Methods and systems for providing synthetic labeled training datasets and applications thereof

Abstract

The invention relates to a computer-implemented method for providing a labeled training dataset, wherein - at least one sub-object (21, 22, 23, 24, 25, 26) is selected in a CAD model (1) of an object (2) comprising a plurality of sub-objects, - a plurality of different rendered images (45, 46, 47, 48) is generated, wherein the different rendered images (45, 46, 47, 48) comprise the at least one selected sub-object (21, 22, 23, 24, 25, 26), - the different rendered images are labeled based on the CAD model in order to provide a training dataset (49) of labeled rendered images.

Description

Technical Field

[0001] This invention relates to a computer-implemented method for providing a labeled training dataset.

[0002] The present invention also relates to a system for providing such a training dataset.

[0003] Furthermore, the present invention relates to a computer-implemented method for providing a training function trained on the aforementioned training dataset, and a computer-implemented image recognition method using such trained function.

[0004] Furthermore, the present invention relates to a system for image recognition, comprising an image recording device and a data processing system designed and configured to perform the image recognition method described above.

[0005] Furthermore, the present invention relates to a computer program comprising a first instruction causing the aforementioned system to execute a method for providing a labeled training dataset, and / or a second instruction causing, when the program is executed by a computer, to execute the aforementioned computer-implemented method for providing trained functionality, and / or a third instruction causing the aforementioned system for image recognition to execute the aforementioned image recognition method. Background Technology

[0006] Creating labeled datasets is a known problem. Large labeled datasets are typically needed to train features or AI algorithms, such as neural networks for image or object recognition applications.

[0007] Especially in industrial environments, it is difficult to generate a balanced database (with an equal proportion of good and bad data). Intentionally creating errors requires more time. Furthermore, unforeseen environmental impacts can only be mapped onto the dataset to a limited extent.

[0008] One possible method for generating labeled datasets is to perform the labeling manually. Manual labeling (e.g., using Google) is very time-consuming and expensive. Furthermore, errors are prone to occur due to lack of focus and monotony. The technical field of labeling requires expertise in components and processes, and is correspondingly more expensive. Summary of the Invention

[0009] Therefore, the object of the present invention can be regarded as providing a method and system for creating labeled datasets of any size with little time and cost.

[0010] According to the present invention, this objective is achieved by the computer-implemented method mentioned at the beginning for providing a labeled training dataset, wherein

[0011] - Select at least one sub-object from a CAD model that includes a large number of sub-objects, wherein the CAD model includes a description of at least one sub-object and the coordinates of the at least one sub-object (in the image of the object).

[0012] - Generate a large number of different rendered images, wherein the (synthesized) different rendered images preferably each include at least one selected sub-object.

[0013] - Label different rendered images based on the description of at least one sub-object contained in the CAD model and the coordinates of at least one sub-object contained in the CAD model, so as to provide a training dataset of labeled rendered images.

[0014] By using the method according to the invention, the overhead involved in manual labeling is eliminated and large labeled datasets can be created quickly and without errors.

[0015] Marking based on CAD models is preferably performed automatically.

[0016] The CAD model of the (overall) object therefore contains information about the coordinates of its individual sub-objects, and preferably also the relationships between them. The coordinates of each sub-object allow us to determine its position within the overall object.

[0017] Using the training dataset labeled as described above, it is possible to train, for example, a function for image recognition methods that can check whether the correct real sub-objects of a real (whole) object are in the correct positions.

[0018] In the context of this invention, the term "rendered image" is understood to be, for example, a virtual (near-realistic) three-dimensional representation of an object generated by means of pre-computation on a computer.

[0019] The aforementioned CAD model can be two-dimensional or three-dimensional (2D or 3D). Here, the (application-related) scene or environment can be used as an object comprising a large number of sub-objects. The application-related environment / scene is understood here as a scene / environment containing sub-objects related to a selected application situation. For example, in the case of industrial facility construction (an example of an application situation), the object can be the entire industrial facility, where sub-objects can be areas of the industrial facility. For example, in automobile production, the object can be the vehicle body, where sub-objects are parts of the vehicle body. In the field of autonomous driving, such as trains, the object can be, for example, the environment, which is typically visible from the train driver's cab or within the driver's field of vision, where sub-objects are instruments / dashboards, front-mounted signaling systems, such as H / V signaling systems, etc.

[0020] Depending on the level of detail in the CAD model, a sub-object can contain even more sub-sub-objects.

[0021] For example, it is also possible to consider selecting two, three, four, five, six, or all sub-objects, such that the selected sub-objects form the entire scene represented by that object. This generates a (labeled) training dataset, which can be used to train a function for classifying images as a whole, as described below. Here, classification can be a good / bad evaluation of the complete image.

[0022] The labeling of rendered images can be performed, for example, based on data available from the CAD model (such as material number, material, related process steps, etc.).

[0023] One design proposal suggests storing rendered images in, for example, a database (such as a cloud database or a cloud data lake).

[0024] Advantageously, it can be proposed that the label of at least one (selected) sub-object on the rendered image has a description of at least one sub-object and the coordinates of at least one sub-object (the position X, Y of the sub-object in the image). The labels can be generated, for example, in the form of a list: object 1 at position (X1, Y1), object 2 at position (X2, Y2), ..., object 3 at position (X4, Y5), ..., object 4 at position (X6, Y7), ..., object 5 at position (X8, Y9), ..., object 6 at position (X1, Y1), ..., object 7 at position (X8, Y9), ..., object 8 at position (X1, Y1), ..., object 9 at position (X1, Y1), ..., object 1 2 at position (X2, Y1), ..., object 1 2 ... N Y N Object N at location ).

[0025] For example, the description of a sub-object can include information about the sub-object's type, properties, functionality, etc.

[0026] Furthermore, the labels can be visualized, for example, in the form of rectangular borders, preferably in preset colors. For example, visualization allows installers to quickly check the results.

[0027] If the training data record includes multiple real images of the object, further advantages can be obtained, wherein the number of real images is smaller than the number of rendered images, and is, for example, about 0.1% to about 5% of the number of rendered images, particularly about 1%.

[0028] Furthermore, it can be proposed that sub-objects are constructed separately from each other.

[0029] Furthermore, it is possible to propose that (all or only some) sub-objects be constructed differently from each other.

[0030] In one embodiment, it can be proposed that the object is designed as an electronic component, particularly a printed circuit board assembly, wherein the sub-objects are preferably designed as structural elements, particularly integrated or discrete structural elements.

[0031] Here, the bill of materials (BOM) for electronic components can be used to mark different rendered images based on the CAD model. Such a BOM can be stored in the CAD model.

[0032] Advantageously, it is possible to propose that the rendered images of at least one sub-object differ from each other in at least one criterion, wherein the at least one criterion is selected from the group consisting of: the size of at least one sub-object (in the image), the exposure of at least one sub-object, the perspective of at least one sub-object that is visible in the rendered image, the background of at least one sub-object, the position, surface, texture, and color of at least one sub-object.

[0033] Furthermore, it is possible to propose selecting multiple sub-objects in a CAD model, wherein the selected sub-objects can be associated with a process sequence.

[0034] For example, it can perform the allocation of actual processes based on production information or other information from the PLM system.

[0035] According to the present invention, the aforementioned objective is also achieved by the aforementioned system for providing labeled training data records, the system comprising:

[0036] - A computer-readable storage device containing a CAD model of an object that includes multiple sub-objects.

[0037] - The first interface is designed and configured to allow selection of at least one sub-object within the CAD model.

[0038] - A computing unit configured to generate multiple distinct rendered images, each comprising at least one selected sub-object, and labeled based on a CAD model.

[0039] - A second interface, which is designed and configured to provide a training dataset based on rendered images with different labels.

[0040] Furthermore, the aforementioned objective is achieved through a computer-implemented method for providing trained functionality, wherein...

[0041] - Provide at least one labeled training dataset obtained as described above.

[0042] - Train a feature based on a training dataset with at least one label to produce a trained feature.

[0043] - Provides trained functionality.

[0044] If the function is a classification function and / or a function for locating objects, especially if it is based on a convolutional neural network, then it can be effective.

[0045] In addition, a training system is disclosed, comprising: a first training interface device configured to acquire at least one labeled training dataset that can be obtained by the above method; a training computing unit configured to train a function based on the labeled training dataset; and a second training interface device configured to provide the trained function.

[0046] In terms of simplification in the field of image recognition, the aforementioned objective is achieved through a computer-implemented image recognition method, wherein...

[0047] - Provide input data, which is a record of images of real objects, where the real objects include a large number of real sub-objects.

[0048] - The trained function is applied to the input data to generate output data, wherein the trained function is provided according to the aforementioned method, and wherein the output data includes the classification and / or location of at least one real sub-object in the input data.

[0049] - Provides output data.

[0050] In other words, the output data includes dividing the real object into sub-objects.

[0051] Advantageously, it is possible to propose repeating the steps of the method for multiple different image records.

[0052] It is valid if the description is assigned to at least one real sub-object.

[0053] It is possible to provide output data in the form of processed image records, wherein the processed image records are image records with at least one tag having at least one real sub-object.

[0054] Here, it can be proposed that a label has a description of at least one real sub-object and / or the coordinates of at least one sub-object.

[0055] For example, the description of a real sub-object can include information about its type, properties, and functions. Markings can also be visualized in the form of rectangular borders, preferably in a preset color. For example, visualization allows installers to quickly check the results.

[0056] This can be effective if the description includes a bill of materials and a confidence level for identifying sub-objects.

[0057] In one implementation, it can be proposed that the real object is designed as electronic firmware, particularly a printed circuit board assembly, and the sub-objects are designed as electronic structural elements, such as integrated or discrete structural elements.

[0058] In one implementation, the system for image recognition is designed as a system for inspecting printed circuit boards, particularly a printed circuit board inspection system.

[0059] This circuit board inspection system can be used very effectively in the manufacture of printed circuit boards. Here, the availability of knowledge in product development (product structure, bill of materials (parts list)) can automatically generate meaningful training data, as this training data can be generated at the individual component level, and trained functions / algorithms can be trained to identify individual components / structural elements.

[0060] Furthermore, CAD models can be used to assign Bills of Materials (BOMs) to Process Operation Sequences (BOPs). This link enables the automatic generation of test programs for specific processes (based on specially designed training data). Similarly, knowledge of hierarchical structures is utilized here. Attached Figure Description

[0061] The invention will now be described and explained in more detail with reference to the embodiments shown in the accompanying drawings. The drawings show:

[0062] Figure 1 A flowchart is shown showing a computer-implemented method for providing a labeled training dataset.

[0063] Figure 2 A CAD model of the printed circuit board assembly is shown.

[0064] Figures 3 to 6 An intermediate image used to generate the label is shown.

[0065] Figures 7 to 10 The generated rendered image is shown.

[0066] Figure 11 A system for providing a training dataset with synthetic labels is shown.

[0067] Figure 12 The rendered image of the marker is shown.

[0068] Figure 13 The training system is shown.

[0069] Figure 14 A flowchart of a computer-implemented image recognition method is shown.

[0070] Figure 15 It shows that according to Figure 14 Image records processed by image recognition methods

[0071] Figure 16 A data processing system for performing an image recognition method is shown, and

[0072] Figure 17 An automated system is shown.

[0073] In the embodiments and figures, identical or equivalent elements may each have the same reference numerals. The dimensions of the elements shown and their ratios to each other should not be considered as true proportions; rather, individual elements may be displayed at a larger scale for better display and / or better understanding. Detailed Implementation

[0074] First refer to Figures 1 to 2 . Figure 1 A flowchart illustrating an embodiment of a computer-implemented method for providing a labeled training dataset is shown.

[0075] Based on the CAD model of the object, which includes multiple sub-objects, select at least one sub-object—step S1. Figure 2 A CAD model 1 of a printed circuit board assembly 2 on background 3 is shown as an example. The printed circuit board assembly 2 has multiple electronic components. Six structural elements 21, 22, 23, 24, 25, and 26 of the printed circuit board assembly 2 are selected.

[0076] As shown in the figure, these structural components can be structurally separated from each other and designed differently from one another. Figure 2 This indicates that the same structural elements can also be used. One of the structural elements of the same design can be selected to generate the rendered image (see below).

[0077] The selection of structural elements 21, 22, 23, 24, 25, and 26 in CAD model 1 can be completed manually or automatically via a computer's manual operation interface, for example, based on a BOM (Bill of Materials) and / or a BOP (Procedure Format).

[0078] Step S2: Generate multiple different rendered images 45, 46, 47, 48 for the selected structural elements 21, 22, 23, 24, 25, 26. The rendered images are, for example, virtual, preferably realistic, three-dimensional illustrations generated pre-computed on a computer, on which at least one of the selected structural elements 21, 22, 23, 24, 25, 26 can be seen. Preferably, at least one of the selected structural elements 21, 22, 23, 24, 25, 26 can be seen in each rendered image.

[0079] In sub-step ( Figure 1 In (not shown), intermediate images 41, 42, 43, and 44 can be generated. Figures 3 to 6For example, these intermediate images 41, 42, 43, and 44 can be used to generate labels (rectangles with coordinates x and y).

[0080] In each of the intermediate images 41, 42, 43, and 44, component 23 can be seen from different positions and from different perspectives. Here, the position of component 23 in the intermediate images 41 to 44 can be randomized, for example, generated using a random number generator. However, the position of structural component 23 in the intermediate images 41 to 44 can also correspond to the position of one of the other structural components of the same design housed on the printed circuit board assembly 2 (see [link to relevant documentation]). Figure 2 and Figures 3 to 6 The middle images 41 to 44 have a white background 30.

[0081] Figures 7 to 10 Examples of different rendered images 45 to 48 are shown. Selected structural elements 21, 22, 23, 24, 25, and 26 are shown on each rendered image.

[0082] able to Figures 7 to 10 In each of the rendered images 45 to 48, the printed circuit board 20 of the printed circuit board assembly 2 is seen against different backgrounds 31 to 34. The color of the circuit board can also be varied (not shown here). Furthermore, each rendered image 45 to 48 shows all selected structural elements 21 to 26, which can be randomly distributed on the visible surface of the printed circuit board 20. The positions of the structural elements 21 to 26 on the surface can also correspond to the positions of elements of the same type (structural elements of the same design) in the printed circuit board assembly 2 (see [link to rendering]). Figure 2 Other components visible in CAD model 1 (see...) Figure 2 These elements are not visible in rendered images 45 to 48. When generating the rendered images, it is possible, but not necessary, to exclude them. To achieve the aforementioned random distribution, random positions can be generated for selected structural elements 21 to 26 on the surface of circuit board 20.

[0083] In addition, Figure 10 It can also be seen that the size and / or position and / or orientation of the printed circuit board 20 can also be changed when generating the rendered image.

[0084] Labeling different rendered images 45 to 48 based on the CAD model—Step S3. For example, labeling can be performed based on corresponding entries in the CAD model. For example, rendered images 45 to 48 can be labeled according to the bill of materials. Rendered images 45 to 48 can be labeled based on generated images / intermediate images 41, 42, 43, 44. For example, the label for structural element 23 can be generated together with intermediate image 43 or 44 and used for rendered image 47 or rendered images 45 and 46. Labeling can therefore be performed automatically.

[0085] Rendered images 45 to 48 can be saved before or after marking.

[0086] After rendering images 45 to 48 are labeled, a labeled training dataset is provided based on this—step S4.

[0087] The above process can exist in the form of computer program instructions. For example, this computer program can be processed on computing unit 101. Figure 11 For example, computer programs can exist / be stored in the volatile or non-volatile memory of the processing unit.

[0088] Figure 11 The system 100 is shown, for example, when the computer program described above is processed on the computing unit 101 included in the system 100, the system 100 is able to provide the training dataset of the (synthetic) tags described above.

[0089] System 100 may also include CAD model 1 and a first interface 102, the first interface 102 being designed and configured to allow selection of structural elements 21 to 26 in CAD model 1.

[0090] The calculation unit 101 is configured to generate different rendered images 45 to 48 and to mark them based on CAD models, such as based on a bill of materials that can exist in the CAD model.

[0091] In addition, system 100 has a second interface 103, which is designed and configured to provide labeled training datasets based on different rendered images 45 to 48.

[0092] Figure 12 An example of a marked rendered image 49 is shown, in which a total of ten sub-objects (structural elements) of an object (printed circuit board assembly) are each labeled with a tag 5. In this case, each tag 5 can include a description of the element and / or its coordinates on the marked rendered image 49. Furthermore, each tag 5 can be designed with a rectangular border, for example, to surround the marked component in a preset color. This visualization allows setup personnel to quickly review the results.

[0093] In short, Figures 7 to 10 and Figure 12 The marked rendered images 45 to 49 differ from each other in at least one criterion, wherein the at least one criterion is selected from the group consisting of: the size of the structural element in the rendered image, exposure, perspective view from which the structural element can be seen, background of the structural element (the circuit board 20 can be included in the background of the structural element), position of the structural element in each rendered image 45 to 49, surface pattern, texture, and color.

[0094] Furthermore, the selection of structural elements 21 to 26 can be associated with process sequences, such as printed circuit board assembly manufacturing processes.

[0095] For example, CAD programs allow the allocation of Bill of Materials (BOM) to process sequences (BOPs). Through such links, it becomes possible to automatically generate test programs for specific processes / process steps based on a specially designed training dataset, which can be provided as described above. Here, knowledge of hierarchical structures can be utilized. For printed circuit board assemblies, the hierarchy can be the division of the PCB assembly into sub-assemblies (which can in turn be divided into individual components) and individual components. Here, knowledge available in product development (product structure, Bill of Materials (parts list)) can be used to automatically generate training datasets corresponding to specific process sequences.

[0096] Such a testing procedure can include, for example, a trained function trained on a training dataset generated as described above, and apply the trained function to image recordings of real printed circuit board assemblies, for example, at a specific manufacturing stage, in order to identify the individual structural elements of these printed circuit board assemblies.

[0097] This function can be a classification function and / or a localization function, especially based on convolutional neural networks. For example, classification can be a good / bad evaluation of a complete image record.

[0098] Figure 13 An implementation of a training system is illustrated. The training system 200 shown includes a first training interface device 202 configured to receive, for example, a training dataset provided by the system 100 described above. Furthermore, the training system 200 includes a training computing unit 201 configured to train functions based on the training dataset. For this purpose, the training computing unit 201 can have a training computer program with corresponding commands. Additionally, the training system 200 includes a second training interface device 203 configured to provide the trained functions.

[0099] If the function is further trained on real labeled images, it can achieve particularly good results in object recognition. The number of real images in the training dataset can be very small compared to the number of synthetic or rendered images. For example, the proportion of real images in the training dataset can be between 0.1% and 5%, especially 1%.

[0100] Figure 14 An embodiment of a computer-implemented image recognition method is illustrated. This embodiment relates to an object detection method, which is illustrated using the detection of individual structural elements of a printed circuit board assembly as an example. It is understood that the image recognition method according to the present invention is not limited to identifying structural elements in a printed circuit board assembly.

[0101] First, provide input data—step B1. This can be done, for example, by means of a first interface device 302 (see also...). Figure 16 This is achieved through [method 301]. The input data is an image record 301 of a real object, in this case, a printed circuit board assembly comprising multiple structural elements.

[0102] To identify the various structural elements of the printed circuit board assembly, trained functions, as described above, can be applied to the input data—step B2. Here, output data is generated, which is a segmentation of the input data and includes the input (e.g., the entire image record) itself or a classification of at least one structural element. This can be performed, for example, by computing device 303.

[0103] Then, output data is provided, for example, via the second interface device 303—step B3.

[0104] The output data can be provided as a processed image record 304, wherein the processed image record 304 can be an image record 301 containing markers of structural elements.

[0105] Figure 15 An example of a processed image record 304 is shown. Figure 15 It is possible to identify five different types of structural elements on image record 301. Therefore, it is possible to identify five different types of markers 3040, 3041, 3042, 3043, 3044, and 3045 on the processed image record 304.

[0106] Figure 15 Each identified actual structural element is also labeled. Each label 3040, 3041, 3042, 3043, 3044, 3045 can contain the coordinates / position of the actual structural element on image 304, and preferably includes a description of the identified structural element. For example, the description can be the structural element's designation E0, E1, E2, E3, E4, E5, which can correspond to the bill of materials (BOM) from CAD model 1, and / or include the confidence level (given in %) of the identification of the corresponding structural element.

[0107] Furthermore, for example, markings 3040, 3041, 3042, 3043, 3044, and 3045 can be visualized with rectangular borders. This visualization allows installers to quickly view the results.

[0108] In addition, each marker 3040, 3041, 3042, 3043, 3044, and 3045 can have a preset color. Different types of markers 3040, 3041, 3042, 3043, 3044, and 3045 can have different colors.

[0109] It goes without saying that the steps of the above method can be repeated for multiple different image records of a real object (such as a printed circuit board assembly).

[0110] Figure 16 An embodiment of a data processing system is illustrated, which is suitable, for example, for performing the image recognition method described above with steps B1 to B3. The illustrated data processing system 300 includes a computing device 303 designed and configured to acquire image records (e.g., image record 301) via a first interface device 302, process these image records as described above, and provide or output them in the form of processed image records, such as processed image record 301, via a second interface device 304.

[0111] The data processing system 300 may have image recognition software, which includes instructions (when the image recognition software is executed) that, for example, cause the data processing system 300 to perform the aforementioned method steps B1 to B3. For example, the image recognition software may be stored on the computing device 303.

[0112] Figure 17 An implementation of the industrial system is shown. The above-described methods and systems can be applied to the example industrial system shown. This industrial system is designed as an automated facility 1000.

[0113] Automated facilities can include multiple software and hardware components. Figure 17 The automated facility 1000 shown includes two levels: a shop floor level 1100 (e.g., a workshop, factory, production site, etc.) and a data processing level designed as a cloud level 1200. It is entirely possible that the data processing level is located at the shop floor level (not shown) and includes, for example, one or more computing units.

[0114] Workshop level 1100, for example, can be set up for manufacturing printed circuit board assemblies.

[0115] For example, workshop level 1100 may include the aforementioned system 100 with a corresponding computer program configured and / or set up to transmit the generated labeled dataset to, for example, a database 1201 arranged in cloud level 1200.

[0116] The cloud layer 1200 can include the aforementioned training system 200, which is configured, for example, to retrieve a labeled dataset from the database 1201, based on which a function is trained and provided. It is entirely possible that the data processing layer is arranged at the workshop level and includes the training system 200 (not shown).

[0117] For example, it can provide trained functions for retrieval from a data processing layer (e.g., from the cloud), or can transfer trained functions to test system 1101 to test printed circuit board assemblies.

[0118] The test system 1101 can be deployed in the workshop level 1100.

[0119] The testing system 1101 includes an image recording device 1102 and the aforementioned data processing system 300. The data processing system 300 is capable of retrieving or transmitting trained functions from the cloud or from a local (high-performance) training calculator (e.g., from the training system 200) as described above.

[0120] The image recording device 1102 is designed to capture images of printed circuit board assemblies after specific process steps and transmit them to the data processing system 300. The data processing system then checks the condition of the corresponding printed circuit board assemblies and provides inspection results in the form of, for example, OK / NOK evaluation (Ok-or-Not-Ok).

[0121] A Bill of Materials (BOM) can be set in System 100 for importing and naming components / structural elements. However, System 100 may not necessarily contain information on the correct quantity and location of components for a specific printed circuit board assembly variant.

[0122] Although the invention has been described and illustrated in more detail through embodiments relating to printed circuit board assemblies and their manufacturing processes, the invention is not limited to the disclosed examples. Those skilled in the art can make modifications without departing from the scope of the invention. The invention can be applied with necessary modifications in other fields where image, particularly object, recognition plays an important role. A non-exhaustive list of application areas of the invention includes: automobile production, aircraft manufacturing, medical technology, packaging processes, picking processes, inventory inspection, robotics, and industrial plant construction. Therefore, the invention can be used in all production areas where CAD data can be used and tested. Thus, the invention is not limited to individual applications. For example, applications in the automatic driving of automobiles and trains are entirely conceivable if CAD models of the corresponding environment and its components are available.

Claims

1. A computer-implemented method for providing a labeled training dataset, wherein, - Select at least one electronic structural element in a CAD model (1) of an electronic assembly that includes multiple electronic structural elements, wherein the CAD model (1) contains a description of the at least one electronic structural element and the coordinates of the at least one electronic structural element, wherein the description of the electronic structural element includes information about the type of the electronic structural element, and the coordinates of the electronic structural element determine the position of the electronic structural element in the electronic assembly. - Generate multiple different rendered images (45, 46, 47, 48), wherein each of the different rendered images (45, 46, 47, 48) includes at least one selected electronic structural element, and the position of the at least one selected electronic structural element is different in at least two of the different rendered images. -Based on the description of the at least one electronic structural element contained in the CAD model (1) and the coordinates of the at least one electronic structural element contained in the CAD model (1), the different rendered images are labeled to provide a training dataset of labeled rendered images, wherein the label (5) of the at least one electronic structural element has the description of the at least one electronic structural element and the position of the at least one electronic structural element in the rendered image.

2. The method according to claim 1, wherein, The label (5) is visualized as a rectangular border.

3. The method according to claim 1 or 2, wherein, The electronic components are designed as printed circuit board assemblies, wherein the electronic structural elements are designed as integrated or discrete structural elements.

4. The method according to claim 1 or 2, wherein, The rendered images (45, 46, 47, 48) of the at least one selected electronic structural element are different from each other in at least one criterion, wherein the at least one criterion is selected from the group: the size of the at least one electronic structural element, the exposure of the at least one electronic structural element, the perspective of the at least one electronic structural element, the background (3, 31, 32, 33, 34) of the at least one electronic structural element, the position, surface, texture, and color of the at least one electronic structural element.

5. The method according to claim 1 or 2, wherein, In the CAD model (1), a plurality of electronic structural elements are selected, wherein the selected electronic structural elements can be associated with a process sequence.

6. A system (100) for providing labeled training datasets, the system comprising: - A computer-readable storage device having a CAD model (1) of an electronic assembly containing multiple electronic structural elements, wherein the CAD model (1) includes a description of at least one electronic structural element and coordinates of the at least one electronic structural element, wherein the description of the electronic structural element includes information about the type of the electronic structural element, and the coordinates of the electronic structural element determine the position of the electronic structural element in the electronic assembly. - A first interface (102), the first interface being designed and configured to enable the selection of at least one electronic structural element in the CAD model (1), - A computing unit (101) configured to generate multiple different rendered images (45, 46, 47, 48), wherein each of the different rendered images (45, 46, 47, 48) contains at least one selected electronic structural element, wherein the position of the at least one selected electronic structural element is different in at least two of the different rendered images. Furthermore, based on the description of the at least one electronic structural element contained in the CAD model (1) and the coordinates of the at least one electronic structural element contained in the CAD model (1), the different rendered images (45, 46, 47, 48) are labeled, wherein the label (5) of the at least one electronic structural element has the description of the at least one electronic structural element and the position of the at least one electronic structural element in the rendered image. - Second interface (103), which is designed and configured to provide a training dataset based on rendered images with different labels.

7. A computer-implemented method for providing trained functionality, wherein, - The method according to any one of claims 1 to 5 provides a training dataset with at least one label. - Train the function based on the training dataset containing at least one labeled element to generate the trained function. - Provides the trained functionality.

8. The method according to claim 7, wherein, The functionality is based on the classification and / or localization functions of a convolutional neural network.

9. A computer-implemented image recognition method, wherein, - Provide input data, wherein the input data is an image record (301) of a real electronic component, wherein the real electronic component includes multiple real electronic structural elements. - The trained function is applied to the input data to produce output data, wherein the trained function is provided by the method according to claim 7 or 8, wherein the output data includes the classification and / or location of at least one real electronic structural element in the input data. - Provide the output data.

10. The image recognition method according to claim 9, wherein, The steps of the image recognition method are repeated for multiple different image records (301).

11. The image recognition method according to claim 9 or 10, wherein, The output data is provided in the form of a processed image record, wherein the processed image record is an image record (301) having at least one mark having at least one real electronic structural element.

12. The image recognition method according to claim 11, wherein, The marker has the coordinates of at least one electronic structural element and / or the mark of at least one actual electronic structural element (E0, E1, E2, E3, E4, E5), and the marker is visualized in the form of a rectangular boundary and in a preset color.

13. The image recognition method according to claim 9 or 10, wherein, The actual electronic components are designed as printed circuit board assemblies, and the actual electronic structural elements are designed as integrated or discrete structural elements.

14. A system comprising an image recording device (1102) and a data processing system (300), wherein, The data processing system (300) is designed and configured to perform the image recognition method according to any one of claims 9 to 13.

15. A computer program product, the computer program product comprising: A first instruction, the first instruction causing the system according to claim 6 to perform the method according to any one of claims 1 to 5; And / or a second instruction, which, when the program is executed by the computer, causes the computer to perform the method according to claim 7 or 8; And / or a third instruction, said third instruction causing the system according to claim 14 to perform the image recognition method according to any one of claims 9 to 13.

Images