Evaluation method and evaluation apparatus

Abstract

To provide an evaluation method and an evaluation apparatus for evaluating the dispersion state of a plurality of types of regions in an image in which regions classified into a plurality of types are dispersed.SOLUTION: An evaluation method that is performed by a computer, comprises: an image capturing step for capturing an image in which regions classified into a plurality of types are dispersed; a segmentation step for segmenting the captured image into the plurality of types of regions; a measurement and recording step for focusing on one region of a first-type regions, measuring the shortest distance r between all pixels in the region and pixels of the second-type region that are closest to the region, and recording the measurement results; and a dispersion index calculation step for calculating, after the measurement and recording step has been performed for all of the first-type regions, the inverse λ of the slope of a histogram of the shortest distance r and a logarithmic scale frequency I as an index of dispersibility of the regions classified into the plurality of types.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a method and an apparatus for evaluating the dispersion state of a plurality of types of regions. [Background technology]

[0002] Particle dispersion technology is used in secondary batteries, multilayer ceramic capacitors, etc. Patent Document 1 discloses an evaluation method that uses the standard deviation of the area of ​​divided regions obtained by Voronoi division of an inspection image as an index representing the dispersibility of microparticles.

[0003] Furthermore, secondary batteries are formed by applying a paste containing dispersed active material to a positive electrode sheet, and the degree of dispersion of the active material affects the output of the secondary battery. According to Non-Patent Document 1, in the slurry used to make positive electrode materials for secondary batteries, such as lithium iron phosphate, the dispersibility of the conductive additive and active material affects battery performance. It has also been pointed out that issues related to this battery material go beyond simple dispersibility; battery performance can also be affected by the shape of the network formed by the active material and conductive additive.

[0004] On the other hand, the internal electrodes of a multilayer ceramic capacitor are formed by applying an internal electrode paste in which nickel particles and dielectric particles are dispersed in a binder resin and a solvent, drying, and firing. If the nickel particles and dielectric particles in the internal electrode paste are not dispersed properly, the multilayer ceramic capacitor will have problems such as voids in the internal electrodes and insufficient capacitance.

[0005] The dispersion state of nickel particles and dielectric particles in the internal electrode paste can be evaluated in a dried film obtained by removing the solvent after application. For example, an SEM image of a dried film of the internal electrode paste shows that the nickel particles and dielectric particles are dispersed, with voids formed between each nickel particle and each dielectric particle. A method for evaluating the dispersion state of particles, including the evaluation of the dispersibility in the dried film after application of such an internal electrode paste, is needed.

[0006] Furthermore, in the image of the crystal structure of a polycrystalline metal material, crystal grains with different orientations are distributed. Such a distribution of crystal grains can cause problems in the properties of the metal material. Therefore, a method for evaluating the distribution of crystal grains in such a crystal structure is also required. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2010-122122 [Non-patent literature]

[0008] [Non-Patent Document 1] Fang-Yuan Su, et al., "Could graphene construct an effective conducting network in a high-power lithium ion battery?", Nano Energy, Volume 1, Issue 3, doi: 10.1016 / j.nanoen.2012.02.004, (2012) Summary of the Invention [Problem to be solved by the invention]

[0009] The technology disclosed in the prior art document is limited to distinguishing one type of background from one type of particle, and is not applicable when there are two types of particles and the mutual dispersibility of those particles needs to be evaluated. When the technology of Patent Document 1 is applied to the above-mentioned internal electrode paste, if this method is applied without distinguishing between nickel particles and dielectric particles, the particle dispersibility will be evaluated as good even if the dielectric particles are strongly agglomerated. Therefore, it is necessary to distinguish between nickel particles and dielectric particles in the internal electrode paste.

[0010] There is also a need for a method for evaluating the distribution of polycrystalline grains.

[0011] An object of one aspect of the present invention is to provide an evaluation method and evaluation device for evaluating the dispersion state of a plurality of types of regions in an image in which regions classified into a plurality of types are dispersed. [Means for solving the problem]

[0012] One aspect of the present invention is an evaluation method in which a computer performs the following steps: an image capture procedure for capturing an image in which regions classified into multiple types are dispersed; a segmentation procedure for segmenting the captured image into the multiple types of regions; a measurement and recording procedure for focusing on one of the regions of a first type, measuring the shortest distance r between all pixels in that region and the pixels of the region of a second type that is closest to that region, and recording the results of the measurement; and a dispersibility index calculation procedure for performing the measurement and recording procedure for all regions of the first type, and then calculating the inverse λ of the slope of a histogram of the shortest distance r versus the logarithmic scale frequency I as an index of dispersibility of the regions classified into multiple types. [Effects of the Invention]

[0013] According to one aspect of the present invention, it is possible to evaluate the distribution state of multiple types of regions in an image in which regions classified into multiple types are dispersed. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is a system configuration diagram of an example of an evaluation device according to an embodiment of the present invention; [Figure 2] FIG. 2 is a diagram illustrating an example of a hardware configuration of a computer according to the present embodiment. [Figure 3] FIG. 2 is a functional configuration diagram of an example of an evaluation device according to the present embodiment. [Figure 4] 1 is a flowchart illustrating an example of a method for generating a trained deep learning model for particle discrimination in this embodiment. [Figure 5] FIG. 1 is a diagram illustrating an example of the structure of a deep neural network for semantic segmentation. [Figure 6]1 is a flowchart showing an example of a method for evaluating the dispersion state of particles in the present embodiment. [Figure 7] FIG. 1 is an example diagram showing an SEM image of Ni paste for MLCC and an image in which the Ni region, the co-material region, and the resin region of the SEM image are color-coded. [Figure 8] FIG. 10 is a diagram illustrating an example of a histogram created by a variance evaluation unit. DETAILED DESCRIPTION OF THE INVENTION

[0015] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0016] [System Configuration] First, an example of an evaluation system 1 according to this embodiment will be described. Fig. 1 is a system configuration diagram of an example of an evaluation device 10 according to this embodiment. Fig. 1(A) is an example of the evaluation system 1 in which a user operates the evaluation device 10 to use the evaluation device 10. Fig. 1(B) is an example of the evaluation system 1 in which a user operates a user terminal 12 to use the evaluation device 10. The evaluation device 10 and the user terminal 12 are connected to each other so as to be able to communicate with each other via a network 18 such as the Internet or a local area network (LAN).

[0017] The evaluation device 10 in Fig. 1(A) is a device that evaluates the dispersion state of multiple types of regions in an image in which regions classified into multiple types are dispersed. The evaluation device 10 is, for example, a PC (personal computer) or a workstation. The evaluation device 10 may also be, for example, a tablet terminal, a smartphone, or a measuring instrument.

[0018] The evaluation device 10 receives operations from a user via a touch panel, controller, mouse, keyboard, etc. The evaluation device 10 performs various processes, as described below, in accordance with the operations received from the user. The evaluation device 10 also outputs the results of the processing in accordance with the operations received from the user.

[0019] The user terminal 12 in FIG. 1(B) is a device that accepts operations from a user using a touch panel, a controller, a mouse, a keyboard, or the like. The user terminal 12 is, for example, a PC, a tablet terminal, a smartphone, or the like. The user terminal 12 communicates with the evaluation device 10 via a network 18. The user terminal 12 requests processing from the evaluation device 10 in accordance with the operation content accepted from the user, and outputs the processing results received from the evaluation device 10. The output of the processing results may be a display output or a print output.

[0020] The evaluation device 10 in FIG. 1(B) performs various processes as described below in accordance with a request received from the user terminal 12, and transmits the results of the processes to the user terminal 12. The evaluation device 10 in FIG. 1(B) may be realized by, for example, a cloud computer. The number of evaluation devices 10 is not limited to one. Two or more evaluation devices 10 may be configured and connected to each other so as to be able to communicate with each other via a network 18.

[0021] The evaluation system 1 in FIG. 1 may include a device that captures an image in which regions classified into multiple types are dispersed. A scanning electron microscope (SEM) is an example of a device that captures an image (SEM image) of a material for evaluating the dispersion state. Below, an example of an SEM will be described as a device that captures an image in which regions classified into multiple types are dispersed, but this is not limiting. An SEM is a device that performs image observation by detecting backscattered electrons or secondary electrons that are generated when electrons irradiated from an electron gun are reflected by the surface of a sample.

[0022] The evaluation device 10 may be communicatively connected to an SEM. The evaluation device 10 communicatively connected to an SEM can receive an SEM image. The evaluation device 10 may receive the SEM image from a database that stores the SEM image, or may read the SEM image from a storage medium, such as a USB (Universal Serial Bus) memory, that stores the SEM image.

[0023] Note that the configuration of the evaluation system 1 shown in Fig. 1 is an example and is not limited to the configuration of Fig. 1. The evaluation device 10 and user terminal 12 of the evaluation system 1 according to this embodiment can be realized, for example, by cooperation between hardware constituting a general computer and a program (software) executed by the computer.

[0024] The evaluation device 10 and user terminal 12 according to this embodiment can realize an evaluation device 10 and an evaluation method for evaluating the dispersion state of multiple types of regions in an image in which regions classified into multiple types are dispersed, for example, by executing a program on the computer 20 of FIG. 2.

[0025] [Hardware configuration] Fig. 2 is a hardware configuration diagram of an example of a computer according to this embodiment. The computer 20 shown in Fig. 2 has a hardware configuration including a processor such as a CPU (Central Processing Unit) 21, memories such as a ROM (Read Only Memory) 22 and a RAM (Random Access Memory) 23, an auxiliary storage device 24, an output device 25 such as a display device, an input device 27 such as an input device, an I / F device 26, and a bus connecting these components. The auxiliary storage device 24, the output device 25, and the input device 27 may be provided outside the housing of the computer 20.

[0026] The program executed by the computer 20 is provided by being recorded on, for example, a magnetic disk, an optical disk, a semiconductor memory, or a similar recording medium, and is stored in the auxiliary storage device 24 or the like. The recording medium on which the program is recorded may be in any storage format as long as it is a recording medium readable by the computer 20. The program may also be configured to be pre-installed on the computer 20. The program may also be configured to be stored in the ROM 22. The program may also be distributed via the network 18 and installed in the computer 20 as appropriate.

[0027] The CPU 21 controls the overall operation of the computer 20. The CPU 21 appropriately reads out and executes programs stored in the ROM 22 and auxiliary storage device 24 onto a memory such as the RAM 23, thereby realizing various processing sections described below.

[0028] The ROM 22 stores programs and data. The RAM 23 is used as a work area for the CPU 21. The RAM 23 may include non-volatile RAM. The auxiliary storage device 24 is a storage device such as an SSD (Solid State Drive) or an HDD (Hard Disk Drive).

[0029] For example, the auxiliary storage device 24 can store various programs, data, files, etc. necessary for evaluating the distribution state of multiple types of regions in an image in which regions classified into multiple types are distributed. The output device 25 is an output device such as a monitor or display. The input device 27 is an input device such as a touch panel, controller, mouse, keyboard, operation button, etc. The I / F device 26 includes an interface for communication via the network 18 and an I / F for reading data from or writing data to a recording medium.

[0030] [Function Configuration] Hereinafter, an example of the evaluation system 1 according to this embodiment will be described with reference to Fig. 1(A). In the case of the evaluation system 1 shown in Fig. 1(B), the evaluation device 10 and the user terminal 12 cooperate with each other to realize the processing described below.

[0031] In this embodiment, the method for evaluating the distribution of multiple types of regions uses image recognition AI (artificial intelligence) based on learning the characteristics of each region that makes up an image to classify the regions by their characteristics, measure the shortest distance between a specific type of region (a first type of region) and another type of region (a second type of region), and process the shortest distance using a histogram to evaluate the distribution of the regions. Image recognition AI is sometimes called an image recognition model.

[0032] For example, polycrystalline metal materials have a distribution of crystal grains with different crystal orientations. The distribution of such crystal grains can sometimes cause problems in terms of the properties of the metal material. The method for evaluating the distribution of multiple types of regions according to this embodiment can classify the regions into those with different crystal orientations and evaluate the distribution of the crystal grains.

[0033] Furthermore, in an image in which particles are dispersed, the method of evaluating the dispersion state of a region of this embodiment can evaluate the dispersion state of particles by defining the region as a region in which one particle (first particle) with an outline that can be distinguished from its surroundings is present, a region of voids formed by the gaps between multiple first particles (regions of voids formed by the spaces between first particles), and a region in which one particle (second particle) with high brightness is present in the void.

[0034] This embodiment will be described using a method for evaluating the dispersion state of particles as an example. FIG. 3 is a functional configuration diagram of an example of an evaluation device 10 according to this embodiment. The evaluation device 10 shown in FIG. 3 includes a collection unit 50, a calculation unit 52, an output unit 54, a storage unit 56, and a learning unit 58. The calculation unit 52 includes a post-processing unit 60 and a dispersion evaluation unit 62. Note that the functional configuration of FIG. 3 is only an example, and the function for performing deep learning and the function for evaluating the dispersion state of particles may be configured in separate devices that can communicate with each other via a network 18. Furthermore, functional configurations not necessary for explaining the evaluation device 10 according to this embodiment will be omitted from the illustration as appropriate.

[0035] The collection unit 50 captures an image of particles classified into multiple types dispersed therein. The storage unit 56 stores the images, such as SEM images, captured by the collection unit 50, as well as weight parameters of a deep learning model that has been trained (described later).

[0036] The learning unit 58 generates a trained deep learning model by performing deep learning using the results of semantic segmentation, as described below. Semantic segmentation is a process of labeling pixels (picture elements) of an image, such as an SEM image. Using the trained deep learning model, the learning unit 58 infers labels for all pixels of an image, such as an SEM image, as the results of semantic segmentation.

[0037] The post-processing unit 60 of the calculation unit 52 performs segmentation by image processing using the brightness difference between pixels of an image such as an SEM image. Segmentation is a process of classifying an image such as an SEM image into multiple types of regions. For example, the post-processing unit 60 performs image processing to segment void regions and particle B regions based on brightness difference.

[0038] The dispersion evaluation unit 62 measures and records the distances between all pixels in one type of region (e.g., region A of particle A) and the pixels in another type of region (e.g., region B of particle B) closest to region A. The dispersion evaluation unit 62 also performs statistical processing using a histogram of the measured distances. The histogram of the measured distances represents the dispersion state between region A and region B. Details of the histogram of the measured distances will be described later. The output unit 54 uses the results of the statistical processing by the dispersion evaluation unit 62 to output an evaluation result of the dispersion state between region A and region B, such as the inverse λ of the slope of the histogram of the dispersion state.

[0039] [Semantic Segmentation Overview] Before evaluating dispersibility, we will first explain how to generate a trained deep learning model for particle discrimination. Here, we will use as an example a system containing a mixture of two types of particles. This system is an SEM image of a dried film of internal electrode paste for a multilayer ceramic capacitor (MLCC). The average particle diameter of nickel particles (powder), an example of the first type of particles, is 100 nm. The average particle diameter of dielectric particles (powder), an example of the second type of particles, is 25 nm.

[0040] In images of samples with three-dimensional dried film surfaces, such as the dried film of the internal electrode paste of this embodiment, the brightness reflects not only the types of particles that make up the film but also the three-dimensional shapes of the particles, making it difficult to distinguish between multiple types of particles. Even if the three-dimensional shape issue does not arise, high-magnification SEM images make it difficult to simply distinguish particles by brightness due to the influence of noise. SEM images of the dried film of the internal electrode paste are composed of images of nickel particles, which do not reflect brightness or are affected by noise during imaging, low-brightness voids (where the binder resin in the internal electrode paste may be present), and high-brightness images of dielectric particles present in the voids.

[0041] On the other hand, there is a method for performing particle discrimination and extraction as a semantic segmentation task using deep learning, which demonstrates high performance in image recognition tasks. This method generally enables particle discrimination and extraction with high accuracy as long as a sufficient amount of training data is created. However, this method has the disadvantage that a large amount of data is required to create the training data, and the labor required for data preparation and annotation work is required. However, the labor required for data preparation and annotation work can be reduced by taking into account the target of deep learning.

[0042] Distinguishing between nickel particles and dielectric particles in the internal electrode paste is achieved by combining semantic segmentation with brightness classification of SEM images of the dried film. Specifically, image recognition AI performs classification inference for nickel particles, which are expected to be difficult to separate based on brightness. In addition, image processing is used to separate dielectric particles, which can be separated from voids based on brightness.

[0043] The flow of generating a learning model in the method for evaluating the particle dispersion state of this embodiment is shown in Fig. 4. Fig. 4 is a flowchart showing an example of a method for generating a trained deep learning model for particle discrimination in this embodiment. Since creating training data for semantic segmentation is extremely costly, in the flowchart of Fig. 4, pre-learning is performed using training data for a classification task, for which training data is easy to prepare.

[0044] In step S10, a deep learning model for the classification task is constructed in the learning unit 58 of the evaluation device 10. The classification task here is to separate nickel particles.

[0045] In step S12, the learning unit 58 performs pre-learning using an SEM image data set that is easy to prepare. The SEM image data set used in step S12 is assumed to be composed of, for example, 8,000 or more SEM images, and to have labels manually assigned to classify particle types as training data. The SEM image data set does not necessarily have to be the same material as the material whose dispersibility is to be evaluated, or a material whose shape is very similar. In addition, two or more types of labels are used.

[0046] The deep neural network used in pre-training consists of convolutional layers, activation layers, dropout layers, batch normalization layers, pooling layers, fully connected layers, and attention mechanisms. It is also acceptable for the deep neural network used in pre-training, or parts of it, to have weight parameters previously trained on other SEM image datasets.

[0047] Deep neural networks are structured so that the resolution of the input data gradually decreases with increasing depth, due to the convolutional and pooling layers. The final output of a deep neural network is a vector that represents the likelihood of classifying the input image using the softmax function, but it is also possible to extract outputs from multiple intermediate layers with different resolutions in preparation for semantic segmentation tasks, which will be described later.

[0048] The pre-learning may be performed by the evaluation device 10 equipped with a GPU (Graphics Processing Unit) for deep learning, or by a computer other than the evaluation device 10 equipped with a GPU.

[0049] In step S14, the collection unit 50 collects SEM images of materials such as powders, pastes, and slurries whose dispersibility is to be evaluated. The collected SEM images are secondary electron images, backscattered electron images, and EDS mapping images, as long as an expert can visually distinguish the particle types. The conditions for capturing the SEM images are not limited as long as they can be visually distinguished. The collection unit 50 collects one or more SEM images for learning and one or more for verification.

[0050] The collection unit 50 also collects mask data in which pixels in the SEM image that show particles that are difficult to extract are filled with 1 and pixels in other areas are filled with 0. The mask data is created by a user operating a computer input interface or by machine processing derivatives obtained by human operation. Furthermore, difficult-to-extract particles refer to particle species that, among two types of particle species, have a brightness intermediate between the background of the SEM image and the other particle region, or particle species that have a large brightness variance in the region in which that particle species appears, making it difficult to distinguish the particle species using rule-based image processing.

[0051] In step S16, the learning unit 58 constructs a deep neural network for semantic segmentation incorporating the pre-trained deep neural network. The learning unit 58 constructs a deep learning model for the semantic segmentation task by using the pre-trained deep learning model.

[0052] The deep learning model for the semantic segmentation task receives an image such as an SEM image as input and outputs a two-dimensional array of the same size and one channel. The deep learning model for the semantic segmentation task is trained using training data composed of SEM images of the material whose variance is to be evaluated.

[0053] The array output by the deep learning model for the semantic segmentation task has floating-point numbers close to 1.0 but less than 1.0 in the region where the particle to be extracted exists, and floating-point numbers close to 0.0 but greater than 0.0 in other regions as components.

[0054] The structure of the deep neural network for semantic segmentation is based on the U-net type structure shown in Figure 5. During training and inference, the deep neural network for semantic segmentation performs image processing such as enlarging, reducing, cropping, or tiling the original image, such as an SEM image, so that it can accommodate the input image size of the trained neural network.

[0055] The deep neural network for semantic segmentation uses a deep learning model pre-trained on an SEM image classification task in the encoder section. The deep neural network for semantic segmentation may, as necessary, insert a convolutional layer or the like between the input layer and the input of the pre-trained deep learning model to adjust the input image size, etc. Furthermore, depending on the amount of semantic segmentation training data available for training and the computer resources for training, it is acceptable to freeze the parameters of the pre-trained encoder section without calculating gradients during training.

[0056] The decoder part of a deep neural network for semantic segmentation comprises a deep neural network model consisting of convolutional layers, activation layers, dropout layers, batch normalization layers, pooling layers, fully connected layers, and attention mechanisms, as well as operations such as deconvolutional layers and upsampling. All or part of the decoder part may be replaced with a deep learning model that has been trained in advance for other tasks such as semantic segmentation and instance segmentation.

[0057] If the size of the SEM image of the material whose variance is to be evaluated is larger than the input image size of the trained deep learning model for semantic segmentation, a region of the entire SEM image equal to the input image size of the deep learning model is cropped to ensure no part is missed, and the deep learning model is used for inference. The training unit 58 adds the inference results to correspond to the original position of the entire SEM image, and finally divides each pixel by the overlap of the crop to achieve high-resolution semantic segmentation. For example, in step S18, the deep learning model for semantic segmentation is trained to extract the region where one of two types of particles (nickel particles) is captured.

[0058] In step S20, the learning unit 58 stores the structure and parameters of the learned deep learning model for semantic segmentation in the memory unit 56, thereby generating a learned model that extracts areas containing one type (nickel particles) from an SEM image.

[0059] The evaluation device 10 according to this embodiment uses the mask data described above to distinguish between regions. The mask data sets an appropriate threshold for the inference results of the trained deep learning model for semantic segmentation, and distinguishes between regions containing particles that are difficult to extract and other regions. The value in the mask data is 1 for pixels containing particles that are difficult to extract, and 1 for pixels in other regions.

[0060] By multiplying this mask data by the original SEM image, the evaluation device 10 according to this embodiment eliminates the cause of the difficulty in distinguishing between the three regions. By setting an appropriate threshold value for brightness from the image obtained by multiplying the mask data by the original SEM image, the evaluation device 10 distinguishes between regions of unextracted particle species (such as regions of the second type) and background regions (such as regions of voids).

[0061] By implementing this series of operations after the output of the trained deep learning model for semantic segmentation, the evaluation device 10 of this embodiment can realize a trained particle discrimination model for evaluating the dispersion state of particles.

[0062] Taking the classification of nickel particles in the SEM image 100 on the left in Figure 5 as an example, the trained particle discrimination model outputs the inference result 102 on the right in Figure 5. The nickel particle region in the SEM image 100 matches the region in the inference result 102. In this way, by combining the brightness of each pixel in the original SEM image 100 with the output of the trained deep learning model for semantic segmentation, it is possible to distinguish between the two types of particle regions and the rest of the background region.

[0063] This discrimination method, which combines discrimination using image recognition AI with discrimination using image processing based on brightness without using image recognition AI, can be applied to images that consist of images of nickel particles, such as dried films of internal electrode paste, where brightness is not reflected or the images are affected by noise during photography, low-brightness voids (which may contain binder resin in the paste), and high-brightness dielectric particles present in the voids. Furthermore, a discrimination method that combines discrimination using image recognition AI with discrimination using image processing based on brightness without using image recognition AI can be used when observing the crystalline structure of polycrystalline materials or sintered ceramics, when there are areas that can be clearly identified using indicators such as brightness. This discrimination method, which combines discrimination using image recognition AI with discrimination using image processing based on brightness without using image recognition AI, is useful for reducing the effort required for preparing training data and annotating it.

[0064] Next, the evaluation of particle dispersibility in this embodiment is carried out, for example, according to the procedure of the flowchart shown in Fig. 6. Fig. 6 is a flowchart showing an example of a method for evaluating the dispersion state of particles in this embodiment.

[0065] The flowchart in Fig. 6 is divided into three steps: area division, distance information collection, and statistical index calculation. Steps S30 to S34 in Fig. 6 are area division processing. Steps S36 to S40 are distance information collection processing. Steps S42 to S44 are statistical index calculation processing.

[0066] In step S30, the collection unit 50 of the evaluation device 10 reads in the SEM image of the dried film of the internal electrode paste, and stores it in the storage unit 56.

[0067] In step S32, the learning unit 58 uses image recognition AI to segment the region of particle A in the SEM image captured in step S30. For example, in the case of an SEM image of a dried film of an internal electrode paste, the learning unit 58 uses image recognition AI to segment region A of nickel particles (particle A) and then forms an image in which region A is discriminated.

[0068] Then, in step S34, the post-processing unit 60 performs image processing to segment the void region and the dielectric particle (particle B) region B based on the brightness of the dielectric particles, thereby forming an image in which region B is differentiated. Here, image recognition AI is not used to differentiate the void region from the dielectric particle region, but the image is determined based on the brightness difference between the dielectric particle (particle B) and the void.

[0069] In the two stages of the distance information collection process in steps S36 to S40 and the statistical index calculation process in steps S42 to S44, the SEM image segmented in steps S30 to S34 is processed. In the two stages of the distance information collection process and the statistical index calculation process, the image recognition AI does not function. Although the description will be given using a dried film of an internal electrode paste as an example, the method for evaluating the particle dispersion state in this embodiment is not limited to the internal electrode paste and can be applied to an image composed of multiple types of particles, for example, three regions of particle A, particle B, and voids.

[0070] The region segmentation process of steps S30 to S34 will now be further explained. First, region A is discriminated from the SEM image using trained image recognition AI. Then, region B and void regions are separated from regions other than region A in the SEM image using image processing. If region B and void regions can be easily distinguished using rule-based image processing or the like, that method may be used. After that, if necessary, post-processing such as noise removal is performed on the separated image. For example, if salt-and-pepper noise occurs, one method is to remove regions of an area of ​​a few pixels or less.

[0071] In the region segmentation stage, in the case of an SEM image of a dried film of internal electrode paste, the developed image recognition AI is used to extract region A of nickel particles, and the regions other than nickel particles are differentiated into region B of dielectric particles and region B of resin (void region) using brightness classification through image processing.

[0072] The distance information collection process of steps S36 to S40 will be further described. In step S36, the dispersion evaluation unit 62 of the evaluation device 10 focuses on one pixel in region A of each nickel particle (particle A). In step S38, the dispersion evaluation unit 62 calculates the distance from the pixel in the region of the nickel particle (particle A) focused on in step S36 to the pixel in region B of the nearest (shortest) dielectric particle (particle B). Here, region A of the nickel particle (particle diameter 100 nm) contains approximately 700,000 pixels. Region B of the dielectric particle (particle diameter 25 nm) contains approximately 200,000 pixels. Specifically, a coordinate list of pixels in region B of the dielectric particle is created in advance, and all distances to the pixels in region A of the nickel particle of interest are calculated. The distance between the pixel in region A of the nickel particle of interest and the pixel in region B of the dielectric particle is calculated, for example, using an image analysis program. The dispersion evaluation unit 62 adopts the smallest distance value among the calculated distances as the shortest distance (shortest dielectric particle distance) between the pixel of the nickel particle and the pixel of the dielectric particle under consideration.

[0073] In step S40, the variance evaluation unit 62 determines whether there are any unfocused pixels in the nickel particle region A, and if there are any unfocused pixels in the nickel particle region A, the process returns to step S36; if there are no unfocused pixels in the nickel particle region A, the process of calculating the statistical index in steps S42 to S44 is performed.

[0074] The variance evaluation unit 62 applies the distance information collection process of steps S36 to S40 to the pixels in region A of all nickel particles, and collects information on the shortest dielectric particle distance to the pixels in region A of all nickel particles. Note that the distance information collection process of steps S36 to S40 does not necessarily require creating a list for the pixels in region B of all dielectric particles. In accordance with the principles of this method, even if there is only a coordinate list of the pixels on the outer periphery of region B of the dielectric particles, the shortest dielectric particle distance to the pixels in region A of the nickel particles will not change, and it will be possible to significantly reduce calculation costs.

[0075] The process of calculating the statistical index in steps S42 to S44 will be further explained. In step S42, the dispersion evaluation unit 62 creates a histogram of the information on the shortest dielectric particle distance to the pixel in region A of the nickel particle recorded in the distance information collection process of steps S36 to S40. This histogram provides a probability distribution of the distance from a pixel in region A of a certain nickel particle to the pixel in region B of the shortest dielectric particle. The short-distance portion of the histogram reflects the size of the nickel particle itself rather than the distance between the nickel particle and the dielectric particle. On the other hand, the long-distance portion of the histogram is strongly influenced by the dispersion. The histogram reflects the fact that a sample with poor dispersion is more likely to have a longer shortest distance (shortest dielectric particle distance) between a pixel in region A of the nickel particle and a pixel in region B of the dielectric particle than a sample with good dispersion.

[0076] In step S44, the dispersion evaluation unit 62 derives an index characterizing the dispersibility from the histogram. Various indexes can be considered as indexes characterizing the dispersibility. For example, the shortest dielectric particle distances of nickel particles can be sorted in descending order, and the distance at the top 0.1% position can be used as the index. By using the distance at the top 0.1% position as the index, the dispersibility can be evaluated based on outliers of the shortest dielectric particle distances present in samples with poor dispersibility.

[0077] In this embodiment, the evaluation index λ_AB was derived using the gradient of the region in the histogram where the shortest dielectric particle distance is long, using the following formula (1): AB is the value r of the shortest dielectric particle distance between the pixel in area A on the surface of the nickel particle and the pixel in area B on the surface of the dielectric particle in the histogram. b and r a It indicates the inverse of the slope of the histogram frequency (logarithmic scale) between .

[0078]

number

[0079] Below, an example is shown in which the dispersibility evaluation index λ is obtained from an observation image by a scanning electron microscope (SEM) of the surface of the coating film of an internal electrode paste composed of an organic resin, a dispersant, nickel particles with a particle size (diameter) of 100 nm, dielectric particles with a particle size (diameter) of 25 nm, a binder resin, and a solvent.

[0080] [Image recognition AI learning] A deep neural network whose weight parameters have been pre-trained using more than 10,000 SEM images is trained using only a portion of the weight parameters using SEM images of Ni paste, an internal electrode paste for multilayer ceramic capacitors (MLCCs).

[0081] The deep neural network learned to classify the imaged objects in the pre-training stage. The output layer outputs a vector with a number of elements corresponding to the number of classes to be classified, and categorical cross-entropy is used as the activation function. In the pre-training stage, SEM images of various observation objects, in addition to Ni paste, were used as the imaged objects.

[0082] A deep neural network with a U-net structure was constructed using a pre-trained deep neural network as the encoder, and semantic segmentation of Ni paste for MLCCs was performed. The weight parameters of the encoder were fixed, and only the weight parameters of the decoder were trained. The decoder output was a vector with the same shape as the input image data. A sigmoid function was used as the activation function.

[0083] The dataset used for this training was data in which an engineer had manually marked the areas of Ni particles in an SEM image of Ni paste for MLCCs. A single SEM image was used to create the training data. When training the weight parameters of the deep neural network, the marked image data was randomly cropped to a size of, for example, 448 x 448 pixels, and the brightness, image orientation, mirror image inversion, and contrast were randomly adjusted to enhance the data.

[0084] After training, the training data was processed using a deep neural network, and the output inference results were denoised using a bilateral filter or other method. Pixels with an output value of 0.5 or greater were determined to be in the Ni particle region, while pixels with an output value of less than 0.5 were determined to be in the co-material and resin regions. Additionally, to accurately evaluate dispersibility, the Ni particle region was enlarged using an expansion process, preventing regions with high brightness around the Ni particles that were not accurately incorporated into the training data annotation from being determined to be the co-material region. Furthermore, because the image brightness differs between the co-material and resin regions, with low density resin and high density co-material, an appropriate threshold was determined from the inference results of the training data to separate the co-material and resin regions.

[0085] Figure 7 shows an example of an SEM image of Ni paste for MLCCs and an image in which the Ni region, co-material region, and resin region of the SEM image are color-coded. The original SEM image shown in Figure 7(A) is an SEM image of Ni paste for MLCCs that was not incorporated into the training data. Figure 7(B) shows an image showing the inference result in which the SEM image of Figure 7(A) is color-coded with the Ni region in a first color (e.g., green), the co-material region in a second color (e.g., pink), and the resin region in a third color (e.g., black), and the image is made semi-transparent and superimposed on the SEM image of Figure 7(A).

[0086] The brightness thresholds for the co-material and resin regions, the noise removal processing parameters, and other parameters were the same as those set at the time of processing the training data, and were determined independently of the inference results from the SEM images, which were not incorporated into the training data.

[0087] [Calculation of evaluation index λ] The collection unit 50 of the evaluation device 10 acquires an SEM image of the Ni paste for MLCCs as an observation image. The learning unit 58 separates the Ni regions from the SEM image of the Ni paste for MLCCs acquired as an observation image using the above-mentioned deep neural network. As described above, the post-processing unit 60 separates the common material region and the resin region based on the brightness difference between the common material region and the resin region.

[0088] The variance evaluation unit 62 calculates the distance from every pixel included in the Ni region to the pixel in the co-material region that is closest to that pixel (shortest co-material particle distance) according to the processing of steps S36 to S42 shown in Fig. 6, and creates a histogram from the information on the shortest co-material particle distance for the pixels in the Ni region. This histogram gives a probability distribution function for the shortest distance to the co-material region when one pixel included in the Ni region is randomly extracted.

[0089] Fig. 8 is a diagram illustrating an example of a histogram created by the dispersion evaluation unit 62. As shown in Fig. 8, the dispersion evaluation unit 62 analyzes the behavior of the histogram using the exponential function model of the above-mentioned formula (1) within a range where a distance is greater than the average particle diameter of the Ni particles and the co-material particles and where sufficient statistics are ensured, and obtains the desired correlation distance ??.

[0090] In Figure 8, a and b represent the start and end of the calculation range of the histogram for which λ is calculated. I represents the frequency of the histogram. r represents the distance. In the histogram shown in Figure 8, the radius of the nickel particle is 50 nm, so r A Since the diameter of the nickel particle is 100 nm, r B is set to 100 nm.

[0091] The histogram shown in Figure 8 shows the evaluation index λ for the dispersibility of Ni particles and co-material particles, measured for two types of Ni paste for MLCCs, labeled "Sample 1" and "Sample 2." The evaluation index λ for the dispersibility of Ni particles and co-material particles for "Sample 1" shown in Figure 8 is 15 nm. The evaluation index λ for the dispersibility of Ni particles and co-material particles for "Sample 2" shown in Figure 8 is 12 nm.

[0092] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0093] 1. Rating System 10 Evaluation equipment 12 User terminal 18 Network 50 Collection Department 52 Arithmetic section 54 Output section 56 Memory section 58 Learning Department 60 Post-processing section 62 Distributed Evaluation Unit

Claims

1. an image capturing step of capturing an image in which regions classified into a plurality of types are dispersed; a segmentation step for segmenting the captured image into a plurality of types of regions; a measurement and recording step of focusing on one of the first-type regions, measuring the shortest distance r between all pixels in the region and the pixels of the second-type region that are closest to the region, and recording the results of the measurement; a dispersibility index calculation step of calculating, after performing the measurement and recording step for all of the first type of regions, an inverse λ of the slope of a histogram of the shortest distance r and the logarithmic scale frequency I as an index of dispersibility of the regions classified into the plurality of types; This is an evaluation method performed by a computer.

2. The segmentation step involves segmenting the first type of region in the captured image using an image recognition model that has been machine-learned for the first type of region. The evaluation method according to claim 1 .

3. The segmentation step performs segmentation based on the machine-learned image recognition model only on the first type of region among a plurality of types of regions included in the image, and performs segmentation on the second type of region by image processing not based on the machine-learned image recognition model. The evaluation method according to claim 2.

4. The segmentation step includes performing segmentation to separate the first type of region included in the image based on the machine-learned image recognition model, and then separating the second type of region from regions other than the first type of region included in the image by image processing using pixel brightness differences. The evaluation method according to claim 3.

5. The first type of region is a region where first particles having an outline that can be distinguished from the surroundings are present, and the regions other than the first type of region include void regions formed by being surrounded by the first particles and the second type of region where second particles are present. The evaluation method according to claim 4.

6. a collection unit that captures an image in which regions classified into a plurality of types are dispersed; a learning unit that segments the first type of region in the captured image using an image recognition model that has been machine-learned for the first type of region; a post-processing unit that segments a second type of region from a region other than the first type of region included in the image by image processing that is not based on the machine-learned image recognition model; a dispersion evaluation unit that focuses on one of the first-type regions, measures the shortest distance r between all pixels in the region and pixels in the second-type region that are closest to the first-type region, and calculates the reciprocal λ of the slope of a histogram of the shortest distance r versus a logarithmic scale frequency I as an index of dispersion of the regions classified into the multiple types; An evaluation device having:

Images