KR102478814B1 - Method and apparatus for improving resolution based on image segmentation - Google Patents

Abstract

An image segmentation-based resolution improvement method according to an embodiment of the present invention includes the steps of obtaining a std value indicating the degree of distribution of data for each image, classifying the images into image groups according to the numerical value of the std value, and Selecting an image group meeting a criterion from among image groups of the image data for training and performing training of an AI image processing model.

Description

Method and apparatus for improving resolution based on image segmentation

The present invention relates to a method and apparatus for improving resolution of an image based on an image segmentation method.

A super resolution image restoration method using deep learning and a convolution neural network (CNN) learns a network through supervised learning. At this time, general clear images are usually used as target images or labels for supervised learning, and images whose resolution is lowered through downsampling are used as input low-resolution images.

The learning process is as follows. At each iteration, the input low-resolution image or video is converted into a super-resolution image or video by passing through a network of convolutional and upsampling layers, comparing the super-resolution image with the target image to measure the difference, and the difference to reduce the difference. Back-propagate to change the weight of the convolution layer. As learning progresses, the difference value between the super-resolution image and the target image gradually decreases for the training data, and the more diverse the data, the more generalized the learning effect is, so that the super-resolution image restoration performance improves for other images that are not used for learning.

In the super-resolution image reconstruction method using such deep learning and CNN, in order to improve resolution restoration performance, there is a method of increasing the number of convolution layers to have a deeper dimensional neural network structure. However, in the case of this method, there is a problem in that the calculation load of the data required in the learning step and the test step is large and the time required for the calculation is long.

Accordingly, a new method for performing restoration to a target resolution while reducing the data load required for calculation and shortening the time required for resolution improvement has been required.

An object of the present invention is to propose a method and apparatus capable of performing high-quality resolution restoration in a simple manner without performing excessive data calculation operations such as using a plurality of GPUs to perform resolution improvement of an image. .

Another object of the present invention is to propose a method and apparatus capable of improving the resolution to a target image quality while reducing the time required to improve the resolution of an image.

In order to achieve the above object, an image segmentation-based resolution improvement method according to an embodiment of the present invention includes the steps of obtaining a std value indicating the degree of distribution of data for each image, and the image according to the numerical value of the std value. It may include classifying them by image group, and selecting an image group that meets a criterion among a plurality of image groups as image data for training and performing learning of an AI image processing model.

And the image segmentation-based resolution improvement method of the present invention further comprises an image processing step of inputting a target image for which resolution improvement is required to the AI image processing model and performing an AI operation when learning of the AI image processing model is completed. can do.

And the image processing step is to divide the target image into an arbitrary patch size in order to input the target image for which resolution improvement is required to the AI image processing model, and input the divided target image to the AI image processing model to AI It may be configured to include performing an operation and generating final result data by joining the division result data calculated after the AI operation.

In the step of dividing the target image into arbitrary patch sizes, the target image may be divided into a size corresponding to the patch size of the image data for training introduced in the learning step of the AI image processing model. In addition, when dividing the target image, it may be set to divide the target image into patch sizes adjusted according to the presence or absence of objects included in the target image and the size of the object.

In the step of performing the learning of the AI image processing model, a range of std values selected as image data for training may be differently set in correspondence to an attribute of an image to be improved in resolution.

In the present invention, in performing image resolution improvement, excessive data calculations such as using a plurality of conventional GPUs are required. Since restoration is performed, the burden of excessive data calculation can be reduced.

In addition, in the present invention, in order to perform image resolution improvement with the same quality, the resolution improvement is performed after dividing the image according to a predetermined criterion, and the resolution-improved divided images are joined later, so that the convolution layer of the AI model There is an advantage of not having to excessively increase the number, and as a result, the time required for AI model learning and resolution improvement calculation can be shortened.

1 is a diagram showing the configuration of a resolution improving device according to an embodiment of the present invention.
2 is a diagram showing the configuration of a learning support unit according to an embodiment of the present invention.
3 is a diagram showing the configuration of an AI image processing unit according to an embodiment of the present invention.
4 is a diagram for helping understanding of an image group according to an embodiment of the present invention.
5 is a diagram illustrating an operation of performing resolution improvement through image segmentation according to an embodiment of the present invention.
6 is a flowchart illustrating a sequence of a learning process of an AI model according to an embodiment of the present invention.
7 is a flowchart illustrating a process of improving image resolution through an AI model according to an embodiment of the present invention.

Advantages and features of the present invention, and methods of achieving them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, only these embodiments are intended to complete the disclosure of the present invention, and are common in the art to which the present invention belongs. It is provided to fully inform the person skilled in the art of the scope of the invention, and the invention is only defined by the scope of the claims.

Terminology used herein is for describing the embodiments and is not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. As used herein, "comprises" and/or "comprising" does not exclude the presence or addition of one or more other elements other than the recited elements. Like reference numerals throughout the specification refer to like elements, and “and/or” includes each and every combination of one or more of the recited elements. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first element mentioned below may also be the second element within the technical spirit of the present invention.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used with meanings commonly understood by those skilled in the art to which the present invention belongs. In addition, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless explicitly specifically defined.

The term "unit" or "module" used in the specification means a hardware component such as software, FPGA or ASIC, and "unit" or "module" performs certain roles. However, "unit" or "module" is not meant to be limited to software or hardware. A “unit” or “module” may be configured to reside in an addressable storage medium and may be configured to reproduce one or more processors. Thus, as an example, a “unit” or “module” may refer to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and variables. Functions provided within components and "units" or "modules" may be combined into smaller numbers of components and "units" or "modules" or may be combined into additional components and "units" or "modules". can be further separated.

The spatially relative terms "below", "beneath", "lower", "above", "upper", etc. It can be used to easily describe a component's correlation with other components. Spatially relative terms should be understood as including different orientations of elements in use or operation in addition to the orientations shown in the drawings. For example, if you flip a component that is shown in a drawing, a component described as "below" or "beneath" another component will be placed "above" the other component. can Thus, the exemplary term “below” may include directions of both below and above. Components may also be oriented in other orientations, and thus spatially relative terms may be interpreted according to orientation.

In this specification, a computer means any kind of hardware device including at least one processor, and may be understood as encompassing a software configuration operating in a corresponding hardware device according to an embodiment. For example, a computer may be understood as including a smartphone, a tablet PC, a desktop computer, a laptop computer, and user clients and applications running on each device, but is not limited thereto.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram showing the configuration of a resolution improving device according to an embodiment of the present invention.

As shown in FIG. 1 , the resolution improvement device 100 according to an embodiment of the present invention may include a learning support unit 110 and an AI image processing unit 120.

The learning support unit 110 may support learning of an AI calculation model for improving the resolution of a target image. The AI operation model may be a deep learning model including at least three or more convolutional layers, and a high-resolution target image defined by GT (ground truth) and a corresponding low-resolution version of the image (LR; Low Resolution image) may include a Super Resolution (SR) model, which is a method of learning to restore a low-resolution image LR to a GT by inputting in pairs.

In addition, the learning support unit 110 may select an image for learning and additionally perform a series of processes to learn an AI calculation model using the selected image, which will be described with reference to FIG. 2 .

2 is a diagram showing the configuration of a learning support unit according to an embodiment of the present invention.

As shown in FIG. 2 , the learning support unit 110 may include a training image classification unit 111 and a training image selection unit 112 .

First, the training image classification unit 111 may classify the training images to be trained with the AI calculation model by group. At this time, the criterion by which the image is classified may be based on the std value, and reference will be made to FIG. 4 for a description of this.

4 is a diagram for helping understanding of an image group according to an embodiment of the present invention.

As shown in FIG. 4, arbitrary images have a std value (meaning standard deviation), which is a value indicating the degree of distribution of data, and the training image classification unit 111 according to an embodiment of the present invention It is possible to check the std value for the image, and perform an operation to classify image groups based on this.

It can be understood that the std value contains an approximate structure of data, and the complexity of the image increases as the size of the value increases. At this time, the learning image classification unit 111 can divide a specific image into a predetermined patch size when a random patch size is determined, and when the image is divided by the patch size, the std value of the divided image can be calculated based on this. there is.

The training image classification unit 111 may classify the image groups for which the std value is calculated according to the size of the value. 4 shows an example, the std value can be calculated in a more diverse range than that shown in FIG. 4 . However, in this specification, for ease of understanding, description will be made with reference to only images whose std value is in the range of 0 to 50.

As shown in FIG. 4, the training image classification unit 111 can classify image groups according to the range of std values. For example, images having std values from 0 to 50 are classified as It can be classified into a total of five image groups, such as 20, 20-30, 30-40, and 40-50.

When the images are classified into groups according to the std value, the training image selector 112 may select image data for training to be input to the AI calculation model. At this time, the training image selection unit 112 may select training images to be input into the AI computation model and input them into the AI computation model to perform a learning operation.

According to various embodiments, when the training image selector 112 performs a learning operation by inputting training image data to an AI computation model, it may train several AI computation models for each pre-classified image group. For example, the training image selector 112 may input a first group (eg, an image group having a std value of 0 to 10) into an AI calculation model for training, and generate a first model as a result of the learning. Similarly, the training image selection unit 112 may generate various types of result models, such as a second model and a third model, as a result of inputting each image group into an AI calculation model and learning it.

Further, when selecting image data for training, the training image selector 112 may differently determine training image selection criteria according to the target image properties of the corresponding AI model. Specifically, for example, when classifying the target image according to its properties, the training image selector 112 may classify the target image into a test image and a daily image, and apply different learning criteria to each.

The images for inspection may refer to images taken for inspections such as X-ray and CT, and images for the purpose of such inspections have a relatively narrow distribution range of std values, and compared to that, a general camera is used. Daily images captured through this can be calculated with a relatively wide distribution range of std values.

In addition, the training image selection unit 112 sets the range of std values selected as training image data to, for example, 10 to 50 when learning the AI model targeting the inspection image, and learning the AI model targeting everyday images. In this case, a range of std values selected as image data for learning may be set to, for example, 0 to 100. At this time, the range of std values selected as the training image data is just an example, and the actual range of std values may be more limited or have a wider range than this.

Alternatively, the learning image selector 112 may select an image to be selected for learning according to the properties of the image to be improved (e.g., bio X-ray, bio ultrasound image, CT image, industrial X-ray, industrial general camera image, etc.) The range of the data structure value (std value) of is changed, but as the average complexity of the target image for image quality improvement increases, the selection of image groups based on the data structure value (std) of the image to be selected for learning can be set to increase. For example, if the AI model to be created is a model for improving the resolution of X-ray images for bio use, the image data to be selected for training is three image groups (e.g., std 20-30, 30-40, 40-50 ), if the AI model to be created is a model for improving the resolution of general industrial camera images, six image groups (e.g., std 20-30, 30-40, 40-50, 50-60, 60~ 70, 70 to 80) may be selected as image data for learning. As such, the range of the data structure value (std) selected as an image for training may be increased or decreased according to the property of the target image for image quality improvement.

According to various embodiments, when the balance of training images introduced in the learning step is broken, the image improvement performance of the corresponding AI model may deteriorate. In order to prevent this problem, the training image selector 112 according to an embodiment of the present invention may select only images whose std value falls within a reference range as image data for training. To this end, the super-resolution model learning unit 121a may exclude data having extreme std values from the learning step. For example, the std values of provided images are checked in the range of 0 to 100, and each image is When it is assumed that 10 groups are designated according to the size of the std value (e.g., group 1: std 0 to 10, group 2: 10 to 20…10 group: 90 to 100), the super-resolution model learning unit 121a sets the extreme value Images corresponding to image groups (eg, group 1, group 10) having a predetermined std value considered to be may be removed in the learning step.

In this way, the training image selector 112 may set the std reference value of an image to be selected as training image data according to the property of the image (whether or not the image is for inspection) to be subject to resolution improvement.

In addition, the AI model provided by the resolution improvement device 100 according to an embodiment of the present invention may be configured to include a plurality of lower AI models corresponding to each std value. For convenience of description, if the upper AI model is referred to as an integrated model and the lower AI model is referred to as a split model, the resolution improvement apparatus 100 according to an embodiment of the present invention may configure one integrated model including a plurality of split models. .

In addition, the resolution improvement device 100 can train each segmentation model as an image model having a specific std value. The number of split models included in the unified model can be adjusted.

For example, the resolution improvement device 100 sets the number of image groups input for learning to a reference value or less (eg, 4) in order to generate an AI model targeting images for inspection such as X-rays, Corresponding to the number of image groups, the number of segmentation models may also be set to four. On the other hand, in order to generate an AI model targeting everyday images, the resolution improvement device 100 may increase the number of image groups input for learning and set it to, for example, 10, and the number of split models corresponding to the number of image groups. It can also be set to 10.

In other words, the resolution improving device 100 increases the std distribution range of the image set for learning, the number of image groups, and the number of segmentation models to be generated correspondingly as the complexity increases due to the nature of the image to be improved. can

Hereinafter, with reference to FIG. 3, an operation of improving the resolution of an image through an operation using an AI model will be described in detail.

3 is a diagram showing the configuration of an AI image processing unit according to an embodiment of the present invention.

As shown in FIG. 3 , the AI image processing unit 120 according to an embodiment of the present invention may include a patch size determining unit 121, an image dividing unit 122, an AI calculation processing unit 123, and an image joining unit 124.

Prior to a description of each component of the AI image processing unit 120, the entire operation of the AI image processing unit 120 upgrading the resolution of a target image will be first described with reference to FIG. 5 for better understanding.

5 is a diagram illustrating an operation of performing resolution improvement through image segmentation according to an embodiment of the present invention.

As shown in FIG. 5, the AI image processing unit 120 divides the original image subject to resolution improvement, classifies the divided images according to each std value, and distributes and inputs them to the corresponding AI model (segmentation model). can do. As shown in FIG. 5, assuming that the AI model according to an embodiment includes 11 segmentation models, the segmented images are classified into 11 types based on the std value, and corresponding to the std value of each image. It is put into the segmentation model and AI calculations for resolution improvement can be processed respectively. Since the result data calculated in each segmentation model will still be in a segmented form, the AI image processing unit 120 can combine the segmentation result data calculated through each segmentation model to calculate final result data.

Referring to FIG. 3 again, the patch size determining unit 121 divides the image before inputting the target image to the AI model. At this time, the target image is divided into preset patch sizes. At this time, the image is The size of the division is determined.

According to various embodiments, the patch size determining unit 121 may determine a patch size of, for example, 24*24 or 48*48.

Also, according to various embodiments, the patch size determining unit 121 may differently set the patch size according to the properties of the target image, the presence or absence of a main object in the target image, and the properties of the main object in the target image.

Due to the nature of the present invention, since the resolution is restored by combining images after segmentation, a decrease in resolution may inevitably occur in the bonding area. Therefore, in order to minimize this problem, it is important to divide the image so that the main object is not divided in the area where the main object is displayed. Accordingly, in consideration of this problem, the patch size determining unit 121 may adjust the patch size in consideration of the existence of a main object, the position of the main object, or the size of the main object.

More specifically, the patch size determiner 121 may first determine whether there is a main object, draw a virtual dividing line in consideration of the location and size of the main object, and determine whether the main object is to be divided. . And then, the patch size determination unit 121 may determine a size that does not divide the main object or minimizes the division of the main object among the settable patch sizes.

Also, the image divider 122 may divide the target image according to the patch size determined by the patch size determiner 121 . The image division unit 122 may further obtain std values of the divided images and classify image groups corresponding to the checked std values.

The AI calculation processing unit 123 may input the classified image groups into a segmentation model corresponding to each std value, perform AI calculation, and calculate result data. At this time, the resultant data calculated from the segmentation model will have an improved resolution, but since it will be a segmented image, a process of rejoining the segmented image to the original image is required.

Finally, the image splicing unit 124 may perform a splicing operation for calculating final result data based on the primary result value (images whose resolution has been improved but are still divided) calculated by the AI arithmetic processing unit 123.

According to an embodiment of the present invention, at the time of image division, a position value required for recognizing a location of a divided image within an original image may be assigned so that the divided images can be joined again later. Accordingly, each divided image may be in a state in which a position value is assigned, and the image joining unit 124 may recombine the divided image based on the position value and restore the same as the original image.

6 is a flowchart illustrating a sequence of a learning process of an AI model according to an embodiment of the present invention.

As shown in FIG. 6 , the resolution improving device 100 according to an embodiment of the present invention may perform operation 605 of checking the std value of an image. At this time, the std value means a standard deviation, and contains a rough structure of the data as a value indicating the degree of distribution of the data.

When the std value is calculated in operation 605, the resolution improving device 100 may perform operation 610 of classifying images into groups according to the calculated std value. For example, the resolution improving device 100 may classify the std value of an image into a preset unit (eg, 10). Accordingly, images having a std value of 0 or more and less than 10 may be classified as a first group, and images having a std value of 10 or more and less than 20 may be classified as a second group.

Thereafter, the resolution improving device 100 may perform operation 615 of selecting an image group meeting the criterion as an image for training. Assuming that, for example, 5 image groups are generated, the resolution improving device 100 may train each segmentation model with 5 image groups.

Thereafter, the resolution improving device 100 may perform operation 620 of training an AI model with the selected training image. For example, the resolution improvement apparatus 100 may obtain 5 trained segmentation models when learning by inputting 5 image groups to each segmentation model.

7 is a flowchart illustrating a process of improving image resolution through an AI model according to an embodiment of the present invention.

As shown in FIG. 7 , when the resolution improving device 100 according to an embodiment of the present invention intends to improve the resolution by inputting the target image to the AI model, operation 705 of dividing the target image into arbitrary patch sizes is first performed. can be done

In addition, the resolution improving device 100 may perform operation 710 of checking the std value of the divided image.

Thereafter, the resolution improving device 100 may perform operation 715 of inputting the divided image into an AI model of a type corresponding to the std value of the divided image and performing calculation. When operation 715 is performed, the divided images may be distributed and input into multiple types of AI models according to the std value. In addition, since the primary result value calculated through the AI model is in the form of a divided image, the divided image needs to be joined again based on information at the time of division.

Accordingly, the resolution improving device 100 may perform operation 720 of generating a final result image by combining the calculated divided result images.

In summary, the resolution improving device 100 divides a target image whose resolution is to be improved into preset sizes, classifies the divided images according to std values, and assigns the divided images to an AI model corresponding to the std values of the divided images. It is possible to increase the resolution by inputting and combining the divided images with improved resolution to calculate the resulting image in a non-segmented state with improved resolution.

In short, the image segmentation-based resolution improvement method according to an embodiment of the present invention includes the steps of obtaining a std value indicating the degree of distribution of data for each image, and classifying the images into image groups according to the numerical value of the std value. and selecting an image group meeting a criterion from among a plurality of image groups as image data for learning and performing learning of the AI image processing model.

And the image segmentation-based resolution improvement method of the present invention further comprises an image processing step of inputting a target image for which resolution improvement is required to the AI image processing model and performing an AI operation when learning of the AI image processing model is completed. can do.

And the image processing step is to divide the target image into an arbitrary patch size in order to input the target image for which resolution improvement is required to the AI image processing model, and input the divided target image to the AI image processing model to AI It may be configured to include performing an operation and generating final result data by joining the division result data calculated after the AI operation.

In the step of dividing the target image into arbitrary patch sizes, the target image may be divided into a size corresponding to the patch size of the image data for training introduced in the learning step of the AI image processing model. In addition, when dividing the target image, it may be set to divide the target image into patch sizes adjusted according to the presence or absence of objects included in the target image and the size of the object.

In the step of performing the learning of the AI image processing model, a range of std values selected as image data for training may be differently set in correspondence to an attribute of an image to be improved in resolution. For example, the attributes of the target image may be classified into examination images having a relatively narrow distribution of std values, such as X-ray and CT images, and everyday images having a relatively wide distribution of std values. In the case of inspection images, the distribution range of std values is more limited than that of normal images, so the range of std values selected as training image data for the AI model dedicated to inspection images corresponds to this in the training stage as well. It may be limited compared to the std standard range of image data for training. For example, assuming that the range of the std value is 0 to 100, and the range of the std value of the training image data set of the AI model dedicated to inspection images is 10 to 40, the std value of the training image data set of the AI model for everyday images The value range can be set to 0 to 100.

In addition, the resolution improvement device 100 according to an embodiment of the present invention may include a learning support unit and an AI image processing unit, wherein the learning support unit obtains a std value indicating the degree of distribution of data for each image, and the image A learning image classification unit for generating an image group classified according to the numerical value of the std value and a learning image selection unit for performing learning of an AI image processing model by selecting images in the image group that meet the criteria as image data for training, , The AI image processing unit may be configured to input a target image for which resolution improvement is required to the AI image processing model to perform an AI operation upon completion of learning of the AI image processing model by the learning support unit.

In addition, according to various embodiments of the present invention, although not shown in the drawings, the resolution improving device 100 includes a communication unit for receiving image data from the outside and a storage unit for storing the image data received from the outside and the result value of which the resolution is improved. can include more.

The communication unit may use a network to transmit/receive data between the user device and the server, and the type of the network is not particularly limited. The network may be, for example, an Internet Protocol (IP) network that provides large data transmission and reception services through Internet Protocol (IP) or an All IP network integrating different IP networks. In addition, the network includes a wired network, a Wibro (Wireless Broadband) network, a mobile communication network including WCDMA, a mobile communication network including a High Speed Downlink Packet Access (HSDPA) network and a Long Term Evolution (LTE) network, LTE advanced (LTE-A ), a mobile communication network including Five Generation (5G), a satellite communication network, and a Wi-Fi network, or a combination of at least one or more of them.

The storage unit may include, for example, an internal memory or an external memory. The built-in memory includes, for example, volatile memory (e.g., DRAM (dynamic RAM), SRAM (static RAM), SDRAM (synchronous dynamic RAM), etc.), non-volatile memory (non-volatile memory) (e.g., OTPROM (one time programmable ROM), PROM (programmable ROM), EPROM (erasable and programmable ROM), EEPROM (electrically erasable and programmable ROM), mask ROM, flash ROM, flash memory (such as NAND flash or NOR flash), hard drive, Or, it may include at least one of a solid state drive (SSD).

The external memory includes a flash drive, for example, compact flash (CF), secure digital (SD), micro secure digital (Micro-SD), mini secure digital (Mini-SD), extreme digital (XD), A multi-media card (MMC) or a memory stick may be further included. The external memory may be functionally and/or physically connected to the electronic device through various interfaces.

Although the present invention has been described in detail with reference to the foregoing examples, those skilled in the art may make alterations, changes, and modifications to the examples without departing from the scope of the present invention. In short, it is not necessary to separately include all the functional blocks shown in the drawings or follow all the steps shown in the drawings in the order shown in order to achieve the intended effect of the present invention, even if not, the technical skills of the present invention described in the claims Note that it may fall within the range.

100: resolution improvement device
110: learning support department
111: learning image classification unit
112: learning image selection unit
120: AI image processing unit
121: patch size determining unit
122: image division unit
123: AI calculation processing unit
124: image junction

Claims (7)

obtaining a std value indicating a degree of distribution of data for each image;
classifying the images into image groups according to the value of the std value; and
Selecting an image group that meets a criterion among a plurality of image groups as image data for learning and performing learning of an AI image processing model; including,
The step of performing learning of the AI image processing model
Selecting an image group whose std value is within a preset reference range as image data for learning;
Corresponding to the properties of the image that is the subject of resolution improvement, the range of std values selected as image data for learning during training of the AI image processing model targeting the inspection image taken through the inspection device, When learning an AI image processing model targeting everyday images, set it narrower than the range of std values selected as image data for training,
When learning the AI image processing model targeting the inspection image, set the number of image groups input as image data for learning to be less than or equal to a preset reference value, but for learning when learning the AI image processing model targeting the everyday image Set the number of image groups input as image data to exceed a preset reference value,
Set the number of image groups to be selected in proportion to the average complexity of images to be improved in resolution;
The complexity of the image is
An image segmentation-based resolution improvement method characterized in that it is proportional to the std value of the image. According to claim 1
When the learning of the AI image processing model is completed, an image processing step of inputting a target image for which resolution improvement is required to the AI image processing model and performing an AI operation; image segmentation-based resolution improvement method further comprising . According to claim 2,
The image processing step is
Dividing the target image into arbitrary patch sizes to input the target image for which resolution improvement is required to the AI image processing model;
Injecting the divided target image into the AI image processing model to perform an AI operation; and
An image segmentation-based resolution improvement method comprising: generating final result data by joining the segmentation result data calculated after the AI operation. According to claim 3,
Dividing the target image into arbitrary patch sizes
Image segmentation-based resolution improvement method, characterized in that for dividing the target image into a size corresponding to the patch size of the training image data introduced in the learning step of the AI image processing model. According to claim 3
Dividing the target image into arbitrary patch sizes
An image division-based resolution improvement method characterized in that a target image is divided into patch sizes adjusted according to the presence or absence of an object and the size of the object. delete In the resolution improvement device comprising a learning support unit and an AI image processing unit,
The learning support department
a training image classifier for obtaining a std value indicating a degree of data distribution for each image, and generating an image group by classifying the images according to the value of the std value; and
Including; a learning image selection unit for performing learning of an AI image processing model by selecting images in the image group that meet the criteria as image data for learning;
The AI image processing unit
As the learning of the AI image processing model is completed by the learning support unit, an AI operation is performed by putting a target image for which resolution improvement is required into the AI image processing model,
The training image selection unit
Selecting an image group whose std value is within a preset reference range as image data for learning;
Corresponding to the properties of images that are subject to resolution improvement, the range of std values selected as image data for learning during training of the AI image processing model targeting images for inspection captured through the inspection device, When learning an AI image processing model targeting everyday images, set it narrower than the range of std values selected as image data for training,
When learning the AI image processing model targeting the inspection image, set the number of image groups input as image data for learning to be less than or equal to a preset reference value, but for learning when learning the AI image processing model targeting the everyday image Set the number of image groups input as image data to exceed a preset reference value,
Set the number of image groups to be selected in proportion to the average complexity of images that are subject to resolution improvement,
The complexity of the image is
An image segmentation-based resolution improvement device characterized in that it is proportional to the std value of the image.

Images