KR102492121B1 - Image classification method using data augmentation technology and computing device for performing the method - Google Patents

Abstract

Disclosed are an image classification method using data augmentation technology and a computing device for performing the method, which can increase the number of training images required to train an artificial neural network model. The image classification method performed by a computing device comprises: a step of generating a gradient energy map for training images included in a training dataset; a step of increasing the number of training images included in the training dataset by adjusting the size of the training images based on the generated gradient energy map; and a step of using the training dataset including the increased training images to train a learning model for image classification.

Description

Image classification method using data augmentation technology and computing device performing the method

The present invention relates to an image classification method, and more particularly, to a method for classifying an image using computer vision technology and image processing technology.

Various types of artificial neural network models are used to classify images using computer vision technology and image processing technology. The artificial neural network model can improve image classification accuracy as the number of training images required for learning increases, and for this purpose, efficient image resizing is required.

However, conventional methods such as scaling and cropping for image resizing may cause loss of important information included in training images, which may negatively affect the learning result of an artificial neural network model.

The present invention provides a method and apparatus capable of increasing the number of training images required for learning an artificial neural network model by performing image resizing without loss of important information included in the training images.

In addition, the present invention provides a method and apparatus capable of improving the classification accuracy of an image through an artificial neural network model using an integrated max-mean pooling layer and an attention-based network node. do.

An image classification method performed by a computing device according to an embodiment of the present invention includes generating a gradient energy map for each training image included in a training data set; increasing the number of training images included in the training data set by adjusting the size of each training image based on the generated gradient energy map; and learning a learning model for image classification using a training data set including the augmented training image.

The gradient energy map may be determined using a summation result of an absolute gradient value in an x direction and an absolute value in a y direction for each color channel with respect to pixels constituting the training image.

In the generating of the gradient energy map, a pixel value for a corresponding pixel may be determined by combining a summation result of an absolute gradient value in an x direction and an absolute absolute gradient value in a y direction determined for each color channel for the pixel.

The increasing of the number of training images may include extracting seam pixels having a minimum pixel value for all rows of a gradient energy map corresponding to the training image when scaling the training image horizontally; and reducing or enlarging the size of the training image in a horizontal direction by removing or copying the extracted seam pixels.

The extracting of the seam pixels having the minimum pixel value may include identifying a pixel having the minimum pixel value as a first seam pixel among pixels located in a first row of a gradient energy map corresponding to the training image; selecting, as a second seam pixel, a neighboring pixel having a minimum pixel value among neighboring pixels of a second row adjacent to the identified first seam pixel of the first row; and iteratively selecting a third seam pixel having a minimum pixel value up to a last row of the gradient energy map based on the method for selecting the second seam pixel.

The increasing of the number of training images may include extracting seam pixels having minimum pixel values for all columns of a gradient energy map corresponding to the training image when scaling the training image in a vertical direction; and reducing or enlarging the size of the training image in a vertical direction by removing or copying the extracted seam pixels.

The extracting of the seam pixels having the minimum pixel value may include identifying a pixel having the minimum pixel value as a first seam pixel among pixels located in a first column of a gradient energy map corresponding to the training image; selecting, as a second seam pixel, a neighboring pixel having a minimum pixel value among neighboring pixels of a second column adjacent to the identified first seam pixel of the first row; and iteratively selecting a third seam pixel having a minimum pixel value up to a last row of the gradient energy map based on the method for selecting the second seam pixel.

An image classification method according to an embodiment of the present invention includes identifying a test data set including a plurality of test images; and classifying test images included in the identified test data set for each defect type by applying the identified test data set to a learning model for image classification, wherein the learning model is a training image based on a gradient energy map. It can be learned through a training data set in which the number of training images is increased by adjusting the size of .

The gradient energy map may be determined using a summation result of an absolute gradient value in an x direction and an absolute value in a y direction for each color channel with respect to pixels constituting the training image.

The gradient energy map may be generated by combining results of summing the gradient absolute values in the x direction and the absolute gradient values in the y direction determined for each color channel for the pixel and determining a pixel value for the corresponding pixel.

When the training image is to be scaled in the horizontal direction, seam pixels having a minimum pixel value are extracted for all rows of the gradient energy map corresponding to the training image, and the extracted seam pixels are removed; or It can be increased by reducing or enlarging the size of the training image in the horizontal direction by copying.

The seam pixels having the minimum pixel value identify a pixel having the minimum pixel value among pixels located in the first row of the gradient energy map corresponding to the training image as a first seam pixel, and 1 A neighboring pixel having a minimum pixel value among neighboring pixels of a second row adjacent to a seam pixel is selected as a second seam pixel, and up to the last row of the gradient energy map based on the method of selecting the second seam pixel. It can be extracted by repeatedly selecting the third seam pixel having the minimum pixel value.

When the training image is to be resized in the vertical direction, seam pixels having a minimum pixel value are extracted for all columns of the gradient energy map corresponding to the training image, and the extracted seam pixels are removed or copied. By doing so, it can be increased by reducing or enlarging the size of the training image in the vertical direction.

The seam pixels having the minimum pixel value identify a pixel having the minimum pixel value among pixels located in a first column of the gradient energy map corresponding to the training image as a first seam pixel, and A neighboring pixel having a minimum pixel value among neighboring pixels in a second column adjacent to a pixel is selected as a second seam pixel, and based on a method of selecting the second seam pixel, the minimum pixel value up to the last row of the gradient energy map. It can be extracted by repeatedly selecting the third seam pixel having .

A computing device that performs an image classification method according to an embodiment of the present invention includes a processor, and the processor includes a gradient energy map for each training image for a training data set including a plurality of training images. ), and increasing the number of training images included in the training data set by adjusting the size of each training image based on the generated gradient energy map, and a training data set including the increased training image. A learning model for image classification can be learned using .

The gradient energy map may be determined using a summation result of an absolute gradient value in an x direction and an absolute value in a y direction for each color channel with respect to pixels constituting the training image.

The processor may generate a gradient energy map by combining a summation result of an absolute gradient value in the x direction and an absolute gradient value in the y direction determined for each color channel for the pixel and determining a pixel value for the corresponding pixel.

When the processor intends to scale the training image horizontally, the processor extracts seam pixels having a minimum pixel value for all rows of the gradient energy map corresponding to the training image, and removes or copies the extracted seam pixels. By doing so, the number of training images can be increased by reducing or enlarging the size of the training images in the horizontal direction.

The processor identifies a pixel having a minimum pixel value among pixels located in a first row of the gradient energy map corresponding to the training image as a first seam pixel, and adjacent to the first seam pixel of the identified first row. Among the neighboring pixels in the second row, a neighboring pixel having a minimum pixel value is selected as a second seam pixel, and based on the method of selecting the second seam pixel, a first row having a minimum pixel value up to the last row of the gradient energy map is selected. By repeatedly selecting 3 seam pixels, it is possible to extract seam pixels having the minimum pixel value.

When the processor intends to resize the training image in the vertical direction, the processor extracts seam pixels having a minimum pixel value for all columns of the gradient energy map corresponding to the training image, and removes or copies the extracted seam pixels. The number of training images may be increased by reducing or enlarging the size of the training images in the vertical direction.

The processor identifies a pixel having a minimum pixel value among pixels located in a first column of the gradient energy map corresponding to the training image as a first seam pixel, and a second seam pixel adjacent to the identified first seam pixel in the first column. A neighboring pixel having a minimum pixel value among neighboring pixels in a row is selected as a second seam pixel, and a third seam pixel having a minimum pixel value up to the last row of the gradient energy map based on the method for selecting the second seam pixel. It is possible to extract seam pixels having the minimum pixel value by repeatedly selecting .

The present invention can increase the number of training images required for learning an artificial neural network model by performing image resizing without loss of important information included in the training images.

In addition, the present invention can improve classification accuracy of images through an artificial neural network model using an integrated max-mean pooling layer and an attention-based network node.

1 is a diagram illustrating a computing device performing an image classification method according to an embodiment of the present invention.
2 is a diagram illustrating a learning method of a learning model for image classification according to an embodiment of the present invention.
3 is a diagram showing samples of concrete images included in a learning data set according to an embodiment of the present invention.
4 is a diagram illustrating an example in which the number of training images is increased by generating duplicate training images by applying an image augmentation method to an original training image according to an embodiment of the present invention.
5 to 7 are views illustrating a method of increasing the number of training images included in a training data set using seam carving according to an embodiment of the present invention.
8 is a diagram illustrating an image processing technique according to an embodiment of the present invention.
9 is a diagram showing the structure of a learning model according to an embodiment of the present invention.

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

1 is a diagram illustrating a computing device performing an image classification method according to an embodiment of the present invention.

As shown in FIG. 1, the computing device 100 includes one or more processors 110, a memory 130 for loading a program 140 executed by the processor 110, and a program 140. It may include a storage 120 for storing. Components included in the computing device 100 of FIG. 1 are only examples, and those skilled in the art to which the present invention pertains may further include other general-purpose components in addition to the components shown in FIG. 1 . Able to know.

The processor 110 controls the overall operation of each component of the computing device 100 . The processor 110 may include at least one of a Central Processing Unit (CPU), a Micro Processor Unit (MPU), a Micro Controller Unit (MCU), a Graphic Processing Unit (GPU), or any type of processor well known in the art. can be configured to include Also, the processor 110 may perform an operation for at least one application or program for executing a method/operation according to various embodiments of the present disclosure. Computing device 100 may include one or more processors.

Memory 130 stores various data, commands and/or information. The memory 130 may load the program 140 stored in the storage 120 to execute methods/operations according to various embodiments of the present invention. An example of the memory 130 may be RAM, but is not limited thereto.

The storage 120 may non-temporarily store one or more programs 140 . The storage 120 may include nonvolatile memory such as read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, a hard disk drive (HDD), a solid state disk (SSD), and a removable memory. It may be configured to include a disk, or any type of computer-readable recording medium well known in the art to which the present invention pertains.

Program 140 may include one or more actions in which methods/acts according to various embodiments of the present invention are implemented. Here, an operation corresponds to an instruction realized in the program 140 . For example, the program 140 generates a gradient energy map for each training image included in the training data set, and adjusts the size of each training image based on the generated gradient energy map. It may include instructions for performing an operation of increasing the number of training images included in the training data set and an operation of learning a learning model for image classification using the training data set including the increased training images.

When the program 140 is loaded into the memory 130, the processor 110 may perform methods/operations according to various embodiments of the present disclosure by executing a plurality of operations for implementing the program.

An execution screen of the program 140 may be displayed through the display 150 . In the case of FIG. 1 , the display 150 is represented as a separate device connected to the computing device 100, but in the case of a computing device 100 such as a user-portable terminal such as a smartphone or a tablet, the display 150 ) may be a component of the computing device 100. The screen displayed on the display 150 may be before information is input to the program or a result of executing the program.

2 is a diagram illustrating a learning method of a learning model for image classification according to an embodiment of the present invention.

At step 210, the processor 110 of the computing device 100 may identify a training data set to be used when training a training model for image classification. For example, in the present invention, concrete images can be used in the training data set, and the concrete images include non-crack concrete images, surface crack concrete images, delamination concrete images, and spalling concrete images. ) may correspond to any one of the concrete images. However, the type and classification of images used in this learning data set is only an example and is not limited to the above example.

3 is a diagram showing samples of concrete images included in a learning data set according to an embodiment of the present invention. 3 (a) shows an image of non-cracked concrete, (b) shows an image of surface cracked concrete, (c) shows an image of peeled concrete, and (d) shows an image of spalled concrete.

The processor 110 may generate a training image set and a test image set by dividing the concrete images included in the identified training data set into training images and test images. For example, the processor 110 may generate a training image set and a test image set by dividing concrete images included in the training data set into training images and test images at a ratio of 7:3. However, the ratio value for classifying concrete images in this way is only an example, and the ratio value can be changed by the user.

In step 220, the processor 110 may increase the number of training images included in the training data set among the training data sets. More specifically, the processor 110 applies at least one image augmentation method of scaling, cropping, and seam carving to each of the training images included in the training data set to obtain the training data set. The number of included training images can be increased.

4 is a diagram illustrating an example in which the number of training images is increased by generating duplicate training images by applying an image augmentation method to an original training image according to an embodiment of the present invention. First, the processor 110 applies scale factors of 0.5 and 2 to the original training image 410, thereby generating a copy training image 420 whose size is reduced by 1/4 compared to the original training image 410 and a copy training image 420 whose size is increased by 4 times. A duplicate training image 430 may be created. The size of this scale factor is only one example and is not limited thereto.

Also, the processor 110 may generate a copy training image 440 including only the specific region by applying cropping to a specific region inside the original training image 410 .

In addition, the processor 110 applies seam carving to the original training image 410 to obtain a copy training image 450 having the same horizontal length but different vertical length and a copy training image 460 having the same vertical length but different horizontal length. can create

In step 230, the processor 110 may apply an image processing technique to the training data set in which the training images are augmented. More specifically, the processor 110 extracts an object of interest by applying image segmentation to a training image, and converts the training image from which the object of interest is extracted into a gray scale, thereby improving performance of a learning model.

In step 240, the processor 110 may learn a learning model for image classification using a training image of a training data set to which an image processing technique is applied. As an example, the present invention may use a Convolutional Neural Network (CNN) architecture as a learning model for image classification.

A CNN is a deep neural network often used for image classification, which can consist of a convolutional layer with an activation function, a pooling layer for analyzing input characteristics, and a connection layer for classification.

As an example, VGG16 is the most frequently used CNN variant. VGG16 consists of a total of 16 layers, 13 of which are convolutional layers, and the remaining 3 layers are fully connected. VGG16 uses ReLU as an activation function to improve the nonlinearity of the learning model, and uses a softmax function to classify images in the final layer.

5 to 7 are views illustrating a method of increasing the number of training images included in a training data set using seam carving according to an embodiment of the present invention.

The processor 110 may generate a gradient energy map for each training image included in the training data set constituting the data set. More specifically, the processor 110 may sum the x-direction gradient absolute value and the y-direction gradient absolute value for each color channel with respect to pixels constituting the training image. Then, the processor 110 combines the sum of the gradient absolute values in the x direction and the absolute gradient values in the y direction for all color channels identified for the corresponding pixel, and obtains the gradient energy with the combined result as the pixel value for the corresponding pixel. You can create maps. As an example, when using the MATLAB platform, the processor 110 may use the built-in function of “gradients(img)” to calculate the absolute value of the gradient in the x direction Fx and the absolute value of the gradient in the y direction Fy of the training image, and the calculated Fx A gradient energy map can be generated through the sum of F and Fy.

As an example, FIGS. 5 to 7 show a method of generating copy training images 530 and 540 having the same vertical length but different horizontal lengths by applying seam carving to an original training image 510 . First, in FIG. 5 , the processor 110 may extract seam pixels having minimum pixel values for all rows of the gradient energy map 520 corresponding to the original training image 510 . To this end, the processor 110 may identify a pixel having a minimum pixel value among pixels included in the first row of the gradient energy map 520 as the first seam pixel 11 .

Thereafter, the processor 110 selects a neighboring pixel having a minimum pixel value among the neighboring pixels 25, 17, and 12 of the second row adjacent to the seam pixel 11 of the first row as the second seam pixel 12. can

In this way, the processor 110 repeats the method of selecting a neighboring pixel having the minimum pixel value among the neighboring pixels of the next row adjacent to the selected seam pixel as a seam pixel up to the last row of the gradient energy map 520, so that the seam pixels It is possible to determine the seam piece consisting of For example, in FIG. 5, the seam piece may be composed of (11, 12, 14, 16, 14, 10). In the above example, the seam pixels are selected from the first row, but this is only an example, and the seam pixels may be selected from the last row.

Meanwhile, the processor 110 may generate copy training images 530 and 540 having the same vertical length but different horizontal lengths from the original training image 510 using the seam pieces determined in this way.

For example, the processor 110 removes seam pixels corresponding to the seam pieces determined in the gradient energy map 520 as shown in (a) of FIG. 6 so that the vertical length is the same but the horizontal length is the same as in (b) of FIG. A short copy training image 530 can be created.

As another example, the processor 110 copies the seam pixels corresponding to the seam pieces determined in the gradient energy map 520 as shown in (a) of FIG. 7 and adds them in the horizontal direction, thereby extending the vertical length as shown in (b) of FIG. may generate a copy training image 540 having the same but long horizontal length.

5 to 7 provide a method of generating copy training images 530 and 540 having the same vertical length but different horizontal lengths by applying seam carving to the original training image 510, but this is only an example of the original training image 510. By applying seam carving to the image 510, copy training images having the same horizontal length but different vertical lengths may be generated.

To this end, the processor 110 determines a seam piece by extracting seam pixels having minimum pixel values for all columns of the gradient energy map 520 corresponding to the original training image 510, and the seam pixel corresponding to the determined seam piece. By removing or copying them, copy training images having the same horizontal length but different vertical lengths can be created.

8 is a diagram illustrating an image processing technique according to an embodiment of the present invention.

As mentioned above, the processor 110 may apply an image processing technique to a training data set in which training images are increased.

To this end, in step 810, the processor 110 may extract an object of interest by applying image segmentation to the training image. Image segmentation can be used to transform training images into more meaningful and understandable images for analysis. For example, the processor 110 may extract an object of interest by dividing a training image using a k-means clustering algorithm. This image segmentation algorithm is only one example and is not limited thereto.

In operation 820, the processor 110 may convert the training image from which the object of interest is extracted into a gray scale. Grayscale levels can be used to remove hue and chroma information from training images while preserving luminance. In this way, the processor 110 converts the training image into a gray scale and increases the depth of change in contrast in pixel values, thereby improving the performance of a later learning model.

In step 830, the processor 110 may generate a clean edge map by identifying the most important edge features of the training image converted to grayscale.

After that, in step 840, the processor 110 may supplement color channels of the training image by performing image binarization on the generated edge map.

Finally, in step 850, the processor 110 may improve the resolution of the training image by removing noise by applying a median filter to the training image on which image binarization has been performed.

9 is a diagram showing the structure of a learning model according to an embodiment of the present invention.

In the learning model for image classification provided by the present invention, the test images included in the test data set among the training data sets are non-crack concrete images, surface crack concrete images, and delamination concrete images. and spalling concrete images.

9 shows the architecture of a convolution-based defect classification neural network corresponding to this learning model. The learning model provided by the present invention, the defect classification neural network, is reconstructed into VGG16 using an integrated maximum-average pooling layer and an attention-based network node, thereby improving the classification accuracy of concrete images.

In general, the method of using the maximum pooling layer and the average pooling layer alone has a disadvantage in that information in the image may be lost. However, the combined maximum-average pooling layer provided by the present invention can avoid loss of important information in the image. To this end, the integrated maximum-average pooling layer provided by the present invention evaluates all components of the pooling area to avoid increasing variance while maintaining background information, and captures only the foreground texture information of the strongest activation as a representative feature of the pooling area. to be.

On the other hand, the main purpose of the attention-based network is to recognize multiple objects in an image, and the attention-based network node provided in the present invention implements a means of connecting the network that determines the average based on the channel axis to achieve the maximum possible performance. A max-average pooling technique can be used.

Meanwhile, the method according to the present invention is written as a program that can be executed on a computer and can be implemented in various recording media such as magnetic storage media, optical reading media, and digital storage media.

Implementations of the various techniques described herein may be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or combinations thereof. Implementations may be a computer program product, i.e., an information carrier, e.g., a machine-readable storage, for processing by, or for controlling, the operation of a data processing apparatus, e.g., a programmable processor, computer, or plurality of computers. It can be implemented as a computer program tangibly embodied in a device (computer readable medium) or a radio signal. A computer program, such as the computer program(s) described above, may be written in any form of programming language, including compiled or interpreted languages, and may be written as a stand-alone program or in a module, component, subroutine, or computing environment. It can be deployed in any form, including as other units suitable for the use of. A computer program can be deployed to be processed on one computer or multiple computers at one site or distributed across multiple sites and interconnected by a communication network.

Processors suitable for processing a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from read only memory or random access memory or both. Elements of a computer may include at least one processor that executes instructions and one or more memory devices that store instructions and data. In general, a computer may include, receive data from, send data to, or both, one or more mass storage devices that store data, such as magnetic, magneto-optical disks, or optical disks. It can also be combined to become. Information carriers suitable for embodying computer program instructions and data include, for example, semiconductor memory devices, for example, magnetic media such as hard disks, floppy disks and magnetic tapes, compact disk read only memory (CD-ROM) ), optical media such as DVD (Digital Video Disk), magneto-optical media such as Floptical Disk, ROM (Read Only Memory), RAM (RAM) , Random Access Memory), flash memory, EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), and the like. The processor and memory may be supplemented by, or included in, special purpose logic circuitry.

In addition, computer readable media may be any available media that can be accessed by a computer, and may include both computer storage media and transmission media.

Although this specification contains many specific implementation details, they should not be construed as limiting on the scope of any invention or what is claimed, but rather as a description of features that may be unique to a particular embodiment of a particular invention. It should be understood. Certain features that are described in this specification in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable subcombination. Further, while features may operate in particular combinations and are initially depicted as such claimed, one or more features from a claimed combination may in some cases be excluded from that combination, and the claimed combination is a subcombination. or sub-combination variations.

Similarly, while actions are depicted in the drawings in a particular order, it should not be construed as requiring that those actions be performed in the specific order shown or in the sequential order, or that all depicted actions must be performed to obtain desired results. In certain cases, multitasking and parallel processing can be advantageous. Further, the separation of various device components in the embodiments described above should not be understood as requiring such separation in all embodiments, and the program components and devices described may generally be integrated together into a single software product or packaged into multiple software products. You need to understand that you can.

On the other hand, the embodiments of the present invention disclosed in this specification and drawings are only presented as specific examples to aid understanding, and are not intended to limit the scope of the present invention. In addition to the embodiments disclosed herein, it is obvious to those skilled in the art that other modified examples based on the technical idea of the present invention can be implemented.

100: computing device
110: processor
120: storage
130: memory
140: program
150: display

Claims (20)

An image classification method performed by a computing device,
generating a gradient energy map for each training image included in the training data set;
extracting seam pixels having minimum pixel values for all rows of a gradient energy map corresponding to the training image when scaling the training image horizontally;
increasing the number of training images by reducing or enlarging the size of the training images in a horizontal direction by removing or copying the extracted seam pixels; and
Learning a learning model for image classification using a training data set including the augmented training image
Image classification method comprising a. According to claim 1,
The gradient energy map,
An image classification method determined by using a summation result of an absolute gradient value in an x direction and an absolute absolute gradient value in a y direction for each color channel with respect to pixels constituting the training image. According to claim 2,
Generating the gradient energy map,
The image classification method of determining a pixel value for a corresponding pixel by combining all results of summing the gradient absolute value in the x direction and the absolute absolute value of the gradient in the y direction determined for each color channel for the pixel. delete According to claim 1,
Extracting the seam pixels having the minimum pixel value,
identifying a pixel having a minimum pixel value among pixels located in a first row of a gradient energy map corresponding to the training image as a first seam pixel;
selecting, as a second seam pixel, a neighboring pixel having a minimum pixel value among neighboring pixels of a second row adjacent to the identified first seam pixel of the first row; and
Repeatedly selecting a third seam pixel having a minimum pixel value until the last row of the gradient energy map based on the method for selecting the second seam pixel.
Image classification method comprising a. An image classification method performed by a computing device,
generating a gradient energy map for each training image included in the training data set;
extracting seam pixels having minimum pixel values for all columns of a gradient energy map corresponding to the training image when scaling the training image in the vertical direction;
increasing the number of training images by reducing or enlarging the size of the training images in a vertical direction by removing or copying the extracted seam pixels; and
Learning a learning model for image classification using a training data set including the augmented training image
Image classification method comprising a. According to claim 6,
Extracting the seam pixels having the minimum pixel value,
identifying a pixel having a minimum pixel value among pixels located in a first column of a gradient energy map corresponding to the training image as a first seam pixel;
selecting, as a second seam pixel, a neighboring pixel having a minimum pixel value among neighboring pixels of a second column adjacent to the identified first seam pixel of the first row; and
Repeatedly selecting a third seam pixel having a minimum pixel value until the last column of the gradient energy map based on the method for selecting the second seam pixel.
Image classification method comprising a. identifying a test data set including a plurality of test images;
Classifying test images included in the identified test data set according to defect types by applying the identified test data set to a learning model for image classification.
including,
The training image used to learn the learning model is
When resizing the training image in the horizontal direction, extracting seam pixels having a minimum pixel value for all rows of the gradient energy map corresponding to the training image, and removing or copying the extracted seam pixels to perform the training Increased by reducing or enlarging the size of the image,
The learning model,
An image classification method in which learning for image classification is performed using a training data set including the augmented training image. According to claim 8,
The gradient energy map,
An image classification method determined by using a summation result of an absolute gradient value in an x direction and an absolute absolute gradient value in a y direction for each color channel with respect to pixels constituting the training image. According to claim 9,
The gradient energy map,
The method of classifying the image by combining the summation results of the gradient absolute value in the x direction and the absolute gradient value in the y direction determined for each color channel for the pixel and determining the pixel value for the corresponding pixel. delete According to claim 8,
The seam pixels having the minimum pixel value,
Among the pixels located in the first row of the gradient energy map corresponding to the training image, a pixel having a minimum pixel value is identified as a first seam pixel, and a second row adjacent to the identified first seam pixel of the first row is identified. A neighboring pixel having a minimum pixel value among neighboring pixels of is selected as a second seam pixel, and a third seam pixel having a minimum pixel value up to the last row of the gradient energy map based on a method for selecting the second seam pixel. Image classification method extracted by repeatedly selecting identifying a test data set including a plurality of test images;
Classifying test images included in the identified test data set by defect type by applying the identified test data set to a learning model for image classification.
including,
The training image used to learn the learning model is
When resizing the training image in the vertical direction, extracting seam pixels having a minimum pixel value for all columns of the gradient energy map corresponding to the training image, and removing or copying the extracted seam pixels to obtain the training image is increased by reducing or enlarging the size of
The learning model,
An image classification method in which learning for image classification is performed using a training data set including the augmented training image. According to claim 13,
The seam pixels having the minimum pixel value,
Among the pixels located in the first column of the gradient energy map corresponding to the training image, a pixel having a minimum pixel value is identified as a first seam pixel, and a neighboring pixel in a second column adjacent to the identified first seam pixel in the first column. Among them, a neighboring pixel having the minimum pixel value is selected as a second seam pixel, and based on the method of selecting the second seam pixel, the third seam pixel having the minimum pixel value is repeated until the last column of the gradient energy map. Image classification method extracted by selecting. A computing device for performing an image classification method,
contains a processor;
the processor,
For a training data set including a plurality of training images, a gradient energy map is generated for each training image, and when scaling the training image horizontally, the gradient corresponding to the training image is desired. Extracting seam pixels having a minimum pixel value for all rows of the energy map, removing or copying the extracted seam pixels to reduce or enlarge the size of the training image in the horizontal direction to increase the number of training images, A computing device that learns a learning model for image classification using a training data set including the augmented training image. According to claim 15,
The gradient energy map,
Computing device generated by combining a summation result of an absolute gradient value in the x direction and an absolute gradient value in the y direction for each color channel for pixels constituting the training image and determining a pixel value for the corresponding pixel. delete According to claim 15,
the processor,
Among the pixels located in the first row of the gradient energy map corresponding to the training image, a pixel having a minimum pixel value is identified as a first seam pixel, and a second row adjacent to the identified first seam pixel of the first row is identified. A neighboring pixel having a minimum pixel value among neighboring pixels of is selected as a second seam pixel, and a third seam pixel having a minimum pixel value up to the last row of the gradient energy map based on a method for selecting the second seam pixel. A computing device for extracting seam pixels having the minimum pixel value by repeatedly selecting . A computing device for performing an image classification method,
contains a processor;
the processor,
For a training data set including a plurality of training images, a gradient energy map for each training image is generated, and when scaling the training image in the vertical direction, the gradient corresponding to the training image is desired. The number of training images is increased by extracting seam pixels having a minimum pixel value for all columns of the energy map, reducing or enlarging the size of the training image in the vertical direction by removing or copying the extracted seam pixels, A computing device that learns a learning model for image classification using a training data set comprising augmented training images. According to claim 19,
the processor,
Among the pixels located in the first column of the gradient energy map corresponding to the training image, a pixel having a minimum pixel value is identified as a first seam pixel, and a neighboring pixel in a second column adjacent to the identified first seam pixel in the first column. Among them, a neighboring pixel having the minimum pixel value is selected as a second seam pixel, and based on the method of selecting the second seam pixel, the third seam pixel having the minimum pixel value is repeated until the last column of the gradient energy map. A computing device for extracting seam pixels having the minimum pixel value by selecting.

Images