Method and system for training image classification model

1. A method of training an image classification model, the method comprising:

maintaining a plurality of ground-penetrating radar (GPR) images captured by a GPR in a database;

establishing a feature value extraction model by primarily learning training data comprising a GPR image in the database and a representative feature value determined as one of feature values of the GPR image, wherein the feature value extraction model is a first model that outputs a feature value set of a new GPR image input to an image classification model;

inputting a GPR image to the feature value extraction model and acquiring a feature value set of the GPR image output from the feature value extraction model; and

establishing a feature value classification model by secondarily learning training data comprising the acquired feature value set of the GPR image and a label value to which the GPR image is classified, wherein the feature value classification model is a second model that outputs a label value of the new GPR image as a result value of the image classification model.

2. The method of claim 1 , further comprising:

learning, before the primarily learning, training data comprising the GPR image and a label value of the GPR image, so that the primarily learning is performed in a state in which the GPR image and a representative feature value of the GPR image are separated based on the label value of the GPR image learned through the learning.

3. The method of claim 1 , further comprising:

initializing the feature value extraction model by training a convolutional neural network (CNN) using training data comprising the GPR image and a label value of the GPR image; and

primarily learning, by the initialized feature value extraction model, training data comprising the GPR image and a representative feature value of the GPR image.

4. The method of claim 3 , wherein the primarily learning comprises:

setting the GPR image in the training data to be an input value of the initialized feature value extraction model, setting the representative feature value of the GPR image in the training data to be an output of the initialized feature value extraction model, and then performing the primarily learning.

5. The method of claim 1 , wherein the establishing of the feature value classification model comprises:

secondarily learning, by a CNN, training data comprising the feature value set of the GPR image and a label value of the GPR image and establishing the feature value classification model.

6. The method of claim 1 , further comprising:

determining, when each of the plurality of GPR images is classified as a label value of one of “Left”, “Peak”, “Right” and “Other” based on a pattern of each image, a representative feature value of each of the plurality of GPR images using different K-means clustering models corresponding to a total number of the label value,

wherein the maintaining in the database comprises:

maintaining, in the database, each of the plurality of GPR images in association with the representative feature value and the label value as data for training.

7. The method of claim 6 , further comprising:

acquiring a feature value set associated with the GPR image by inputting the new GPR image to the feature value extraction model;

inputting the acquired feature value set to the feature value classification model:

presenting, by the feature value classification model, a distribution of a plurality of feature values included in the feature value set in a distribution map that shows a central feature value allocated to each label value for classifying an entire GPR image and identifying, in the distribution map, a first central feature value having a minimum distance from a representative feature value of the new GPR image distributed relatively at a center among the plurality of feature values;

outputting a label value allocated to the first central feature value from the feature value classification model as a classification result of the new GPR image; and

adding the new GPR image in association with the representative feature value of the new GPR image and a label value to which the new GPR image is classified, to the database as data for training.

8. The method of claim 7 , further comprising:

applying, when a representative feature value of a GPR image maintained in the database is updated in response to the data for training related to the new GPR image being added in the database, the updated representative feature value to the feature value extraction model by performing the primarily learning using the updated representative feature value within a predetermined update count.

9. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1 .

10. A system for training an image classification model, the system comprising:

a database in which a plurality of ground-penetrating radar (GPR) images captured by a GPR is maintained:

an extraction model establisher configured to establish a feature value extraction model by primarily learning training data comprising a GPR image in the database and a representative feature value determined as one of feature values of the GPR image, wherein the feature value extraction model is a first model that outputs a feature value set of a new GPR image input to an image classification model;

an acquirer configured to input a GPR image to the feature value extraction model and acquire a feature value set of the GPR image output from the feature value extraction model; and

a classification model establisher configured to establish a feature value classification model by secondarily learning training data comprising the acquired feature value set of the GPR image and a label value to which the GPR image is classified, wherein the feature value classification model is a second model that outputs a label value of the new GPR image as a result value of the image classification model.

11. The system of claim 10 , wherein the extraction model establisher is configured to learn, before the primarily learning, training data comprising the GPR image and a label value of the GPR image, so that the primarily learning is performed in a state in which the GPR image and a representative feature value of the GPR image are separated based on the label value of the GPR image learned through the learning.

12. The system of claim 10 , wherein the extraction model establisher is configured to initialize the feature value extraction model by training a convolutional neural network (CNN) using training data comprising the GPR image and a label value of the GPR image, and primarily learn, by the initialized feature value extraction model, training data comprising the GPR image and a representative feature value of the GPR image.

13. The system of claim 12 , wherein the extraction model establisher is configured to set the GPR image in the training data to be an input value of the initialized feature value extraction model, set the representative feature value of the GPR image in the training data to be an output of the initialized feature value extraction model, and then perform the primarily learning.

14. The system of claim 10 , wherein the classification model establisher is configured to establish the feature value classification model by allowing a CNN to secondarily learn the feature value set of the GPR image and training data comprising a label value of the GPR image.

15. The system of claim 10 , further comprising:

a determiner configured to determine, when each of the plurality of GPR images is classified as a label value of one of “Left”, “Peak”, “Right” and “Other” based on a pattern of each image, a representative feature value of each of the plurality of GPR images using different K-means clustering models corresponding to a total number of the label value,

wherein the database is configured to maintain each of the plurality of GPR images in association with the representative feature value and the label value as data for training.

16. The system of claim 15 , wherein when the acquirer acquires a feature value set associated with the GPR image by inputting the new GPR image to the feature value extraction model, the classification model establisher is configured to:

establish a feature value classification model that presents a distribution of a plurality of feature values included in the feature value set in a distribution map that shows a central feature value allocated to each label value for classifying an entire GPR image, identifies, in the distribution map, a first central feature value having a minimum distance from a representative feature value of the new GPR image distributed relatively at a center among the plurality of feature values, and outputs a label value allocated to the first central feature value as a classification result of the new GPR image; and

add the new GPR image in association with the representative feature value of the new GPR image and a label value to which the new GPR image is classified by the feature value classification model, to the database as data for training.

17. The system of claim 16 , wherein when the classification model establisher updates a representative feature value of a GPR image maintained in the database in response to the data for training related to the new GPR image being added in the database, the extraction model establisher is configured to apply the updated representative feature value to the feature value extraction model by performing the primarily learning using the updated representative feature value within a predetermined update count.