CN115495608A - Defect detection method and system based on model - Google Patents
Defect detection method and system based on model Download PDFInfo
- Publication number
- CN115495608A CN115495608A CN202211432680.5A CN202211432680A CN115495608A CN 115495608 A CN115495608 A CN 115495608A CN 202211432680 A CN202211432680 A CN 202211432680A CN 115495608 A CN115495608 A CN 115495608A
- Authority
- CN
- China
- Prior art keywords
- model
- labeling
- data set
- current
- marking
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/5866—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using information manually generated, e.g. tags, keywords, comments, manually generated location and time information
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/10—File systems; File servers
- G06F16/18—File system types
- G06F16/182—Distributed file systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Library & Information Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a defect detection method and a system based on a model, which comprises the following steps: step S1: acquiring defect data of a historical product, and establishing a defect detection model; step S2: training a defect detection model through a defect sample product; and step S3: and carrying out defect detection on the product to be detected according to the trained defect detection model. According to the invention, through the defect detection model, the labeling speed of the industrial quality inspection picture data set sample is increased, and the calculation of the comparison results of the test data sets under different confidence coefficients is automatically realized, so that an implementation engineer can quickly and comprehensively judge the defect condition of a product, and the defect detection rate is improved.
Description
Technical Field
The invention relates to the technical field of defect detection, in particular to a defect detection method and system based on a model. The method can be applied to the quality inspection link in the manufacturing process of high-end equipment and new energy vehicles.
Background
With the continuous increase of the quality inspection requirements of industrial fields, massive data reasoning requests can be generated, and reasoning tasks are deployed on a single computing node and cannot meet the requirements, so that reasoning services need to be deployed on a plurality of computing nodes to realize distributed reasoning. Since the inference calculation nodes are distributed on a plurality of servers, and a plurality of nodes infer part of data respectively, the problem of data deployment and loading of the plurality of nodes needs to be considered.
Patent document CN112734742A (application number: CN 202110071345.6) discloses a method and system for improving the accuracy of industrial quality inspection, comprising the steps of: training to obtain a detection model; step B: and reasoning by using the detection model.
The existing model defect detection mode comprises data marking and marking comparison, wherein the data marking comprises manual marking and model marking, namely, a part of representative data is selected as a test set, the data is marked manually, and then the model is used for reasoning to generate marking information of the model. And the marking comparison is to compare the manual marking information with the model marking information. The defect detection of the model is carried out once, so that the period is long, and a large amount of manpower is consumed.
Therefore, a new technical solution is needed to improve the above technical problems.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a defect detection method and system based on a model.
The defect detection method based on the model provided by the invention comprises the following steps:
step S1: acquiring defect data of a historical product, and establishing a defect detection model;
step S2: training a defect detection model through a defect sample product;
and step S3: and performing defect detection on the product to be detected according to the trained defect detection model.
Preferably, the step S1 includes: carrying out distributed reasoning on the defective product, including parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition, specifically comprising the following steps:
setting an inference model, batch processing size and picture resolution, covering the value of a default parameter file, generating a complete parameter file, storing the parameter file on a distributed file storage S3 server, writing the address of the parameter file into an environment variable of a container, acquiring the storage address of the parameter file in the distributed storage system from the environment variable after the container is started, reading the file content in a code through the file address, and analyzing and converting the file content into all parameters required by inference;
mounting a distributed storage directory for storing a test data set on each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in the path corresponding to the distributed server;
acquiring storage paths of the model and the test data set from all analyzed parameters, loading the simplex model from the storage paths of the model, converting the simplex model into a distributed model, loading the data set from the storage paths of the test data set, converting the data set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
setting a storage path of the inference result of the test data set in the parameter file, storing the inference result of each picture in the set path of the distributed server in the inference process, loading the inference result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the inference marking information of each picture on a browser page.
Preferably, the step S1 further includes: and performing auxiliary marking on the defective product, wherein the auxiliary marking comprises reasoning of the test data set, loading a reasoning result as an auxiliary marking result, and checking the marking result, and the method specifically comprises the following steps:
carrying out model first training by using a training data set labeled manually, storing the model trained for the first time in a distributed storage server, setting a model storage path into a reasoning parameter file, reasoning the expanded training data set by using the model, and storing a reasoning result according to a reasoning parameter;
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing next training.
Preferably, the step S2 includes: carrying out defect model test, including carrying out manual labeling and model reasoning labeling on a test data set, then comparing the manual labeling and model labeling under different confidence coefficients of each picture in the test data set, summarizing comparison data, and specifically:
manually marking pictures of the test data set one by one on a browser page, and selecting the same test data set again to use the model for distributed inference marking;
performing traversal circulation on all pictures of the current data set, respectively acquiring artificial labeling information and model labeling information for each picture in a test set, and the confidence of the current test, and removing model labeling information which does not meet the preset requirement according to the confidence;
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect names, and collecting all artificial labeling information lists and model labeling information lists of different defect names;
and circularly comparing the manual marking information and the model marking information of the grouped defects, wherein the grouped defects include the following conditions:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, judging that the current picture has no current type defects, and judging that the killing number and the loss number are both equal to 0;
b: the number of the artificial marking frames of the current defects is 0, but the number of the model marking frames is not 0, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model marking frames, and the leakage number is equal to 0;
c: the number of the artificial marking frames of the current defects is not 0 but the number of the model marking frames is 0, at the moment, the model has a leakage phenomenon on the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames with the current defects and the number of the model labeling frames are not 0, firstly, missing frame labels are marked on all the artificial labeling frames with the current defects of the current picture, then, killing frame labels are marked on all the model labeling frames with the current defects of the current picture, then, the artificial labeling frame list is traversed in a circulating mode, the area of the current artificial labeling frame is calculated, meanwhile, the model labeling frame list is traversed in a circulating mode, the area of the model labeling frame is calculated, whether the overlapping degree IOU of the current artificial labeling frame and the model labeling frame is larger than 0 or not is judged, if the overlapping degree IOU is larger than 0, the missing labels of the current artificial labeling frame are removed, the killing labels of the current model labeling frame are removed, after circulation is completed, the counting result shows that the number of the killing labels is equal to the number of the model labeling frames with the killing labels, and the number of the missing labels is equal to the number of the artificial labeling frames with the missing labels;
and circularly comparing all defects under the current confidence coefficient, acquiring the missing labeling quantity and the over-killing labeling quantity of all types of defects of the current data set under the current confidence coefficient, acquiring the missing labeling quantity and the over-killing labeling quantity of all the defects of the current data set under all the set confidence coefficients, and summarizing comparison data.
Preferably, the model test further comprises a picture classification test, a statistical result object initialization, a labeling information acquisition, a result object confusion matrix updating and a statistical value calculation;
the initializing statistical result object includes: initializing a confusion matrix, a label mapping relation and the number of picture categories for storing the statistical result;
the acquiring of the labeling information includes: loading a data set to be tested, performing cyclic traversal on the data set, and respectively acquiring artificial labeling information and model labeling information of a current picture aiming at each picture in the data set;
updating the confusion matrix in the result object comprises: respectively marking corresponding values of the manual marking label and the model marking label in the label mapping relation, forming an (x, y) coordinate according to the manual marking mapping value and the model marking mapping value, namely (the manual marking mapping value and the model marking mapping value), and updating the coordinate into a confusion matrix;
the calculating the statistical value comprises: and counting the correct prediction proportion of the test set according to the accuracy rate, the recall rate and the precision rate.
The model-based defect detection system provided by the invention comprises:
a module M1: acquiring defect data of historical products, and establishing a defect detection model;
a module M2: training a defect detection model through a defect sample product;
a module M3: and performing defect detection on the product to be detected according to the trained defect detection model.
Preferably, the module M1 comprises: performing distributed reasoning on defective products, including parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition, specifically:
setting an inference model, a batch processing size and a picture resolution, covering the value of a default parameter file to generate a complete parameter file, storing the parameter file to a distributed file storage S3 server, writing the address of the parameter file into an environment variable of a container, acquiring the storage address of the parameter file in the distributed storage system from the environment variable after the container is started, reading the file content in a code through the file address, and analyzing and converting the file content into all parameters required by inference;
mounting a distributed storage directory for storing a test data set on each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in the path corresponding to the distributed server;
acquiring storage paths of the model and the test data set from all analyzed parameters, loading the single machine model from the storage paths of the model, converting the single machine model into a distributed model, loading the data set from the storage paths of the test data set, converting the data set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
setting a storage path of the inference result of the test data set in the parameter file, storing the inference result of each picture in the set path of the distributed server in the inference process, loading the inference result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the inference marking information of each picture on a browser page.
Preferably, the module M1 further comprises: and performing auxiliary marking on the defective product, wherein the auxiliary marking comprises reasoning of the test data set, loading a reasoning result as an auxiliary marking result, and checking the marking result, and the method specifically comprises the following steps:
carrying out model first training by using a training data set labeled manually, storing the model trained for the first time in a distributed storage server, setting a model storage path into a reasoning parameter file, reasoning the expanded training data set by using the model, and storing a reasoning result according to a reasoning parameter;
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing the next training.
Preferably, the module M2 includes: carrying out defect model test, including carrying out manual labeling and model reasoning labeling on a test data set, then comparing the manual labeling and model labeling under different confidence coefficients of each picture in the test data set, summarizing comparison data, and specifically:
manually labeling the pictures of the test data set one by one on a browser page, and selecting the same test data set again to perform distributed inference labeling by using the model;
performing traversal circulation on all pictures of the current data set, respectively acquiring artificial labeling information and model labeling information for each picture in a test set, and the confidence of the current test, and removing model labeling information which does not meet the preset requirement according to the confidence;
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect names, and collecting all artificial labeling information lists and model labeling information lists of different defect names;
and circularly comparing the grouped artificial marking information and the model marking information of the defects, wherein the grouped defects comprise the following conditions:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, judging that the current picture has no current type defects, and judging that the killing number and the loss number are both equal to 0;
b: the number of the artificial marking frames of the current defects is 0, but the number of the model marking frames is not 0, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model marking frames, and the leakage number is equal to 0;
c: the number of the artificial marking frames of the current defects is not 0 but the number of the model marking frames is 0, at the moment, the model has a leakage phenomenon on the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames with the current defects and the number of the model labeling frames are not 0, firstly, missing frame labels are marked on all the artificial labeling frames with the current defects of the current picture, then, killing frame labels are marked on all the model labeling frames with the current defects of the current picture, then, the artificial labeling frame list is traversed in a circulating mode, the area of the current artificial labeling frame is calculated, meanwhile, the model labeling frame list is traversed in a circulating mode, the area of the model labeling frame is calculated, whether the overlapping degree IOU of the current artificial labeling frame and the model labeling frame is larger than 0 or not is judged, if the overlapping degree IOU is larger than 0, the missing labels of the current artificial labeling frame are removed, the killing labels of the current model labeling frame are removed, after circulation is completed, the counting result shows that the number of the killing labels is equal to the number of the model labeling frames with the killing labels, and the number of the missing labels is equal to the number of the artificial labeling frames with the missing labels;
and circularly comparing all defects under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all kinds of defects of the current data set under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all defects of the current data set under all set confidence coefficients, and summarizing comparison data.
Preferably, the model test further comprises a picture classification test, a statistical result object initialization, a labeling information acquisition, a result object confusion matrix updating and a statistical value calculation;
the initializing statistical result object includes: initializing a confusion matrix, a label mapping relation and the number of picture categories for storing a statistical result;
the acquiring of the labeling information includes: loading a data set to be tested, performing cyclic traversal on the data set, and respectively acquiring artificial labeling information and model labeling information of a current picture aiming at each picture in the data set;
updating the confusion matrix in the result object comprises: respectively labeling corresponding values of the manual labeling label and the model labeling label in the label mapping relation, forming an (x, y) coordinate according to the manual labeling mapping value and the model labeling mapping value, namely (the manual labeling mapping value and the model labeling mapping value), and updating the coordinate into a confusion matrix;
the calculating the statistical value comprises: and counting the correct prediction proportion of the test set according to the accuracy rate, the recall rate and the precision rate.
Compared with the prior art, the invention has the following beneficial effects: according to the invention, through the defect detection model, the labeling speed of the industrial quality inspection picture data set sample is increased, the labeling process is simplified for an industrial field labeling engineer, and the calculation of the comparison results of the test data sets under different confidence coefficients is automatically realized, so that the implementing engineer can quickly and comprehensively judge the defect condition of the product, and the defect detection rate is increased.
Drawings
Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:
FIG. 1 is a flowchart of a defect detection method according to the present invention;
FIG. 2 is a flow chart of a distributed reasoning, model-assisted labeling and model testing method in the field of industrial quality inspection.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that variations and modifications can be made by persons skilled in the art without departing from the concept of the invention. All falling within the scope of the present invention.
Example 1:
as shown in fig. 1, the present invention provides a defect detection method based on a model, comprising: step S1: acquiring defect data of historical products, and establishing a defect detection model; step S2: training a defect detection model through a defect sample product; and step S3: and performing defect detection on the product to be detected according to the trained defect detection model.
The step S1 includes: performing distributed reasoning on defective products, including parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition, specifically:
setting an inference model, a batch processing size and a picture resolution, covering the value of a default parameter file to generate a complete parameter file, storing the parameter file to a distributed file storage S3 server, writing the address of the parameter file into an environment variable of a container, acquiring the storage address of the parameter file in the distributed storage system from the environment variable after the container is started, reading the file content in a code through the file address, and analyzing and converting the file content into all parameters required by inference;
mounting a distributed storage directory for storing a test data set to each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in a path corresponding to the distributed server;
acquiring storage paths of the model and the test data set from all analyzed parameters, loading the single machine model from the storage paths of the model, converting the single machine model into a distributed model, loading the data set from the storage paths of the test data set, converting the data set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
setting a storage path of the inference result of the test data set in the parameter file, storing the inference result of each picture in the set path of the distributed server in the inference process, loading the inference result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the inference marking information of each picture on a browser page.
The step S1 further includes: and performing auxiliary marking on the defective product, wherein the auxiliary marking comprises reasoning of the test data set, loading a reasoning result as an auxiliary marking result, and checking the marking result, and the method specifically comprises the following steps:
carrying out model first training by using a manually labeled training data set, storing the model which is trained for the first time in a distributed storage server, setting a model storage path into a reasoning parameter file, reasoning the expanded training data set by using the model, and storing a reasoning result according to a reasoning parameter;
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing next training.
The step S2 includes: carrying out defect model test, including carrying out manual labeling and model reasoning labeling on a test data set, then comparing the manual labeling and the model labeling under different confidence degrees of each picture in the test data set, summarizing comparison data, and specifically:
manually marking pictures of the test data set one by one on a browser page, and selecting the same test data set again to use the model for distributed inference marking;
performing traversal circulation on all pictures of the current data set, respectively acquiring artificial labeling information and model labeling information for each picture in a test set, and the confidence of the current test, and removing model labeling information which does not meet the preset requirement according to the confidence;
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect names, and collecting all artificial labeling information lists and model labeling information lists of different defect names;
and circularly comparing the grouped artificial marking information and the model marking information of the defects, wherein the grouped defects comprise the following conditions:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, judging that the current picture has no current type defects, and judging that the killing number and the loss number are both equal to 0;
b: the number of the artificial marking frames of the current defects is 0, but the number of the model marking frames is not 0, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model marking frames, and the leakage number is equal to 0;
c: the number of the artificial marking frames with the current defects is not 0 but the number of the model marking frames is 0, at the moment, the model has a leakage phenomenon on the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames with the current defects and the number of the model labeling frames are not 0, firstly, missing frame labels are marked on all the artificial labeling frames with the current defects of the current picture, then, killing frame labels are marked on all the model labeling frames with the current defects of the current picture, then, the artificial labeling frame list is traversed in a circulating mode, the area of the current artificial labeling frame is calculated, meanwhile, the model labeling frame list is traversed in a circulating mode, the area of the model labeling frame is calculated, whether the overlapping degree IOU of the current artificial labeling frame and the model labeling frame is larger than 0 or not is judged, if the overlapping degree IOU is larger than 0, the missing labels of the current artificial labeling frame are removed, the killing labels of the current model labeling frame are removed, after circulation is completed, the counting result shows that the number of the killing labels is equal to the number of the model labeling frames with the killing labels, and the number of the missing labels is equal to the number of the artificial labeling frames with the missing labels;
and circularly comparing all defects under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all kinds of defects of the current data set under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all defects of the current data set under all set confidence coefficients, and summarizing comparison data.
The model test also comprises the steps of carrying out classification test on the pictures, initializing a statistical result object, acquiring marking information, updating a confusion matrix in the result object and calculating a statistical value;
the initializing statistical result object includes: initializing a confusion matrix, a label mapping relation and the number of picture categories for storing a statistical result;
the acquiring of the labeling information includes: loading a data set to be tested, performing cyclic traversal on the data set, and respectively acquiring artificial labeling information and model labeling information of a current picture aiming at each picture in the data set;
updating the confusion matrix in the result object comprises: respectively marking corresponding values of the manual marking label and the model marking label in the label mapping relation, forming an (x, y) coordinate according to the manual marking mapping value and the model marking mapping value, namely (the manual marking mapping value and the model marking mapping value), and updating the coordinate into a confusion matrix;
the calculating the statistical value comprises: and counting the proportion of correct prediction of the test set according to the accuracy rate, the recall rate and the precision rate.
Example 2:
example 2 is a preferred example of example 1.
The invention provides a distributed reasoning, model auxiliary labeling and model testing method in the field of industrial quality inspection, wherein the distributed reasoning comprises parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition; the auxiliary marking comprises reasoning of a test data set, loading a storage format of a JSON file of a reasoning result as an auxiliary marking result, and correcting the marking result; the model test comprises the steps of manually marking a test data set, carrying out model reasoning marking on the test data set, then comparing the manual marking and the model marking under different confidence degrees of each picture in the test data set, and summarizing comparison data.
The distributed reasoning comprises the following steps:
step A: the acquisition of the distributed inference parameters comprises the generation, transmission and acquisition of the distributed parameters, and specifically comprises the following steps:
an engineer is implemented to set adjustable parameters such as a model of inference, batch processing size batch _ size, picture resolution and the like, the parameters are used for covering the value of a default parameter file to generate a complete parameter file, the parameter file is stored in a distributed file storage S3 server, the address of the parameter file is written into an environment variable of a container, after the container is started, the storage address of the parameter file in the distributed storage system is obtained from the environment variable, the file content is read from a code through the file address, and the file content is analyzed and converted into all parameters required by inference;
and B, step B: the distributed loading of the test data set specifically comprises the following steps:
firstly, mounting a distributed storage directory for storing a test data set on each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in the path corresponding to the distributed server.
And C: the distributed reasoning specifically comprises the following steps:
acquiring storage paths of a model and a test data set from all analyzed parameters, loading the model from the model storage path as a single machine model, converting the single machine model into a distributed model by using a method for training a frame of a current model, loading data from the storage path of the test data set as a data set data _ set, converting the data set data _ set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
the method for training the framework of the current model is to find the framework corresponding to the current model from the parameters, such as Pytroch and tensrflow, and convert the single model into the distributed model by using the method corresponding to the framework.
The reasoning process is as follows: firstly, loading a trained model into a video memory of a computer display card, then loading a picture, loading the picture into a binary file stream and transmitting the binary file stream to a program, calling the display card by the program for calculation to obtain a calculation result machine probability value, wherein the calculation result can correspond to one type of labeling information, for example: the result value is 1, the probability value is 90%, and the dog is in a state of no error; the result value is a value of 2, a probability value of 80%, cat.
Step D: acquiring an inference result specifically comprises the following steps:
setting a storage path of a test data set reasoning result json file storage format in the parameter file, storing the reasoning result of each picture in the set path of the distributed server in the reasoning process, loading the reasoning result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the reasoning marking information of each picture on a WEB page of the browser.
The auxiliary labeling comprises the following steps:
step A: reasoning of the training data set specifically includes:
the method comprises the steps of training a model for the first time by using a training data set labeled manually, storing the model trained for the first time in a distributed storage server, setting a model storage path in an inference parameter file, reasoning the extended training data set by using the model, and storing an inference result according to inference parameters.
The model is a weight file obtained after artificial intelligence training, the input is a file stream of pictures read by a program, and the output is the labeled information and the probability value of the labeled information after the pictures are taken in the current picture.
And B: loading a storage format of a JSON file of the inference result as an auxiliary annotation result, and checking the annotation result, wherein the method specifically comprises the following steps:
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing the next training.
The model test comprises the following steps:
step A: carrying out manual labeling and model labeling on the test data set, which specifically comprises the following steps:
and selecting a set test data set, manually marking pictures of the test data set one by one on a WEB page of the browser, and then selecting the same test data set again to use the model to carry out distributed inference marking.
The selection standard of the test data set is that the test data set can cover the label information to be detected by the model, and the number of the labels can be uniformly distributed, wherein the proportion of a general test set, a verification set and a training set is 7.
The manual labeling is to label features in the picture, for example, to train a model capable of identifying the dog, the dog in the picture is framed by a closed graph.
And B: traversing the current test data set specifically comprises:
traversing and circulating all pictures of the current data set, and respectively acquiring manual labeling information, model labeling information and confidence of the current test for each picture in the test set;
the confidence is to filter the model label which does not conform to the current probability value, and it may happen that although an object is labeled on the picture by the model, the probability value of the label is too small, so that the confidence of the current test is compared with the probability value of the model label, labels which are greater than the confidence are retained, and labels which are smaller than the confidence are ignored.
And C: grouping the defect information of the pictures, specifically:
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect name, and collecting all artificial labeling information lists and model labeling information lists of the defects with different names;
step D: the manual marking information and the model marking information of the grouped defects are circularly compared, and four conditions can occur after single defects are grouped, specifically:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, the current picture is considered to have no current type defects, and the result is that the killing number and the loss number are both equal to 0;
b: if the number of the manual labeling frames with the current defects is 0 but the number of the model labeling frames is not 0, then the labeling personnel think that the current picture has no current type defects, but the model reasoning thinks that the current picture has the current type defects, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model labeling frames, and the leakage number is equal to 0;
c: if the number of the artificial marking frames with the current defects is not 0 but the number of the model marking frames is 0, then the marking personnel think that the current picture has the current type defects, but the model reasoning thinks that the current picture does not have the current type defects, at the moment, the model has a leakage phenomenon for the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames of the current defects and the number of the model labeling frames are not 0, firstly, all the artificial labeling frames of the current defects of the current picture are marked with missing frame labels, then all the model labeling frames of the current defects of the current picture are marked with killing frame labels, then, a first layer is circulated and traversed through an artificial labeling frame list, the area of the current artificial labeling frames is calculated, a second layer is circulated and traversed through a model labeling frame list, the area of the model labeling frames is calculated, whether the IOU (input output Unit) of the current artificial labeling frames and the IOU of the model labeling frames is larger than 0 is judged, if the IOU is larger than 0, the missing labels of the current artificial labeling frames are removed, the killing labels of the current model labeling frames are removed, after two layers of circulation are completed, the statistical result is that: the killing number is equal to the number of model marking frames with killing labels, and the missing number is equal to the number of artificial marking frames with missing labels;
step E: comparing different defects under different confidence coefficients, and summarizing comparison data, specifically:
and D, performing step D on all defects under the current confidence coefficient, acquiring the missing labeling quantity and the over-killing labeling quantity of all types of defects of the current data set under the current confidence coefficient, and performing step B, step C and step D on the current data set under different confidence coefficients. And acquiring the number of all defective missed labels and the number of overdischarged labels of the current data set under all set confidence degrees, and summarizing and comparing data.
The invention provides a model test system in the field of industrial quality inspection, which comprises a test for target detection and a test for picture classification, wherein the test data sets of the target detection and the picture classification are respectively subjected to manual labeling, model reasoning labeling is carried out, then manual labeling and model labeling under different confidence coefficients of each picture in the test data sets are compared, and comparison data are summarized.
The model test system comprises the steps of testing image classification, initializing a statistical result object, acquiring marking information, updating a confusion matrix in the result object and calculating a statistical value.
The expression form of the confusion matrix is an n-x-n matrix, wherein n is the number of the categories of the objects to be identified by the current model, a zero matrix is initially arranged, and the final confusion matrix can be generated after the current model is completely tested under the current confidence coefficient; calculating the statistical value means that after the confusion matrix is generated, the accuracy, the recall rate and the accuracy rate under the current confidence coefficient are settled according to the confusion matrix.
The method specifically comprises the following steps:
step A: initializing a statistical result object, specifically:
initializing a confusion matrix, a label mapping relation and the number of picture categories for storing the statistical result;
and B: acquiring the labeling information, specifically:
loading a data set to be tested, performing cycle traversal on the data set, and respectively acquiring manual labeling information and model labeling information of a current picture aiming at each picture in the data set;
and C: updating the confusion matrix in the result object, specifically:
and respectively marking the corresponding values of the label and the model label in the label mapping relation, forming an (x, y) coordinate according to the manual marking mapping value and the model marking mapping value, namely (the manual marking mapping value and the model marking mapping value), updating the coordinate into the confusion matrix, and updating by increasing the value a by 1, namely a +1, covering the value a of the corresponding position of the original confusion matrix by using the value a +1 to complete the updating.
Step D: calculating a statistical value, specifically:
the Accuracy is a statistical index for a test set, and refers to the proportion of a correct sample in a total sample, namely the proportion of a true class, a true negative class and an overall prediction sample found by a model, and is expressed by a formula as follows: accuracy = (TP + TN)/(TP + TN + FP + FN), the range of Accuracy is [0,1], generally, the larger the value is, the better the model prediction capability is represented, wherein:
the TP (True) sample is positive, and the prediction result is positive;
FP (False positive) sample is negative, and the prediction result is positive;
the TN (True negative) sample is negative, and the prediction result is negative;
FN (False Negatives) samples are positive, and the prediction result is negative.
Recall is an indicator for the original sample under different categories that indicates how many positive examples in the original sample are predicted correctly. The positive examples in the original sample have two cases, one is to predict the original positive class as positive class (TP) and the other is to predict the original positive class as negative class (FN), which constitute all the positive examples of the original sample. The calculation formula is as follows: recall = TP/(TP + FN).
Precision is an index for the prediction result, which indicates how many pairs are in the samples predicted as positive classes. The prediction result is positive, and there are two cases, namely, the positive class is predicted as the positive class (TP), and the negative class is predicted as the positive class (FP). Therefore, the calculation formula of the accuracy rate is as follows: precision = TP/(TP + FP).
As shown in fig. 2, a task is started first, parameters are acquired through a web page of a browser, a parameter file is generated and stored, a parameter file and a data set in a distributed storage server are loaded in a second step, a test data set is loaded in a distributed manner in a third step, inference calculation nodes are distributed in a fourth step and distributed inference is performed, an inference result is acquired in a fifth step and stored in a distributed file system, manual labeling and model inference results are compared in a sixth step, results under different confidence degrees are collected in a seventh step, and model quality is judged comprehensively in an eighth step.
Those skilled in the art will appreciate that, in addition to implementing the systems, apparatus, and various modules thereof provided by the present invention in purely computer readable program code, the same procedures can be implemented entirely by logically programming method steps such that the systems, apparatus, and various modules thereof are provided in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system, the device and the modules thereof provided by the present invention can be considered as a hardware component, and the modules included in the system, the device and the modules thereof for implementing various programs can also be considered as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.
Claims (8)
1. A method for model-based defect detection, comprising:
step S1: acquiring defect data of a historical product, and establishing a defect detection model;
step S2: training a defect detection model through a defect sample product;
and step S3: performing defect detection on a product to be detected according to the trained defect detection model;
the step S1 includes: carrying out distributed reasoning on the defective product, including parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition, specifically comprising the following steps:
setting an inference model, a batch processing size and a picture resolution, covering the value of a default parameter file to generate a complete parameter file, storing the parameter file to a distributed file storage S3 server, writing the address of the parameter file into an environment variable of a container, acquiring the storage address of the parameter file in the distributed storage system from the environment variable after the container is started, reading the file content in a code through the file address, and analyzing and converting the file content into all parameters required by inference;
mounting a distributed storage directory for storing a test data set on each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in the path corresponding to the distributed server;
acquiring storage paths of the model and the test data set from all analyzed parameters, loading the single machine model from the storage paths of the model, converting the single machine model into a distributed model, loading the data set from the storage paths of the test data set, converting the data set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
setting a storage path of the inference result of the test data set in the parameter file, storing the inference result of each picture in the set path of the distributed server in the inference process, loading the inference result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the inference marking information of each picture on a browser page.
2. The model-based defect detection method of claim 1, wherein the step S1 further comprises: and performing auxiliary labeling on the defective product, wherein the auxiliary labeling comprises reasoning of a test data set, loading a reasoning result as an auxiliary labeling result, and checking the labeling result, and specifically comprises the following steps:
carrying out model first training by using a training data set labeled manually, storing the model trained for the first time in a distributed storage server, setting a model storage path into a reasoning parameter file, reasoning the expanded training data set by using the model, and storing a reasoning result according to a reasoning parameter;
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing the next training.
3. The model-based defect detection method of claim 2, wherein the step S2 comprises: carrying out defect model test, including carrying out manual labeling and model reasoning labeling on a test data set, then comparing the manual labeling and the model labeling under different confidence degrees of each picture in the test data set, summarizing comparison data, and specifically:
manually marking pictures of the test data set one by one on a browser page, and selecting the same test data set again to use the model for distributed inference marking;
performing traversal circulation on all pictures of the current data set, respectively acquiring artificial labeling information and model labeling information for each picture in a test set, and the confidence of the current test, and removing model labeling information which does not meet the preset requirement according to the confidence;
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect names, and collecting all artificial labeling information lists and model labeling information lists of different defect names;
and circularly comparing the grouped artificial marking information and the model marking information of the defects, wherein the grouped defects comprise the following conditions:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, judging that the current picture has no current type defects, and judging that the killing number and the loss number are both equal to 0;
b: the number of the artificial marking frames of the current defects is 0, but the number of the model marking frames is not 0, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model marking frames, and the leakage number is equal to 0;
c: the number of the artificial marking frames of the current defects is not 0 but the number of the model marking frames is 0, at the moment, the model has a leakage phenomenon on the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames with the current defects and the number of the model labeling frames are not 0, firstly, missing frame labels are marked on all the artificial labeling frames with the current defects of the current picture, then, killing frame labels are marked on all the model labeling frames with the current defects of the current picture, then, the artificial labeling frame list is traversed in a circulating mode, the area of the current artificial labeling frame is calculated, meanwhile, the model labeling frame list is traversed in a circulating mode, the area of the model labeling frame is calculated, whether the overlapping degree IOU of the current artificial labeling frame and the model labeling frame is larger than 0 or not is judged, if the overlapping degree IOU is larger than 0, the missing labels of the current artificial labeling frame are removed, the killing labels of the current model labeling frame are removed, after circulation is completed, the counting result shows that the number of the killing labels is equal to the number of the model labeling frames with the killing labels, and the number of the missing labels is equal to the number of the artificial labeling frames with the missing labels;
and circularly comparing all defects under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all kinds of defects of the current data set under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all defects of the current data set under all set confidence coefficients, and summarizing comparison data.
4. The model-based defect detection method of claim 3, wherein the model test further comprises classifying the pictures for testing, initializing the statistical result object, obtaining the labeling information, updating the confusion matrix in the result object, and calculating the statistical value;
the initializing statistical result object includes: initializing a confusion matrix, a label mapping relation and the number of picture categories for storing the statistical result;
the acquiring of the labeling information includes: loading a data set to be tested, performing cycle traversal on the data set, and respectively acquiring manual labeling information and model labeling information of a current picture aiming at each picture in the data set;
updating the confusion matrix in the result object comprises: respectively marking corresponding values of the manual marking label and the model marking label in the label mapping relation, forming an (x, y) coordinate according to the manual marking mapping value and the model marking mapping value, namely (the manual marking mapping value and the model marking mapping value), and updating the coordinate into a confusion matrix;
the calculating the statistical value comprises: and counting the proportion of correct prediction of the test set according to the accuracy rate, the recall rate and the precision rate.
5. A model-based defect detection system, comprising:
a module M1: acquiring defect data of historical products, and establishing a defect detection model;
a module M2: training a defect detection model through a defect sample product;
a module M3: performing defect detection on a product to be detected according to the trained defect detection model;
the module M1 comprises: carrying out distributed reasoning on the defective product, including parameter acquisition, distributed loading of a test data set, distributed reasoning and reasoning result acquisition, specifically comprising the following steps:
setting an inference model, a batch processing size and a picture resolution, covering the value of a default parameter file to generate a complete parameter file, storing the parameter file to a distributed file storage S3 server, writing the address of the parameter file into an environment variable of a container, acquiring the storage address of the parameter file in the distributed storage system from the environment variable after the container is started, reading the file content in a code through the file address, and analyzing and converting the file content into all parameters required by inference;
mounting a distributed storage directory for storing a test data set to each server where a container is located, loading and analyzing a parameter file after the container is started, acquiring a storage path of the test data set from all analyzed parameters, and loading pictures in a path corresponding to the distributed server;
acquiring storage paths of the model and the test data set from all analyzed parameters, loading the single machine model from the storage paths of the model, converting the single machine model into a distributed model, loading the data set from the storage paths of the test data set, converting the data set into a distributed test set by using a distributed sampler, and performing distributed reasoning after acquiring the distributed model and the distributed test set;
setting a storage path of the inference result of the test data set in the parameter file, storing the inference result of each picture in the set path of the distributed server in the inference process, loading the inference result of the test data set stored in the distributed server in a back-end service, providing a query interface, and displaying the inference marking information of each picture on a browser page.
6. The model-based defect detection system of claim 5, wherein the module M1 further comprises: and performing auxiliary marking on the defective product, wherein the auxiliary marking comprises reasoning of the test data set, loading a reasoning result as an auxiliary marking result, and checking the marking result, and the method specifically comprises the following steps:
carrying out model first training by using a training data set labeled manually, storing the model trained for the first time in a distributed storage server, setting a model storage path into a reasoning parameter file, reasoning the expanded training data set by using the model, and storing a reasoning result according to a reasoning parameter;
and loading each picture of the training data set on a browser page, taking the inference result as an auxiliary label, manually correcting each defect of the auxiliary label, converting the inferred auxiliary label result into a manual label result after correction, generating a new training data set, and performing next training.
7. The model-based defect detection system of claim 6, wherein the module M2 comprises: carrying out defect model test, including carrying out manual labeling and model reasoning labeling on a test data set, then comparing the manual labeling and the model labeling under different confidence degrees of each picture in the test data set, summarizing comparison data, and specifically:
manually labeling the pictures of the test data set one by one on a browser page, and selecting the same test data set again to perform distributed inference labeling by using the model;
performing traversal circulation on all pictures of the current data set, respectively acquiring artificial labeling information and model labeling information for each picture in a test set, and the confidence coefficient of the current test, and removing model labeling information which does not meet preset requirements according to the confidence coefficient;
for each picture in the traversal cycle, grouping the acquired artificial labeling information and model labeling information according to the defect names, and collecting all artificial labeling information lists and model labeling information lists of different defect names;
and circularly comparing the grouped artificial marking information and the model marking information of the defects, wherein the grouped defects comprise the following conditions:
a: if the number of the artificial marking frames and the number of the model marking frames with the current defects are both 0, judging that the current picture has no current type defects, and judging that the killing number and the loss number are both equal to 0;
b: the number of the artificial marking frames of the current defects is 0, but the number of the model marking frames is not 0, at the moment, the model has over-killing phenomena on the current type defects of the current picture, and the result is that the over-killing number is equal to the number of the model marking frames, and the leakage number is equal to 0;
c: the number of the artificial marking frames of the current defects is not 0 but the number of the model marking frames is 0, at the moment, the model has a leakage phenomenon on the current type defects of the current picture, and the result is that the killing number is equal to 0 and the leakage number is equal to the number of the artificial marking frames;
d: the number of the artificial labeling frames with the current defects and the number of the model labeling frames are not 0, firstly, missing frame labels are marked on all the artificial labeling frames with the current defects of the current picture, then, killing frame labels are marked on all the model labeling frames with the current defects of the current picture, then, the artificial labeling frame list is traversed circularly, the area of the current artificial labeling frame is calculated, meanwhile, the model labeling frame list is traversed circularly, the area of the model labeling frame is calculated, whether the overlapping degree IOU (input object) of the current artificial labeling frame and the model labeling frame is larger than 0 or not is judged, if the overlapping degree IOU is larger than 0, the missing labels of the current artificial labeling frame are removed, the killing labels of the current model labeling frame are removed, after the circulation is completed, the counting result shows that the killing number is equal to the number of the model labeling frames with the killing labels, and the missing number is equal to the number of the artificial labeling frames with the missing labels;
and circularly comparing all defects under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all kinds of defects of the current data set under the current confidence coefficient, acquiring the missing labeling quantity and the overdividing quantity of all defects of the current data set under all set confidence coefficients, and summarizing comparison data.
8. The model-based defect detection system of claim 7, wherein the model testing further comprises classifying the pictures for testing, initializing statistical result objects, obtaining labeling information, updating confusion matrices in the result objects, and calculating statistical values;
the initializing statistical result object includes: initializing a confusion matrix, a label mapping relation and the number of picture categories for storing the statistical result;
the acquiring of the labeling information includes: loading a data set to be tested, performing cycle traversal on the data set, and respectively acquiring manual labeling information and model labeling information of a current picture aiming at each picture in the data set;
updating the confusion matrix in the result object comprises: respectively labeling corresponding values of the manual labeling label and the model labeling label in the label mapping relation, forming an (x, y) coordinate according to the manual labeling mapping value and the model labeling mapping value, namely (the manual labeling mapping value and the model labeling mapping value), and updating the coordinate into a confusion matrix;
the calculating the statistical value comprises: and counting the proportion of correct prediction of the test set according to the accuracy rate, the recall rate and the precision rate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211432680.5A CN115495608B (en) | 2022-11-16 | 2022-11-16 | Defect detection method and system based on model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211432680.5A CN115495608B (en) | 2022-11-16 | 2022-11-16 | Defect detection method and system based on model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115495608A true CN115495608A (en) | 2022-12-20 |
| CN115495608B CN115495608B (en) | 2023-04-07 |
Family
ID=85115857
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211432680.5A Active CN115495608B (en) | 2022-11-16 | 2022-11-16 | Defect detection method and system based on model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115495608B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116399873A (en) * | 2023-06-07 | 2023-07-07 | 杭州百子尖科技股份有限公司 | Machine vision-based sheet defect labeling method, device, equipment and medium |
| CN117973945A (en) * | 2024-03-29 | 2024-05-03 | 南京认知物联网研究院有限公司 | Quality inspection model continuous optimization method and system for personalized intelligent manufacturing |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109446369A (en) * | 2018-09-28 | 2019-03-08 | 武汉中海庭数据技术有限公司 | The exchange method and system of the semi-automatic mark of image |
| CN112799850A (en) * | 2021-02-26 | 2021-05-14 | 重庆度小满优扬科技有限公司 | Model training method, model prediction method, and model control system |
| CN112966772A (en) * | 2021-03-23 | 2021-06-15 | 之江实验室 | Multi-person online image semi-automatic labeling method and system |
-
2022
- 2022-11-16 CN CN202211432680.5A patent/CN115495608B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109446369A (en) * | 2018-09-28 | 2019-03-08 | 武汉中海庭数据技术有限公司 | The exchange method and system of the semi-automatic mark of image |
| CN112799850A (en) * | 2021-02-26 | 2021-05-14 | 重庆度小满优扬科技有限公司 | Model training method, model prediction method, and model control system |
| CN112966772A (en) * | 2021-03-23 | 2021-06-15 | 之江实验室 | Multi-person online image semi-automatic labeling method and system |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116399873A (en) * | 2023-06-07 | 2023-07-07 | 杭州百子尖科技股份有限公司 | Machine vision-based sheet defect labeling method, device, equipment and medium |
| CN116399873B (en) * | 2023-06-07 | 2023-09-26 | 杭州百子尖科技股份有限公司 | Machine vision-based sheet defect labeling method, device, equipment and medium |
| CN117973945A (en) * | 2024-03-29 | 2024-05-03 | 南京认知物联网研究院有限公司 | Quality inspection model continuous optimization method and system for personalized intelligent manufacturing |
| CN117973945B (en) * | 2024-03-29 | 2024-05-31 | 南京认知物联网研究院有限公司 | Quality inspection model continuous optimization method and system for personalized intelligent manufacturing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115495608B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115495608B (en) | Defect detection method and system based on model | |
| US11763443B2 (en) | Method for monitoring manufacture of assembly units | |
| CN110221975B (en) | Method and device for creating interface case automation test script | |
| Kuric et al. | Visual product inspection based on deep learning methods | |
| CN111401418A (en) | Employee dressing specification detection method based on improved Faster r-cnn | |
| CN114414935A (en) | Automatic positioning method and system for feeder fault area of power distribution network based on big data | |
| CN111124863B (en) | Intelligent device performance testing method and device and intelligent device | |
| CN110969600A (en) | Product defect detection method and device, electronic equipment and storage medium | |
| CN112527676A (en) | Model automation test method, device and storage medium | |
| CN116823793A (en) | Device defect detection method, device, electronic device and readable storage medium | |
| CN111311546A (en) | Container detection method, device and computer readable storage medium | |
| CN111311545A (en) | Container detection method, device and computer readable storage medium | |
| CN115184193A (en) | Automatic testing method, device and equipment for cable and storage medium | |
| CN111275682A (en) | Container detection method, device and computer readable storage medium | |
| CN117495786A (en) | Defect detection meta-model construction method, defect detection method, device and medium | |
| CN113408630A (en) | Transformer substation indicator lamp state identification method | |
| CN112836724A (en) | Object defect recognition model training method and device, electronic equipment and storage medium | |
| CN115861305A (en) | Flexible circuit board detection method and device, computer equipment and storage medium | |
| CN114882242A (en) | Violation image identification method and system based on computer vision | |
| CN116229355A (en) | Image detection method and device, electronic equipment and storage medium | |
| CN113238956B (en) | Fault analysis method, device, equipment and storage medium for abnormal application | |
| CN113240628A (en) | Method, device and system for judging quality of steel coil | |
| CN112149546A (en) | Information processing method and device, electronic equipment and storage medium | |
| CN111124862A (en) | Intelligent equipment performance testing method and device and intelligent equipment | |
| CN112529039B (en) | Method and device for checking material information of main board and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |