Method and device for generating training data and computer program stored in computer-readable recording medium

1. A method for generating defect data of a target domain using defect data of a source domain, executed on one or more processors of a computing device, the method comprising:

inputting a defect data of the source domain, to which a first mask is applied and a defect data of the source domain, to which the first mask is not applied, to a reconstruction algorithm;

first training the reconstruction algorithm to generate a defect data of the source domain, to which the first mask is reconstructed;

inputting a normal data of the source domain, to which a second mask is applied and a normal data of the source domain, to which the second mask is not applied, to the reconstruction algorithm;

second training the reconstruction algorithm to generate a normal data of the source domain, to which the second mask is reconstructed;

inputting a normal data of the target domain, to which the second mask is applied and a normal data of the target domain, to which the second mask is not applied, to the reconstruction algorithm; and

third training the reconstruction algorithm to generate a normal data of the target domain, to which the second mask is reconstructed;

inputting a normal data of the target domain, to which the first mask is applied, to the reconstruction algorithm which is trained to generate a mask reconstructed data from a mask applied data; and

generating a defect data of the target domain using a final-trained reconstruction algorithm.

2. The method for generating defect data of a target domain using defect data of a source domain according to claim 1 , wherein the defect data and the normal data is image data.

3. The method for generating defect data of a target domain using defect data of a source domain according to claim 1 , wherein the first mask masks a defect part of a defect data, and the first mask and the second mask are distinguished by different colors.

4. The method for generating defect data of a target domain using defect data of a source domain according to claim 1 , wherein the first mask is a shape for masking a defect part of a defect data, and the second mask is a same or different shape as the first mask.

5. The method for generating defect data of a target domain using defect data of a source domain according to claim 1 , wherein the source domain and the target domain have different patterns or shapes.

6. A method for generating defect data of a target domain using defect data of a source domain, executed on one or more processors of a computing device, the method comprising:

inputting a defect data of the source domain, to which a first mask is applied and a defect data of the source domain, to which the first mask is not applied, to a reconstruction algorithm;

first training the reconstruction algorithm to generate a defect data of the source domain, to which the first mask is reconstructed;

inputting a normal data of the source domain, to which a second mask is applied and a normal data of the source domain, to which the second mask is not applied, to the reconstruction algorithm;

second training the reconstruction algorithm to generate a normal data of the source domain, to which the second mask is reconstructed;

inputting a normal data of the target domain, to which the second mask is applied and a normal data of the target domain, to which the second mask is not applied, to the reconstruction algorithm; and

third training the reconstruction algorithm to generate a normal data of the target domain, to which the second mask is reconstructed, wherein the reconstruction algorithm comprises:

a generating network; and

an identification network, and

wherein the first training comprises:

inputting the defect data of the source domain, to which the first mask is applied, to the generating network;

generating the defect data of the source domain, to which the first mask is reconstructed, from the defect data of the source domain, to which the first mask is applied, using the generating network;

inputting the defect data of the source domain, to which the first mask is not applied and the defect data of the source domain, to which the first mask is reconstructed, to the identification network;

outputting distinguishing information by comparing the defect data of the source domain, to which the first mask is not applied, and the defect data of the source domain, to which the first mask is reconstructed, using the identification network;

training the generating network and the identification network based on the distinguishing information; and

regenerating the defect data of the source domain, to which the first mask is reconstructed, from the defect data of the source domain, to which the first mask is applied, using the trained generating network.

7. The method for generating defect data of a target domain using defect data of a source domain according to claim 6 , wherein the outputting of distinguishing information by comparing a data, to which the mask is not applied, and a data, to which the mask is reconstructed, using the identification network comprises:

dividing the data, to which the mask is reconstructed and the data to which the mask is not applied, into image patches of a constant size, and comparing each of an image patches between each image patch.

8. The method for generating defect data of a target domain using defect data of a source domain according to claim 7 , wherein the image patch in the size of 1 pixel.

9. The method for generating defect data of a target domain using defect data of a source domain according to claim 6 , wherein the training the generating network and the identification network based on the distinguishing information comprises:

calculating a value for a loss function of the reconstruction algorithm from the distinguishing information.

10. The method for generating defect data of a target domain using defect data of a source domain according to claim 6 , wherein the generating network is composed of n layers, the first layer to the nth layer of the generating network are sequentially connected, an ith layer and a n-i+1th layer of the generating network are connected, and wherein i>0 and i<n/2.

11. A non-transitory computer readable medium storing a computer program, wherein when the computer program is executed by one or more processors of a computing device, the computer program performs an operation, and the operations comprise:

inputting a defect data of a source domain, to which a first mask is applied and a defect data of the source domain, to which the first mask is not applied, to a reconstruction algorithm;

first training the reconstruction algorithm to generate a defect data of the source domain, to which the first mask is reconstructed;

inputting a normal data of the source domain, to which a second mask is applied and a normal data of the source domain, to which the second mask is not applied, to the reconstruction algorithm;

second training the reconstruction algorithm to generate a normal data of the source domain, to which the second mask is reconstructed;

inputting a normal data of a target domain, to which the second mask is applied and a normal data of the target domain, to which the second mask is not applied, to the reconstruction algorithm;

third training the reconstruction algorithm to generate a normal data of the target domain, to which the second mask is reconstructed;

inputting a normal data of the target domain, to which the first mask is applied, to the reconstruction algorithm which is trained to generate a mask reconstructed data from a mask applied data; and

generating a defect data of the target domain using a final-trained reconstruction algorithm.

12. A computing device for generating defect data of a target domain using defect data of a source domain, including:

one or more processors; and

a memory storing program codes executable in the one or more processors, and

wherein the one or more processors are configured to:

input a defect data of the source domain, to which a first mask is applied and a defect data of the source domain, to which the first mask is not applied, to a reconstruction algorithm;

first train the reconstruction algorithm to generate a defect data of the source domain, to which the first mask is reconstructed;

input a normal data of the source domain, to which a second mask is applied and a normal data of the source domain, to which the second mask is not applied, to the reconstruction algorithm;

second train the reconstruction algorithm to generate a normal data of the source domain, to which the second mask is reconstructed;

input a normal data of the target domain, to which the second mask is applied and a normal data of the target domain, to which the second mask is not applied, to the reconstruction algorithm; and

third train the reconstruction algorithm to generate a normal data of the target domain, to which the second mask is reconstructed,

wherein the reconstruction algorithm comprises:

a generating network; and

an identification network, and

wherein the first training comprises:

inputting the defect data of the source domain, to which the first mask is applied, to the generating network;

generating the defect data of the source domain, to which the first mask is reconstructed, from the defect data of the source domain, to which the first mask is applied, using the generating network;

inputting the defect data of the source domain, to which the first mask is not applied and the defect data of the source domain, to which the first mask is reconstructed, to the identification network;

outputting distinguishing information by comparing the defect data of the source domain, to which the first mask is not applied, and the defect data of the source domain, to which the first mask is reconstructed, using the identification network;

training the generating network and the identification network based on the distinguishing information; and

regenerating the defect data of the source domain, to which the first mask is reconstructed, from the defect data of the source domain, to which the first mask is applied, using the trained generating network.