Granted Patent B2

US 12,632,951 · App. 17/764,707 · Granted May 19, 2026

Inventors: Yaoping Wang (Beijing, CN); Zhaoyue Li (Beijing, CN); Haodong Yang (Beijing, CN); Meijuan Zhang (Beijing, CN); Dong Chai (Beijing, CN); Hong Wang (Beijing, CN)

Assignees: Beijing Zhongxiangying Technology Co., Ltd.; BOE TECHNOLOGY GROUP CO., LTD.

G06T7/001G06T2207/20081G06T2207/20084

Abstract

Claims (78)

1 . A method for detecting a defect, comprising:

acquiring, by a terminal device, a sample data set comprising defective product data, and identifying feature information of the sample data set, the feature information comprising a number of samples in the sample data set;

transmitting, by the terminal device, the sample data set and the feature information to a server;

acquiring, by the server, an initial model, the initial model being a neural network model;

setting, by the server, a learning rate descent strategy, a total number of training rounds and a test strategy in a training parameter for the initial model based on the number of samples in the feature information;

obtaining, by the server, a target model by inputting the sample data set into the initial model and training, according to the training parameter, the initial model with the sample data set; and

obtaining, by the server, defect information of a product corresponding to the sample data set by inputting real data of the product into the target model,

wherein the learning rate descent strategy comprises a number of learning rate descents and a round number when a learning rate descends, and the test strategy comprises a number of tests and a round number at which a test is performed on a trained model,

wherein the server configures the total number of training rounds according to the number of samples under a preset rule, comprising:

configuring, in response to the number of samples being less than or equal to 10000, the total number of training rounds to 300000; and

configuring, in response to the number of samples being greater than 10000, the total number of training rounds with a formula of:

Y= 300000+INT( X/ 10000)× b

where Y represents the total number of training rounds, X represents the number of samples and is greater than or equal to 10000, INT is a rounding function, and b represents a growth factor having a fixed value and is greater than or equal to 30000 and less than or equal to 70000,

wherein the server configures the round number when the learning rate descends to be proportional to the total number of training rounds, and the round number at which the test is performed on the trained model to be greater than or equal to the round number when the learning rate descends for a first time and less than or equal to the total number of training rounds,

wherein the method further comprises:

generating, by the server, and displaying a training parameter display interface, wherein the training parameter display interface displays the learning rate descent strategy, the total number of training rounds and the test strategy set by the server.

2 . The method according to claim 1 , wherein the learning rate descends a plurality of times, and at least two tests are performed within a preset number of rounds around the round number when the learning rate descends for a second time.

3 . The method according to claim 2 , wherein the learning rate descends three times, and at least three tests are performed within the preset number of rounds around the round number when the learning rate descends for the second time.

4 . The method according to claim 1 , wherein the learning rate descent strategy comprises a learning rate descent manner and a learning rate descent magnitude.

5 . The method according to claim 1 , wherein the defective product data comprises a defective product image, and the feature information comprises a size and a type of the defective product image in the sample data set, and

wherein the method further comprises:

adjusting a size of an input image input into the initial model according to the size and the type of the defective product image.

6 . The method according to claim 5 , wherein adjusting the size of the input image input into the initial model according to the size and the type of the defective product image comprises:

adjusting the size of the input image to be a first preset multiple of the size of the defective product image in response to that the type of the defective product image indicates an AOI color image or a DM image; and

adjusting the size of the input image to be a second preset multiple of the size of the defective product image in response to that the type of the defective product image indicates a TDI image,

wherein the first preset multiple is less than or equal to 1, and the second preset multiple is greater than or equal to 1.

7 . The method according to claim 6 , wherein the input image comprises images with a plurality of sizes corresponding to a same defective product image.

8 . The method according to claim 5 , wherein the feature information further comprises a defect level of the defective product, and wherein the method further comprises:

configuring a confidence level in a training process according to defect levels corresponding to respective defects.

9 . The method according to claim 8 , wherein the defect level comprises a first defect level and a second defect level, and configuring the confidence level in the training process according to the defect levels corresponding to the respective defects comprises:

configuring the confidence level as a first confidence level in response to the defect level being the first defect level; and

configuring the confidence level as a second confidence level in response to the defect level being the second defect level,

wherein the second confidence level is greater than the first confidence level.

10 . The method according to claim 1 , further comprising acquiring the initial model according to a type of a defective product image.

11 . The method according to claim 1 ,

wherein the training parameter display interface comprises a parameter modification identifier; and

wherein the method further comprises:

updating the training parameter in response to a triggering operation of a user on the parameter modification identifier.

12 . The method according to claim 1 , further comprising:

acquiring a loss curve in a training process; and

updating the training parameter according to the loss curve.

13 . The method according to claim 1 , wherein obtaining, by the server, the target model by inputting the sample data set into the initial model and training, according to the training parameter, the initial model comprises:

acquiring a plurality of reference models according to the test strategy, and acquiring an accuracy rate and a recall rate of each of the reference models; and

determining the target model from the reference models according to the accuracy rate and the recall rate of each of the reference models.

14 . The method according to claim 1 , wherein obtaining, by the server, the target model by inputting the sample data set into the initial model and training, according to the training parameter, the initial model comprises:

acquiring a plurality of reference models according to the test strategy, and determining a confusion matrix of each of the reference models; and

determining the target model from the reference models according to the confusion matrix.

15 . The method according to claim 14 , further comprising: updating a confidence level according to the confusion matrix.

16 . A method for training a model, comprising:

acquiring, by a terminal device, a sample data set comprising defective product data, and identifying feature information of the sample data set, the feature information comprising a number of samples in the sample data set;

transmitting, by the terminal device, the sample data set and the feature information to a server;

acquiring, by the server, an initial model, the initial model being a neural network model;

setting, by the server, a learning rate descent strategy, a total number of training rounds and a test strategy in a training parameter for the initial model based on the number of samples in the feature information; and

obtaining, by the server, a target model by inputting the sample data set into the initial model and training, according to the training parameter, the initial model with the sample data set, the target model being configured to perform a defect detection on real data of a product corresponding to the sample data set,

wherein the learning rate descent strategy comprises a number of learning rate descents and a round number when a learning rate descends, and the test strategy comprises a number of tests and a round number at which a test is performed on a trained model,

wherein the server configures the total number of training rounds according to the number of samples under a preset rule, comprising:

configuring, in response to the number of samples being less than or equal to 10000, the total number of training rounds to 300000; and

configuring, in response to the number of samples being greater than 10000, the total number of training rounds with a formula of:

Y= 300000+INT( X/ 10000)× b

where Y represents the total number of training rounds, X represents the number of samples and is greater than or equal to 10000, INT is a rounding function, and b represents a growth factor having a fixed value and is greater than or equal to 30000 and less than or equal to 70000,

wherein the server configures the round number when the learning rate descends to be proportional to the total number of training rounds, and the round number at which the test is performed on the trained model to be greater than or equal to the round number when the learning rate descends for a first time and less than or equal to the total number of training rounds,

wherein the method further comprises:

generating, by the server, and displaying a training parameter display interface, wherein the training parameter display interface displays the learning rate descent strategy, the total number of training rounds and the test strategy set by the server.

17 . A method for training a model, comprising:

acquiring, by a terminal device, a sample data set comprising defective product data in response to a configuration operation of a user on a parameter of the sample data set, and identifying feature information of the sample data set, the feature information comprising a number of samples in the sample data set;

transmitting, by the terminal device, the sample data set and the feature information to a server;

acquiring, by the server, an initial model, the initial model being a neural network model;

setting, by the server, a learning rate descent strategy, a total number of training rounds and a test strategy in a training parameter for the initial model based on the number of samples in the feature information; and

obtaining, by the server, a target model by inputting the sample data set into the initial model and training, according to the training parameter, the initial model with the sample data set, the target model being configured to perform a defect detection on real data of a product corresponding to the sample data set,

wherein the learning rate descent strategy comprises a number of learning rate descents and a round number when a learning rate descends, and the test strategy comprises a number of tests and a round number at which a test is performed on a trained model,

wherein the server configures the total number of training rounds according to the number of samples under a preset rule, comprising:

configuring, in response to the number of samples being less than or equal to 10000, the total number of training rounds to 300000; and

configuring, in response to the number of samples being greater than 10000, the total number of training rounds with a formula of:

Y= 300000+INT( X/ 10000)× b

where Y represents the total number of training rounds, X represents the number of samples and is greater than or equal to 10000, INT is a rounding function, and b represents a growth factor having a fixed value and is greater than or equal to 30000 and less than or equal to 70000,

wherein the server configures the round number when the learning rate descends to be proportional to the total number of training rounds, and the round number at which the test is performed on the trained model to be greater than or equal to the round number when the learning rate descends for a first time and less than or equal to the total number of training rounds,

wherein the method further comprises:

generating, by the server, and displaying a training parameter display interface, wherein the training parameter display interface displays the learning rate descent strategy, the total number of training rounds and the test strategy set by the server.

Assignments (1)

Continuity (1)

References Cited (37)