks. This method is particularly suitable for the detection of industrial defects, whose aim is to identify irregularities or relative standard deviations in visual data. PaDiM models the distribution of features in the image. Subsequently, features extracted by the CNN are collected for each patch. For each patch, the Mahalanobis distance between the features of the patch and a normal distribution derived from the training data is calculated. This step determines how “abnormal” or unusual each patch is compared to the normal training data. The calculated Mahalanobis distance is used as an anomaly score, where higher values indicate a greater deviation from the normality. A threshold is set based on the anomaly score. Patches whose score exceeds this threshold are considered to be abnormal. The anomaly is localized by marking the location of the patch classified as abnormal in the image data portion, which enables the localization of the anomaly in the respective image data portion.

[0077] The anomaly score is calculated individually for each patch by computing its features’ Mahalanobis distance with respect to an expected normal distribution, represented by a mean and a covariance matrix from the training data. A large Mahalanobis distance indicates that the features of the patch strongly deviate from the normal distribution, suggesting a potential anomaly. Mathematically, the Mahalanobis distance D of a point x to a distribution with mean μ and covariance matrix ∑ is computed as follows:

[0078]

[0079] D(x) refers to the Mahalanobis distance of that point.

[0080] x refers to the vector of observed values.

[0081] μ is the mean vector based on the amount of training data.

[0082] ∑ is the covariance matrix of the training data.

[0083] ∑ -1 is the inverse of the covariance matrix.

[0084] T denotes the transpose of a vector.

[0085] For each block, the anomaly score leads to an assessment of the severity of the defects present in the respective block.

[0086] Applying the second NN algorithm (CNN algorithm, Mask RCNN algorithm) to the corresponding image data portion also leads to the detection of defects, defect types, and the severity of the detected defects, including defective corners, of the respective defect (or defective corner). Furthermore, as described above, defects in the respective image / portion are also detected, analyzed, and characterized by using defect detection of the maximum image and / or the absorption image and / or the topology image and / or the image data portion based on the row-wise determined values. The defect type is also specified and the severity of the respective defect is determined. For example, the respective severity level can be assessed based on the size of the respective defect, wherein small defects are assigned a severity level of a small category of severity levels (e.g., category 1) and more extensive defects are assigned a severity level of a larger category (e.g., category 4). Alternatively or additionally, other parameters of the defect, such as the position of the defect on the object, can be used to assess the severity.

[0087] For example, the size of the defect can be determined by the data processing unit after taking into account the perspective and / or optical distortion of the matrix camera. Similarly, the dimensions of the object (e.g., the edge length of the housing) can be determined.

[0088] Once the above information about the detected defects and their severity has been determined, the data processing unit performs an overall assessment of each object, for example, using the quality score. All identified defects and their severity are included, wherein different defects can be weighted differently depending on their severity. Thereby, it is determined, for example, by using a corresponding table provided in the data processing unit, whether the object meets a specified quality requirement, i.e., whether the determined quality score is less than or greater than or equal to a quality score threshold. In both cases, the determined data (defects, severity of defects, size of defects, position / positioning of defects) can be output at an interface of the data processing unit for further processing of the object and can be used for further processing.

[0089] To capture the image data, in one embodiment the matrix camera is configured (e.g. controlled by the data processing unit in such a way) that a line-by-line capture and a matrix capture of the image data take place in order to record a sequence (a time sequence of recording sequences over the entire field of view of the matrix camera). This can be synchronized with a corresponding control of the illumination (i.e. the line illumination unit and / or the area illumination unit). In one embodiment, the image data to be captured line-by-line of a first object (in a state of motion) can be captured at least partially simultaneously with the image data to be captured in matrix fashion of a second object (in a state of rest), which is different from the first object. This design of the recording sequence can shorten the total time required for the quality assessment of the object. This means that the matrix camera is configured such that at least one of its recording sequences (i.e. in the same capture) contains:

[0090] in the state of motion of the first object, a line-by-line capture of light reflected from a linear area of the top side into the matrix camera by the line illumination unit, and

[0091] in the state of rest of the second object, a matrix capture of light reflected upwards from the top side (in one embodiment, the entire top side and / or optionally an upper tab surface including the first tab and / or the second tab) by the area illumination unit, optionally including light reflected from a side surface section of the second object, if applicable, reflected via the first deflection mirror to the matrix camera, wherein the second object is different from the first object.

[0092] The capture and illumination sequence may, for example, include a plurality of line-by-line captures (e.g. between 50 and 120 line-by-line captures) of light reflected from a linear area of the top side (optionally including an upper tab surface of the first tab and / or the second tab), and wherein one of some (between 5 and 20) matrix captures of the top side are included partially in the same capture. Alternatively, the line-by-line captured image data and the matrix captured image data of two different objects can be recorded sequentially by the matrix camera in a capture sequence. In this case, in order to save time, only a section of the entire pixel matrix of the matrix camera can be read out, for example a corresponding section of the line-by-line captured and a corresponding section of the matrix captured.

[0093] When capturing the image data, the matrix camera is stationary (i.e. it does not move, nor do its parts), and the size of the field of view of the matrix camera is such that both the image data to be captured line by line and the image data to be captured matrix-wise are contained in the same field of view. The object to be inspected is in motion during the line-by-line capturing, i.e. the object to be inspected continues to move while the image data is created. In contrast, during the matrix-wise capturing, the object to be inspected is stationary (i.e. in a stationary state) in a predetermined position and for a predetermined period of time, so that the image data captured matrix-wise can be determined accurately. Furthermore, at least two first deflection mirrors can be arranged next to the object to be inspected, which first deflection mirrors can also be captured by the field of view of the matrix camera and provide further image data of the side surface sections of the top side of the object to be inspected. In this embodiment, these further image data are captured together with (simultaneously, i.e. in the same recording) the matrix-wise capturing of the object to be inspected. The image data captured line by line and the image data captured matrix-wise, including the image data transmitted via the first deflection mirrors if applicable, are assigned to the respective inspected object and are included in the determination of the presence and / or quality score of at least one defect type of defects.

[0094] In the above-described embodiments, the arrangement and inclination of the first deflection mirrors are such that the matrix camera receives light reflected from the largest possible area of the respective side surface section of the top side. In one embodiment of the device, at least two, in particular four, first deflection mirrors are provided, wherein each first deflection mirror is arranged next to a respective side of the housing in the stationary state of the object. With four first deflection mirrors, it is possible to capture reflected light of the side surface sections of all sides of the housing. As an example, each first deflection mirror is designed in such a way that its length (dimension parallel to the respective side next to which it is arranged) corresponds at least to the length of the respective side of the housing. Furthermore, in one embodiment, each first deflection mirror is arranged at a distance of at least 30 mm from the respective side of the housing in the horizontal direction. In another embodiment, the width (dimension perpendicular to the respective side next to which the respective first deflection mirror is arranged) of each first deflection mirror is at least 20 mm. The angle of inclination of the first deflection mirrors is, for example, at least 30° with respect to the horizontal direction. Furthermore, it is advantageous for the accuracy of the inspection if the deflection mirrors achieve a very good optical imaging quality to avoid distortions in the image of the matrix camera.

[0095] The illuminated linear area can extend over the entire length of the top side (optionally including the protruding first and second tabs). Here, the length of the top side is the dimension of the housing in its largest dimension. In this embodiment, the illuminated linear area can be used to obtain image data about the entire top side when the entire object is moved past the line illumination unit.

[0096] The matrix camera can be calibrated in such a way that the data processing unit can take into account the perspective and / or optical distortions from the image data captured in matrix fashion. For this calibration, for example, the method described in the article "Digital camera self-calibration" by C. S. Fraser, published in 1997 in the journal ISPRS Photogrammetry & Remote Sensing 52, pages 149-159, is used. In one embodiment of the device, the data processing unit is configured to determine at least one dimension of the object and / or at least one size of the detected defects after taking into account the perspective and optical distortions. To this end, for example, a look-up table is predetermined on the basis of the calibration, by means of which a conversion of the number of pixels into length or area units is provided. The look-up table is stored, for example, in a memory unit of the data processing unit.

[0097] In one embodiment, the position correction can additionally be performed using the calibration and by using fixed points, for example the corner points of the housing, by means of the data processing unit. Here, the coordinates of the four corner points of the housing are determined, for example, by a software-based "sounding" of the housing in horizontal and vertical directions. The sounding involves checking the intensity variations of the respective rows and columns of the image data matrix (large increases or decreases in intensity from one pixel to the next). The position correction facilitates the comparison of the captured image data of the matrix with the corresponding target values in order to determine defects or to determine the quality score, since the object can not always be in exactly the same position in the stationary state. In one embodiment, the position correction can also be used to determine the positioning (position) of each detected defect, in particular on the top side, optionally including the upper tab surface. On the basis of this positioning information, a marking device downstream of the inspection device can mark the defects, for example by applying (e.g. spraying) water-soluble paint around the defects on the object surface. Alternatively or additionally, knowledge of the positioning of the defects can make it easier to control the device for removing the defects.

[0098] In one embodiment, image data from at least one further camera, for example attached to a frame below the matrix camera, can be included in the inspection method by means of the data processing unit. In this way, it observes the top side of the object, for example the upper tab surface of the tab, and optionally adjacent parts, from above. The at least one further camera can generate, for example, images with a higher resolution in specified sections of the object. The image data obtained from the corresponding field of view are transmitted to the data processing unit and thereby yield further information about smaller defects from these parts.

[0099] The method for inspecting an object can be implemented as a computer-implemented method based on the captured image data, i.e. as a method executed with a data processing unit (computer). The method can also comprise controlling the line illumination unit and / or the area illumination unit and / or the matrix camera such that a predetermined recording and / or illumination sequence is implemented. For this purpose, the data processing unit and the line illumination unit and / or the area illumination unit are connected to each other by wired or wireless means. The matrix camera is also connected to the data processing unit by wired or wireless connection, also for transmitting the image data captured by the matrix camera to the data processing unit.

[0100] The data processing unit for processing the image data and determining whether at least one defect type of a defect is present and / or determining which quality score can be assigned to the object comprises a processor which is a functional module that interprets and executes instructions / commands of an algorithm and comprises a command control unit as well as an arithmetic unit and a logic unit. The processor can comprise at least a microprocessor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA - a digital integrated circuit into which logic circuits can be programmed), a discrete logic circuit, or any combination of these components. The data processing unit can also comprise a memory unit, an input module (e.g. a keyboard or touchpad), a power module (e.g. a battery) and a display module (e.g. a display). The data processing unit can be configured as a real hardware resource, e.g. a smartphone, a desktop computer, a server, a notebook, a cluster / warehouse scale computer, an embedded system, etc. or as a virtualized computer resource. Furthermore, the data processing unit can comprise a transmitter / receiver (transceiver) for exchanging data / image data with a display device (display). The data processing unit further comprises an interface for exchanging data with the line illumination unit and / or the area illumination unit and / or the matrix camera and / or a control device for the drive unit.

[0101] As already indicated above, the above explained method can for example be implemented as a computer program or computer-implemented method comprising instructions which, when executed, cause the processor of the data processing unit to perform the steps of the above described method, wherein the computer program comprises a combination of the above described steps and data definitions which enable the computer hardware to perform a computing or control function and / or are syntactic units which are in compliance with the rules of a specific programming language and consist of declarations and statements or instructions which are required by the above described function, task or problem solution.

[0102] It is further disclosed a computer program product comprising instructions which, when executed by the processor of the data processing unit, cause the apparatus to perform the steps of any or all of the above defined methods. Thus, a computer readable medium storing such a computer program product is disclosed. The computer program product can be a software routine.

[0103] The above objects are also solved by a device for inspecting three-dimensional objects, such as pouch cells, wherein each object comprises a substantially pouch-like or cuboid-like housing having a top side and a bottom side,

[0104] wherein the top side of the housing consists of at least one upper surface section and a plurality of side surface sections which extend obliquely, parallel or perpendicularly with respect to the at least one upper surface section or represent corner sections,

[0105] wherein the bottom side of the housing consists of at least one lower surface section and a plurality of side surface sections which extend obliquely, parallel or perpendicularly with respect to the at least one lower surface section or represent corner sections,

[0106] The device comprises a matrix camera which generates, for each object, in a stationary state of the object to be inspected, image data of the light reflected from the top side of the area illumination unit in a matrix-like manner and transfers the image data to a data processing unit, wherein the image data captured in a matrix-like manner comprises light reflected from the side surface portions,

[0107] wherein the data processing unit is configured to further process the image data captured in a matrix-like manner as a first overall matrix and to perform the following steps:

[0108] segmenting the first overall matrix into:

[0109] a first image data portion comprising image data of the upper surface section, and

[0110] at least one second image data portion, wherein each second image data portion comprises image data of at least one predetermined section of the side surface sections and / or at least one predetermined corner section,

[0111] subdividing the first image data portion into a plurality of individual blocks,

[0112] identifying a defect type of the detected defect and / or a severity of the detected defect and / or determining a quality score allowing to assess a quality of the object based on:

[0113] determining for each block of the plurality of blocks, respectively, whether the respective block of the first image data portion comprises one or more anomalies by a correspondingly trained first NN algorithm, wherein a defect is identified if there are anomalies, and

[0114] identifying, by a correspondingly trained second NN algorithm different from the first NN algorithm, whether there are defects in the at least one second image data portion and classifying the respective second image data portion accordingly.

[0115] In one embodiment of the above-described apparatus, a camera, for example a matrix camera, is provided, which is configured to generate a plurality of line-by-line captured image data of the reflected light of the line illumination unit at the linear area of the top side in the state of motion of the object to be inspected and to transmit it to the data processing unit,

[0116] wherein the data processing unit is configured to perform the following steps:

[0117] merging the line-by-line captured image data into a second overall matrix of image data comprising the top side of the object,

[0118] identifying a defect type of the detected defect and / or a severity of the detected defect and / or determining a quality score allowing to assess a quality of the object based on the image data of the second overall matrix.

[0119] In another embodiment of the apparatus, the data processing unit is configured to perform the following steps:

[0120] segmenting the second overall matrix into:

[0121] a third image data portion comprising image data of the upper surface section, and / or

[0122] at least one fourth image data portion, wherein each fourth image data portion comprises image data of at least one predetermined section of the side surface sections and / or at least one predetermined corner section,

[0123] subdividing the third image data portion into a plurality of individual blocks,

[0124] wherein a defect type of the detected defect and / or a severity of the detected defect and / or a quality score allowing to assess a quality of the object is identified based on:

[0125] determining by a first NN algorithm for each block of the plurality of blocks individually whether the respective block of the third image data portion comprises one or more anomalies, wherein a defect is detected if there are anomalies, and / or

[0126] identifying by a second NN algorithm whether a defect is present in the at least one fourth image data portion and classifying the respective fourth image data portion accordingly.

[0127] In one embodiment of the apparatus, the data processing unit is configured to determine at least one dimension of the object and / or at least one size of the detected defect after taking into account a perspective and / or optical distortion of the matrix camera.

[0128] With regard to the above-described embodiments of the apparatus, reference is made to the above description of the method and apparatus. Further embodiments and advantages thereof are given here, which are also considered to be disclosed with regard to the respective apparatus. BRIEF DESCRIPTION OF DRAWINGS

[0129] Further advantages, features and possible applications of the present application are described below with reference to embodiments and the accompanying drawings. All features described and / or illustrated are subject matter of the present application, even independently of their summary in the claims and references thereof.

[0130] is schematically shown:

[0131] Figure 1 is a first perspective view of a side of an embodiment of the device for the method of the present application,

[0132] Figure 2 is a second perspective view of a side of the device according to Figure 1 is a side view of the device according to

[0133] Figure 3 is a side view of the device according to Figure 1 with the edge rays of the illumination device and the edge rays of the field of view of the matrix camera,

[0134] Figure 4 is a front view of the device according to Figure 1 with the edge rays of the illumination device and the central rays of the field of view of the matrix camera,

[0135] Figure 5 is a front view of the device according to Figure 1 with the edge rays of the illumination device and the edge rays of the field of view of the matrix camera,

[0136] Figure 6 is a lower cross-section of the device according to Figure 1 with the field of view of the matrix camera viewed from above,

[0137] Figure 7 is a perspective side view of an embodiment of the system consisting of two consecutively arranged devices according to Figure 1

[0138] Figure 8 is a perspective side view of a first example of a design of an object (soft pack battery),

[0139] Figure 9 is a perspective side view of a second example of a design of an object (soft pack battery), and

[0140] Figure 10 is an embodiment of the method of the present application for inspection as a flow chart.

[0141] The device for inspection shown in Figures 1 to 6 is used for inspecting an object in the form of a soft pack battery, for example. ​

[0142] In Figure 8 and Figure 9 two examples of pouch cells 11, 111 are shown.

[0143] The pouch cell 11 (see Figure 8 ) comprises a substantially bag-like housing 12. The housing 12 comprises a first connecting tab (short: tab) 14 protruding from one short side and a second connecting tab (short: tab) 15 protruding from an opposite second short side. The bag-like housing 12 comprises a top side with a substantially horizontally extending upper surface section 13. The first tab 14 has an upper tab surface 17 and the second tab 15 has a corresponding upper tab surface 18. Corresponding lower tab surfaces of the tabs 14, 15 are not visible in Figure 8 . The shape of the housing 12 is substantially cuboid. The housing 12 comprises side surfaces 21, 22 on the short sides and side surfaces 23, 24 on the long sides. The side surfaces 21, 22, 23, 24 extend approximately perpendicular to the horizontal upper surface section 13. The horizontal upper surface section 13, the upper tab surfaces 17, 18 of the tabs 14, 15 and the side surfaces 21, 22, 23, 24 together form a top side of the housing 12. A bottom side is correspondingly shaped and comprises a horizontal lower surface section, the lower tab surfaces of the tabs 14, 15 and the side surfaces 21, 22, 23, 24. In this embodiment, the side surfaces 21, 22, 23, 24 belong to both the top side and the bottom side, as these can also be detected when inspecting the top side and the bottom side, respectively. In Figure 8 , the length of the housing 12 is denoted as L (not including the tabs 14, 15) and the width as B (see double arrowed dashed line). Due to the simple design of the pouch cell 11, this pouch cell is used to illustrate the operational mode of the inspection device 1 (see Figures 1 to 6 ). However, correspondingly, the inspection device 1 can also be used for other types of objects, in particular objects in the form of pouch cells.

[0144] Figure 9 A second example of a pouch cell 111 is shown, which comprises a substantially bag-like housing 112 with a horizontal upper surface section 113. The housing 112 further comprises a first connecting tab 114 and a second connecting tab 115, which tabs are adjacently arranged and protrude from a single short side. The first tab 114 has an upper tab surface 117 and the second tab 115 has an upper tab surface 118.

[0145] The pouch cell 111 further comprises side edges 121, 122, 123, 124 on the housing 112 surrounding the horizontal upper surface section 113, which extend at an angle to the horizontal upper surface section 113 and merge into curves therein. Corners 125 are formed at the transition from one side edge to the adjacent side edge 121, 122, 123, 124. Furthermore, the housing 112 comprises platform sections 126, 127, 128, 129, each of which subsequently adjoins a side edge 121, 122, 123, 124 and extends substantially parallel to the horizontal upper surface section 113. The top side of the housing 112 is formed by the horizontal upper surface section 113, the side edges 121, 122, 123, 124, the corners 125 and the platform sections 126, 127, 128, 129.

[0146] Figures 1 to 6 The shown device 1 for inspecting a pouch cell 11 has a support frame 3 on which a base plate 5 is arranged, which has a first through-going opening 7 and a second through-going opening 8 (see Figure 1 、 Figure 2 and Figure 6 ). Figures 1 to 6 Two of the plurality of pouch cells to be inspected 11, 31 are shown, which are guided through the inspection device 1 below the base plate 5, as indicated by the arrows 11a and 31a. The pouch cell 31 is another pouch cell having a structure as shown in Figure 8 .

[0147] Above the base plate 5, a matrix camera 40 is arranged on the frame 3, which observes the pouch cells 11, 31 through the openings 7, 8 from above. Here, the pouch cells 11, 31 are arranged in such a way that the top side of the housing 12 is in each case at the top and can be observed from above with the matrix camera 40 on the horizontal upper surface section 13. The upper tab surface 17 of the first tab 14 and the upper tab surface 18 of the second tab 15 are also captured by the matrix camera 40. Here, the field of view 42 of the matrix camera 40 is so large (see in particular Figure 6 ) that it extends over both openings 7, 8, so that not only the pouch cells 11, 31 arranged below the respective opening 7, 8 are covered by the field of view 42 of the matrix camera 40 over their entire length and width (viewed from above), but also the deflection mirrors 65, 67 arranged next to the pouch cell 11 are covered.

[0148] Furthermore, a line illumination unit 51 is provided on the frame 3, which illuminates the linear region 31b of the top side of the housing and the upper tab surface of the tab. From Figure 4It can be seen that light reflected from the top side of the housing (including the upper tab surface) reaches the matrix camera 40 via the observation light ray 41 and is captured there. Thus, the matrix camera 40 captures respective illuminated linear areas 31b of the soft-pack battery 31 row by row, wherein the soft-pack battery 31 moves during the capturing transversely to the length of the opening 8 (see arrow 31A in Figure 6 Thus, a plurality of recordings is generated by the matrix camera 40. Each recording comprises a respective capturing of a respective illuminated linear area of the top side of the housing of the soft-pack battery 31 (including the upper tab surface) that moves past. The soft-pack battery is moved in the moving state by a drive unit described in more detail below and is brought into a stationary state for a predetermined period of time and moved out of the inspection device from the stationary state.

[0149] In addition, four area illumination units 52, 53, 55, 56 are provided on the frame. As can be seen from Figure 2 , Figure 3 , Figure 5 and Figure 6 , and as indicated by the edge light rays 52a and 52b or 53a, 53b of the area illumination units 52, 53, the area illumination units 52, 53 obliquely illuminate the top side of the housing 12 of the soft-pack battery 11 (including the upper tab surfaces 17, 18 of the tabs 14, 15) from above over the entire length L + tab length, so that in particular the side opposite the respective area illumination unit 52, 53 (which is arranged below the opening 7 in the base plate 5) is captured with respect to the width of the soft-pack battery 11 (in particular compared to Figure 5 ). From Figure 3It can be seen that the area illumination units 55, 56 illuminate the top side of the housing 12 (including the upper tab surfaces 17, 18 of the tabs 14, 15) via the mirrors 61, 62. Here, the light emitted by the area illumination unit 55 via the mirror 62 illuminates essentially the opposite first end of the pouch cell 11, and the light emitted by the area illumination unit 56 via the mirror 61 illuminates essentially the second end of the pouch cell 11 opposite the first end of the pouch cell (in the longitudinal direction). This can be understood by following the edge rays 55a, 55b, 56a and 56b. The area illumination units 52, 53, 55, 56 also partially illuminate the side surfaces 21, 22, 23, 24, so that reflections from these side surfaces and from the horizontally extending upper surface section 13 of the housing 12 are captured by the matrix camera 40. The light reflected from the side surfaces 21, 22, 23, 24 by the area illumination units 52, 53, 55, 56 is captured, inter alia, via the deflection mirrors 65, 67, which are arranged next to the pouch cell 11 below the opening 7, so that the side surfaces 23, 24 on the long sides of the housing 12 of the pouch cell 11 are observed by the deflection mirror 65, while the deflection mirror 67 serves to capture the side surfaces 21, 22 on the short sides of the housing 12 of the pouch cell 11.

[0150] The illumination by the area illumination units 52, 53, 55, 56 now takes place in such a way that these area illumination units are each switched on one after the other, so that one area illumination unit respectively obliquely illuminates the pouch cell 11 from above, while the other three area illumination units are each switched off. For example, the illumination is first provided by the area illumination unit 52, then by the area illumination unit 55, then by the area illumination unit 53 and finally by the area illumination unit 56. The matrix camera 40 captures the reflected light in each of the four illumination states, in which the pouch cell 11 is in a stationary state in the same predetermined position below the opening 7 in the base plate 5 in all four illumination states. Thus, four recordings of the entire top side of the housing 12 (including the upper tab surfaces 17, 18) are generated by the matrix camera 40, which captures these areas four times in a matrix, namely once when the area illumination unit 52, the area illumination unit 55, the area illumination unit 53 and the area illumination unit 56 are switched on, with the pouch cell 11 in each case being in the same position.

[0151] In the four matrix-like captures of the entire top side of the housing 12 (including the upper tab surfaces 17, 18), the side surfaces 21, 22, 23, 24 are also captured via the deflection mirrors 65, 67, since the light reflected from these side surfaces 21, 22, 23, 24 reaches the matrix camera 40 via the deflection mirrors 65, 67, since the field of view 42 of the matrix camera 40 includes these areas, as is Figure 6 shown.

[0152] The multiple line-by-line capturing and matrix-by-matrix capturing of the pouch cells 11, 31 by the matrix camera 40 are transmitted to a data processing unit (computer) 70 (see Fig. 1) after they have been captured. The data processing unit 70 receives the image data from the line-by-line capturing and matrix-by-matrix capturing of the respective pouch cell 11, 31. Figure 1 ). The data processing unit 70 receives the image data from the line-by-line capturing and matrix-by-matrix capturing of the respective pouch cell 11, 31.

[0153] In this regard, the motion state and the rest state of the pouch cell 11, 31 are captured by a motion detection unit of the inspection device. For example, by monitoring the movement of a predetermined marker (e.g. a barcode) on the pouch cell 11, 31 or a corresponding signal from the drive unit, motion data are generated which, in particular, contain information about the respective motion state in which the respective pouch cell 11, 31 is located. For example, the motion unit can transmit information to the motion detection unit that the pouch cell is ready for inspection (start signal). From this point in time, the motion detection unit can continuously capture motion data of the motion unit and thus of the respective pouch cell (e.g. the movement speed of the motion unit). Alternatively or additionally, the motion detection unit receives a signal when the line-by-line capturing of the respective pouch cell is complete. Subsequently, when the respective pouch cell is in a predetermined position in the rest state and remains stationary there, a signal is generated by the motion unit or by means of a marker of the pouch cell and transmitted to the motion detection unit. As soon as the matrix capturing of the respective pouch cell has been completed, the motion detection unit then generates a further signal indicating that the respective pouch cell can be transported out of the inspection device. These signals also represent important motion data required for processing the image data.

[0154] These image data and motion data are further processed and analyzed by the data processing unit 70 and, as will be shown in more detail hereinafter, the object is evaluated considering the presence of at least one defect type of defects and / or determining a quality score, which allows to evaluate the quality of the pouch cells 11, 31. In this context, the image data determined at different points in time from the line-by-line capturing and matrix-by-matrix capturing are assigned to the respective pouch cell 11, 31 or to the motion state and the rest state. This can be done, for example, using the motion data transmitted by the motion detection unit for the motion state and the rest state of the respective pouch cell 11, 31. In this process, the recorded image data are generated with regard to image cropping, mirror distortion, line-by-line capturing (so-called line scan correction) and with regard to position correction associated with the respective motion state or rest state.

[0155] The image data of the line-by-line capturing of the respective pouch cell 11, 31 can be used, for example, to search for defects using bright field illumination (reflected bright field (RBF) illumination) or to determine a quality score, which in particular relates to the evaluation of the tab surface and possible contamination of the electrolyte.

[0156] For example, as explained in more detail below, the data processing unit 70 generates a quality rating (quality score) from the number, type and size / size of the detected defects.

[0157] Further cameras 45, 46 are also arranged on the frame 3, which cameras are attached to the frame 3 below the matrix camera. These cameras observe the top side of the pouch cell 11 from above in such a way (see edge light 45a, 45b, 46a, 46b) that they observe the top tab surface 17 of the first tab 14 and the top tab surface 18 of the second tab 15 and, if applicable, the adjacent portion of the upper surface section of the top side 13. The further cameras 45, 46 generate images with a higher resolution in the indicated portion of the pouch cell 11. The image data obtained from the corresponding field of view are transmitted to the data processing unit 70 and from there further information about the presence of smaller defects is generated.

[0158] Figure 7 A system for inspecting pouch cells according to the application is shown. The top side of a pouch cell (e.g. pouch cell 11) is first inspected by the inspection device 1. Subsequently, the pouch cell (e.g. pouch cell 11) reaches a flipping unit 180 with a flipping device and grippers with suction cups, which flips the pouch cell (e.g. pouch cell 11) so that the bottom side is now on top. Subsequently, the pouch cell (e.g. pouch cell 11), i.e. the bottom side of the cell (here above), is inspected by an inspection device 101 which is identical in construction to the inspection device 1. The processing of the image data obtained by the inspection devices 1, 101 about the pouch cell (e.g. pouch cell 11) and the determination from the image data about the presence of defects / numerous defects of at least one defect type and / or the determination of a quality score which allows an assessment of the quality of the object is performed by a data processing unit 170 which is connected with a display 172 for displaying the inspection results. The pouch cell (e.g. pouch cell 11) is transported from the first inspection device 1 to the flipping device 180 and the second inspection device 101 by a drive unit which comprises e.g. a slide displaceable on a linear unit. The respective pouch cell is attached to the slide by a suction cup. After the capture of the image data by the matrix camera 40 and optionally by the further cameras 45, 46, in particular after the matrix-like capture of the respective pouch cell is completed, a corresponding signal is generated by the inspection device, which signal is transmitted to a control device of the drive unit. The drive unit is then controlled in such a way that it moves the respective pouch cell out of the respective inspection device 1, 101 and, if necessary, moves the pouch cell to the flipping device 180 in order to be flipped subsequently and then transported to the second inspection device 101.

[0159] For example, the analysis of the captured image data to determine the presence of defects and / or to determine the quality score can be performed as follows. Using Figure 10 The process is illustrated using the flow chart shown in Fig. 2.

[0160] The starting point for the analysis of the captured image data is the four matrix-wise captured image data 200 of the top side of the object generated by illumination from different illumination directions and the line-by-line captured image data 201 of the top side of the object using a pouch cell 111 as an example.

[0161] As described above, first in step 202 the perspective and / or optical distortion of the matrix camera is compensated for each of the four matrix-wise captured image data. Subsequently, if necessary, in step 204 (position correction) the current position of the object during the optical capturing by the matrix camera is corrected, i.e. the object is rotated and / or shifted such that the image data occupies a predetermined position of the object in the field of view of the matrix camera. At the same time, in step 203 the plurality of single line-by-line captured image data 201 is merged into a single image (data) by the data processing unit 70, wherein the single image contains the matrix of image data and, as described above, inconsistencies are corrected during the merging of the image of the top side of the object from the line-by-line captured image data. Subsequently, as described above, this data is also corrected with respect to its position in step 204. The merged and corrected line-by-line captured image data forms a second overall matrix.

[0162] For the four matrix-wise captured, compensated and corrected image data, subsequently in step 206 the maximum image, the absorption image and / or the topological image of the top side of the object is determined. The calculation from the four matrices of image data is described in detail above. The maximum image is also referred to as the first overall matrix.

[0163] Now, in step 210 the segmentation is performed. As explained in detail above, a layout recipe can be used to extract the desired image data portions from the respective (corrected) image data matrices of the top side determined by the matrix-wise capturing or the line-by-line capturing. For example, a first image data portion in the form of the upper surface section 113 (directly recorded image data from the camera / matrix camera) is extracted, a second image data portion in the form of two corner sections 125 from the recording area via a mirror on the short side is extracted, and a further second image data portion in the form of four platform sections 126, 127, 128, 129 (directly recorded image data from the camera / matrix camera) is extracted. The segmentation is performed for the corrected and merged line-by-line captured image data (i.e. from the second overall matrix) and for the matrix-wise captured, compensated and corrected image data and the maximum image (first overall matrix) and / or the absorption image and / or the topological image. Subsequently, the image data portions obtained by the segmentation are processed in parallel and finally fed to the overall evaluation of the object.

[0164] In step 230, the result of the segmentation is, for example, the image data portion of the upper surface in the maximum image, the absorption image, the topography image and the second overall matrix, respectively. In step 232, the data processing unit 70 checks each of these image data portions for each pixel to determine whether they exceed a predetermined threshold. Such a threshold can be 240 for the image data portion of the maximum image, 220 for the image data portion of the topography image, 203 for the image data portion of the absorption image and 120 for the image data portion of the second overall matrix. If the image data value of the respective pixel is equal to or higher than the respective threshold, a defect is detected. Subsequently, in step 234, further characteristics of the detected defect are determined, for example its pixel size (by analyzing whether a defect is also detected in neighboring pixels), a histogram of the image data values in the area of the respective defect and / or the shape and orientation of the defect. The characteristics of the found defect and any other defects found in the image data portion are then used in step 236 to determine the defect type, wherein different images / matrices related to the same location of the image data portion can be considered for this purpose. For example, based on the absence of a defect at the location in the absorption image, based on a determined aspect ratio in the absorption image being greater than 5 and based on an average value of the image data histogram along the defect in the topography image being greater than 200, the defect type “scratch” can be inferred together. In step 238, the severity of the detected scratch defect is then determined, wherein the determination can be based on, for example, an assignment of a severity level to the size of the scratch. In this context, a scratch defect can be assigned a severity level of zero if the defect is smaller than 2 pixels, a severity level of 1 if the defect is greater than or equal to 2 pixels and smaller than 4 pixels, a severity level of 2 if the defect is greater than or equal to 4 pixels and smaller than 6 pixels, and so on.

[0165] The result of the segmentation in step 240 is, for example, image data portions in the form of four “corner segments” (e.g. 128x128 pixels) from the maximum image determined in step 206, wherein the “corner segments” are obtained, for example, from the image data generated via the mirror 67 on the short side of the soft pack. As described above, a CNN algorithm with a binary classifier is now applied to these image data portions in step 242. Accordingly, the attribute “defective corner” (step 244) or “complete corner” (step 246) is determined for each corner image data portion and assigned to the respective corner in step 248.

[0166] The segmentation result in step 250 is, for example, an image data portion in the form of a table-like portion (tables 126, 127, 128, 129) of the maximum image determined in step 206, wherein the image data has been generated by direct capturing from above by means of the matrix camera 40. Now, the mask RCNN algorithm is applied to these table image data portions (described in more detail above, step 252). Thus, defects in the table image data portions of different defect types are identified and provided with a bounding box (step 254). Subsequently, the severity of the respective defect is also determined with respect to the defects detected in the table image data portion, for example based on the detected defect type, the size of the bounding box, the shape of the bounding box, etc. (step 256).

[0167] For example, in step 260, the result of the segmentation is an image data portion in the form of an upper surface section 113, for example a maximum image, and in the form of a second overall matrix (from the line-by-line capturing). In step 262, these image data portions are divided into blocks, wherein, if applicable, the resolution of the respective portion has previously been reduced to a predefined value (see above, for example). Subsequently, in step 264, the image data portions can be analyzed using the pre-trained CNN algorithm "Wide ResNet-50" as described above, and, if applicable, one or more anomalies in some of the blocks can be detected. In step 266, the Mahalanobis distance from the relative normal distribution is determined for each block in which an anomaly was detected, and each detected anomaly is determined. Thereby, in step 268, the severity of the anomaly and thus of the respective defect is determined.

[0168] In all of the above cases, the severity of the defect is expressed in predetermined classes.

[0169] Subsequently, in step 270, the data processing unit 70 evaluates the overall quality of the soft pack battery 111 based on all defects determined in the four analysis chains, the respective defect types and the respective defect severities. It is assessed whether the soft pack battery 111 as a whole meets the specified quality requirements. In step 280, the result of the overall evaluation is provided at the interface of the data processing unit, optionally together with a list of the determined defects and their properties. For example, a soft pack battery with two defects of the "dimple" type with a severity rating of 5 is judged to be sufficient to meet the quality requirements. In contrast, for example, a soft pack battery with a defect of the defect type "dimple" with a severity rating of 7 can be classified as not meeting the quality requirements.

[0170] The above-described method can also be carried out analogously for the bottom side of the soft pack battery.

[0171] As mentioned above, the method according to the present application can be used to perform an inspection of a three-dimensional object (e.g. a pouch battery) in a simple and fast manner, wherein various properties of the segments of the object can be taken into account during the analysis.

Claims

1. A method of inspecting three-dimensional objects, such as pouch batteries (11, 31, 111), wherein each object comprises a substantially bag-like or cuboid-like housing having a top side and a bottom side, wherein the top side of the housing (12, 112) consists of at least one upper surface section (13, 113) and a plurality of side surface sections (17, 18, 21, 22, 23, 24, 117, 118, 121, 122, 123, 124, 125, 126, 127, 128, 129) extending obliquely, parallel or perpendicularly with respect to the at least one upper surface section (13, 113) or forming a corner section (125), wherein for each object, image data captured in a matrix fashion by a matrix camera is generated from light reflected from the top side in accordance with an area illumination unit in a stationary state of the object to be inspected and is transmitted to a data processing unit (70), wherein the image data captured in a matrix fashion comprises light reflected from the side surface sections, wherein the image data captured in a matrix fashion is further processed by the data processing unit into a first overall matrix, wherein the following steps are further performed by the data processing unit (70): segmenting the first overall matrix into: a first image data portion comprising image data of the upper surface section, and at least one second image data portion, wherein each second image data portion comprises image data of at least one predetermined portion of the side surface sections and / or at least one predetermined corner section, subdividing the first image data portion into a plurality of individual blocks, identifying a defect type of a detected defect and / or a severity of a detected defect and / or determining a quality score allowing to assess a quality of the object based on: a specific determination of whether a respective block of the first image data portion comprises one or more anomalies for each block of the plurality of blocks by a correspondingly trained first NN algorithm, wherein a defect is identified if there are anomalies, and an identification of whether there are defects in the at least one second image data portion and a corresponding classification of the respective second image data portion by a correspondingly trained second NN algorithm different from the first NN algorithm.

2. The method of claim 1, wherein, The classification of the at least one second image data portion is performed by a classifier having two states or by a classifier having at least three states, wherein the classifier having at least three states allows for example to assign different types of defects.

3. The method according to any of the preceding claims, characterized in that, for each object, a plurality of line-wise captured image data of reflected light of an area illumination unit at a linear area of the top side in a moving state of the object to be inspected is generated by a camera, such as by the matrix camera, and is transmitted to a data processing unit (70), wherein the following steps are further performed by the data processing unit (70): merging the line-wise captured image data into a second overall matrix comprising image data of the top side of the object, Additionally based on the image data of the second overall matrix, a defect type of the detected defect and / or a severity of the detected defect and / or a quality score is determined which allows to assess the quality of the object.

4. The method of claim 3, wherein, The following further steps are performed by the data processing unit (70): The second overall matrix is split into: a third image data portion comprising image data of the upper surface section, and / or at least one fourth image data portion, wherein each fourth image data portion comprises image data of at least one predetermined section of the side surface section and / or at least one predetermined corner section, the third image data portion is subdivided into a plurality of individual blocks, wherein, based on the following, a defect type of the detected defect and / or a severity of the detected defect and / or a quality score is determined which allows to assess the quality of the object: by means of the first NN algorithm, it is specifically determined for each block of the plurality of blocks whether the respective block of the third image data portion comprises one or more anomalies, wherein, if there are anomalies, a defect is identified, and / or by means of the second NN algorithm, it is identified whether a defect is present in the at least one fourth image data portion and the respective fourth image data portion is classified accordingly.

5. The method according to any of the preceding claims, characterized in that, For each object, at least n (n > 2) recordings of the top side of the matrix- wise captured image data are generated by capturing the reflected light consecutively in time and are transmitted to the data processing unit (70), wherein the matrix- wise captured image data is further processed by the data processing unit into n first overall matrices, wherein the reflected light is generated by obliquely illuminating the top side of the shell of the object to be inspected from above separately in time from n different directions in a stationary state of the object.

6. The method of claim 5, wherein, The following further steps are performed by the data processing unit (70): The n first overall matrices are split into: n fifth image data portions comprising image data of the upper surface section of each of the n first overall matrices, and / or n sixth image data portions and optionally further image data portions of each of the n first overall matrices, wherein each sixth image data portion and optionally further image data portion comprises image data of at least one predetermined portion of the side surface section and / or at least one predetermined corner section, in each case, a maximum image and / or an absorption image and / or a topological image is determined from the image data of the fifth image data portion and / or the sixth image data portion and / or the possible further image data portion, if applicable, defects are determined in the maximum image and / or the absorption image and / or the topological image of the fifth image data portion and / or the sixth image data portion and / or the further image data portion and the defects are analyzed and characterized, wherein the identification of a defect type of the detected defect and / or a severity of the detected defect and / or the determination of a quality score which allows to assess the quality of the object is based on the results of the analysis and / or the characterization of the respective detected defect.

7. The method according to any of the preceding claims, characterized in that, A block-wise distribution modeling framework for anomaly detection is used to detect anomalies in blocks, wherein the degree of anomaly is determined by Mahalanobis distance with respect to an expected normal distribution in the respective block.

8. The method according to any of the preceding claims, characterized in that, The localization of anomalies in the respective block is determined and used to localize possible defects in / on the object.

9. The method according to any of the preceding claims, characterized in that, At least one dimension of the object and / or at least one size of the detected defects is determined by the data processing unit after taking into account the perspective distortion and / or the optical distortion of the matrix camera.

10. The method according to any of the preceding claims, characterized in that, A position correction is performed by the data processing unit by means of a predetermined reference point of the object.

11. The method according to any of the preceding claims, characterized in that, The resolution of the first image data portion and / or the third image data portion is reduced before subdividing the first image data portion and / or the third image data portion into a plurality of individual blocks.

12. A device (1, 101) for inspecting three-dimensional objects, such as pouch cells (11, 31, 111), wherein each object comprises a substantially bag-like or cuboid-like housing having a top side and a bottom side, wherein The top side of the housing (12, 112) consists of at least one upper surface section (13, 113) and a plurality of side surface sections (17, 18, 21, 22, 23, 24, 117, 118, 121, 122, 123, 124, 125, 126, 127, 128, 129) extending obliquely, parallel or perpendicularly with respect to the at least one upper surface section (13, 113) or presenting a corner section (125), The device comprises a matrix camera generating, for each object, image data of the light reflected from the top side of a regional illumination unit in a stationary state of the object to be inspected in a matrix-like manner and transmitting this image data to a data processing unit (70), wherein the image data captured in a matrix-like manner comprises light reflected from the side surface sections, wherein the data processing unit (70) is configured to further process the image data captured in a matrix-like manner as a first overall matrix and to perform the following steps: segmenting the first overall matrix into: a first image data portion comprising image data of the upper surface section, and at least one second image data portion, wherein each second image data portion comprises image data of at least one predetermined section of the side surface sections and / or at least one predetermined corner section, subdividing the first image data portion into a plurality of individual blocks, identifying a defect type of the detected defects and / or a severity of the detected defects and / or determining a quality score allowing to assess the quality of the object based on: determining, by a correspondingly trained first NN algorithm, for each block of the plurality of blocks, whether the respective block of the first image data portion comprises one or more anomalies, respectively, wherein a defect is identified if there is an anomaly, and identifying, by a correspondingly trained second NN algorithm different from the first NN algorithm, whether a defect is present in the at least one second image data portion and classifying the respective second image data portion accordingly.

13. The apparatus of claim 12, wherein, A camera, e.g. the matrix camera, is provided, which is configured to generate a plurality of line-by-line captured image data of the reflected light of a line lighting unit at a linear area of the top side in a motion state of an object to be inspected and to transfer this image data to a data processing unit (70), wherein the data processing unit is configured to perform the following steps: merging the line-by-line captured image data into a second overall matrix of image data comprising the top side of the object, identifying a defect type of the detected defect and / or a severity of the detected defect and / or determining a quality score allowing to assess a quality of the object based on the image data of the second overall matrix.

14. The apparatus of claim 13, wherein, The data processing unit is configured to perform the following steps: segmenting the second overall matrix into: a third image data portion comprising image data of the upper surface section, and / or at least one fourth image data portion, wherein each fourth image data portion comprises image data of at least one predetermined section of the side surface section and / or at least one predetermined corner section, subdividing the third image data portion into a plurality of individual blocks, wherein a defect type of the detected defect and / or a severity of the detected defect and / or a quality score allowing to assess a quality of the object is identified based on: determining, by means of the first NN algorithm, for each block of the plurality of blocks, whether the respective block of the third image data portion comprises one or more anomalies, wherein a defect is detected if there are anomalies, and / or identifying, by means of the second NN algorithm, whether a defect is present in the at least one fourth image data portion and classifying the respective fourth image data portion accordingly.

15. The apparatus of any one of claims 12-14, wherein, The data processing unit is configured to determine at least one dimension of the object and / or at least one size of the detected defect after taking into account a perspective distortion and / or an optical distortion of the matrix camera.

Images