Image processing device and method of operating same

Abstract

An image processing device according to one embodiment disclosed in this document may include an image acquisition unit that acquires active material images for a plurality of active materials; an artificial intelligence model learning unit that trains an artificial intelligence model using learning data including a plurality of reference active material images and a plurality of reference binary images or a plurality of reference boundary images corresponding to the plurality of reference active material images; an image generation unit that inputs the active material images to the artificial intelligence model to generate a binary image or a boundary image; an object identification unit that identifies a plurality of objects based on the binary image or the boundary image; and an image segmentation unit that segment the plurality of active materials included in the active material image based on the plurality of objects to obtain a segmentation image.

Description

[Technical Field]

[0001] The present invention claims the benefit of priority based on Korean Patent Application Nos. 10-2022-0175984, 10-2022-0175985, and 10-2022-0175986, filed December 15, 2022, and all contents disclosed in the documents of these Korean patent applications are incorporated herein by reference. FIELD OF THE INVENTION The embodiments disclosed herein relate to image processing devices and methods of operation thereof. [Background technology]

[0002] With technological developments and increasing demand for mobile devices and electric vehicles, the demand for lithium secondary batteries as an energy source is rapidly increasing. Lithium secondary batteries are fabricated by placing an electrode assembly including a positive electrode, a separator, and a negative electrode in a battery case and injecting an electrolyte. The positive electrode is fabricated by coating a positive electrode current collector with a composition for forming a positive electrode active material layer, including a positive electrode active material, a conductive material, and a binder, followed by drying and rolling. The negative electrode is fabricated by coating a negative electrode current collector with a composition for forming a negative electrode active material layer, including a negative electrode active material, a conductive material, and a binder, followed by drying and rolling.

[0003] Meanwhile, the electrochemical performance of lithium secondary batteries is affected not only by the composition of the positive electrode active material and negative electrode active material particles used, but also by the shape of the active material particles. This is because electrochemical properties such as electrode density, electrical conductivity, and resistance characteristics vary depending on the shape and particle size of the active material particles. Therefore, in order to manufacture lithium secondary batteries with desired performance, it is necessary to accurately analyze the shape of the active material particles.

[0004] Previously, the shape of positive electrode active material particles was analyzed using images of the particles obtained through a scanning electron microscope (SEM). However, while the conventional method of analyzing the shape of positive electrode active material using SEM images allows for qualitative analysis, quantitative analysis requires a person to manually select and measure individual active material particles from the SEM image, making automated quantitative analysis impossible. This not only makes quantitative analysis of large quantities of positive electrode active material particles difficult, but also reduces analytical accuracy due to the possibility of measurement values ​​varying depending on the person making the measurement. Summary of the Invention [Problem to be solved by the invention]

[0005] The present invention is intended to solve the above problems, and has an object to provide a method capable of automatically quantitatively analyzing the shape characteristics of a large amount of positive electrode active material and / or precursor particles.

[0006] Another object of the present invention is to provide a quantitative analysis method that minimizes user involvement in order to solve the problem of decreased analytical accuracy due to different measured values ​​depending on the user. [Means for solving the problem]

[0007] An image processing device according to one embodiment disclosed in this document may include an image acquisition unit that acquires active material images for a plurality of active materials; an artificial intelligence model learning unit that trains an artificial intelligence model using learning data including a plurality of reference active material images and a plurality of reference binary images or a plurality of reference boundary images corresponding to the plurality of reference active material images; an image generation unit that inputs the active material images to the artificial intelligence model to generate a binary image or a boundary image; an object identification unit that identifies a plurality of objects based on the binary image or the boundary image; and an image segmentation unit that segment the plurality of active materials included in the active material image based on the plurality of objects to obtain a segmentation image.

[0008] In one embodiment, the artificial intelligence model learning unit can train the artificial intelligence model so that the difference between a plurality of output images obtained by inputting the plurality of reference active material images into the artificial intelligence model and the plurality of reference binary images is less than or equal to a specified reference value.

[0009] In one embodiment, the plurality of reference active material images may include one or more first reference active material images and one or more second reference active material images, the plurality of reference binary images may include one or more first reference binary images and one or more second reference binary images, the first reference active material image may be obtained by photographing active material powder, the second reference active material image may be generated by applying a specified image processing algorithm to the first reference active material image, the first reference binary image may be a binary image of the first reference active material image, and the second reference binary image may be a binary image of the second reference active material image.

[0010] In one embodiment, the artificial intelligence model includes a plurality of layers, and at least one layer included in the encoding domain and at least one layer included in the decoding domain can be connected by a skip connection.

[0011] In one embodiment, among the plurality of layers, the layer included in the decoding region may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and an upsampling layer are sequentially connected.

[0012] In one embodiment, among the plurality of layers, the layer included in the encoding domain may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and a max pooling layer are sequentially connected.

[0013] In one embodiment, the active material image may be an image based on a scanning electron microscope (SEM), a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0014] In an embodiment, the image segmentation unit may segment the active materials included in the active material image based on a watershed algorithm.

[0015] In one embodiment, the method may further include an information extraction unit that extracts information about the plurality of active materials based on the segmentation image, and the information about the plurality of active materials may include information about particle size, perimeter, sphericity, aspect ratio, convexity, solidity, distribution, quantile, or a combination thereof of the active materials.

[0016] In one embodiment, the image processing apparatus further includes a boundary removal unit that removes boundaries from the active material image based on the boundary image to generate a boundary-removed image, and the object identification unit can identify a plurality of objects included in the boundary-removed image.

[0017] In one embodiment, the artificial intelligence model learning unit can train the artificial intelligence model so that the difference between a plurality of output images obtained by inputting the plurality of reference active material images into the artificial intelligence model and the plurality of reference boundary images is less than or equal to a specified reference value.

[0018] In one embodiment, the artificial intelligence model learning unit may calculate the differences between the plurality of output images and the plurality of reference boundary images based on a mean absolute error function, a root mean square error, a mean square error function, or a binary cross entropy loss function. In one embodiment, the binary cross-entropy loss function may assign a specified weight to a specified color.

[0019] In one embodiment, the image processing apparatus further includes a distance transform unit that transforms the binary image into a distance transformed image based on a distance transform algorithm, and a binary image filtering unit that filters the binary image using a threshold value that is set based on the distance transformed image, and the object identification unit can identify a plurality of objects included in the filtered binary image.

[0020] In one embodiment, the distance transform unit may generate the distance transformed image by calculating a distance from each pixel included in the binary image to the nearest black pixel and normalizing the calculated distance for each pixel.

[0021] In one embodiment, the distance transform unit may identify a maximum distance among the calculated distances for each of the pixels included in the binary image, and generate the distance-transformed image by max-min normalizing the calculated distance for each of the pixels based on the maximum distance.

[0022] In one embodiment, the binary image filtering unit may filter the binary image by selecting pixels having values ​​equal to or less than a specified threshold from among pixels included in the distance transformed image, and setting values ​​of pixels corresponding to the selected pixels from among pixels included in the binary image to a specified value. In one embodiment, the threshold may be the maximum normalized distance of the distance transform image multiplied by a specified percentage. [Effects of the Invention]

[0023] The image processing device and its operating method according to various embodiments disclosed herein can automatically quantitatively analyze the shape characteristics of a large number of positive electrode active material particles by acquiring active material images for a plurality of active materials, inputting the active material images into an artificial intelligence model to generate a binary image, identifying a plurality of objects included in the binary image, and then segmenting the plurality of active materials included in the active material image based on the plurality of objects.

[0024] The image processing device and its operating method according to various embodiments disclosed herein can automatically quantitatively analyze the shape characteristics of a large number of positive electrode active material particles by acquiring active material images for a plurality of active materials, inputting the active material images into an artificial intelligence model to generate a boundary image, removing the boundaries from the active material image based on the boundary image to generate a boundary-removed image, identifying a plurality of objects included in the boundary-removed image, and then segmenting the plurality of active materials included in the active material image based on the plurality of objects.

[0025] The image processing device and its operating method according to various embodiments disclosed herein acquire active material images for a plurality of active materials, input the active material images into an artificial intelligence model to generate a binary image, convert the binary image into a distance-transformed image based on a distance transformation algorithm, filter the binary image using a threshold set based on the distance-transformed image, identify a plurality of objects included in the filtered binary image, and then segment the plurality of active material objects included in the active material image based on the plurality of objects, thereby enabling automatic quantitative analysis of the shape characteristics of a large number of positive electrode active material particles.

[0026] Additionally, the image processing device and its operating method according to various embodiments disclosed herein can minimize the problem of different measurements depending on the user by utilizing an artificial intelligence model that minimizes user input.

[0027] The effects of the image processing device and its operating method disclosed in this document are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the disclosure of this document. [Brief explanation of the drawings]

[0028] [Figure 1] 1 is a block diagram of an image processing device according to an embodiment of the present disclosure. [Figure 2] 1 illustrates training data according to an embodiment of the present disclosure. [Figure 3] 1 illustrates an artificial intelligence model according to an embodiment of the present disclosure. [Figure 4] 1 illustrates an image obtained by image processing of an image processing device according to an embodiment of the present disclosure. [Figure 5] 1 is a flowchart illustrating an active material image processing method of an image processing device according to an embodiment of the present disclosure. [Figure 6]1 is a flowchart illustrating an artificial intelligence model learning method for an image processing device according to an embodiment of the present disclosure. [Figure 7] FIG. 10 is a block diagram of an image processing device according to another embodiment of the present disclosure. [Figure 8] 10 illustrates training data according to another embodiment of the present disclosure. [Figure 9] 1 illustrates an artificial intelligence model according to another embodiment of the present disclosure. [Figure 10a] 10 shows an image obtained by image processing of an image processing device according to another embodiment of the present disclosure. [Figure 10b] 10 illustrates an image difference due to image processing of an image processing device according to another embodiment of the present disclosure. [Figure 11] 10 is a flowchart illustrating an active material image processing method of an image processing device according to another embodiment of the present disclosure. [Figure 12] 10 is a flowchart illustrating an artificial intelligence model learning method for an image processing device according to another embodiment of the present disclosure. [Figure 13] FIG. 10 is a block diagram of an image processing device according to still another embodiment of the present disclosure. [Figure 14] 10 illustrates training data according to yet another embodiment of the present disclosure. [Figure 15] 1 illustrates an artificial intelligence model according to yet another embodiment of the present disclosure. [Figure 16a] 10 illustrates an image obtained by image processing of an image processing device according to still another embodiment of the present disclosure. [Figure 16b] 10 illustrates an image difference due to image processing by an image processing device according to still another embodiment of the present disclosure. [Figure 17] 10 is a flowchart illustrating an active material image processing method of an image processing device according to still another embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0029] Embodiments of the present invention will now be described with reference to the accompanying drawings, although it should be understood that this is not intended to limit the present invention to the particular embodiments, but rather to include various modifications, equivalents, and / or alternatives to the embodiments of the present invention.

[0030] The embodiments and terms used in this document are not intended to limit the technical features described in this document to a specific embodiment, but should be understood to include various modifications, equivalents, or alternatives to the embodiment. In connection with the description of the drawings, similar reference numerals may be used for similar or related components. The singular form of a noun corresponding to an item may include one or more of the item, unless the relevant context clearly dictates otherwise.

[0031] In this document, each of the phrases "A or B," "at least one of A and B," "at least one of A or B," "A, B, or C," "at least one of A, B, and C," and "at least one of A, B, or C" may include any one of the items listed therein or all possible combinations thereof. Terms such as "first," "second," "first," "second," "A," "B," "(a)," or "(b)" may be used merely to distinguish one element from another element, and do not limit the element in other respects (e.g., importance or order) unless specifically stated to the contrary.

[0032] In this document, when a (e.g., first) component is referred to as being "coupled," "coupled," or "connected" to another (e.g., second) component, with or without the terms "functionally" or "communicatively," or when a reference is made to "coupled" or "connected," this means that the component may be coupled to the other component directly (e.g., by wire or wirelessly) or indirectly (e.g., via a third component).

[0033] Methods according to various embodiments disclosed herein may be provided in a computer program product. The computer program product may be traded between a seller and a buyer as a commodity. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., Compact Disc Read Only Memory, CD-ROM) or distributed online (e.g., downloaded or uploaded) via an application store or directly between two user devices. In the case of online distribution, at least a portion of the computer program product may be at least temporarily stored or temporarily generated on a machine-readable storage medium, such as the memory of a manufacturer's server, an application store server, or an intermediary server.

[0034] According to the embodiments disclosed herein, each of the aforementioned components (e.g., modules or programs) may include one or more entities, and some of the entities may be located separately in other components. According to the embodiments disclosed herein, one or more of the aforementioned components or operations may be omitted, or one or more other components or operations may be added. Alternatively or additionally, multiple components (e.g., modules or programs) may be integrated into a single component. In this case, the integrated component may perform one or more functions of each of the multiple components in the same or similar manner to those performed by the respective components of the multiple components before the integration. According to the embodiments disclosed herein, the operations performed by a module, program, or other component may be performed sequentially, in parallel, iteratively, or heuristically, or one or more of the operations may be performed in a different order, omitted, or one or more other operations may be added.

[0035] FIG. 1 is a block diagram of an image processing device 101 according to one embodiment of the present disclosure. Referring to FIG. 1, an image processing device 101 can be coupled to an image acquisition device 103 via wires and / or wirelessly.

[0036] In one embodiment, the connection 105 between the image processing device 101 and the image acquisition device 103 may be a communication connection via a wired and / or wireless network. In one embodiment, the wired network may be based on a local area network (LAN) communication or a power line communication. In one embodiment, the wireless network may be based on a short-range communication network (e.g., Bluetooth, Wireless Fidelity (WiFi), or Infrared Data Association (IrDA)) or a long-range communication network (e.g., a cellular network, a 4G network, a 5G network).

[0037] In other embodiments, the connection 105 between the image processing device 101 and the image acquisition device 103 may be a connection via an inter-device communication method (e.g., a bus, a General Purpose Input and Output (GPIO), a Serial Peripheral Interface (SPI), or a Mobile Industry Processor Interface (MIPI)).

[0038] In one embodiment, the image acquisition device 103 may be a microscope (e.g., a scanning electron microscope) that acquires an image of the sample surface by scanning a focused electron beam over the sample surface and converting secondary electrons generated by the interaction between the electron beam and the sample into an image signal.

[0039] In one embodiment, the image acquisition device 103 can acquire an image of the active material and / or precursor. Hereinafter, an image of the active material and / or precursor may be referred to as an active material image. However, even if an image is referred to as an active material image, the present disclosure does not exclude an image of the precursor.

[0040] In one embodiment, the image acquisition device 103 may acquire an active material image of the active material. For example, the image acquisition device 103 may acquire a scanning electron microscope (SEM) image by scanning an electron beam on a positive or negative electrode active material powder. That is, the SEM image may include a positive electrode active material image and a negative electrode active material image. According to an embodiment, the SEM image may be replaced by an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0041] In one embodiment, the image acquisition device 103 can transmit an active material image of the active material to the image processing device 101. For example, the image acquisition device 103 can transmit the active material image of the active material to the image processing device 101 via the connection 105.

[0042] In one embodiment, the image processing device 101 may be a mobile device (eg, a mobile phone, a laptop computer, a smart phone, a smart pad), a computer (eg, a general-purpose computer, a special-purpose computer).

[0043] In one embodiment, the image processing device 101 may include a communication circuit 110, a memory 120, and a processor 130. According to an embodiment, the image processing device 101 shown in Figure 1 may further include at least one component (e.g., a display, an input device, or an output device) other than the components shown in Figure 1.

[0044] In one embodiment, the communication circuitry 110 can establish a wired and / or wireless communication channel between the image processing device 101 and / or the image acquisition device 103 and transmit and receive data to and from the image acquisition device 103 via the established communication channel.

[0045] In one embodiment, memory 120 may include volatile memory and / or non-volatile memory. In one embodiment, memory 120 may store data used by at least one component (e.g., processor 130) of image processing device 101. For example, the data may include program 125 (or instructions associated therewith), input data, or output data. In one embodiment, the instructions, when executed by processor 130, may cause image processing device 101 to perform the operations defined by the instructions.

[0046] In one embodiment, the memory 120 may include programs 125 (e.g., an artificial intelligence model training unit 141, an artificial intelligence model 145, an image acquisition unit 150, an image generation unit 160, an object identification unit 170, an image segmentation unit 180, and / or an information extraction unit 190).

[0047] In one embodiment, processor 130 may include a central processing unit, an application processor, a graphics processing unit, a neural processing unit (NPU), an image signal processor, a sensor hub processor, or a communication processor.

[0048] In one embodiment, the processor 130 can execute the program 125 (e.g., the artificial intelligence model training unit 141, the artificial intelligence model 145, the image acquisition unit 150, the image generation unit 160, the object identification unit 170, the image segmentation unit 180, and / or the information extraction unit 190), control at least one other component (e.g., hardware or software component) of the image processing device 101 coupled to the processor 130, and perform various data processing or calculations.

[0049] In one embodiment, the AI ​​model training unit 141 may train the AI ​​model 145 based on the training data. In one embodiment, the AI ​​model 145 may be a model trained to convert an active material image into a binary image. In one embodiment, the image acquisition unit 150 may acquire an active material image of the active material from the image acquisition device 103. In one embodiment, the image generation unit 160 may generate a binary image by inputting the active material image into the AI ​​model 145. In one embodiment, the object identification unit 170 may identify multiple objects included in the binary image. In one embodiment, the image segmentation unit 180 may segment multiple active material objects included in the active material image based on the multiple objects and acquire a segmentation image. In one embodiment, the information extraction unit 190 may extract information about the active material based on the active material objects included in the segmentation image.

[0050] Hereinafter, with reference to Figures 2 to 4, a method in which the image processing device 101 processes an image acquired from the image acquisition device 103 via the artificial intelligence model learning unit 141, the artificial intelligence model 145, the image acquisition unit 150, the image generation unit 160, the object identification unit 170, the image segmentation unit 180, and / or the information extraction unit 190 will be described in detail.

[0051] Training data FIG. 2 illustrates training data according to one embodiment of the present disclosure. Referring to FIG. 2, the training data may include reference active material images 211, 213, 215, 221, 223, 225, 231, 233, 235, and reference binary images 251, 253, 255, 261, 263, 265, 271, 273, 275.

[0052] The reference active material images 211, 213, 215, 221, 223, 225, 231, 233, and 235 included in the training data may have the same size as the reference binary images 251, 253, 255, 261, 263, 265, 271, 273, and 275. In one embodiment, the image size may be defined as the number of horizontal pixels x the number of vertical pixels. The number of horizontal pixels of the images included in the training data may be an integer between 32 and 4096, and the number of vertical pixels may be an integer between 32 and 4096. For example, the image size may be 256 x 256.

[0053] In one embodiment, the plurality of reference active material images 211, 213, 215, 221, 223, 225, 231, 233, 235 can include one or more first reference active material images and one or more second reference active material images.

[0054] In one embodiment, the first reference active material image can be acquired by photographing the active material powder. In one embodiment, the first reference active material image can be an SEM image acquired directly via the image acquisition device 103.

[0055] In one embodiment, the second reference active material image may be an image transformed from the first reference active material image. In one embodiment, the second reference active material image may be generated by applying a specified first image processing algorithm to the first reference active material image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0056] In one embodiment, the plurality of reference binary images 251, 253, 255, 261, 263, 265, 271, 273, 275 may include one or more first reference binary images and one or more second reference binary images.

[0057] In one embodiment, the first reference binary image may be generated by applying a second image processing algorithm to the first reference active material image, where the second image processing algorithm may include a mean shift filter, a boundary extraction algorithm, a boundary removal algorithm, a binarization algorithm, or a combination thereof.

[0058] In one embodiment, the second reference binary image may be an image transformed from the first reference binary image. In one embodiment, the second reference binary image may be generated by applying a specified first image processing algorithm to the first reference binary image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0059] In one embodiment, the first reference binary image may be a binary image of the first reference active material image, and the second reference binary image may be a binary image of the second reference active material image. For example, reference binary image 251 may be a binary image of reference active material image 211, reference binary image 263 may be a binary image of reference active material image 223, and reference binary image 275 may be a binary image of reference active material image 235. Therefore, the multiple reference active material images 211, 213, 215, 221, 223, 225, 231, 233, and 235 and the multiple reference binary images 251, 253, 255, 261, 263, 265, 271, 273, and 275 may be grouped into image sets, with corresponding images. For example, reference active material image 211 and reference binary image 251 may be grouped into a single image set.

[0060] In one embodiment, the training data can be used to train the artificial intelligence model 145. In one embodiment, the training data can be used to train the artificial intelligence model 145 for a specified number of epochs. Here, the specified number can be set between 100 and 10,000. For example, the specified number can be 3,000.

[0061] In one embodiment, the training data may be divided into mini-batches, each divided by a batch size. Here, the batch size may be determined between 1 and 512. For example, the batch size may be 3. When the batch size is 3, the mini-batches may be divided into three image sets (i.e., three reference active material images and three reference binary images). For example, the reference active material images 211, 213, and 215 and the reference binary images 251, 253, and 255 may constitute a first mini-batch; the reference active material images 221, 223, and 225 and the reference binary images 261, 263, and 265 may constitute a second mini-batch; and the reference active material images 231, 233, and 235 and the reference binary images 271, 273, and 275 may constitute a third mini-batch.

[0062] Artificial Intelligence Model 3 illustrates an example of training the artificial intelligence model 145 using a mini-batch consisting of reference active material images 301, 303, and 305 and reference binary images 391, 393, and 395, according to an embodiment of the present disclosure.

[0063] In one embodiment, the artificial intelligence model 145 may be a model based on a convolutional neural network (CNN) or a U-NET. In one embodiment, the artificial intelligence model 145 may be a model trained to convert an active material image into a binary image.

[0064] 3, the artificial intelligence model 145 may include multiple layers 310, 320, 330, 340, 350, 360, and 370. The multiple layers 310, 320, 330, 340, 350, 360, and 370 may be sequentially connected. The input of a sequentially connected layer may be the output of the layer immediately preceding it, and the output of a sequentially connected layer may be the input of the layer immediately following it. For example, the output of layer 310 may be the input of layer 320.

[0065] At least two layers (e.g., layers 310, 320, 330, 340, 350, 360, 370) among the plurality of layers 310, 320, 330, 340, 350, 360, 370 may be connected by skip connections 315, 325, 335. In one embodiment, the skip connections 315, 325, 335 may connect a layer included in the encoding domain (e.g., layers 310, 320, 330) to a layer included in the decoding domain (e.g., layers 350, 360, 370) among the plurality of layers 310, 320, 330, 340, 350, 360, 370. Here, the skip connections 315, 325, 335 may refer to connections between layers for inputting the outputs of the layers 310, 320, 330 to the layers 350, 360, 370. For example, skip connection 335 allows layer 350 to receive the output of layer 340 and the output of layer 330 as inputs.

[0066] In one embodiment, the layers 310, 320, 330, 340, 350, 360, and 370 may include an input layer, a batch normalization layer, a 2D convolution layer, an activation layer, a max pooling layer, an upsampling layer, a concatenation layer, or a combination thereof. For example, the layers included in the encoding domain (e.g., layers 310, 320, and 330) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and a max pooling layer are sequentially connected. As another example, the layers included in the decoding domain (e.g., layers 350, 360, and 370) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and an upsampling layer are sequentially connected.

[0067] In one embodiment, the input layer may be a layer that receives input of bit values ​​included in the reference active material images 301, 303, and 305. In one embodiment, the input layer may acquire input consisting of bit values ​​equal to the number of horizontal pixels x the number of vertical pixels x the number of channels (or depth). For example, if the reference active material image has an N x N size, the input layer may receive input of N x N x 1 bit values, where N may be an integer between 32 and 4096 (e.g., 256). According to an embodiment, the number of channels may also be referred to as depth.

[0068] In one embodiment, the batch normalization layer may be a layer for normalizing the output value of a previous layer in batch units. According to an embodiment, the batch normalization layer may be disposed immediately before the activation layer. Here, a batch may include reference active material images input during one iteration to train the artificial intelligence model 145. For example, if the batch size is 3, the number of reference active material images input to the artificial intelligence model 145 during one iteration may be 3. Here, the batch size may be an integer between 1 and 512 (e.g., 3).

[0069] In one embodiment, the 2D convolution layer may be a layer for obtaining an output by performing a convolution operation between an input and a filter of a specified size. In one embodiment, the 2D convolution layer included in the artificial intelligence model 145 assumes that the output image size (number of horizontal pixels × number of vertical pixels) is the same as the input image size (number of horizontal pixels × number of vertical pixels). In one embodiment, the number of output channels can be changed depending on the number of filters (or filter depth) of the 2D convolution layer. For example, if the number of filters is two, the number of output channels can be increased by two times compared to the number of input channels. In one embodiment, the number of output channels can be changed depending on the depth stride of the 2D convolution layer. For example, if the depth stride is two, the number of output channels can be reduced by half compared to the number of input channels.

[0070] In one embodiment, an activation layer may be a layer that applies a specified activation function to an input to obtain an output. For example, the specified activation function may include a step, a sigmoid, a rectifier linear unit (ReLU), an exponential linear unit (ELU), a softmax, or a combination thereof.

[0071] In one embodiment, the max pooling layer may be a layer for selecting the largest value from each pooling region of the input to obtain an output. In one embodiment, the output image size (i.e., the number of horizontal and vertical pixels) can be changed depending on the size of the pooling region of the max pooling layer. For example, if the size of the pooling region is 2x2, the output image size (e.g., 128x128) can be reduced by half compared to the input image size (e.g., 256x256). According to an embodiment, the artificial intelligence model 145 may include a pooling layer of a different type instead of a max pooling layer. For example, the pooling layer of a different type may include an average pooling layer.

[0072] In one embodiment, the upsampling layer may be a layer for increasing the resolution of an image, hi one embodiment, the upsampling layer may be a layer for increasing the output image size compared to the input image size using a specified interpolation algorithm.

[0073] In one embodiment, the concatenation layer may concatenate two or more inputs and output the concatenated inputs. Here, the concatenation may be depthwise concatenation. For example, if a first input is 128x128x128 and a second input is 128x128x128, the output of the concatenation layer may be 128x128x256. If another first input is 256x256x64 and a second input is 128x128x128, the output of the concatenation layer may be 256x256x128.

[0074] 3 shows seven layers 310, 320, 330, 340, 350, 360, and 370, this is for illustrative purposes only and the number of layers is not limited to seven. For example, the number of layers included in the artificial intelligence model 145 may be 32.

[0075] When the artificial intelligence model 145 includes 32 layers, the artificial intelligence model 145 may have a structure in which an input layer, a first batch normalization layer, a first 2D convolution layer, and a second 2D convolution layer are sequentially connected. The input of the sequentially connected layer may be the output of the layer immediately preceding it, and the output of the sequentially connected layer may be the input of the layer immediately following it. Here, the input of the input layer may be 256×256×1, and the output may be 256×256×1. The output of the first batch normalization layer may be 256×256×1. The output of the first 2D convolution layer may be 256×256×64. The output of the second 2D convolution layer may be 256×256×64.

[0076] Following the second 2D convolutional layer, the artificial intelligence model 145 may have a structure in which a second batch normalization layer, a first activation layer, and a first max pooling layer are sequentially connected. Here, the output of the second batch normalization layer may be 256×256×64, the output of the first activation layer may be 256×256×64, and the output of the first max pooling layer may be 128×128×64.

[0077] Following the first max pooling layer, the artificial intelligence model 145 may have a structure in which a third 2D convolutional layer, a fourth 2D convolutional layer, a third batch normalization layer, a second activation layer, and a second max pooling layer are sequentially connected. Here, the output of the third 2D convolutional layer may be 128×128×128. The output of the fourth 2D convolutional layer may be 128×128×128. The output of the third batch normalization layer may be 128×128×128. The output of the second activation layer may be 128×128×128. The output of the second max pooling layer may be 64×64×128.

[0078] Following the second max pooling layer, the artificial intelligence model 145 may have a structure in which a fifth 2D convolutional layer, a sixth 2D convolutional layer, a fourth batch normalization layer, a third activation layer, and a first upsampling layer are sequentially connected. Here, the output of the fifth 2D convolutional layer may be 64×64×256. The output of the sixth 2D convolutional layer may be 64×64×256. The output of the fourth batch normalization layer may be 64×64×256. The output of the third activation layer may be 64×64×256. The output of the first upsampling layer may be 128×128×256.

[0079] Following the first upsampling layer, the artificial intelligence model 145 may have a structure in which a seventh 2D convolutional layer, a first concatenation layer, an eighth 2D convolutional layer, a ninth 2D convolutional layer, a fifth batch normalization layer, a fourth activation layer, and a second upsampling layer are sequentially connected. Of the two inputs of the first concatenation layer, the first input may be the output of the seventh 2D convolutional layer, and the second input may be the output of the fourth 2D convolutional layer. Here, the output of the seventh 2D convolutional layer may be 128×128×128. The output of the first concatenation layer may be 128×128×256. The output of the eighth 2D convolutional layer may be 128×128×128. The output of the ninth 2D convolutional layer may be 128×128×128. The output of the fifth batch normalization layer may be 128×128×128. The output of the fourth activation layer may be 128x128x128. The output of the second upsampling layer may be 256x256x128.

[0080] Following the second upsampling layer, the artificial intelligence model 145 may have a structure in which a tenth 2D convolutional layer, a second concatenation layer, an eleventh 2D convolutional layer, a twelfth 2D convolutional layer, a sixth batch normalization layer, a fifth activation layer, a thirteenth 2D convolutional layer, and a fourteenth 2D convolutional layer are sequentially concatenated. Of the two inputs of the second concatenation layer, the first input may be the output of the tenth 2D convolutional layer, and the second input may be the output of the second 2D convolutional layer. Here, the output of the tenth 2D convolutional layer may be 256×256×64. The output of the second concatenation layer may be 256×256×128. The output of the eleventh 2D convolutional layer may be 256×256×64. The output of the twelfth 2D convolutional layer may be 256×256×64. The output of the sixth batch normalization layer may be 256x256x64. The output of the fifth activation layer may be 256x256x64. The output of the thirteenth 2D convolutional layer may be 256x256x2. The output of the fourteenth 2D convolutional layer may be 256x256x2.

[0081] AI model learning Referring to FIG. 3, the operation of the AI ​​model learning unit 141 to learn the AI ​​model 145 using images 301, 303, 305, 391, 393, and 395 included in the mini-batch will be described.

[0082] In one embodiment, the AI ​​model learning unit 141 can sequentially input the reference activation images 301 , 303 , and 305 to the AI ​​model 145 .

[0083] In one embodiment, the artificial intelligence model learning unit 141 can train the artificial intelligence model 145 so that the difference between the output images of the artificial intelligence model 145 acquired sequentially and the reference binary images 391, 393, 395 is less than or equal to a specified reference value.

[0084] For example, the artificial intelligence model learning unit 141 may calculate the difference between the output image and the reference binary images 391, 393, and 395 based on a loss function (or cost function). Here, the loss function may include a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function. According to an embodiment, the binary cross entropy loss function may be a function in which a weight between 1 and 1000 is applied to one color (e.g., white) of two colors (or classes) (e.g., black and white).

[0085] Thereafter, the artificial intelligence model learning unit 141 may adjust the weights of the artificial intelligence model 145 so that the difference is equal to or less than a specified reference value (or has a minimum value). In one embodiment, the artificial intelligence model learning unit 141 may adjust the weights of the artificial intelligence model 145 using a gradient descent-based algorithm (e.g., Adam, SGD, or Momentum) so that the difference is equal to or less than a specified reference value (or has a minimum value). Here, the adjusted weights may be determined according to a learning rate. The learning rate may be determined between 0.00000001 and 0.1. For example, the learning rate may be 0.0001.

[0086] Thereafter, the AI ​​model learning unit 141 can train the AI ​​model 145 using images included in the next mini-batch. After that, when the AI ​​model 145 has been trained using images included in all mini-batches, the AI ​​model learning unit 141 can train the AI ​​model 145 again using images included in the mini-batch according to the specified number of epochs.

[0087] Image processing using artificial intelligence models FIG. 4 shows images 410, 430, and 450 resulting from image processing by the image processing device 101 according to one embodiment of the present disclosure.

[0088] In one embodiment, the image acquisition unit 150 may acquire the active material image 410 for the active materials 411, 413, and 415. In one embodiment, the image acquisition unit 150 may acquire the active material image 410 via the image acquisition device 103.

[0089] In one embodiment, the image generation unit 160 may generate the binary image 430. In one embodiment, the image generation unit 160 may generate the binary image 430 by inputting the active material image 410 into the artificial intelligence model 145.

[0090] In one embodiment, the object identification unit 170 may identify a plurality of objects 431, 433, and 435 included in the binary image 430. Here, the plurality of objects 431, 433, and 435 included in the binary image 430 may correspond to the active materials 411, 413, and 415 in the active material image 410. In one embodiment, the plurality of objects 431, 433, and 435 may be configured as regions having a specified value (e.g., a value indicating white). In one embodiment, the plurality of objects 431, 433, and 435 may be separated by regions having another specified value (e.g., a value indicating black). Although only three objects are labeled in FIG. 4, it can be seen that other objects exist in the binary image 430 by referring to FIG. 4.

[0091] In one embodiment, the image segmentation unit 180 may segment the plurality of active materials 411, 413, and 415 included in the active material image 410 based on the plurality of objects 431, 433, and 435 to obtain a segmentation image 450. In one embodiment, the image segmentation unit 180 may segment the plurality of active materials 411, 413, and 415 included in the active material image 410 based on a watershed algorithm.

[0092] In one embodiment, the information extraction unit 190 can extract information about the active materials 411, 413, and 415 based on the active material objects 451, 453, and 455 included in the segmentation image 450. In one embodiment, the information about the active materials 411, 413, and 415 can include information about the average particle size, average perimeter, average sphericity, average aspect ratio, average convexity, and average solidity of the active materials, or distribution values ​​thereof. The information about the active materials 411, 413, and 415 can also include information about the individual particle size, perimeter, sphericity, aspect ratio, convexity, and / or solidity of each active material. The information about the active materials 411, 413, and 415 can also include information about the quantile values ​​(e.g., 1% to 99%, or D5, D50, D95, etc.) of each active material.

[0093] FIG. 5 is a flowchart illustrating an active material image processing method of the image processing device 101 according to one embodiment of the present disclosure. 5, in operation 510, the image processing device 101 may acquire an active material image 410 for the active materials 411, 413, and 415. In one embodiment, the image processing device 101 may acquire the active material image 410 via the image acquisition device 103.

[0094] In operation 520, the image processing device 101 can generate a binary image 430 by inputting the active material image 410 into the artificial intelligence model 145. Here, the artificial intelligence model 145 may be a model trained by the artificial intelligence model training method shown in FIG.

[0095] In operation 530 , the image processing device 101 may identify a number of objects 431 , 433 , 435 contained in the binary image 430 .

[0096] In operation 540, the image processing device 101 may segment the plurality of active materials 411, 413, 415 included in the active material image 410 based on the plurality of objects 431, 433, 435 to obtain a segmentation image 450. In one embodiment, the image processing device 101 may segment the plurality of active materials 411, 413, 415 included in the active material image 410 based on a watershed algorithm.

[0097] 6 is a flowchart showing an artificial intelligence model training method for the image processing device 101 according to an embodiment of the present disclosure. The operation of FIG. 6 can be applied to the first epoch of training data.

[0098] 6, in operation 610, the image processing device 101 may set parameters of the artificial intelligence model 145. In one embodiment, the image processing device 101 may set parameters of the artificial intelligence model 145 based on initial parameter values.

[0099] In operation 615, the image processing device 101 may set i to 1. i may indicate an index of an image set included in the training data, where the image set may include a reference active material image and its corresponding reference binary image.

[0100] In operation 620, the image processing device 101 may obtain an output image based on the i-th reference active material image. In one embodiment, the image processing device 101 may generate the output image by inputting the i-th reference active material image into the artificial intelligence model 145.

[0101] In operation 630, the image processing device 101 may determine whether the difference between the output image and the i-th reference binary image is equal to or smaller than a reference value. In one embodiment, the image processing device 101 may calculate the difference between the output image and the i-th reference binary image based on a loss function (or cost function). Here, the loss function may include a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function. According to an embodiment, the binary cross entropy loss function may be a function in which a weight between 1 and 1000 is applied to one color (e.g., white) of two colors (or classes) (e.g., black and white).

[0102] If the result of the determination in operation 630 is that the difference is equal to or less than the standard (YES determination), the image processing device 101 can proceed to operation 650. If the result of the determination in operation 630 is that the difference exceeds the standard (NO determination), the image processing device 101 can proceed to operation 640.

[0103] In operation 640, the image processing device 101 may adjust parameters of the artificial intelligence model 145. In one embodiment, the image processing device 101 may adjust parameters of the artificial intelligence model 145 so that the difference is equal to or less than a specified reference value (or has a minimum value). In one embodiment, the image processing device 101 may adjust parameters of the artificial intelligence model 145 using a gradient descent-based algorithm (e.g., Adam, SGD, Momentum) so that the difference is equal to or less than a specified reference value (or has a minimum value). Here, the adjusted weights may be determined according to a learning rate. The learning rate may be determined between 0.00000001 and 0.1. For example, the learning rate may be 0.0001.

[0104] In operation 650, the image processing device 101 may determine whether training is complete. In one embodiment, training may be determined to be complete when i indicates the end of the image set.

[0105] If the result of the determination in operation 650 is that learning is complete (YES determination), the image processing device 101 can end the operation according to Fig. 6. If the result of the determination in operation 650 is that learning is not complete (NO determination), the image processing device 101 can proceed to operation 655. In operation 655, the image processing device 101 may increment the value of i by 1. Operation 620 may then be performed again.

[0106] Although the present disclosure illustrates an example in which the image processing device 101 performs an image processing method based on an active material image, this is merely an example. According to an embodiment, the image processing device 101 can separate and / or analyze particles by performing similar image processing on not only the active material image but also the precursor image. The image acquisition device 103 can acquire images of the active material and / or the precursor.

[0107] Furthermore, although the active material image is illustrated as an SEM image in this disclosure, this is merely an example. According to embodiments, the active material image may be replaced with an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0108] The image processing device 101 and its operating method as described above acquire an active material image 410 for a plurality of active materials 411, 413, and 415, input the active material image 410 into an artificial intelligence model 145 to generate a binary image 430, identify a plurality of objects 431, 433, and 435 included in the binary image 430, and then segment the plurality of active materials 411, 413, and 415 included in the active material image 410 based on the plurality of objects 431, 433, and 435, thereby enabling automatic quantitative analysis of the shape characteristics of a large number of positive electrode active material particles.

[0109] Furthermore, the image processing device 101 and its operating method as described above can minimize the problem of different measurements depending on the user by utilizing an artificial intelligence model 145 that minimizes user input.

[0110] Meanwhile, although the above description has been given as an example of a case where the active material image is a positive electrode active material image, the technical concepts disclosed in this document can be substantially similarly applied to a case where the active material image is a negative electrode active material image.

[0111] FIG. 7 is a block diagram of an image processing device 701 according to another embodiment of the present disclosure. Referring to FIG. 7, an image processing device 701 can be coupled to an image acquisition device 703 via wires and / or wirelessly.

[0112] In one embodiment, the connection 705 between the image processing device 701 and the image acquisition device 703 may be a communication connection via a wired and / or wireless network. In one embodiment, the wired network may be based on a local area network (LAN) communication or a power line communication. In one embodiment, the wireless network may be based on a short-range communication network (e.g., Bluetooth, Wireless Fidelity (WiFi), or Infrared Data Association (IrDA)) or a long-range communication network (e.g., a cellular network, a 4G network, a 5G network).

[0113] In other embodiments, the connection 705 between the image processing device 701 and the image acquisition device 703 may be a connection via an inter-device communication method (e.g., a bus, a General Purpose Input and Output (GPIO), a Serial Peripheral Interface (SPI), or a Mobile Industry Processor Interface (MIPI)).

[0114] In one embodiment, the image acquisition device 703 can acquire an image of the active material and / or precursor. Hereinafter, an image of the active material and / or precursor may be referred to as an active material image. However, even if an image is referred to as an active material image, the present disclosure does not exclude an image of the precursor.

[0115] In one embodiment, the image acquisition device 703 may be a microscope (e.g., a scanning electron microscope) that acquires an image of the sample surface by scanning a focused electron beam over the sample surface and converting secondary electrons generated by the interaction between the electron beam and the sample into image signals.

[0116] In one embodiment, the image acquisition device 703 may acquire an active material image of the active material. For example, the image acquisition device 703 may scan an electron beam on the positive or negative electrode active material powder to acquire a scanning electron microscope (SEM) image. That is, the SEM image may include a positive electrode active material image and a negative electrode active material image. According to an embodiment, the SEM image may be replaced by an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0117] In one embodiment, the image capture device 703 can transmit an active material image of the active material to the image processing device 701. For example, the image capture device 703 can transmit the active material image of the active material to the image processing device 701 via connection 705.

[0118] In one embodiment, the image processing device 701 may be a mobile device (eg, a mobile phone, a laptop computer, a smart phone, a smart pad), a computer (eg, a general-purpose computer, a special-purpose computer).

[0119] In one embodiment, the image processing device 701 may include a communication circuit 710, a memory 720, and a processor 730. According to an embodiment, the image processing device 701 shown in Figure 7 may further include at least one component (e.g., a display, an input device, or an output device) other than the components shown in Figure 7.

[0120] In one embodiment, the communication circuitry 710 can establish a wired and / or wireless communication channel between the image processing device 701 and / or the image acquisition device 703 and transmit and receive data to and from the image acquisition device 703 via the established communication channel.

[0121] In one embodiment, memory 720 may include volatile memory and / or non-volatile memory. In one embodiment, memory 720 may store data used by at least one component (e.g., processor 730) of image processing device 701. For example, the data may include program 725 (or instructions associated therewith), input data, or output data. In one embodiment, the instructions, when executed by processor 730, may cause image processing device 701 to perform the operations defined by the instructions.

[0122] In one embodiment, the memory 720 may include programs 725 (e.g., an artificial intelligence model training unit 741, an artificial intelligence model 745, an image acquisition unit 750, an image generation unit 760, a boundary removal unit 765, an object identification unit 770, an image segmentation unit 780, and / or an information extraction unit 790).

[0123] In one embodiment, processor 730 may include a central processing unit, an application processor, a graphics processing unit, a neural processing unit (NPU), an image signal processor, a sensor hub processor, or a communication processor.

[0124] In one embodiment, the processor 730 can execute the program 725 (e.g., the artificial intelligence model training unit 741, the artificial intelligence model 745, the image acquisition unit 750, the image generation unit 760, the boundary removal unit 765, the object identification unit 770, the image segmentation unit 780, and / or the information extraction unit 790), control at least one other component (e.g., hardware or software component) of the image processing device 701 coupled to the processor 730, and perform various data processing or calculations.

[0125] In one embodiment, the AI ​​model training unit 741 may train the AI ​​model 745 based on the training data. In one embodiment, the AI ​​model 745 may be a model trained to convert an active material image into a binary image. In one embodiment, the image acquisition unit 750 may acquire an active material image of the active material from the image acquisition device 703. In one embodiment, the image generation unit 760 may generate a boundary image by inputting the active material image into the AI ​​model 745. In one embodiment, the boundary removal unit 765 may remove boundaries from the active material image based on the boundary image to generate a boundary-removed image. In one embodiment, the object identification unit 770 may identify multiple objects included in the boundary-removed image. In one embodiment, the image segmentation unit 780 may segment multiple active materials included in the active material image based on the multiple objects to obtain a segmentation image. In one embodiment, the information extraction unit 790 may extract information about the active material based on the active material objects included in the segmentation image.

[0126] Hereinafter, with reference to Figures 8, 9, 10a, and 10b, a detailed description will be given of how the image processing device 701 processes images acquired from the image acquisition device 703 via the artificial intelligence model learning unit 741, the artificial intelligence model 745, the image acquisition unit 750, the image generation unit 760, the boundary removal unit 765, the object identification unit 770, the image segmentation unit 780, and / or the information extraction unit 790.

[0127] Training data FIG. 8 shows training data according to another embodiment of the present disclosure. Referring to FIG. 8, the training data may include reference active material images 811, 812, 813, 814, 815, 816, 817, 818, 819, 820, 821, and reference edge images 851, 852, 853, 854, 855, 856, 857, 858, 859, 860, 861.

[0128] The reference active material images 811, 812, 813, 814, 815, 816, 817, 818, 819, 820, and 821 included in the training data may have the same size as the reference boundary images 851, 852, 853, 854, 855, 856, 857, 858, 859, 860, and 861. In one embodiment, the image size may be defined as the number of horizontal pixels x the number of vertical pixels. The number of horizontal pixels of the images included in the training data may be an integer between 32 and 4096, and the number of vertical pixels may be an integer between 32 and 4096. For example, the image size may be 256 x 256.

[0129] In one embodiment, the plurality of reference active material images 811, 812, 813, 814, 815, 816, 817, 818, 819, 820, 821 can include one or more first reference active material images and one or more second reference active material images.

[0130] In one embodiment, the first reference active material image can be acquired by photographing the active material powder. In one embodiment, the first reference active material image can be an SEM image acquired directly via the image acquisition device 703.

[0131] In one embodiment, the second reference active material image may be an image transformed from the first reference active material image. In one embodiment, the second reference active material image may be generated by applying a specified first image processing algorithm to the first reference active material image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0132] In one embodiment, the plurality of reference boundary images 851, 852, 853, 854, 855, 856, 857, 858, 859, 860, 861 may include one or more first reference boundary images and one or more second reference boundary images.

[0133] In one embodiment, the first reference boundary image may be generated by applying a second image processing algorithm to the first reference active material image, where the second image processing algorithm may include a mean shift filter and / or a boundary extraction algorithm.

[0134] In one embodiment, the second reference boundary image may be an image transformed from the first reference boundary image. In one embodiment, the second reference boundary image may be generated by applying a specified first image processing algorithm to the first reference boundary image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0135] In one embodiment, the first reference boundary image may be a boundary image of the first reference active material image, and the second reference boundary image may be a boundary image of the second reference active material image. For example, reference boundary image 851 may be a boundary image for reference active material image 811, reference boundary image 856 may be a boundary image for reference active material image 816, and reference boundary image 861 may be a boundary image for reference active material image 821. Therefore, the multiple reference active material images 811, 812, 813, 814, 815, 816, 817, 818, 819, 820, and 821 and the multiple reference boundary images 851, 852, 853, 854, 855, 856, 857, 858, 859, 860, and 861 may be grouped into image sets, with corresponding images. For example, reference active material image 811 and reference boundary image 851 may be grouped into a single image set.

[0136] In one embodiment, the training data can be used to train the artificial intelligence model 745. In one embodiment, the training data can be used to train the artificial intelligence model 745 for a specified number of epochs. Here, the specified number can be set between 100 and 10,000. For example, the specified number can be 3,000.

[0137] In one embodiment, the training data may be divided into mini-batches, each divided by a batch size. Here, the batch size may be determined between 1 and 512. For example, the batch size may be 4. When the batch size is 4, the mini-batches may be divided into four image sets (i.e., four reference active material images and four reference boundary images). For example, reference active material images 811, 812, 813, and 814 and reference boundary images 851, 852, 853, and 854 may constitute a first mini-batch, and reference active material images 815, 816, 817, and 818 and reference boundary images 855, 856, 857, and 858 may constitute a second mini-batch.

[0138] Artificial Intelligence Model 9 illustrates an artificial intelligence model 745 according to another embodiment of the present disclosure. Specifically, FIG. 9 illustrates an example in which the artificial intelligence model 745 is trained using a mini-batch consisting of reference active material images 901, 905 and reference boundary images 991, 995.

[0139] In one embodiment, the artificial intelligence model 745 may be a model based on a convolutional neural network (CNN) or a U-NET. In one embodiment, the artificial intelligence model 745 may be a model trained to convert an active material image into a boundary image.

[0140] 9, the artificial intelligence model 745 may include multiple layers 910, 920, 930, 940, 950, 960, and 970. The multiple layers 910, 920, 930, 940, 950, 960, and 970 may be sequentially connected. The input of a sequentially connected layer may be the output of the immediately preceding layer, and the output of a sequentially connected layer may be the input of the immediately following layer. For example, the output of layer 910 may be the input of layer 920.

[0141] At least two layers (e.g., layers 910, 970, 920, 960, or 930, 950) among the plurality of layers 910, 920, 930, 940, 950, 960, and 970 may be connected by skip connections 915, 925, and 935. In one embodiment, the skip connections 915, 925, and 935 may connect a layer included in the encoding domain (e.g., layers 910, 920, and 930) among the plurality of layers 910, 920, 930, 940, 950, 960, and 970 to a layer included in the decoding domain (e.g., layers 950, 960, and 970). Here, the skip connections 915, 925, and 935 may refer to connections between layers for inputting the outputs of the layers 910, 920, and 930 to the layers 950, 960, and 970. For example, skip connection 935 allows layer 950 to receive the output of layer 940 and the output of layer 930 as inputs.

[0142] In one embodiment, the layers 910, 920, 930, 940, 950, 960, and 970 may include an input layer, a batch normalization layer, a 2D convolution layer, an activation layer, a max pooling layer, an upsampling layer, a concatenation layer, or a combination thereof. For example, the layers included in the encoding domain (e.g., layers 910, 920, and 930) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and a max pooling layer are sequentially connected. As another example, the layers included in the decoding domain (e.g., layers 950, 960, and 970) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and an upsampling layer are sequentially connected.

[0143] In one embodiment, the input layer may be a layer that receives input of bit values ​​included in the reference active material images 901 and 905. In one embodiment, the input layer may acquire input consisting of bit values ​​equal to the number of horizontal pixels x the number of vertical pixels x the number of channels (or depth). For example, if the reference active material image has an N x N size, the input layer may receive input of N x N x 1 bit values, where N may be an integer between 32 and 4096 (e.g., 256). According to an embodiment, the number of channels may also be referred to as depth.

[0144] In one embodiment, the batch normalization layer may be a layer for normalizing the output value of a previous layer in batch units. According to an embodiment, the batch normalization layer may be disposed immediately before the activation layer. Here, a batch may include reference active material images input during one iteration to train the artificial intelligence model 745. For example, if the batch size is 4, the number of reference active material images input to the artificial intelligence model 745 during one iteration may be 4. Here, the batch size may be an integer between 1 and 512 (e.g., 4).

[0145] In one embodiment, the 2D convolution layer may be a layer for obtaining an output by performing a convolution operation between an input and a filter of a specified size. In one embodiment, the 2D convolution layer included in the artificial intelligence model 745 assumes that the output image size (number of horizontal pixels × number of vertical pixels) is the same as the input image size (number of horizontal pixels × number of vertical pixels). In one embodiment, the number of output channels can be changed depending on the number of filters (or filter depth) of the 2D convolution layer. For example, if the number of filters is two, the number of output channels can be increased by two times compared to the number of input channels. In one embodiment, the number of output channels can be changed depending on the depth stride of the 2D convolution layer. For example, if the depth stride is two, the number of output channels can be reduced by half compared to the number of input channels.

[0146] In one embodiment, an activation layer may be a layer that applies a specified activation function to an input to obtain an output. For example, the specified activation function may include a step, a sigmoid, a rectifier linear unit (ReLU), an exponential linear unit (ELU), a softmax, or a combination thereof.

[0147] In one embodiment, the max pooling layer may be a layer for selecting the largest value for each pooling region of the input to obtain an output. In one embodiment, the output image size (i.e., the number of horizontal and vertical pixels) can be changed depending on the size of the pooling region of the max pooling layer. For example, if the size of the pooling region is 2x2, the output image size (e.g., 128x128) can be reduced by half compared to the input image size (e.g., 256x256). According to an embodiment, the artificial intelligence model 745 may include a pooling layer of a different type instead of a max pooling layer. For example, the pooling layer of a different type may include an average pooling layer.

[0148] In one embodiment, the upsampling layer may be a layer for increasing the resolution of an image, hi one embodiment, the upsampling layer may be a layer for increasing the output image size compared to the input image size using a specified interpolation algorithm.

[0149] In one embodiment, the concatenation layer may concatenate two or more inputs and output the concatenated inputs. Here, the concatenation may be depthwise concatenation. For example, if a first input is 128x128x128 and a second input is 128x128x128, the output of the concatenation layer may be 128x128x256. If another first input is 256x256x64 and a second input is 128x128x128, the output of the concatenation layer may be 256x256x128.

[0150] 9 shows seven layers 910, 920, 930, 940, 950, 960, and 970, this is for illustrative purposes only and the number of layers is not limited to seven. For example, the number of layers included in the artificial intelligence model 745 may be 32.

[0151] If the artificial intelligence model 745 includes 32 layers, the artificial intelligence model 745 may have a structure in which an input layer, a first batch normalization layer, a first 2D convolutional layer, and a second 2D convolutional layer are sequentially connected. The input of the sequentially connected layer may be the output of the layer immediately preceding it, and the output of the sequentially connected layer may be the input of the layer immediately following it. Here, the input of the input layer may be 256×256×1, and the output may be 256×256×1. The output of the first batch normalization layer may be 256×256×1. The output of the first 2D convolutional layer may be 256×256×64. The output of the second 2D convolutional layer may be 256×256×64.

[0152] Following the second 2D convolutional layer, the artificial intelligence model 745 may have a structure in which a second batch normalization layer, a first activation layer, and a first max pooling layer are sequentially connected. Here, the output of the second batch normalization layer may be 256×256×64, the output of the first activation layer may be 256×256×64, and the output of the first max pooling layer may be 128×128×64.

[0153] Following the first max pooling layer, the artificial intelligence model 745 may have a structure in which a third 2D convolutional layer, a fourth 2D convolutional layer, a third batch normalization layer, a second activation layer, and a second max pooling layer are sequentially connected. Here, the output of the third 2D convolutional layer may be 128×128×128. The output of the fourth 2D convolutional layer may be 128×128×128. The output of the third batch normalization layer may be 128×128×128. The output of the second activation layer may be 128×128×128. The output of the second max pooling layer may be 64×64×128.

[0154] Following the second max pooling layer, the artificial intelligence model 745 may have a structure in which a fifth 2D convolutional layer, a sixth 2D convolutional layer, a fourth batch normalization layer, a third activation layer, and a first upsampling layer are sequentially connected. Here, the output of the fifth 2D convolutional layer may be 64×64×256. The output of the sixth 2D convolutional layer may be 64×64×256. The output of the fourth batch normalization layer may be 64×64×256. The output of the third activation layer may be 64×64×256. The output of the first upsampling layer may be 128×128×256.

[0155] Following the first upsampling layer, the artificial intelligence model 745 may have a structure in which a seventh 2D convolutional layer, a first concatenation layer, an eighth 2D convolutional layer, a ninth 2D convolutional layer, a fifth batch normalization layer, a fourth activation layer, and a second upsampling layer are sequentially connected. Of the two inputs of the first concatenation layer, the first input may be the output of the seventh 2D convolutional layer, and the second input may be the output of the fourth 2D convolutional layer. Here, the output of the seventh 2D convolutional layer may be 128×128×128. The output of the first concatenation layer may be 128×128×256. The output of the eighth 2D convolutional layer may be 128×128×128. The output of the ninth 2D convolutional layer may be 128×128×128. The output of the fifth batch normalization layer may be 128×128×128. The output of the fourth activation layer may be 128x128x128. The output of the second upsampling layer may be 256x256x128.

[0156] Following the second upsampling layer, the artificial intelligence model 745 may have a structure in which a tenth 2D convolutional layer, a second concatenation layer, an eleventh 2D convolutional layer, a twelfth 2D convolutional layer, a sixth batch normalization layer, a fifth activation layer, a thirteenth 2D convolutional layer, and a fourteenth 2D convolutional layer are sequentially concatenated. Of the two inputs of the second concatenation layer, the first input may be the output of the tenth 2D convolutional layer, and the second input may be the output of the second 2D convolutional layer. Here, the output of the tenth 2D convolutional layer may be 256×256×64. The output of the second concatenation layer may be 256×256×128. The output of the eleventh 2D convolutional layer may be 256×256×64. The output of the twelfth 2D convolutional layer may be 256×256×64. The output of the sixth batch normalization layer may be 256x256x64. The output of the fifth activation layer may be 256x256x64. The output of the thirteenth 2D convolutional layer may be 256x256x2. The output of the fourteenth 2D convolutional layer may be 256x256x2.

[0157] AI model learning Referring to FIG. 9, the operation of the AI ​​model learning unit 741 to train the AI ​​model 745 using images 901, 905, 991, and 995 included in the mini-batch will be described.

[0158] In one embodiment, the AI ​​model learning unit 741 can sequentially input the reference activation images 901 and 905 to the AI ​​model 745 . In one embodiment, the artificial intelligence model learning unit 741 can train the artificial intelligence model 745 so that the difference between the output images of the artificial intelligence model 745 acquired sequentially and the reference boundary images 991, 995 is less than or equal to a specified reference value.

[0159] For example, the artificial intelligence model learning unit 741 may calculate the difference between the output image and the reference boundary images 991, 995 based on a loss function (or cost function). Here, the loss function may include a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function. According to an embodiment, the binary cross entropy loss function may be a function in which a weight between 1 and 1000 is applied to one color (e.g., white) of two colors (or classes) (e.g., black and white).

[0160] Thereafter, the artificial intelligence model learning unit 741 may adjust the weights of the artificial intelligence model 745 so that the difference is equal to or less than a specified reference value (or has a minimum value). In one embodiment, the artificial intelligence model learning unit 741 may adjust the weights of the artificial intelligence model 745 using a gradient descent-based algorithm (e.g., Adam, SGD, Momentum) so that the difference is equal to or less than a specified reference value (or has a minimum value). Here, the adjusted weights may be determined according to a learning rate. The learning rate may be determined between 0.00000001 and 0.1. For example, the learning rate may be 0.0001.

[0161] Thereafter, the AI ​​model learning unit 741 can train the AI ​​model 745 using the images included in the next mini-batch. After that, when the AI ​​model 745 has been trained using the images included in all mini-batches, the AI ​​model learning unit 741 can train the AI ​​model 745 again using the images included in the mini-batch according to the specified number of epochs.

[0162] Image processing using artificial intelligence models FIG. 10a shows images 1010, 1020, 1030, 1040 resulting from image processing by the image processing device 701 according to another embodiment of the present disclosure.

[0163] In one embodiment, the image acquisition unit 750 can acquire the active material image 1010 for the active materials 1011, 1013, and 1015. In one embodiment, the image acquisition unit 750 can acquire the active material image 1010 via the image acquisition device 703.

[0164] In one embodiment, the image generation unit 760 can generate the boundary image 1020. In one embodiment, the image generation unit 760 can generate the boundary image 1020 by inputting the active material image 1010 into the artificial intelligence model 745.

[0165] In one embodiment, the boundary removal unit 765 may remove the boundary from the active material image 1010 based on the boundary image 1020 to generate the boundary-removed image 1030. In one embodiment, the boundary removal unit 765 may remove the boundary image 1020 from the active material image 1010 to generate the boundary-removed image 1030. In one embodiment, the boundary removal unit 765 may remove the boundary image 1020 from the active material image 1010 and then binarize the image to generate the boundary-removed image 1030. According to an embodiment, the boundary-removed image may also be referred to as a binary image.

[0166] In one embodiment, the object identification unit 770 may identify a plurality of objects 1031, 1033, and 1035 included in the boundary-removed image 1030. Here, the plurality of objects 1031, 1033, and 1035 included in the boundary-removed image 1030 may correspond to the active materials 1011, 1013, and 1015 in the active material image 1010. In one embodiment, the plurality of objects 1031, 1033, and 1035 may be configured as regions having a specified value (e.g., a value indicating white). In one embodiment, the plurality of objects 1031, 1033, and 1035 may be separated by regions having another specified value (e.g., a value indicating black). Although only three objects are labeled in FIG. 10a, it can be seen from FIG. 10a that other objects are present in the boundary-removed image 1030.

[0167] In one embodiment, the image segmentation unit 780 may segment the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 based on the plurality of objects 1031, 1033, and 1035 to obtain the segmentation image 1040. In one embodiment, the image segmentation unit 780 may segment the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 based on a watershed algorithm. As a result, the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 may correspond to the plurality of active material objects 1041, 1043, and 1045 included in the segmentation image 1040, respectively.

[0168] In one embodiment, the information extraction unit 790 can extract information about the active materials 1011, 1013, and 1015 based on the active material objects 1041, 1043, and 1045 included in the segmentation image 1040. In one embodiment, the information about the active materials 1011, 1013, and 1015 can include information about the average particle size, average perimeter, average sphericity, average aspect ratio, average convexity, average solidity, or distribution values ​​thereof. The information about the active materials 1011, 1013, and 1015 can also include information about the individual particle size, perimeter, sphericity, aspect ratio, convexity, and / or solidity of each active material. The information about the active materials 1011, 1013, and 1015 can also include information about the quantile values ​​(e.g., 1% to 99%, or D5, D50, D95, etc.) of each active material.

[0169] FIG. 10b shows the image difference due to image processing by the image processing device 701 according to another embodiment of the present disclosure. 10b, a user can visually recognize the difference between the two active material images 1051 and 1053, but it is difficult to quantify. However, by using the image processing device 701 to segment the active material into two segmentation images 1061 and 1063, information about the active material can be obtained, allowing the active material to be quantitatively measured.

[0170] FIG. 11 is a flowchart illustrating an active material image processing method of the image processing device 101 according to another embodiment of the present disclosure. 11 , in operation 1110, the image processing device 701 may acquire an active material image 1010 for the active materials 1011, 1013, and 1015. In one embodiment, the image processing device 701 may acquire the active material image 1010 via the image acquisition device 703.

[0171] In operation 1120, the image processing device 701 can generate the boundary image 1020 by inputting the active material image 1010 into the artificial intelligence model 745. Here, the artificial intelligence model 745 may be a model trained by the artificial intelligence model training method shown in FIG.

[0172] In operation 1130, the image processing device 701 can remove the boundary from the active material image 1010 based on the boundary image 1020 to generate the boundary-removed image 1030. In one embodiment, the image processing device 701 can generate the boundary-removed image 1030 by removing the boundary image 1020 from the active material image 1010. In one embodiment, the image processing device 701 can generate the boundary-removed image 1030 by removing the boundary image 1020 from the active material image 1010 and then binarizing it.

[0173] In operation 1140 , the image processor 701 may identify a number of objects 1031 , 1033 , 1035 contained in the boundary-removed image 1030 .

[0174] In operation 1150, the image processing device 701 may segment the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 based on the plurality of objects 1031, 1033, and 1035 to obtain a segmentation image 1040. In one embodiment, the image processing device 701 may segment the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 based on a watershed algorithm. Thus, the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 may correspond to the plurality of active material objects 1041, 1043, and 1045 included in the segmentation image 1040, respectively.

[0175] 12 is a flowchart showing an artificial intelligence model training method for the image processing device 701 according to another embodiment of the present disclosure. The operation of FIG. 12 can be applied to the first epoch of the training data.

[0176] 12, in operation 1210, the image processing device 701 may set parameters of the artificial intelligence model 745. In one embodiment, the image processing device 701 may set parameters of the artificial intelligence model 745 based on initial parameter values.

[0177] In operation 1215, the image processing device 701 may set i to 1. i may indicate an index of an image set included in the training data, where the image set may include a reference active material image and its corresponding reference boundary image.

[0178] In operation 1220, the image processing device 701 can obtain an output image based on the i-th reference active material image. In one embodiment, the image processing device 701 can generate the output image by inputting the i-th reference active material image into the artificial intelligence model 745.

[0179] In operation 1230, the image processing device 701 may determine whether the difference between the output image and the i-th reference boundary image is equal to or smaller than a reference value. In one embodiment, the image processing device 701 may calculate the difference between the output image and the i-th reference boundary image based on a loss function (or cost function). Here, the loss function may include a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function. According to an embodiment, the binary cross entropy loss function may be a function in which a weight between 1 and 1000 is applied to one color (e.g., white) of two colors (or classes) (e.g., black and white).

[0180] If the result of the determination in operation 1230 is that the difference is equal to or less than the standard (YES determination), the image processing device 701 can proceed to operation 1250. If the result of the determination in operation 1230 is that the difference exceeds the standard (NO determination), the image processing device 701 can proceed to operation 1240.

[0181] In operation 1240, the image processing device 701 may adjust parameters of the artificial intelligence model 745. In one embodiment, the image processing device 701 may adjust parameters of the artificial intelligence model 745 so that the difference is equal to or less than a specified reference value (or has a minimum value). In one embodiment, the image processing device 701 may adjust parameters of the artificial intelligence model 745 using a gradient descent-based algorithm (e.g., Adam, SGD, Momentum) so that the difference is equal to or less than a specified reference value (or has a minimum value). Here, the adjusted weights may be determined according to a learning rate. The learning rate may be determined between 0.00000001 and 0.1. For example, the learning rate may be 0.0001.

[0182] In operation 1250, the image processing device 701 may determine whether training is complete. In one embodiment, training may be determined to be complete when i indicates the end of the image set.

[0183] If the result of the determination in operation 1250 is that learning is complete (YES determination), the image processing device 701 can end the operation according to Fig. 12. If the result of the determination in operation 1250 is that learning is not complete (NO determination), the image processing device 701 can proceed to operation 655. In operation 1255, the image processor 701 may increment the value of i by 1. Operation 1220 may then be performed again.

[0184] Although the present disclosure illustrates an example in which the image processing device 701 performs an image processing method based on an active material image, this is merely an example. According to an embodiment, the image processing device 701 can separate and / or analyze particles by performing similar image processing on not only the active material image but also the precursor image. The image acquisition device 703 can acquire images of the active material and / or the precursor.

[0185] Furthermore, although the active material image is illustrated as an SEM image in this disclosure, this is merely an example. According to embodiments, the active material image may be replaced with an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0186] The image processing device 701 and its operating method as described above acquire an active material image 1010 for a plurality of active materials 1011, 1013, and 1015, input the active material image 1010 into an artificial intelligence model 745 to generate a boundary image 1020, remove the boundary from the active material image 1010 based on the boundary image 1020 to generate a boundary-removed image 1030, identify a plurality of objects 1031, 1033, and 1035 included in the boundary-removed image 1030, and then segment the plurality of active materials 1011, 1013, and 1015 included in the active material image 1010 based on the plurality of objects 1031, 1033, and 1035, thereby enabling automatic quantitative analysis of the shape characteristics of a large number of positive electrode active material particles.

[0187] Furthermore, the image processing device 701 and its operating method as described above can minimize the problem of different measurements depending on the user by utilizing an artificial intelligence model 745 that minimizes user input.

[0188] Meanwhile, although the above description has been given as an example of a case where the active material image is a positive electrode active material image, the technical concepts disclosed in this document can be substantially similarly applied to a case where the active material image is a negative electrode active material image.

[0189] FIG. 13 is a block diagram of an image processing device 1301 according to yet another embodiment of the present disclosure. Referring to FIG. 13, an image processing device 1301 can be coupled to an image acquisition device 1303 via wires and / or wirelessly.

[0190] In one embodiment, the connection 1305 between the image processing device 1301 and the image acquisition device 1303 may be a communication connection via a wired and / or wireless network. In one embodiment, the wired network may be based on a local area network (LAN) communication or a power line communication. In one embodiment, the wireless network may be based on a short-range communication network (e.g., Bluetooth, Wireless Fidelity (WiFi), or Infrared Data Association (IrDA)) or a long-range communication network (e.g., a cellular network, a 4G network, a 5G network).

[0191] In other embodiments, the connection 1305 between the image processing device 1301 and the image acquisition device 1303 may be a connection via an inter-device communication method (e.g., a bus, a General Purpose Input and Output (GPIO), a Serial Peripheral Interface (SPI), or a Mobile Industry Processor Interface (MIPI)).

[0192] In one embodiment, the image acquisition device 1303 can acquire an image of the active material and / or precursor. Hereinafter, an image of the active material and / or precursor may be referred to as an active material image. However, even if an image is referred to as an active material image, the present disclosure does not exclude an image of the precursor.

[0193] In one embodiment, the image acquisition device 1303 may be a microscope (e.g., a scanning electron microscope) that acquires an image of the sample surface by scanning a focused electron beam over the sample surface and converting secondary electrons generated by the interaction between the electron beam and the sample into an image signal.

[0194] In one embodiment, the image acquisition device 1303 may acquire an active material image of the active material. For example, the image acquisition device 1303 may acquire a scanning electron microscope (SEM) image by scanning an electron beam on the positive or negative electrode active material powder. That is, the SEM image may include a positive electrode active material image and a negative electrode active material image. According to an embodiment, the SEM image may be replaced by an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0195] In one embodiment, the image capture device 1303 can transmit an active material image of the active material to the image processing device 1301. For example, the image capture device 1303 can transmit the active material image of the active material to the image processing device 1301 via the connection 1305.

[0196] In one embodiment, the image processing device 1301 may be a mobile device (eg, a mobile phone, a laptop computer, a smart phone, a smart pad), a computer (eg, a general-purpose computer, a special-purpose computer).

[0197] In one embodiment, the image processing device 1301 may include a communication circuit 1310, a memory 1320, and a processor 1330. According to an embodiment, the image processing device 1301 shown in Figure 13 may further include at least one component (e.g., a display, an input device, or an output device) other than the components shown in Figure 13.

[0198] In one embodiment, the communication circuitry 1310 can establish a wired and / or wireless communication channel between the image processing device 1301 and / or the image acquisition device 1303 and transmit and receive data to and from the image acquisition device 1303 via the established communication channel.

[0199] In one embodiment, the memory 1320 may include volatile and / or non-volatile memory. In one embodiment, memory 1320 may store data used by at least one component (e.g., processor 1330) of image processing device 1301. For example, the data may include program 1325 (or instructions associated therewith), input data, or output data. In one embodiment, the instructions, when executed by processor 1330, may cause image processing device 1301 to perform the operations defined by the instructions.

[0200] In one embodiment, the memory 1320 may include programs 1325 (e.g., an artificial intelligence model training unit 1341, an artificial intelligence model 1345, an image acquisition unit 1350, an image generation unit 1360, a distance transformation unit 1361, a binary image filtering unit 1365, an object identification unit 1370, an image segmentation unit 1380, and / or an information extraction unit 1390).

[0201] In one embodiment, the processor 1330 may include a central processing unit, an application processor, a graphics processing unit, a neural processing unit (NPU), an image signal processor, a sensor hub processor, or a communication processor.

[0202] In one embodiment, the processor 1330 can execute the program 1325 (e.g., the artificial intelligence model training unit 1341, the artificial intelligence model 1345, the image acquisition unit 1350, the image generation unit 1360, the distance transformation unit 1361, the binary image filtering unit 1365, the object identification unit 1370, the image segmentation unit 1380, and / or the information extraction unit 1390), control at least one other component (e.g., hardware or software component) of the image processing device 1301 coupled to the processor 1330, and perform various data processing or calculations.

[0203] In one embodiment, the AI ​​model training unit 1341 may train the AI ​​model 1345 based on the training data. In one embodiment, the AI ​​model 1345 may be a model trained to convert an active material image into a binary image. In one embodiment, the image acquisition unit 1350 may acquire an active material image of an active material from the image acquisition device 1303. In one embodiment, the image generation unit 1360 may generate a binary image by inputting the active material image into the AI ​​model 1345. In one embodiment, the distance transformation unit 1361 may convert the binary image into a distance transformed image based on a distance transformation algorithm. In one embodiment, the binary image filtering unit 1365 may filter the binary image using a threshold set based on the distance transformed image. In one embodiment, the object identification unit 1370 may identify multiple objects included in the filtered binary image. In one embodiment, the image segmentation unit 1380 may segment multiple active materials included in the active material image based on the multiple objects and acquire a segmentation image. In one embodiment, the information extractor 1390 can extract information about the active material based on the active material object included in the segmentation image.

[0204] Below, with reference to Figures 14, 15, 16a, and 16b, we will explain in detail how the image processing device 1301 processes images acquired from the image acquisition device 1303 via the artificial intelligence model learning unit 1341, the artificial intelligence model 1345, the image acquisition unit 1350, the image generation unit 1360, the distance transformation unit 1361, the binary image filtering unit 1365, the object identification unit 1370, the image segmentation unit 1380, and / or the information extraction unit 1390.

[0205] Training data FIG. 14 shows training data according to yet another embodiment of the present disclosure. Referring to FIG. 14, the training data may include reference active material images 1411, 1421, 1431 and reference binary images 1451, 1461, 1471.

[0206] The reference active material images 1411, 1421, and 1431 and the reference binary images 1451, 1461, and 1471 included in the training data may have the same size. In one embodiment, the image size may be defined as the number of horizontal pixels x the number of vertical pixels. The number of horizontal pixels of the images included in the training data may be an integer between 32 and 4096, and the number of vertical pixels may be an integer between 32 and 4096. For example, the image size may be 256 x 256.

[0207] In one embodiment, the plurality of reference active material images 1411, 1421, 1431 can include one or more first reference active material images and one or more second reference active material images.

[0208] In one embodiment, the first reference active material image can be acquired by photographing the active material powder. In one embodiment, the first reference active material image can be an SEM image acquired directly via the image acquisition device 1303.

[0209] In one embodiment, the second reference active material image may be an image transformed from the first reference active material image. In one embodiment, the second reference active material image may be generated by applying a specified first image processing algorithm to the first reference active material image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0210] In one embodiment, the plurality of reference binary images 1451, 1461, 1471 may include one or more first reference binary images and one or more second reference binary images.

[0211] In one embodiment, the first reference binary image may be generated by applying a second image processing algorithm to the first reference active material image, where the second image processing algorithm may include a mean shift filter, a boundary extraction algorithm, a boundary removal algorithm, a binarization algorithm, or a combination thereof.

[0212] In one embodiment, the second reference binary image may be an image transformed from the first reference binary image. In one embodiment, the second reference binary image may be generated by applying a specified first image processing algorithm to the first reference binary image. Here, the first image processing algorithm may include rotation, tilt, shear, brightness adjustment, contrast adjustment, zoom in, zoom out, or a combination thereof.

[0213] In one embodiment, the first reference binary image may be a binary image of the first reference active material image, and the second reference binary image may be a binary image of the second reference active material image. For example, reference binary image 1451 may be a binary image for reference active material image 1411, reference binary image 1461 may be a binary image for reference active material image 1421, and reference binary image 1471 may be a binary image for reference active material image 1431. Therefore, the multiple reference active material images 1411, 1421, and 1431 and the multiple reference binary images 1451, 1461, and 1471 can be divided into image sets, with corresponding images. For example, reference active material image 1411 and reference binary image 1451 can be divided into a single image set.

[0214] In one embodiment, the training data can be used to train the artificial intelligence model 1345. In one embodiment, the training data can be used to train the artificial intelligence model 1345 for a specified number of epochs. Here, the specified number can be set between 100 and 10,000. For example, the specified number can be 1,500.

[0215] In one embodiment, the training data may be divided into mini-batches, each divided by a batch size. Here, the batch size may be determined between 1 and 512. For example, the batch size may be 16. When the batch size is 4, the training data may be divided into mini-batches of 16 image sets (i.e., 16 reference active material images and 16 reference binary images).

[0216] Artificial Intelligence Model 15 illustrates an artificial intelligence model 1345 according to another embodiment of the present disclosure. Specifically, FIG. 15 illustrates an example in which the artificial intelligence model 1345 is trained using a mini-batch consisting of reference active material images 1501, 1503, and 1505 and reference binary images 1591, 1593, and 1595.

[0217] In one embodiment, the artificial intelligence model 1345 may be a model based on a convolutional neural network (CNN) or a U-NET. In one embodiment, the artificial intelligence model 1345 may be a model trained to convert an active material image into a binary image.

[0218] 15, the artificial intelligence model 1345 may include multiple layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570. The multiple layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570 may be sequentially connected. The input of a sequentially connected layer may be the output of the immediately preceding layer, and the output of a sequentially connected layer may be the input of the immediately following layer. For example, the output of layer 1510 may be the input of layer 1520.

[0219] At least two layers (e.g., layers 1510, 1570, layers 1520, 1560, or layers 1530, 1550) of the plurality of layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570 may be connected by skip connections 1515, 1525, and 1535. In one embodiment, the skip connections 1515, 1525, and 1535 may connect a layer included in the encoding region (e.g., layers 1510, 1520, and 1530) of the plurality of layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570 to a layer included in the decoding region (e.g., layers 1550, 1560, and 1570). Here, skip connections 1515, 1525, and 1535 may refer to connections between layers for inputting the outputs of layers 1510, 1520, and 1530 to layers 1550, 1560, and 1570. For example, skip connection 1535 allows the output of layer 1540 and the output of layer 1530 to be input to layer 1550.

[0220] In one embodiment, the layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570 may include an input layer, a batch normalization layer, a 2D convolution layer, an activation layer, a max pooling layer, an upsampling layer, a concatenation layer, or a combination thereof. For example, the layers included in the encoding domain (e.g., layers 1510, 1520, and 1530) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and a max pooling layer are sequentially connected. As another example, the layers included in the decoding domain (e.g., layers 1550, 1560, and 1570) may have a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and an upsampling layer are sequentially connected.

[0221] In one embodiment, the input layer may be a layer that receives input of bit values ​​included in the reference active material images 1501, 1503, and 1505. In one embodiment, the input layer may acquire input consisting of bit values ​​equal to the number of horizontal pixels x the number of vertical pixels x the number of channels (or depth). For example, if the reference active material image has an N x N size, the input layer may receive input of N x N x 1 bit values, where N may be an integer between 32 and 4096 (e.g., 256). According to an embodiment, the number of channels may also be referred to as depth.

[0222] In one embodiment, the batch normalization layer may be a layer for normalizing the output value of a previous layer in batch units. According to an embodiment, the batch normalization layer may be disposed immediately before the activation layer. Here, a batch may include reference active material images input during one iteration to train the artificial intelligence model 1345. For example, if the batch size is 16, the number of reference active material images input to the artificial intelligence model 1345 during one iteration may be 16. Here, the batch size may be an integer between 1 and 512 (e.g., 16).

[0223] In one embodiment, the 2D convolution layer may be a layer for obtaining an output by performing a convolution operation between an input and a filter of a specified size. In one embodiment, the 2D convolution layer included in the artificial intelligence model 1345 assumes that the output image size (number of horizontal pixels × number of vertical pixels) is the same as the input image size (number of horizontal pixels × number of vertical pixels). In one embodiment, the number of output channels can be changed depending on the number of filters (or filter depth) of the 2D convolution layer. For example, if the number of filters is two, the number of output channels can be increased by two times compared to the number of input channels. In one embodiment, the number of output channels can be changed depending on the depth stride of the 2D convolution layer. For example, if the depth stride is two, the number of output channels can be reduced by half compared to the number of input channels.

[0224] In one embodiment, an activation layer may be a layer that applies a specified activation function to an input to obtain an output. For example, the specified activation function may include a step, a sigmoid, a rectifier linear unit (ReLU), an exponential linear unit (ELU), a softmax, or a combination thereof.

[0225] In one embodiment, the max pooling layer may be a layer for selecting the largest value from each pooling region of the input to obtain an output. In one embodiment, the output image size (i.e., the number of horizontal and vertical pixels) can be changed depending on the size of the pooling region of the max pooling layer. For example, if the size of the pooling region is 2x2, the output image size (e.g., 128x128) can be reduced by half compared to the input image size (e.g., 256x256). According to an embodiment, the artificial intelligence model 1345 may include a pooling layer of a different type instead of a max pooling layer. For example, the pooling layer of a different type may include an average pooling layer.

[0226] In one embodiment, the upsampling layer may be a layer for increasing the resolution of an image, hi one embodiment, the upsampling layer may be a layer for increasing the output image size compared to the input image size using a specified interpolation algorithm.

[0227] In one embodiment, the concatenation layer may concatenate two or more inputs and output the concatenated inputs. Here, the concatenation may be depthwise concatenation. For example, if a first input is 128x128x128 and a second input is 128x128x128, the output of the concatenation layer may be 128x128x256. If another first input is 256x256x64 and a second input is 128x128x128, the output of the concatenation layer may be 256x256x128.

[0228] 15 shows seven layers 1510, 1520, 1530, 1540, 1550, 1560, and 1570, but this is for illustrative purposes only and the number of layers is not limited to seven. For example, the number of layers included in the artificial intelligence model 1345 may be 32.

[0229] If the artificial intelligence model 1345 includes 32 layers, the artificial intelligence model 1345 may have a structure in which an input layer, a first batch normalization layer, a first 2D convolution layer, and a second 2D convolution layer are sequentially connected. The input of the sequentially connected layer may be the output of the layer immediately preceding it, and the output of the sequentially connected layer may be the input of the layer immediately following it. Here, the input of the input layer may be 256×256×1, and the output may be 256×256×1. The output of the first batch normalization layer may be 256×256×1. The output of the first 2D convolution layer may be 256×256×64. The output of the second 2D convolution layer may be 256×256×64.

[0230] Following the second 2D convolutional layer, the artificial intelligence model 1345 may have a structure in which a second batch normalization layer, a first activation layer, and a first max pooling layer are sequentially connected. Here, the output of the second batch normalization layer may be 256×256×64, the output of the first activation layer may be 256×256×64, and the output of the first max pooling layer may be 128×128×64.

[0231] Following the first max pooling layer, the artificial intelligence model 1345 may have a structure in which a third 2D convolutional layer, a fourth 2D convolutional layer, a third batch normalization layer, a second activation layer, and a second max pooling layer are sequentially connected. Here, the output of the third 2D convolutional layer may be 128×128×128. The output of the fourth 2D convolutional layer may be 128×128×128. The output of the third batch normalization layer may be 128×128×128. The output of the second activation layer may be 128×128×128. The output of the second max pooling layer may be 64×64×128.

[0232] Following the second max pooling layer, the artificial intelligence model 1345 may have a structure in which a fifth 2D convolutional layer, a sixth 2D convolutional layer, a fourth batch normalization layer, a third activation layer, and a first upsampling layer are sequentially connected. Here, the output of the fifth 2D convolutional layer may be 64×64×256. The output of the sixth 2D convolutional layer may be 64×64×256. The output of the fourth batch normalization layer may be 64×64×256. The output of the third activation layer may be 64×64×256. The output of the first upsampling layer may be 128×128×256.

[0233] Following the first upsampling layer, the artificial intelligence model 1345 may have a structure in which a seventh 2D convolutional layer, a first concatenation layer, an eighth 2D convolutional layer, a ninth 2D convolutional layer, a fifth batch normalization layer, a fourth activation layer, and a second upsampling layer are sequentially connected. Of the two inputs of the first concatenation layer, the first input may be the output of the seventh 2D convolutional layer, and the second input may be the output of the fourth 2D convolutional layer. Here, the output of the seventh 2D convolutional layer may be 128×128×128. The output of the first concatenation layer may be 128×128×256. The output of the eighth 2D convolutional layer may be 128×128×128. The output of the ninth 2D convolutional layer may be 128×128×128. The output of the fifth batch normalization layer may be 128×128×128. The output of the fourth activation layer may be 128x128x128. The output of the second upsampling layer may be 256x256x128.

[0234] Following the second upsampling layer, the artificial intelligence model 1345 may have a structure in which a tenth 2D convolutional layer, a second concatenation layer, an eleventh 2D convolutional layer, a twelfth 2D convolutional layer, a sixth batch normalization layer, a fifth activation layer, a thirteenth 2D convolutional layer, and a fourteenth 2D convolutional layer are sequentially concatenated. Of the two inputs of the second concatenation layer, the first input may be the output of the tenth 2D convolutional layer, and the second input may be the output of the second 2D convolutional layer. Here, the output of the tenth 2D convolutional layer may be 256×256×64. The output of the second concatenation layer may be 256×256×128. The output of the eleventh 2D convolutional layer may be 256×256×64. The output of the twelfth 2D convolutional layer may be 256×256×64. The output of the sixth batch normalization layer may be 256x256x64. The output of the fifth activation layer may be 256x256x64. The output of the thirteenth 2D convolutional layer may be 256x256x2. The output of the fourteenth 2D convolutional layer may be 256x256x2.

[0235] AI model learning Referring to FIG. 15, the operation of the artificial intelligence model learning unit 1341 to learn the artificial intelligence model 1345 using images 1501, 1503, 1505, 1591, 1593, and 1595 included in the mini-batch will be described.

[0236] In one embodiment, the AI ​​model training unit 1341 can sequentially input the reference activation images 1501 , 1503 , and 1505 to the AI ​​model 1345 .

[0237] In one embodiment, the artificial intelligence model learning unit 1341 can train the artificial intelligence model 1345 so that the difference between the output images of the artificial intelligence model 1345 acquired sequentially and the reference binary images 1591, 1593, 1595 is less than or equal to a specified reference value.

[0238] For example, the artificial intelligence model learning unit 1341 may calculate the difference between the output image and the reference binary images 1591, 1593, and 1595 based on a loss function (or cost function). Here, the loss function may include a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function. According to an embodiment, the binary cross entropy loss function may be a function in which a weight between 1 and 1000 is applied to one color (e.g., white) of two colors (or classes) (e.g., black and white).

[0239] Thereafter, the artificial intelligence model learning unit 1341 may adjust the weights of the artificial intelligence model 1345 so that the difference is equal to or less than a specified reference value (or has a minimum value). In one embodiment, the artificial intelligence model learning unit 1341 may adjust the weights of the artificial intelligence model 1345 using a gradient descent-based algorithm (e.g., Adam, SGD, or Momentum) so that the difference is equal to or less than a specified reference value (or has a minimum value). Here, the adjusted weights may be determined according to a learning rate. The learning rate may be determined between 0.00000001 and 0.1. For example, the learning rate may be 0.00001.

[0240] Thereafter, the AI ​​model learning unit 1341 can train the AI ​​model 1345 using the images included in the next mini-batch. After that, when the AI ​​model 1345 has been trained using the images included in all mini-batches, the AI ​​model learning unit 1341 can train the AI ​​model 1345 again using the images included in the mini-batch according to the specified number of epochs.

[0241] Image processing using artificial intelligence models FIG. 16a shows images 1610, 1620, 1630, 1640, 1650, and 1660 resulting from image processing by the image processing device 1301 according to yet another embodiment of the present disclosure.

[0242] In one embodiment, the image acquisition unit 1350 can acquire the active material image 1610 for the active material. In one embodiment, the image acquisition unit 1350 can acquire the active material image 1610 via the image acquisition device 1303.

[0243] In one embodiment, the image generation unit 1360 can generate the binary image 1620. In one embodiment, the image generation unit 1360 can generate the binary image 1620 by inputting the active material image 1610 into the artificial intelligence model 1345.

[0244] In one embodiment, the distance transform unit 1361 can transform the binary image 1620 into a distance transformed image 1640 based on a distance transform algorithm.

[0245] In one embodiment, the distance transform unit 1361 may calculate the distance from each pixel included in the binary image 1620 to the nearest black pixel (or pixel indicating a boundary) and normalize the calculated distance for each pixel to generate the distance transformed image 1640. Here, the image 1630 may indicate the distance from each pixel included in the binary image 1620 to the nearest black pixel (or pixel indicating a boundary). Therefore, the values ​​of pixels included in the image 1630 may be distance values ​​before normalization, and the values ​​of pixels included in the distance transformed image 1640 may be distance values ​​after normalization. According to an embodiment, the distance between pixels may be calculated based on the L1 distance or the L2 distance.

[0246] In one embodiment, the distance transform unit 1361 may identify the maximum distance among the distances calculated for each pixel included in the binary image 1620, and perform max-min normalization on the distances calculated for each pixel based on the maximum distance, thereby generating the distance transformed image 1640. According to an embodiment, the distance transform unit 1361 may generate the distance transformed image 1640 based on Z-score normalization, L1 normalization, or L2 normalization instead of max-min normalization.

[0247] In one embodiment, the binary image filtering unit 1365 may filter the binary image 1620 using a threshold value that is set based on the distance transformed image 1640. Here, the threshold value may be a value obtained by multiplying the normalized maximum distance of the distance transformed image 1640 by a specified ratio. For example, the specified ratio may be a value between 0 and 0.1.

[0248] In one embodiment, the binary image filter 1365 may generate a filtered binary image 1650 by filtering the binary image 1620 based on the distance transform image 1640 .

[0249] In one embodiment, the binary image filtering unit 1365 can filter the binary image 1620 by selecting pixels from the distance transformed image 1640 that have distance values ​​below a specified threshold, and setting the color values ​​of pixels from the binary image 1620 that correspond to the pixels selected from the distance transformed image 1640 to a specified color value (i.e., a color value corresponding to black).

[0250] In one embodiment, the object identification unit 1370 may identify a plurality of objects included in the filtered binary image 1650. Here, the plurality of objects included in the filtered binary image 1650 may correspond to active materials in the active material image 1610. In one embodiment, the plurality of objects may be composed of regions having a specified value (e.g., a value indicating white). In another embodiment, the plurality of objects may be separated by regions having another specified value (e.g., a value indicating black).

[0251] In one embodiment, the image segmentation unit 1380 may segment the plurality of active materials included in the active material image 1610 based on the plurality of objects to obtain the segmentation image 1660. In one embodiment, the image segmentation unit 1380 may segment the plurality of active materials included in the active material image 1610 based on a watershed algorithm. As a result, the plurality of active materials included in the active material image 1610 may correspond to the plurality of active material objects included in the segmentation image 1660, respectively.

[0252] In one embodiment, the information extractor 1390 can extract information about the active materials based on the active material objects included in the segmentation image 1660. In one embodiment, the information about the active materials can include information about the average particle size, average perimeter, average sphericity, average aspect ratio, average convexity, average solidity, or distribution values ​​thereof. The information about the active materials can also include information about the individual particle size, perimeter, sphericity, aspect ratio, convexity, and / or solidity of each active material. The information about the active materials 1611, 1613, and 1615 can also include quantile information about the quantile values ​​(e.g., 1 to 99%, or D5, D50, D95, etc.) of each active material information.

[0253] FIG. 16b shows an image difference due to image processing by the image processing device 1301 according to yet another embodiment of the present disclosure. 16b, three filtered binary images 1651, 1653, and 1655 are generated by applying different thresholds to the binary images. For example, filtered binary image 1651 may be filtered using a threshold value of 0 multiplied by the normalized maximum distance, filtered binary image 1653 may be filtered using a threshold value of 0.05 multiplied by the normalized maximum distance, and filtered binary image 1655 may be filtered using a threshold value of 0.1 multiplied by the normalized maximum distance.

[0254] Looking at regions 1652, 1654, 1656 of the filtered binary images 1651, 1653, 1655 and regions 1662, 1664, 1666 of the segmentation images 1661, 1663, 1665, it can be seen that even with a higher threshold, small particles are prevented from disappearing and the separation between adjacent particles becomes clearer.

[0255] FIG. 17 is a flowchart illustrating an active material image processing method of the image processing device 1301 according to still another embodiment of the present disclosure. 17, in operation 1710, the image processing device 1301 may acquire an active material image 1610 for the active material. In one embodiment, the image processing device 1301 may acquire the active material image 1610 via the image acquisition device 1303.

[0256] In operation 1720, the image processing device 1301 can generate a binary image 1620 by inputting the active material image 1610 into the artificial intelligence model 1345. Here, the artificial intelligence model 1345 may be a model trained by the artificial intelligence model training method shown in FIG.

[0257] In operation 1730, the image processor 1301 may transform the binary image 1620 into a distance transformed image 1640 based on a distance transform algorithm.

[0258] In one embodiment, the image processing device 1301 may calculate the distance from each pixel included in the binary image 1620 to the nearest black pixel (or pixel indicating a boundary) and normalize the calculated distance for each pixel to generate the distance transformed image 1640. Here, the image 1630 may indicate the distance from each pixel included in the binary image 1620 to the nearest black pixel (or pixel indicating a boundary). Therefore, the values ​​of the pixels included in the image 1630 may be distance values ​​before normalization, and the values ​​of the pixels included in the distance transformed image 1640 may be distance values ​​after normalization. According to an embodiment, the distance between pixels may be calculated based on the L1 distance or the L2 distance.

[0259] In one embodiment, the image processing device 1301 may identify the maximum distance among the distances calculated for each pixel included in the binary image 1620, and perform max-min normalization on the distances calculated for each pixel based on the maximum distance, thereby generating the distance transformed image 1640. According to an embodiment, the distance transform unit 1361 may generate the distance transformed image 1640 based on Z-score normalization, L1 normalization, or L2 normalization instead of max-min normalization.

[0260] In operation 1740, the image processing device 1301 may filter the binary image 1620 using a threshold value that is set based on the distance transformed image 1640. Here, the threshold value may be a value obtained by multiplying the normalized maximum distance of the distance transformed image 1640 by a specified ratio. For example, the specified ratio may be a value between 0 and 0.1.

[0261] In one embodiment, the image processor 1301 may generate a filtered binary image 1650 by filtering the binary image 1620 based on the distance transform image 1640 .

[0262] In one embodiment, the image processing device 1301 can filter the binary image 1620 by selecting pixels from the distance transformed image 1640 that have distance values ​​less than a specified threshold, and setting the color values ​​of pixels from the binary image 1620 that correspond to the pixels selected from the distance transformed image 1640 to a specified color value (i.e., a color value corresponding to black).

[0263] In operation 1750 , the image processor 1301 may identify multiple objects contained in the filtered binary image 1650 .

[0264] In operation 1760, the image processing device 1301 may segment the plurality of active materials included in the active material image 1610 based on the plurality of objects to obtain the segmentation image 1660. In one embodiment, the image processing device 1301 may segment the plurality of active materials included in the active material image 1610 based on a watershed algorithm. Thus, the plurality of active materials included in the active material image 1610 may respectively correspond to the plurality of active material objects included in the segmentation image 1660.

[0265] Although the present disclosure illustrates an example in which the image processing device 1301 performs an image processing method based on an active material image, this is merely an example. According to an embodiment, the image processing device 1301 can separate and / or analyze particles by performing similar image processing on not only the active material image but also the precursor image. The image acquisition device 1303 can acquire images of the active material and / or the precursor.

[0266] Furthermore, although the active material image is illustrated as an SEM image in this disclosure, this is merely an example. According to embodiments, the active material image may be replaced with an image based on a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

[0267] The image processing device 1301 and its operating method as described above acquires active material images 1610 for a plurality of active materials, inputs the active material images 1610 into an artificial intelligence model 1345 to generate a binary image 1620, converts the binary image 1620 into a distance-transformed image 1640 based on a distance transformation algorithm, filters the binary image 1620 using a threshold set based on the distance-transformed image 1640, identifies a plurality of objects included in the filtered binary image 1650, and then segments a plurality of active material objects included in the active material image based on the plurality of objects, thereby enabling automatic quantitative analysis of the shape characteristics of a large number of positive electrode active material particles.

[0268] Furthermore, the image processing device 1301 and its operating method as described above can minimize the problem of different measurements depending on the user by utilizing an artificial intelligence model 1345 that minimizes user input.

[0269] Meanwhile, although the above description has been given as an example of a case where the active material image is a positive electrode active material image, the technical concepts disclosed in this document can be substantially similarly applied to a case where the active material image is a negative electrode active material image.

Claims

1. an image acquisition unit that acquires active material images for a plurality of active materials; an artificial intelligence model learning unit that learns an artificial intelligence model using learning data including a plurality of reference active material images and a plurality of reference binary images or a plurality of reference boundary images corresponding to the plurality of reference active material images; an image generation unit that generates a binary image or a boundary image by inputting the active material image into the artificial intelligence model; an object identification unit that identifies a plurality of objects based on the binary image or the boundary image; an image segmentation unit that segments the active materials included in the active material image based on the objects and acquires a segmentation image; 12. An image processing device comprising:

2. The artificial intelligence model learning unit The image processing device according to claim 1, wherein the artificial intelligence model is trained so that the differences between the plurality of output images obtained by inputting the plurality of reference active material images into the artificial intelligence model and the plurality of reference binary images are equal to or less than a specified reference value.

3. the plurality of reference active material images include one or more first reference active material images and one or more second reference active material images; the plurality of reference binary images includes one or more first reference binary images and one or more second reference binary images; the first reference active material image is obtained by photographing an active material powder, and the second reference active material image is generated by applying a designated image processing algorithm to the first reference active material image; The image processing device of claim 2 , wherein the first reference binary image is a binary image of the first reference active material image, and the second reference binary image is a binary image of the second reference active material image.

4. The artificial intelligence model includes a plurality of layers; The image processing device according to claim 1 , wherein at least one layer included in the encoding area and at least one layer included in the decoding area are connected by a skip connection among the plurality of layers.

5. The image processing device of claim 4 , wherein the layer included in the decoding region among the plurality of layers has a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and an upsampling layer are sequentially connected.

6. The image processing device of claim 4 , wherein the layer included in the encoding region among the plurality of layers has a structure in which a 2D convolution layer, a batch normalization layer, an activation layer, and a max pooling layer are sequentially connected.

7. 2. The image processing device of claim 1, wherein the active material image is an image based on a scanning electron microscope (SEM), a transmission electron microscope (TEM), an optical microscope (OM), a structured illumination microscope (SIM), or a fused ion beam (FIB).

8. The image segmentation unit The image processing device of claim 1 , wherein the segmentation of the plurality of active materials included in the active material image is performed based on a watershed algorithm.

9. further comprising an information extraction unit; the information extraction unit extracts information about the plurality of active materials based on the segmentation image; The image processing device of claim 1 , wherein the information about the plurality of active materials includes information about particle size, perimeter, sphericity, aspect ratio, convexity, solidity, distribution, quantile, or a combination thereof of the active materials.

10. a boundary removal unit that removes a boundary from the active material image based on the boundary image to generate a boundary-removed image; The image processing device of claim 1 , wherein the object identification unit identifies a plurality of objects included in the boundary-removed image.

11. The artificial intelligence model learning unit The image processing device according to claim 1, wherein the artificial intelligence model is trained so that the difference between a plurality of output images obtained by inputting the plurality of reference active material images into the artificial intelligence model and the plurality of reference boundary images is equal to or less than a specified reference value.

12. The artificial intelligence model learning unit 12. The image processing apparatus of claim 11, wherein the differences between the plurality of output images and the plurality of reference boundary images are calculated based on a mean absolute error function, a root mean square error function, a mean square error function, or a binary cross entropy loss function.

13. The image processing device of claim 12 , wherein the binary cross-entropy loss function assigns a specified weight to a specified color.

14. a distance transform unit for transforming the binary image into a distance transformed image based on a distance transform algorithm; a binary image filtering unit that filters the binary image using a threshold value that is set based on the distance transformed image, The image processing device of claim 1 , wherein the object identification unit identifies a plurality of objects included in the filtered binary image.

15. The distance conversion unit Calculating the distance from each pixel in the binary image to the nearest black pixel; The image processing apparatus of claim 14 , wherein the distance transformed image is generated by normalizing the calculated distance for each of the pixels.

16. The distance conversion unit identifying a maximum distance among the calculated distances for each of the pixels in the binary image; The image processing apparatus of claim 15, wherein the distance transformed image is generated by max-min normalizing the calculated distance for each of the pixels based on the maximum distance.

17. The binary image filtering unit selecting pixels from the distance transformed image that have a value equal to or less than a specified threshold; The image processing apparatus according to claim 14 , wherein the binary image is filtered by setting values ​​of pixels included in the binary image that correspond to the selected pixels to specified values.

18. The image processing apparatus of claim 17 , wherein the threshold is a value obtained by multiplying a normalized maximum distance of the distance transformed image by a specified percentage.

Images