WO2024106682A1

WO2024106682A1 - Device and method for analyzing average surface roughness by extracting feature from membrane image

Info

Publication number: WO2024106682A1
Application number: PCT/KR2023/010717
Authority: WO
Inventors: 강현욱; 강동희; 김나경
Original assignee: 전남대학교산학협력단
Priority date: 2022-11-14
Filing date: 2023-07-25
Publication date: 2024-05-23
Also published as: KR20240070310A

Abstract

Disclosed are a device and a method for analyzing average surface roughness by extracting a feature from membrane image. The analysis device comprises: an input vector generation unit which generates multiple processed images by processing a membrane image related to a nanofiber when the membrane image is input, groups the multiple generated processed images into an image group, and preprocesses the processed images included in the grouped image group to generate input vectors; and an analysis unit which applies the input vectors included in the image group to a pre-trained prediction model to output average logarithmic surface roughness values, and calculates statistical average surface roughness values of the membrane image by using the multiple output average surface roughness values.

Description

Surface average roughness analysis device and method using feature extraction of membrane images

The present invention relates to a surface average roughness analysis device, and more specifically, to a surface average roughness analysis device and method using feature extraction of a nanofiber membrane image to predict surface average roughness from a nanofiber-based membrane image.

With the development of the semiconductor manufacturing industry, demands for yield management, especially metrology and inspection systems, are increasing.

To meet this need, one detection algorithm implemented in scanning electron microscopy (SEM) tools exclusively uses deep learning to detect defects-of-interest (DOI). However, these conventional methods are disadvantageous for various reasons or have problems in that they do not provide optimal performance.

Therefore, research on technologies with improved defect detection performance is needed.

The problem to be solved by the present invention is to provide a surface average roughness analysis device and method using feature extraction of a membrane image to predict the surface average roughness of a nanofiber-based membrane image using a prediction model including a convolutional neural network. will be.

In order to solve the above problem, the analysis device according to the present invention processes the membrane image when a membrane image related to nanofibers is input, generates a plurality of processed images, and groups the generated plurality of processed images into image groups, An input vector generator that preprocesses each of the processed images included in the grouped image group to generate an input vector, and applies each of the input vectors included in the image group to a previously learned prediction model to generate a surface average roughness value in log units. It includes an analysis unit that outputs each surface average roughness value and calculates a statistical average surface roughness value of the membrane image using the plurality of output average surface roughness values.

In addition, it further includes a learning unit that trains a prediction model including a convolutional neural network consisting of a convolution layer, a pooling layer, and a fully connected layer, and the prediction model includes the When 3-channel data that merges the learning CLAHE (Contrast-Limited Adaptive Histogram Equalization) image, the binarized image for learning, and the learning spectrum data are input as the learning input vector, the surface average roughness predicted value in log units corresponding to the learning input vector is calculated. It is characterized by output.

In addition, the learning unit trains the predicted surface average roughness value to be closer to the correct surface average roughness value corresponding to the learning input vector through the convolution operation and pooling operation process of the prediction model, and performs the learning a preset number of times. It is characterized by repeated performance.

In addition, the input vector generator generates a plurality of processed images by cutting the membrane image into a preset size, randomly selects a preset number of processed images among the generated plurality of processed images, and It is characterized by grouping the processed images into one image group.

In addition, the input vector generator is characterized by cutting the membrane image to a preset size using a sliding-window method.

In addition, the input vector generator is characterized in that the membrane image is resized to a preset resolution before cutting the membrane image.

In addition, the input vector generator generates a CLAHE image by normalizing the pixel values of each processed image based on the CLAHE technique, and generates a binarized image by binarizing the CLAHE image based on a global thresholding method, Spectral data is generated by spectralizing the CLAHE image based on a two-dimensional discrete Fourier transform, and the input vector is generated by merging the CLAHE image, the binarized image, and the spectral data into three channels.

In the analysis method according to the present invention, when an analysis device inputs a membrane image related to a nanofiber, the membrane image is processed to generate a plurality of processed images, the generated plurality of processed images are grouped into image groups, and the grouped Preprocessing each of the processed images included in the image group to generate an input vector, and the analysis device applies each of the input vectors included in the image group to a previously learned prediction model to output a surface average roughness value in log units. and calculating a statistical average surface roughness value of the membrane image using the plurality of output average surface roughness values.

It further includes the step of training a prediction model including a convolutional neural network consisting of a convolution layer, a pooling layer, and a fully connected layer, wherein the prediction model includes the When three-channel data that is a merged learning CLAHE image, learning binarized image, and learning spectrum data is input as a learning input vector for learning, the surface average roughness predicted value in log units corresponding to the learning input vector is output.

In addition, the learning step includes training the predicted surface average roughness value to be closer to the correct surface average roughness value corresponding to the learning input vector through the convolution operation and pooling operation process of the prediction model, and performing the learning. It is characterized by including the step of repeating the process a set number of times.

In addition, the step of generating the input vector includes cutting the membrane image to a preset size to generate a plurality of processed images, and randomly selecting a preset number of processed images among the plurality of generated processed images. and grouping the selected plurality of processed images into one image group.

Additionally, the step of generating the input vector is characterized by cutting the membrane image to a preset size using a sliding window method.

Additionally, the step of generating the input vector is characterized by resizing the membrane image to a preset resolution before cutting the membrane image.

In addition, the step of generating the input vector includes generating a CLAHE image by normalizing the pixel value of each processed image based on the CLAHE technique, and generating a binarized image by binarizing the CLAHE image based on a global binarization method. , generating spectral data by spectralizing the CLAHE image based on a two-dimensional discrete Fourier transform, and generating the input vector by merging the CLAHE image, the binarized image, and the spectral data into three channels. It is characterized by

The analysis system according to the present invention includes an analysis device that analyzes the surface average roughness of a membrane image related to nanofibers, receives information related to the analyzed surface average roughness from the analysis device, and provides information related to the received surface average roughness. It includes a user terminal that outputs, wherein when a membrane image related to nanofibers is input, the analysis device processes the membrane image to generate a plurality of processed images, and groups the generated plurality of processed images into image groups, An input vector generator that preprocesses each of the processed images included in the grouped image group to generate an input vector, and applies each of the input vectors included in the image group to a previously learned prediction model to generate a surface average roughness value in log units. It is characterized in that it includes an analysis unit that outputs each surface average roughness value and calculates a statistical average surface roughness value of the membrane image using the plurality of output average surface roughness values.

According to an embodiment of the present invention, the surface average roughness of a nanofiber-based membrane image can be predicted using a prediction model including a convolutional neural network, and a statistical surface average roughness value for the corresponding membrane image can be calculated.

This allows users to quickly and accurately recognize meaningful information related to the average surface roughness of the membrane image.

1 is a configuration diagram for explaining an analysis system according to an embodiment of the present invention.

Figure 2 is a block diagram for explaining an analysis device according to an embodiment of the present invention.

Figure 3 is a block diagram for explaining a control unit according to an embodiment of the present invention.

Figure 4 is a diagram for explaining the structure of a prediction model including a convolutional neural network according to an embodiment of the present invention.

Figure 5 is a diagram for explaining a process of processing an original image according to an embodiment of the present invention.

Figure 6 is a diagram for explaining the process of generating an input vector according to an embodiment of the present invention.

Figure 7 is a diagram for explaining a global binarization image according to an embodiment of the present invention.

Figure 8 is a flowchart for explaining an analysis method for predicting average surface roughness according to an embodiment of the present invention.

Figure 9 is a diagram for explaining performance evaluation of an analysis device through continuous probability distribution according to an embodiment of the present invention.

Figure 10 is a diagram for explaining the performance evaluation of an analysis device for prediction accuracy according to an embodiment of the present invention.

Figure 11 is a block diagram for explaining a computing device according to an embodiment of the present invention.

Below, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily implement the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. In order to clearly explain the present invention in the drawings, parts unrelated to the description are omitted, and similar parts are given similar reference numerals throughout the specification.

In this specification and drawings (hereinafter referred to as “this specification”), duplicate descriptions of the same components are omitted.

Also, in this specification, when a component is mentioned as being 'connected' or 'connected' to another component, it may be directly connected or connected to the other component, but may be connected to the other component in the middle. It should be understood that may exist. On the other hand, in this specification, when it is mentioned that a component is 'directly connected' or 'directly connected' to another component, it should be understood that there are no other components in between.

Additionally, the terms used in this specification are merely used to describe specific embodiments and are not intended to limit the present invention.

Also, in this specification, singular expressions may include plural expressions, unless the context clearly dictates otherwise.

In addition, in this specification, terms such as 'include' or 'have' are only intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, and one or more It should be understood that this does not exclude in advance the presence or addition of other features, numbers, steps, operations, components, parts, or combinations thereof.

Also, in this specification, the term 'and/or' includes a combination of a plurality of listed items or any of the plurality of listed items. In this specification, 'A or B' may include 'A', 'B', or 'both A and B'.

Additionally, in this specification, detailed descriptions of well-known functions and configurations that may obscure the gist of the present invention will be omitted.

Referring to FIG. 1, the analysis system 300 predicts the average surface roughness of a nanofiber-based membrane image using a prediction model including a convolution neural network (CNN). The analysis system 300 includes an analysis device 100 and a user terminal 200.

The analysis device 100 analyzes the surface average roughness of the membrane image related to the nanofiber. Here, the membrane image may be a SEM (Scanning Electron Microscope) image. When a membrane image is input, the analysis device 100 processes the membrane image to generate a plurality of processed images. Here, the processed image refers to an image obtained by cutting one membrane image to a preset size. The analysis device 100 selects some of the plurality of generated processed images and groups them into one image group. The analysis device 100 generates an input vector by preprocessing each processed image included in the grouped image group. The analysis device 100 applies each input vector included in the image group to a previously learned prediction model and outputs each surface average roughness value in log units. Here, the prediction model may be a convolutional neural network consisting of a convolution layer, a pooling layer, and a fully connected layer, but is not limited to this. The analysis device 100 calculates a statistical average surface roughness value of the membrane image using a plurality of output average surface roughness values. Here, the statistical surface average roughness value represents the surface average roughness for the entire image, not the surface average roughness for a part of the membrane image, and means a value that reflects statistical concepts such as average value and mode. In other words, the statistical average surface roughness value may be a meaningful value that statistically represents the average surface roughness of the entire membrane image.

The user terminal 200 is a terminal used by the user and communicates with the analysis device 100. The user terminal 200 receives information related to the statistical average surface roughness calculated from the analysis device 100. At this time, the information related to the statistical average surface roughness received may be an analysis result of the membrane image transmitted from the user terminal 200, but is not limited thereto. The user terminal 200 outputs information related to the received statistical surface average roughness to help the user intuitively recognize the surface average roughness of the membrane image remotely. The user terminal 200 may be a computer system such as a desktop, laptop, smartphone, handheld PC, etc.

In the drawing, the user terminal 200 is shown as a separate configuration from the analysis device 100, but it is not limited to this and may be implemented as a single configuration depending on the situation.

Meanwhile, the analysis system 300 may establish a communication network 350 between the analysis device 100 and the user terminal 200 to support communication between them. The communication network 350 may be composed of a backbone network and a subscriber network. The backbone network may be composed of one or more integrated networks among the X.25 network, Frame Relay network, ATM network, MPLS (Multi-Protocol Label Switching) network, and GMPLS (Generalized Multi-Protocol Label Switching) network. Subscriber networks include FTTH (Fiber To The Home), ADSL (Asymmetric Digital Subscriber Line), cable network, zigbee, Bluetooth, and Wireless LAN (IEEE 802.11b, IEEE 802.11a, IEEE 802.11g, IEEE 802.11n). ), Wireless Hart (ISO/IEC62591-1), ISA100.11a (ISO/IEC 62734), CoAP (Constrained Application Protocol), MQTT (Message Queuing Telemetry Transport), WIBro (Wireless Broadband), Wimax, 3G, HSDPA (High Speed Downlink Packet Access), 4G, 5G and 6G, etc. In some embodiments, the communication network 350 may be an Internet network or a mobile communication network. Additionally, the communication network 350 may include any wireless or wired communication method that is widely known or will be developed in the future.

Figure 2 is a block diagram for explaining an analysis device according to an embodiment of the present invention, Figure 3 is a block diagram for explaining a control unit according to an embodiment of the present invention, and Figure 4 is a synthesis according to an embodiment of the present invention. Figure 5 is a diagram for explaining the structure of a prediction model including a product neural network, Figure 5 is a diagram for explaining the process of processing an original image according to an embodiment of the present invention, and Figure 6 is an input according to an embodiment of the present invention. This is a diagram for explaining the process of generating a vector, and Figure 7 is a diagram for explaining a global binarization image according to an embodiment of the present invention.

Referring to FIGS. 1 to 7 , the analysis device 100 includes a communication unit 10, an input unit 30, a control unit 50, an output unit 70, and a storage unit 90.

The communication unit 10 performs communication with the user terminal 200. The communication unit 10 may receive a membrane image related to nanofibers from the user terminal 200. Here, the membrane image may be an SEM image. Additionally, the communication unit 10 may transmit information about the statistical average surface roughness value of the membrane image to the user terminal 200.

The input unit 30 receives a learning input vector for learning a prediction model. Here, the input vector for learning may be three-channel data that merges a CLAHE (Contrast-Limited Adaptive Histogram Equalization) image for learning, a binarized image for learning, and spectrum data for learning. Additionally, the input unit 30 can input membrane images related to nanofibers.

The control unit 50 performs overall control of the analysis device 100. The control unit 50 may include an input vector generation unit 52 and an analysis unit 53, and may further include a learning unit 51.

The learning unit 51 trains a prediction model that predicts the average surface roughness of the membrane image. Here, the prediction model derives the surface average roughness (log R _a ) value in log units from the nanofiber membrane image through a series of convolution and pooling operations that extract spatial patterns. The prediction model includes a convolutional neural network consisting of a convolutional layer, a pooling layer, and a fully connected layer. At this time, the convolutional neural network may include four convolutional layers, four pooling layers, and two fully connected layers, including a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a first convolutional layer. It may be a neural network connected in the following order: 3 convolutional layers, a 3rd pooling layer, a 4th convolutional layer, a 4th pooling layer, a first fully connected layer, and a second fully connected layer.

The convolution layer performs convolution operations using convolution filters and operations using activation functions. Here, the activation function may be a Rectified Linear Unit (ReLU) that can describe non-linear relationships. The pooling layer performs pooling (or sub-sampling) operations using a pooling filter. Here, the feature map of the convolution layer is subsampled with a 2x2 max pooling operation to reduce the number of feature map coefficients, and the number of hidden neurons can be determined by K-fold cross validation, which helps avoid overfitting problems. there is.

When the prediction model receives three-channel data that merges the CLAHE image for learning, the binarized image for learning, and the spectrum data for learning as the learning input vector, it outputs the surface average roughness prediction value in log units corresponding to the learning input vector. Based on this prediction model structure, the learning unit 51 performs learning so that the surface average roughness predicted value output through the convolution operation and pooling operation process is close to the surface average roughness answer value (known in advance) corresponding to the learning input vector. I order it. To this end, the learning unit 51 can learn the algorithm through the Adam optimizer so that the difference between the predicted value and the correct value is minimized. Additionally, the learning unit 51 can use Mean square error as a loss function to check convergence through a change in the size of the loss value (difference between the correct answer and the predicted value).

When a membrane image related to nanofibers is input through the communication unit 10 or the input unit 30, the input vector generator 52 generates an input vector to predict a value close to the average surface roughness value. That is, the input vector generator 52 processes the membrane image and generates a plurality of processed images. The input vector generator 52 groups the generated plurality of processed images into image groups and preprocesses each processed image included in the grouped image groups to generate an input vector.

In detail, when a membrane image, which is a SEM image, is input (61), the input vector generator 52 cuts the input membrane image to a preset size to generate a plurality of processed images (63). At this time, the input vector generator 52 may cut the membrane image to a preset size using a sliding-window method. That is, the input vector generator 52 can move the window from the upper left to the lower right of the image and cut it to fit the window size (62). Here, the input vector generator 52 resizes the membrane image to a preset resolution before cutting the membrane image. Preferably, the input vector generator 52 resizes the resolution of the image to 51.2 pixel/um and then cuts the image width to 10 um. The input vector generator 52 increases the number of processed images by flipping or rotating the generated plurality of processed images 63 (64). The input vector generator 52 randomly selects a preset number of processed images from among the increased plurality of processed images (64), and groups the selected plurality of processed images into one image group (65). For example, the input vector generator 52 may randomly select 16 images from among images cut to a certain size from an SEM image with one average surface roughness, and group the selected images into one image group. .

Meanwhile, the input vector generator 52 may preprocess each of the plurality of processed images to generate a plurality of input vectors. The input vector generator 52 separates each processed image into RGB channels and converts the processed image into a grayscale image based on the separated RGB channels (66). The input vector generator 52 generates a CLAHE image by normalizing the processed image converted to a gray scale image. That is, the input vector generator 52 normalizes the pixel values of each processed image based on the CLAHE technique in order to minimize noise depending on the resolution between images and maximize the pixel value characteristics of the fiber boundary portion, which is an important factor in the average surface roughness. You can. The input vector generator 52 generates a binarized image by binarizing the CLAHE image based on a global thresholding method. Here, the global binarization method leaves only the shape information of the fiber portion and the pore portion, reveals the distribution characteristics of the boundary, which is the main influencing variable of roughness (Figure 7), and improves the pixel value, thereby improving the kernel layer of the prediction model. Even if you pass, the characteristic remains. The input vector generator 52 spectralizes the CLAHE image based on two-dimensional discrete Fourier transform (2d-DFT) to generate spectral data. Here, the frequency domain through two-dimensional discrete Fourier transform represents the size spectrum according to the thickness, distribution and direction of the fiber, the frequency band is different depending on the position on the image, and the brightness difference is shown for each position, so the characteristics of roughness can be expressed as fiber distribution. The difference in frequency can be expressed in pixels. The input vector generator 52 sequentially stacks the arrays of the CLAHE image, binarized image, and spectrum data (67), and merges the stacked arrays into three channels to generate an input vector (68).

The analysis unit 53 applies each input vector included in the image group to the prediction model learned in the learning unit 51 and outputs each surface average roughness value in log units. The analysis unit 53 calculates a statistical average surface roughness value of the membrane image using the plurality of output average surface roughness values. Through this, users can obtain meaningful statistical average surface roughness values. In detail, the analysis unit 53 can express the surface average roughness value in log units output from the prediction model on a continuous probability distribution and output statistical values through the mode and average values appearing on the continuous probability distribution.

The output unit 70 outputs the membrane image transmitted through the communication unit 10 or the input unit 30. The output unit 70 outputs the input vector generated by the control unit 50 and outputs the statistical average surface roughness value calculated by the control unit 50. The output unit 70 includes a liquid crystal display (LCD), a thin film transistor-liquid crystal display (TFT LCD), an organic light-emitting diode (OLED), and a flexible display (flexible display). It may include at least one of a display and a 3D display.

The storage unit 90 stores a program or algorithm for driving the guide device 100. The storage unit 90 stores the membrane image transmitted through the communication unit 10 or the input unit 30. The storage unit 90 stores the input vector generated by the control unit 50 and the statistical average surface roughness value calculated by the control unit 50. The storage unit 90 includes a flash memory type, hard disk type, multimedia card micro type, card type memory (for example, SD or XD memory, etc.), RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, It may include at least one storage medium of a magnetic disk and an optical disk.

Referring to Figures 1 and 8, the analysis method predicts the surface average roughness of a nanofiber-based membrane image using a prediction model including a convolutional neural network to calculate a statistical average surface roughness value for the membrane image. You can. Through this, the analysis method allows users to quickly and accurately recognize meaningful information related to the average surface roughness of the membrane image.

In step S110, a membrane image related to the nanofiber is input to the analysis device 100. The analysis device 100 may receive and input a membrane image from the user terminal 200, or the membrane image may be directly input by the user.

In step S120, the analysis device 100 generates an input vector for the membrane image. The analysis device 100 processes the membrane image to generate a plurality of processed images. The analysis device 100 groups the generated plurality of processed images into image groups and preprocesses each processed image included in the grouped image group to generate an input vector. At this time, the input vector may be data obtained by merging the CLAHE image, binarized image, and spectrum data into three channels.

In step S130, the analysis device 100 calculates a statistical average surface roughness value. The analysis device 100 applies each input vector included in the image group to a previously learned prediction model and outputs a surface average roughness value in log units. The analysis device 100 can calculate statistical average surface roughness values, such as the mode and average value of the membrane image, using the plurality of output average surface roughness values.

Figure 9 is a diagram for explaining the performance evaluation of an analysis device through continuous probability distribution according to an embodiment of the present invention, and Figure 10 is a diagram for explaining the performance evaluation of an analysis device for prediction accuracy according to an embodiment of the present invention. It is a drawing.

Referring to FIGS. 1, 9, and 10, the analysis device 100 can perform performance evaluation through various methods.

For example, the analysis device 100 may express the learning predicted value and the test predicted value, which are logarithmic average surface roughness values output through the prediction model, on a continuous probability distribution. Through this, the user can confirm that the mode and average values of the corresponding predicted values appear in the experimental values through measurement (Figure 9). In addition, it can be confirmed that the prediction accuracy of the analysis device 100 is improved through an improvement in the coefficient of determination according to the input vector and a decrease in the average absolute ratio error (FIG. 10).

Referring to FIG. 11, the computing device TN100 may be a device described herein (eg, an analysis device, a user terminal, etc.).

The computing device TN100 may include at least one processor TN110, a transceiver device TN120, and a memory TN130. Additionally, the computing device TN100 may further include a storage device TN140, an input interface device TN150, an output interface device TN160, etc. Components included in the computing device TN100 may be connected by a bus TN170 and communicate with each other.

The processor TN110 may execute a program command stored in at least one of the memory TN130 and the storage device TN140. The processor TN110 may refer to a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to embodiments of the present invention are performed. Processor TN110 may be configured to implement procedures, functions, and methods described in connection with embodiments of the present invention. The processor TN110 may control each component of the computing device TN100.

Each of the memory TN130 and the storage device TN140 can store various information related to the operation of the processor TN110. Each of the memory TN130 and the storage device TN140 may be comprised of at least one of a volatile storage medium and a non-volatile storage medium. For example, the memory TN130 may be comprised of at least one of read only memory (ROM) and random access memory (RAM).

The transceiving device TN120 can transmit or receive wired signals or wireless signals. The transmitting and receiving device (TN120) can be connected to a network and perform communication.

Meanwhile, the embodiments of the present invention are not only implemented through the apparatus and/or method described so far, but may also be implemented through a program that realizes the function corresponding to the configuration of the embodiment of the present invention or a recording medium on which the program is recorded. This implementation can be easily implemented by anyone skilled in the art from the description of the above-described embodiments.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by those skilled in the art using the basic concept of the present invention defined in the following claims are also possible. It falls within the scope of invention rights.

Claims

When a membrane image related to nanofibers is input, the membrane image is processed to generate a plurality of processed images, the generated plurality of processed images are grouped into image groups, and each processed image included in the grouped image group is preprocessed. an input vector generator that generates an input vector; and

Each of the input vectors included in the image group is applied to the previously learned prediction model to output logarithmic surface average roughness values, and the plurality of output surface average roughness values are used to calculate the statistical surface average of the membrane image. An analysis unit that calculates roughness values;

An analysis device comprising:
According to clause 1,

It further includes a learning unit that trains a prediction model including a convolutional neural network consisting of a convolution layer, a pooling layer, and a fully connected layer,

The prediction model is,

When 3-channel data that merges the learning CLAHE (Contrast-Limited Adaptive Histogram Equalization) image, the binarized image for learning, and the learning spectrum data are input as the learning input vector for the learning, the surface average roughness predicted value in log units corresponding to the learning input vector An analysis device characterized in that output.
According to clause 2,

The learning department,

Through the convolution operation and pooling operation process of the prediction model, the predicted surface average roughness is trained so that the predicted surface average roughness value is close to the correct surface average roughness value corresponding to the learning input vector, and the learning is repeated a preset number of times. An analysis device that does.
According to clause 1,

The input vector generator,

The membrane image is cut to a preset size to generate a plurality of processed images, a preset number of processed images are randomly selected from among the generated plurality of processed images, and the selected plurality of processed images are converted into one image. An analysis device characterized by grouping into groups.
According to clause 4,

The input vector generator,

An analysis device characterized in that the membrane image is cut to a preset size using a sliding-window method.
According to clause 4,

The input vector generator,

An analysis device characterized in that the membrane image is resized to a preset resolution before cutting the membrane image.
According to clause 1,

The input vector generator,

Based on the CLAHE technique, a CLAHE image is generated by normalizing the pixel values of each processed image, and the CLAHE image is binarized based on a global thresholding method to generate a binarized image, based on a two-dimensional discrete Fourier transform. Spectralizing the CLAHE image to generate spectral data, and generating the input vector by merging the CLAHE image, the binarized image, and the spectral data into three channels.
When a membrane image related to a nanofiber is input to the analysis device, the membrane image is processed to generate a plurality of processed images, the generated plurality of processed images are grouped into image groups, and the processed images included in the grouped image group are processed. Preprocessing each to generate an input vector; and

The analysis device applies each of the input vectors included in the image group to a previously learned prediction model to output surface average roughness values in log units, respectively, and uses the plurality of output surface average roughness values to determine the value of the membrane image. Calculating a statistical average surface roughness value;

Analysis method including.
According to clause 8,

It further includes the step of training a prediction model including a convolutional neural network consisting of a convolution layer, a pooling layer, and a fully connected layer,

The prediction model is,

When three-channel data that merges the learning CLAHE image, the learning binarization image, and the learning spectrum data are input as the learning input vector for the learning, the analysis is characterized in that it outputs the surface average roughness prediction value in log units corresponding to the learning input vector. method.
According to clause 9,

The learning step is,

Learning the predicted surface average roughness value to be closer to the correct surface average roughness value corresponding to the learning input vector through a convolution operation and a pooling operation process of the prediction model; and

Repeating the learning a preset number of times;

An analysis method comprising:
According to clause 8,

The step of generating the input vector is,

generating a plurality of processed images by cutting the membrane image into a preset size;

Randomly selecting a preset number of processed images from among the plurality of generated processed images; and

Grouping the selected plurality of processed images into one image group;

An analysis method comprising:
According to claim 11,

The step of generating the input vector is,

An analysis method characterized by cutting the membrane image to a preset size using a sliding window method.
According to clause 11,

The step of generating the input vector is,

An analysis method characterized in that the membrane image is resized to a preset resolution before cutting the membrane image.
According to clause 8,

The step of generating the input vector is,

Generating a CLAHE image by normalizing the pixel values of each processed image based on the CLAHE technique;

Binarizing the CLAHE image based on a global binarization method to generate a binarized image;

generating spectral data by spectralizing the CLAHE image based on two-dimensional discrete Fourier transform; and

generating the input vector by merging the CLAHE image, the binarized image, and the spectrum data into three channels;

An analysis method comprising:
An analysis device that analyzes the average surface roughness of membrane images related to nanofibers; and

A user terminal that receives information related to the analyzed average surface roughness from the analysis device and outputs information related to the received average surface roughness,

The analysis device is,

When a membrane image related to nanofibers is input, the membrane image is processed to generate a plurality of processed images, the generated plurality of processed images are grouped into image groups, and each processed image included in the grouped image group is preprocessed. an input vector generator that generates an input vector; and

Each of the input vectors included in the image group is applied to the previously learned prediction model to output logarithmic surface average roughness values, and the plurality of output surface average roughness values are used to calculate the statistical surface average of the membrane image. An analysis unit that calculates roughness values;

An analysis system comprising: