A method for training a neural network to classify the microstructure of a material and a computer program

Abstract

A method for training a neural network to classify the microstructure of a material (200), comprising: obtained a labeled image of the microstructure of the material with an original resolution (201); extracting a first section with a reduced resolution from the labeled image of the structure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image (202); rotating the first slice at a fixed angle around a fixed axis; extracting a second slice from a center of the first slice; and training the neural network using the first slice and the second slice (203).

Description

[0001] A method for training a neural network to classify the microstructure of a material and a computer program Technical field

[0002] Examples of implementation include a method for classifying the microstructure of a material, a method for training a neural network to classify the microstructure of a material, and a computer program. background

[0003] The microstructure of a material forms during the solidification process of a melt after a critical temperature is reached. Crystallization nuclei serve as centers where crystals can form and grow. In the further solidification process, grain boundaries form through the collision of individual crystals and characterize the microstructure of the material. The microstructure of the material is characterized, among other things, by the type, size, distribution, and orientation of the crystals within it. The mechanical properties, such as the strength and toughness of the material, can depend significantly on the solidification process and the resulting microstructure. In industry, it is of great importance to know the properties of materials precisely, as they serve as the material for workpieces and must therefore meet important criteria and minimum requirements.

[0004] Determining the microstructure requires a precise examination of the material, for example, using a light microscope. A suitable section of the material can be prepared through mechanical processing, such as milling, grinding, and polishing, and / or chemical processing, such as etching. Microscopic examination of this section allows for the creation of micrographs that reveal the material's microstructure. The material can then be classified by analyzing and evaluating these micrographs.

[0005] The processing of the material into a polished section and the classification of its microstructure are generally carried out by qualified experts. Since the classification of the material can differ, for example, due to the experts' professional experience, the evaluation of the material is partly subjective. Consequently, the microstructure cannot be adequately analyzed by untrained employees or external suppliers without expert supervision. Furthermore, manually comparing polished sections with reference series is often insufficient for a proper and efficient classification of the material.

[0006] The paper by AZIMI, Seyed Majid et al., entitled "Advanced Steel Microstructural Classification by Deep Learning Methods," proposes artificial neural networks (NNs) for classifying steel microstructures. To generate a larger dataset for training the NNs, microscope images acquired with a scanning electron microscope and a light microscope are divided into multiple sections and used for training.

[0007] The document by DECOST, Brian L. et al. entitled “High throughput quantitative metallography for complex microstructures using deep learning: A case study in ultrahigh carbon steel” uses the PixelNet architecture to automatically segment complex microstructure compositions of proeutectoid cementite, ferrite and Widmannstätten cementite.

[0008] The document by Aguiar, JA et al., entitled "Decoding crystallography from high-resolution electron imaging and diffraction datasets with deep learning," describes the determination of crystal structures using artificial neural networks based on a diffraction profile. The diffraction profile is determined using scanning transmission electron microscopy (SEM) and fast Fourier transformation (FFT).

[0009] This creates a need to classify the structure of materials better and more objectively. Summary

[0010] This task can be solved by objects of independent claims.

[0011] One embodiment of a method for classifying the microstructure of a material comprises receiving an image of the material's surface and classifying the microstructure using a trained neural classification network. The surface image can be acquired, for example, using a light microscope, scanning electron microscope, and / or other microscopic imaging devices. This image serves as the basis for the neural network's classification of the microstructure and should therefore be of high quality. The trained neural network can perform an objective classification of the material with high accuracy and efficiency. By using the neural network, untrained employees or external suppliers can appropriately and independently determine microstructures.

[0012] According to some implementation examples, the trained neural classification network corresponds to AlexNet, VGG, ResNet, SqueezeNet, DenseNet, Inception v3, GoogLeNet, ShuffleNet v2, MobileNet v2, ResNeXt, Wide ResNet, or MNASNet. These examples of classification networks can be suitable for obtaining reliable results in the classification of structures. The type of classification method is determined by the network and corresponds, for example, to tasks such as image classification, pixel-wise semantic segmentation, object detection, instance segmentation, or others.

[0013] One embodiment of a method for training a neural network to classify the microstructure of a material comprises obtaining a labeled image of the material's microstructure at an initial resolution. The method further comprises extracting a first section at a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image. The method further comprises training the neural network using this first section. The reliable classification of materials using the neural network can depend significantly on the training data. The training of the neural network can be based on labeled images that have been created and classified, for example, by experts.The original resolution determines the size of the labeled image and is, for example, defined by the device used to capture the material. For training the neural network, it can be advantageous to use a significantly smaller section of the labeled image of the material. This smaller section, for instance, already contains sufficient information to identify a specific crystal type within the microstructure. Furthermore, by reducing the size of the section used for training, multiple sections can be created from the labeled image. From an initially single labeled image, extracting lower-resolution sections can generate multiple labeled training datasets for the neural network, making it possible to train the network to a high recognition rate even with a relatively small number of labeled images.

[0014] According to some embodiments, the method further comprises rotating the first section by a defined angle around a defined axis. The method also includes extracting a second section from the center of the first section and training the neural network with the second section. Rotating the first section allows for the generation of any additional data for training the neural network. Furthermore, existing patterns in the orientation of the material or in the orientation of the labeled image can be excluded by rotation, thus ensuring that the training of the neural network is not affected.

[0015] According to some embodiments, the method further includes transforming the second section into a training section with a predefined format, which corresponds to a format specification of the neural network for classifying the microstructure of the material. The second section can be transformed into the training section, for example, by means of transformations such as blurring, mirroring, and color changes. This transformation can generate modified training data in which, for example, the microstructure is not clearly recognizable, so that the neural network is trained to abstract appropriately, for example, to prevent overfitting. Since the format of the section to be analyzed may be predetermined depending on the type and topology of the neural network, the format specification of the neural network can be fulfilled, for example, by cropping the second section.

[0016] According to some embodiments, the method further includes transforming the first section into a training section with a predefined format, which corresponds to a format specification of the neural network for classifying the microstructure of the material. The first section can be further transformed into the training section by means of transformations such as blurring, mirroring, and color changes.

[0017] One embodiment of the present invention comprises a computer program with program code that executes a method according to the preceding description using a programmable processor. The method for classifying structures and / or training the neural network can be efficiently adapted and extended by modifying the program code in the computer program. Furthermore, the method can function by implementing the computer program in any compatible device.

[0018] One embodiment of a device for classifying the microstructure of a material comprises an interface for receiving an image of the material's surface and an evaluation unit configured to classify the material's microstructure using a trained neural classification network. The surface of the material can be imaged, for example, using a light microscope, scanning electron microscope, and / or other microscopic imaging devices. The image serves as the basis for the classification of the microstructure by the neural network and should therefore be able to represent the material's microstructure with high quality. The trained neural network can perform an objective classification of the material with high accuracy and efficiency. By using the neural network, untrained employees or external suppliers can determine microstructures appropriately and independently.

[0019] According to some implementation examples, the trained neural classification network corresponds to AlexNet, VGG, ResNet, SqueezeNet, DenseNet, Inception v3, GoogLeNet, ShuffleNet v2, MobileNet v2, ResNeXt, Wide ResNet, or MNASNet. These examples of classification networks can be suitable for obtaining reliable results in the classification of structures. The type of classification method is determined by the network and corresponds, for example, to tasks such as image classification, pixel-wise semantic segmentation, object detection, instance segmentation, or others.

[0020] An embodiment of a device for training a neural network to classify the microstructure of a material comprises an interface for obtaining a labeled image of the material's microstructure at an initial resolution. The device further comprises a training device configured to extract a first section with a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image. The training device is further configured to train the neural network using this first section. The reliable classification of materials using the neural network can depend significantly on the training data. The training of the neural network can be based on labeled images that have been created and classified, for example, by experts.The original resolution determines the size of the labeled image and is, for example, defined by the device used to capture the material. For training the neural network, it can be advantageous to use a significantly smaller section of the labeled image of the material. This smaller section, for instance, already contains sufficient information to identify a specific crystal type within the microstructure. Furthermore, by reducing the size of the section used for training, multiple sections can be created from the labeled image. From an initially single labeled image, extracting lower-resolution sections can generate multiple labeled training datasets for the neural network, making it possible to train the network to a high recognition rate even with a relatively small number of labeled images.

[0021] According to some embodiments, the training device is further configured to rotate the first section at a defined angle around a defined axis, to extract a second section from the center of the first section, and to train the neural network with the second section. The rotation of the first section allows for the generation of any additional data for training the neural network. Furthermore, existing patterns in the orientation of the material or in the orientation of the labeled image can be excluded by rotation, so that the training of the neural network is not affected.

[0022] According to some embodiments, the training device is further configured to transform the second section into a training section with a predefined format, where the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. The second section can be transformed into the training section, for example, by means of transformations such as blurring, mirroring, and color changes. This transformation can generate modified training data in which, for example, the microstructure is not clearly identifiable, thus training the neural network to abstract appropriately, for instance, to prevent overfitting. Since the format of the section to be analyzed may be predetermined depending on the type and topology of the neural network, the format specification of the neural network can be fulfilled, for example, by cropping the second section.

[0023] According to some embodiments, the training device is further configured to transform the first section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. The first section can also be transformed into the training section by means of transformations such as blurring, mirroring, and color changes. Character description

[0024] Some examples of procedures are explained in more detail below with reference to the accompanying figures. These show: Fig. 1. A flowchart of an exemplary embodiment of a method for classifying the microstructure of a material; and Fig. 2 a flowchart of an embodiment of a method for training a neural network to classify the microstructure of a material. Description

[0025] Several examples will now be described in more detail with reference to the accompanying figures, which illustrate some of these examples. The thickness of lines, layers, and / or areas may be exaggerated in the figures for clarity.

[0026] While further examples of various modifications and alternative forms are suitable, some specific examples are accordingly shown in the figures and are described in detail below. However, this detailed description does not limit further examples to the specific forms described. Further examples may encompass all modifications, correspondences, and alternatives that fall within the scope of revelation. Equal or similar reference signs throughout the description of the figures refer to identical or similar elements that, upon comparison, may be implemented identically or in a modified form while providing the same or a similar function.

[0027] It is understood that when an element is described as "connected" or "coupled" to another element, the elements may be connected or coupled directly or via one or more intermediate elements. When two elements A and B are combined using "or," this is to be understood as revealing all possible combinations, i.e., only A, only B, and A and B, unless explicitly or implicitly defined otherwise. An alternative formulation for the same combinations is "at least one of A and B" or "A and / or B." The same applies, mutatis mutandis, to combinations of more than two elements.

[0028] The terminology used here to describe certain examples is not intended to be limiting for other examples. Where a singular form, e.g., "a" and "the," "a," "a," is used, and the use of only a single element is neither explicitly nor implicitly defined as mandatory, further examples may also use plural elements to implement the same function. Similarly, where a function is subsequently described as being implemented using multiple elements, further examples may implement the same function using a single element or a single processing entity.It is further understood that the terms “include”, “comprehensive”, “exhibit” and / or “exhibit” when used specify the presence of the indicated features, integers, steps, operations, processes, elements, components and / or a group thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, processes, elements, components and / or a group thereof.

[0029] Unless otherwise defined, all terms (including technical and scientific terms) are used herein in their usual sense within the field to which examples belong.

[0030] Fig. Figure 1 shows a flowchart of an embodiment of a method for classifying the microstructure of a material 100. For example, the method determines the homogeneous, heterogeneous, or multiphase microstructure of materials, such as metals, surface hardenings, coatings, composites, or materials in general, whose internal structure can be determined by the microstructure. The internal structure can, for example, include the spatial arrangement of different constituents of the material within the microstructure. For the classification of the microstructure, an image of a surface of the material 101 is received. For example, the image can show a micrograph obtained after metallographic preparation of a material. Materials can be, for example, materials of a crankshaft, clutch components, screws, nuts, wires, gears, and other components of a workpiece. The image of the material can be, for example, a micrograph of a surface of the material.at 1000x magnification, so that the microstructure can be examined with regard to its components, crystal formation, grain sizes and grain boundaries.

[0031] Based on the recording, the microstructure of the material can be classified using a trained neural classification network. For example, depending on the composition and orientation of its internal structure, steel is classified by the neural network into a martensite, pearlite-ferrite, or bainite microstructure. The method of classification can vary depending on the network topology.

[0032] In one embodiment, the trained neural classification network corresponds to AlexNet, VGG, ResNet, SqueezeNet, DenseNet, Inception v3, GoogLeNet, ShuffleNet v2, MobileNet v2, ResNeXt, Wide ResNet, or MNASNet. These networks are suitable for reliably classifying the microstructure of materials, as they can solve various tasks such as image classification, pixel-wise semantic segmentation, object recognition, or instance segmentation. The topology of the neural network is characterized by the type and arrangement of different layers within the network. The aforementioned classification networks use convolutional and pooling layers, allowing the image to be analyzed to be examined for specific properties such as edges or colors. Unlike pooling layers, convolutional layers can be trained.

[0033] The training is based on trainable parameters that are adjusted during training using backpropagation. The weight adjustments depend on the training data the neural network receives. Once training is complete and the parameters are determined, the neural network can be tested for the desired application and further trained depending on the results.

[0034] For classifying the microstructure of a material, pre-trained neural classification networks can be used, for example. These networks can already have initial parameters and be adapted through further training. Iterative changes to the initial parameters enable transfer learning during subsequent training. To improve the efficiency of microstructure classification, the topology of neural networks can be modified, for example, by removing and / or adding layers.

[0035] The result of the classification network is the assignment of the input image to a class. Classification can be performed using a final classification layer, such as the softmax function, which returns a probability distribution for different possible classes. Various types of image classification can be considered. A simple binary classification method assigns an image to a first or second class. For example, the trained neural network can classify an image of steel as having a martensite or non-martensite microstructure. In multi-class classification, the image is assigned to the class with the highest probability. For example, if steel is identified as having 80% bainite, 15% martensite, and 5% pearlite microstructure, the microstructure is ultimately classified as bainite.

[0036] The trained neural networks can, for example, output an image matrix instead of a class value. This matrix can contain the individual classifications of the grains or crystals within the microstructure. To achieve this, the neural network can analyze the microstructure at the pixel level. After determining the pixels based on their maximum probability, a set of neighboring pixels can be considered, for example, as a unit crystal. In this variant, the neural network can perform the classification, for instance, by color-coding the crystals within the microstructure.

[0037] Fig. Figure 2 shows a flowchart for an embodiment of a method for training a neural network to classify the microstructure of a material 200. Training the neural network can significantly influence the accuracy and reliability of the microstructure classification. In one embodiment, training the neural network includes obtaining a labeled image of the material's microstructure with an initial resolution 201. The training of the neural network can be tailored to the type of material to be classified and can be achieved by providing microstructure images in which the depicted microstructure is already known and which are referred to as labeled images. The evaluation (labeling) of the microstructure images for training the neural network can, for example, be performed by expert image analysis.For an independent evaluation of the microstructure, the image can be examined separately by several experts. By incorporating testing procedures into the evaluation process, such as image reference sequences, the images can be evaluated as objectively and uniformly as possible. Systematic description and tagging of microstructure images allows for the creation of an image database for training the neural network. Furthermore, deviations from standards and specifications can be recorded. For the microstructure evaluation by the experts, multiple images from different imaging devices, such as a light microscope and a scanning electron microscope, can be used, which can reduce the risk of network failure.

[0038] The original resolution of the labeled image of the material's microstructure can depend on the type of imaging device. The original resolution might be defined, for example, by the horizontal and vertical pixel count of the labeled image, such as 2080x1542.

[0039] In one embodiment, the method comprises extracting a first section with a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image 202. Furthermore, the method comprises training the neural network using the first section 203. For training the neural network, it may be advantageous to use significantly smaller image sections compared to the labeled image of the microstructure. By extracting smaller image sections, multiple training images, representing, for example, different sub-regions of the microstructure, can be generated from a single labeled image of the microstructure. This allows the neural network to be adequately trained with fewer labeled images, since multiple training datasets can be generated from one labeled image.

[0040] With the reduced resolution, which is, for example, 60% lower than the original resolution, the image quality of the first section remains unchanged compared to the labeled image, since the first section simply represents a smaller area of ​​the labeled image. Extracting the first section at a reduced resolution therefore eliminates transformations that involve information loss, such as resampling or downsampling.

[0041] Extracting the smaller image section can be done randomly or according to a predefined criterion. For example, the criterion might be that the image section is selected in such a way that the entire microstructure is best represented in terms of the orientation and distribution of the crystals. Another predefined criterion could be to define image sections in such a way that conspicuous features in the microstructure, which could indicate material failure, are displayed.

[0042] The reduced resolution can, for example, be decreased from an initial resolution of 2080x1542 to 520x385. The first image section can be used directly for training the neural network or transformed into a new image section through further processing. In methods where a section with a reduced resolution is extracted from the labeled image of the structure, the reduced resolution can alternatively be at least 60%, 70%, 80%, or 90% lower than the previously mentioned minimum of 50%.

[0043] In some embodiments, the method involves rotating the first section by a defined angle about a defined axis. The rotation of the first section can be by any angle, such as 30°, 90°, or 105.6°, about any axis, such as a perpendicular axis through the center or geometric centroid of the first section. For example, the first section can also be transformed by rotating it 180° about a vertical or horizontal axis of the labeled image. Rotating the first section allows images to be duplicated, enabling the creation of new, distinct training images from a single labeled image. Furthermore, rotating the image can eliminate any random or existing pattern in the labeled image of the structure.For example, patterns can emerge in the image if the original image was captured according to the orientation of the microstructure. For instance, the vertical axis of the image might be aligned with the main fiber direction of the microstructure, as this makes the image easier for experts or according to a template to evaluate. If training images were too similar, for example, due to recurring patterns in the microstructure, it could lead to overfitting of the neural network. With overly similar training images, the neural network might not be able to adequately abstract learned information for an application. Consequently, microstructure images that deviate from the training data might not be correctly classified.

[0044] In one embodiment, the method further comprises extracting a second section from the center of the first section and training the neural network with this second section. Extracting the second image section from the first allows smaller crystals or objects to be magnified and viewed independently of the surrounding microstructure. Using the smaller, second image section, the neural network can, for example, be trained to recognize microstructures with fewer grain boundaries. Furthermore, the format of the rotated first section, e.g., after a 45° rotation, may not be suitable for training the neural network. Extracting the second section can, for example, restore a rectangular format. The second section can then be used for training the neural network or for further processing.

[0045] In some embodiments, the method includes transforming the second section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. In some embodiments, the method further includes transforming the first section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. The neural network can be trained using the training section. The transformation of the first or second section into a training section may involve further image processing methods. Training images can also be duplicated for data augmentation through transformation.Furthermore, overfitting of the neural network can be avoided or reduced by introducing an artificial disturbance into the training sample, such as blurring or changing the exposure. Possible transformations include translation, reflection off a vertical or horizontal axis, rotation, grayscale, blurring, contrast adjustment, color change, intensity change, and many more. By transforming the first or second sample, the training sample can have a predefined format that matches the format specifications of the neural network. For example, training samples for pre-trained neural classification networks, such as ResNet34, can be adapted to the predefined 224x224 format.

[0046] Using only a few labeled microstructure images, the described method for training the neural network can generate a large number of training fragments. Despite the limited number of labeled microstructure images, the neural network can be efficiently trained using the multitude of different training fragments, resulting in the same or even better results regarding the correct classification of microstructures. In a proof of concept, the neural network achieved a recognition probability of over 95% using approximately one hundred labeled images for two different microstructures.

[0047] The neural network can be used, for example, in the automotive industry, information technology, medical technology, telecommunications, and many other sectors across entire companies, such as in materials laboratories at the company site or at suppliers. Furthermore, the network can be used for training and further education, for example at universities or vocational schools, or to support inexperienced employees by automatically interpreting microstructure images. Other areas of application for microstructure classification using the neural network include metallography, materials science, material failure analysis, quality assurance, and other fields that analyze and / or evaluate materials with regard to their microstructure or internal structure. The efficiency of microstructure classification can be increased through automatic comparison with standards.

[0048] Fig. Figure 3 shows an embodiment of a device 300 for classifying the microstructure of a material with an interface 301 for receiving an image of a surface of the material and an evaluation unit 302, which is configured to classify the microstructure of the material using a trained neural classification network.

[0049] Fig.Figure 4 shows an embodiment of a device 400 for training a neural network to classify the microstructure of a material, with an interface 401 for obtaining a labeled image of the microstructure of the material at an initial resolution. The device 400 further comprises a training device 402 configured to extract a first section with a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image. The training device 402 is further configured to train the neural network using the first section.

[0050] The aspects and features described together with one or more of the previously detailed examples and figures can also be combined with one or more of the other examples to replace an identical feature of the other example or to additionally introduce the feature into the other example.

[0051] Examples may also include a computer program with program code for performing one or more of the above procedures, or refer to the execution of the computer program on a computer or processor. Steps, operations, or processes of various procedures described above may be performed by programmed computers or processors. Examples may also include program storage devices, such as digital data storage media, that are machine-, processor-, or computer-readable and encode machine-executable, processor-executable, or computer-executable programs of instructions. The instructions perform or cause some or all of the steps of the procedures described above. The program storage devices may, for example,Digital storage media, magnetic storage media such as magnetic disks and magnetic tapes, hard disk drives, or optically readable digital data storage media. Further examples may also include computers, processors, or control units programmed to perform the steps of the procedures described above, or (field) programmable logic arrays (PLAs) or (field) programmable gate arrays (PGAs) programmed to perform the steps of the procedures described above.

[0052] The descriptions and drawings only illustrate the principles of revelation. Furthermore, all examples presented here are expressly intended for illustrative purposes only, to assist the reader in understanding the principles of revelation and the concepts contributed by the inventor(s) to the advancement of technology. All statements made here regarding principles, aspects, and examples of revelation, as well as specific examples thereof, include their corresponding references.

[0053] A functional block designated as a "means for..." performing a specific function can refer to a circuit configured to perform that function. Thus, a "means for something" can be implemented as a "means configured for or suitable for something," e.g., a component or circuit configured for or suitable for the specific task.

[0054] The functions of various elements shown in the figures, including each functional block designated as "means," "means of providing a signal," "means of generating a signal," etc., can be implemented in the form of dedicated hardware, e.g., "a signal provider," "a signal processing unit," "a processor," "a controller," etc., as well as in hardware capable of executing software in conjunction with associated software. When provided by a processor, the functions can be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some or all of which can be shared.However, the term "processor" or "controller" is by no means limited to hardware capable solely of executing software, but can include digital signal processor hardware (DSP hardware; DSP = Digital Signal Processor), network processors, application-specific integrated circuits (ASICs = Application Specific Integrated Circuits), field-programmable gate arrays (FPGAs = Field Programmable Gate Arrays), read-only memory (ROMs = Read Only Memory) for storing software, random-access memory (RAMs = Random Access Memory), and non-volatile storage devices. Other hardware, both conventional and / or custom-designed, may also be included.

[0055] A block diagram, for example, can represent a rough circuit diagram that implements the principles of the disclosure. Similarly, a flowchart, a process flowchart, a state transition diagram, pseudocode, and the like can represent various processes, operations, or steps that are, for example, substantially depicted in a computer-readable medium and thus executed by a computer or processor, regardless of whether such a computer or processor is explicitly shown. Methods disclosed in the description or in the claims can be implemented by a component that includes means for performing each of the respective steps of these methods.

[0056] It is understood that the disclosure of multiple steps, processes, operations, or functions in the description or claims should not be interpreted as being in a specific order unless explicitly or implicitly stated otherwise, for example, for technical reasons. Therefore, the disclosure of multiple steps or functions does not restrict them to a specific order unless these steps or functions are not interchangeable for technical reasons. Furthermore, in some examples, a single step, function, process, or operation may include and / or be broken down into multiple sub-steps, sub-functions, sub-processes, or sub-operations. Such sub-steps may be included and form part of the disclosure of that single step unless explicitly excluded.

[0057] Furthermore, the following claims are hereby included in the detailed description, where each claim can stand alone as a separate example. While each claim can stand alone as a separate example, it should be noted that—although a dependent claim may refer in the claims to a specific combination with one or more other claims—other examples may also include a combination of the dependent claim with the subject matter of any other dependent or independent claim. Such combinations are explicitly suggested here unless it is stated that a particular combination is not intended. Furthermore, features of a claim for any other independent claim are also to be included, even if that claim is not directly dependent on the independent claim. Reference symbol list 100 An exemplary embodiment of a method for classifying the microstructure of a material 101 Receiving an image of a material surface 102 Classifying the microstructure of the material using a trained neural classification network 200 An embodiment of a method for training a neural network to classify the microstructure of a material 201 Obtaining a labeled image of the microstructure of the material at an original resolution 202 Extracting a first section with a reduced resolution from the labeled image of the structure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image 203 Training the neural network using the first excerpt 300 Device for classifying the structure of a material 301 Interface for receiving an image of a material surface 302 Evaluation unit 400 Device for training a neural network 401 Interface for obtaining a labeled image of the material's microstructure 402 Training device

Claims

A method for training a neural network to classify the microstructure of a material (200), comprising: obtaining a labeled image of the microstructure of the material with an original resolution (201); extracting a first section with a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image (202); rotating the first section at a defined angle about a defined axis; extracting a second section from a center of the first section; and training the neural network using the first section and the second section (203). The method (200) according to claim 1, further comprising: transforming the second section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. The method (200) according to claim 1 or 2, further comprising: transforming the first section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. A computer program with program code that executes a method according to claims 1 to 3 using a programmable processor. A device (400) for training a neural network for classifying the microstructure of a material, comprising: an interface (401) for obtaining a labeled image of the microstructure of the material with an original resolution; a training device (402) configured to extract a first section with a reduced resolution from the labeled image of the microstructure, wherein the reduced resolution is at least 50% lower than the original resolution of the labeled image; to rotate the first section about a defined axis at a defined angle; to extract a second section from a center of the first section; and to train the neural network using the first section and the second section. The device (400) according to claim 5, wherein the training device (402) is further configured to transform the second section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material. The device (400) according to claim 5 or 6, wherein the training device (402) is further configured to transform the first section into a training section with a predefined format, wherein the predefined format corresponds to a format specification of the neural network for classifying the microstructure of the material.

Images