CN112926654A - Pre-labeling model training and certificate pre-labeling method, device, equipment and medium - Google Patents

Pre-labeling model training and certificate pre-labeling method, device, equipment and medium Download PDF

Info

Publication number
CN112926654A
CN112926654A CN202110211382.2A CN202110211382A CN112926654A CN 112926654 A CN112926654 A CN 112926654A CN 202110211382 A CN202110211382 A CN 202110211382A CN 112926654 A CN112926654 A CN 112926654A
Authority
CN
China
Prior art keywords
model
target
migrated
labeling
certificate
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
Application number
CN202110211382.2A
Other languages
Chinese (zh)
Other versions
CN112926654B (en
Inventor
王晟宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Bank Co Ltd
Original Assignee
Ping An Bank Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Ping An Bank Co Ltd filed Critical Ping An Bank Co Ltd
Priority to CN202110211382.2A priority Critical patent/CN112926654B/en
Publication of CN112926654A publication Critical patent/CN112926654A/en
Application granted granted Critical
Publication of CN112926654B publication Critical patent/CN112926654B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to the field of artificial intelligence classification models, and provides a method, a device, equipment and a medium for pre-labeling model training and certificate pre-labeling, wherein the method comprises the following steps: obtaining a target marking category, a target description, a model performance parameter and an image sample set; crawling categories to be migrated in a target classification recognition library by using a text similarity technology; searching a model to be migrated and a target area of each identified image sample from a target classification identification library by a simulated target identification technology; performing target fine adjustment to obtain a fine adjustment area, and inputting the image sample, the fine adjustment area and the target mark type into a model to be migrated; acquiring a marked target marking area by applying a transfer learning technology; determining a loss value according to the target labeling area and the fine tuning area; and training the model to be migrated until the training is finished to obtain the pre-labeled model. The invention realizes the automatic training of the image sample set with zero annotation, obtains the pre-annotation model and reduces the manual annotation time and workload.

Description

Pre-labeling model training and certificate pre-labeling method, device, equipment and medium
Technical Field
The invention relates to the field of artificial intelligence classification models, in particular to a method and a device for pre-labeling model training and certificate pre-labeling, a computer device and a storage medium.
Background
The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like. Among them, Computer Vision technology (CV) is a science that studies how to "see" a machine, and generally includes technologies such as image processing, image Recognition, image semantic understanding, image retrieval, and Optical Character Recognition (OCR).
As the artificial intelligence technology is gradually mature, the image recognition technology is more and more widely applied to daily life. In order to obtain an image recognition model with higher recognition accuracy, the image recognition model needs to be trained through a large number of labeled samples, and in the prior art, when a training sample is constructed, manual labeling such as manual input is usually adopted, so that not only is the labor cost consumed, but also the labeling efficiency of the sample is greatly reduced, and great difficulty is brought to model training.
Disclosure of Invention
The invention provides a pre-labeling model training method, a certificate pre-labeling device, a computer device and a storage medium, which can realize automatic training of a zero-labeled image sample set through a text similarity technology, a crawling technology, a simulated target recognition technology and a transfer learning technology, reduce the manual labeling time and workload, improve the labeling efficiency, save the investment cost and improve the accuracy of pre-labeling.
A pre-labeling model training method comprises the following steps:
acquiring a target labeling type, target description corresponding to the target labeling type, model performance parameters and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation category;
crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated;
searching a model to be migrated matched with the category to be migrated and the model performance parameter from the target classification recognition library and a target area of each image sample recognized by the model to be migrated by a simulation target recognition technology;
performing target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples, and inputting the image samples, the fine adjustment areas corresponding to the image samples and the target labeling types into the model to be migrated; the model to be migrated comprises migration parameters;
acquiring a target marking area corresponding to the image sample marked by the model to be migrated by extracting the class characteristics of the model to be migrated in a self-adaptive learning target by applying a migration learning technology;
determining a loss value according to the target labeling area and the fine adjustment area corresponding to the image sample;
and when the loss value does not reach a preset convergence condition, iteratively updating the migration parameters of the model to be migrated until the loss value reaches the preset convergence condition, and recording the converged model to be migrated as a pre-labeled model after training.
A method of pre-labeling a document, comprising:
receiving an image annotation instruction, and acquiring a certificate image in the image annotation instruction;
the certificate image is input into a certificate pre-labeling model trained by the pre-labeling model training method, the learned certificate features are extracted through the certificate pre-labeling model, and a labeling result output by the certificate pre-labeling model according to the certificate features is obtained; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
A pre-labeling model training device, comprising:
the acquisition module is used for acquiring a target labeling type, a target description corresponding to the target labeling type, a model performance parameter and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation category;
the crawling module is used for crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated;
the identification module is used for searching a model to be migrated which is matched with the category to be migrated and the model performance parameters from the target classification identification library through a simulation target identification technology, and searching a target area of each image sample identified through the model to be migrated;
the fine tuning module is used for performing target fine tuning on all the target areas to obtain fine tuning areas corresponding to the image samples, and inputting the image samples, the fine tuning areas corresponding to the image samples and the target labeling types into the model to be migrated; the model to be migrated comprises migration parameters;
the migration module is used for acquiring a target labeling area corresponding to the image sample marked by the model to be migrated through the extraction of the self-adaptive learning target class characteristics of the model to be migrated by applying a migration learning technology;
the loss module is used for determining a loss value according to the target labeling area and the fine adjustment area corresponding to the image sample;
and the training module is used for iteratively updating the migration parameters of the model to be migrated when the loss value does not reach a preset convergence condition, and recording the converged model to be migrated as a pre-labeled model after training until the loss value reaches the preset convergence condition.
A document pre-labeling apparatus comprising:
the receiving module is used for receiving the image annotation instruction and acquiring the certificate image in the image annotation instruction;
the marking module is used for inputting the certificate image into the certificate pre-marking model trained by the pre-marking model training method, extracting the learned certificate characteristics through the certificate pre-marking model and acquiring the marking result output by the certificate pre-marking model according to the certificate characteristics; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the steps of the aforementioned pre-labeling model training method when executing the computer program, or implementing the steps of the aforementioned certificate pre-labeling method when executing the computer program.
A computer-readable storage medium, in which a computer program is stored, which computer program, when being executed by a processor, carries out the steps of the aforementioned pre-annotation model training method, or which computer program, when being executed by a processor, carries out the steps of the aforementioned certificate pre-annotation method.
The invention provides a pre-labeling model training method, a device, computer equipment and a storage medium, which are characterized in that a target labeling type, a target description corresponding to the target labeling type, model performance parameters and an image sample set are obtained; crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated; searching a model to be migrated matched with the category to be migrated and the model performance parameter from the target classification recognition library and a target area of each image sample recognized by the model to be migrated by a simulation target recognition technology; performing target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples, and inputting the image samples, the fine adjustment areas corresponding to the image samples and the target labeling types into the model to be migrated; acquiring a target labeling area marked by the model to be migrated by extracting the class characteristics of the model to be migrated in a self-adaptive learning mode by applying a migration learning technology; determining a loss value according to the target labeling area and the fine adjustment area; when the loss value does not reach the preset convergence condition, iteratively updating the migration parameters of the model to be migrated, and recording the converged model to be migrated as a trained pre-labeled model when the loss value reaches the preset convergence condition, so that the class to be migrated is crawled from a target classification recognition library by using a text similarity technology and a crawling technology through a target labeling class, a target description, a model performance parameter and an image sample set, the most matched model to be migrated is found from the target classification recognition library by simulating a target recognition technology, a fine tuning region is obtained by using a target fine tuning technology, the characteristics of the target class characteristics are adaptively learned and a target labeling region is recognized by using a migration learning technology, and the model to be migrated is iteratively updated through the loss value, so that the problems that the model to be migrated is continuously updated through the text similarity technology and the image sample set are solved, The crawling technology, the simulated target recognition technology and the migration learning technology can automatically train the image sample set with zero labeling, and train to obtain the pre-labeling model, so that the manual labeling time and workload are reduced, the labeling efficiency is improved, the investment cost is saved, and the accuracy of pre-labeling is improved.
According to the certificate pre-labeling method, the certificate pre-labeling device, the computer equipment and the storage medium, the certificate pre-labeling model is obtained by training in the pre-labeling model training mode under the condition that a small number of certificate photos are obtained, so that the labeling time of manual certificate photos can be reduced, automatic certificate pre-labeling is realized, the cost is saved, the accuracy of pre-labeling is improved, and the recognition reliability is improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments of the present invention will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without inventive labor.
FIG. 1 is a schematic diagram of an application environment of a pre-labeling model training method or a certificate pre-labeling method according to an embodiment of the present invention;
FIG. 2 is a flow chart of a method for training a pre-annotation model according to an embodiment of the invention;
FIG. 3 is a flow chart of a method for pre-labeling documents in accordance with an embodiment of the present invention;
FIG. 4 is a schematic block diagram of a pre-labeling model training apparatus according to an embodiment of the present invention;
FIG. 5 is a schematic block diagram of a credential pre-labeling device in an embodiment of the invention;
FIG. 6 is a schematic diagram of a computer device in an embodiment of the invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The pre-labeling model training method provided by the invention can be applied to the application environment shown in fig. 1, wherein a client (computer device) communicates with a server through a network. The client (computer device) includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, cameras, and portable wearable devices. The server may be implemented as a stand-alone server or as a server cluster consisting of a plurality of servers.
In an embodiment, as shown in fig. 2, a method for training a pre-labeled model is provided, which mainly includes the following steps S10-S70:
s10, acquiring a target annotation category, a target description corresponding to the target annotation category, model performance parameters and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation class.
Understandably, after a certain number of images of a category are collected, the images need to be pre-labeled, a model capable of identifying the category is constructed for subsequent training, the collected images of the category are very huge, in order to automatically label a region triggering labeling request required by the category in each image, the target labeling category, the target description, the model performance parameters and the image sample set are obtained from the labeling request, the target labeling category is the collected category name belonging to the same category, the target description is the description of the relevant features of the target labeling category, the model performance parameters are the relevant performance parameters of the automatically pre-labeled model preset according to an application scene, the model performance parameters comprise preset time consumption, preset capacity and the like, and the application scene can be set according to requirements, for example: the application scene can be a face identification scene, the performance requirement under the application scene is that the identification is fast, and the capacity of the model is not considered; the application scenario may be an identification card recognition scenario, where performance requirements under the application scenario are that the application scenario can be implanted into a mobile device to run, so that a requirement on capacity of a model is low, recognition speed is not considered, and the like, where the image sample set is a set of collected image samples, the image samples are historically collected images or a small number of images are randomly extracted from the historically collected images, the image samples include target objects to be labeled, and the target objects belong to the target labeling category, and the image samples correspond to the target labeling category one to one, that is, the image samples belong to the target labeling category, for example: the collected 1000 identity card pictures need to be labeled, 100 identity card pictures are randomly extracted from the 1000 identity card pictures to be determined as an image sample set, the target labeling type is determined as the identity card, the target description is a rectangular certificate containing a face head portrait, a national emblem and 18-bit characters, the preset time consumption in the model performance parameters is less than 1 second, the preset accuracy is higher than 98%, and the preset capacity is less than 1.5M, so that a labeling request is triggered.
S20, crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated.
Understandably, the crawling process is to crawl out the relevant description of the target labeling category by using a web crawler technology, collecting the crawled related descriptions to extract keywords and the occurrence times of each keyword, using the keywords as network descriptions, performing keyword aggregation on the network descriptions and the target descriptions, namely, the target description is given with keywords consistent with the network description for weighting, and the important words in the target description can be compared with each other with emphasis by weighting, the text similarity technology is applied, a process of comparing, in the target classification recognition library, the history description under each history category with the weighted target description, and determining a history category with the highest similarity value after comparison as the history category similar to the target description, for example: the target labeling type is an identity card, the target description is a rectangular certificate containing a face head portrait, a national emblem and 18-bit characters, the face head portrait, the 18-bit characters and a rectangle are weighted, the image of a rectangular photo frame with historical description of a photo frame as scenery containing a person is crawled from the target classification recognition library by using a text similarity technology, and the image is determined to be the historical type similar to the identity card if the similarity value of the image and the weighted target description is the highest.
The target classification recognition library stores all history categories which are provided with history descriptions and have been trained, the history descriptions are descriptions of relevant features of the history categories, the text similarity technology is a technology for comparing similarity degrees between two texts by using a text similarity algorithm, the text similarity algorithm is an algorithm for performing word embedding (word embedding) conversion processing on the two texts, performing equal weight conversion according to keywords with weights in the process of word embedding conversion processing, and performing similarity calculation on the two processed texts to obtain a similarity value between the two texts, and the web crawler technology is a technology for automatically capturing programs or scripts of web information according to certain rules so as to obtain required information.
In an embodiment, in the step S20, that is, the crawling a history category similar to the target description in a target classification recognition library by using a text similarity technique, and determining the crawled history category as the category to be migrated includes:
s201, crawling the network description matched with the target labeling type by using a web crawler technology.
Understandably, the web crawler technology is a technology for automatically capturing a program or a script of world wide web information according to a certain rule so as to obtain required information, crawling descriptions related to the target labeling category from the internet, summarizing and refining all the crawled related descriptions, extracting keywords by using a TF-IDF algorithm, and determining the occurrence frequency of each keyword of the extracted keywords as network description.
The TF-IDF algorithm is a weighting technique for information retrieval (information retrieval) and text mining (text mining), and is a statistical method for evaluating the importance degree of a word in a text, and the importance of a word or word increases in proportion to the number of times it appears in a document, but decreases in inverse proportion to the frequency of occurrence in a corpus.
S202, according to the network description, carrying out keyword weighting on the target description to obtain a focused description.
Understandably, words which are consistent with the keywords in the network description are found in the target description, and the found words are weighted according to the occurrence frequency of each keyword to obtain the focusing description.
S203, comparing the history description and the focusing description under each history category in the target classification recognition library by using a text similarity technology to obtain a similarity value corresponding to each history description.
Understandably, the text similarity technology is a technology for comparing the similarity between two texts by using a text similarity algorithm, the text similarity algorithm is an algorithm for performing word embedding (word embedding) conversion processing on the two texts, performing equal weight conversion according to keywords with weights in the process of word embedding conversion processing, performing similarity calculation on the two processed texts, and obtaining a similarity value between the two texts, the word embedding (word embedding) conversion processing is also called word2vec, namely, converting a word (word) into a vector (vector) for representation, giving equal weights to the vectors after converting the keywords with weights in the process of conversion, comparing the historical description and the focused description under each historical category by using the text similarity algorithm in the target classification recognition library, and adding equal weights to the vectors corresponding to the keywords with weights in the output similarity values, that is, the influence of the vector corresponding to the weighted keyword on the similarity value as a whole has a higher weight ratio than others.
S204, determining the history category corresponding to the history description corresponding to the maximum similarity value as the category to be migrated.
The invention realizes that the network description matched with the target labeling category is crawled by applying a web crawler technology; according to the network description, carrying out keyword weighting on the target description to obtain a focused description; comparing the historical description and the focused description under each historical category in the target classification recognition library by using a text similarity technology to obtain a similarity value corresponding to each historical description; the history category corresponding to the history description corresponding to the maximum similarity value is determined as the category to be migrated, so that the category to be migrated is automatically identified from a target classification identification library by applying a web crawler technology, a keyword weighting technology and a text similarity technology, the cost of manual identification is reduced, the category to be migrated can be quickly and accurately found, and the identification accuracy and reliability are improved.
S30, searching a model to be migrated matched with the category to be migrated and the model performance parameters from the target classification recognition library through a simulation target recognition technology, and searching a target area of each image sample recognized through the model to be migrated.
Understandably, the object classification recognition library further stores an object detection depth model set which comprises object detection depth models of various network structures under each history category, and further comprises object detection depth models of various network structures, the object detection depth models of different network structures are embodied as models which accord with different performance parameters, because the complexity and the hierarchical structure of the network structures directly influence the performance parameters, the object detection depth model set comprises at least one object detection depth model, the simulation object recognition technology is a technology which inputs each image sample into each object detection depth model, performs object recognition through each object detection depth model, acquires the object region recognized by each object detection depth model, predicts the performance parameters, and determines a model to be migrated which is matched with the model performance parameters according to all the object regions and the predicted performance parameters, by applying the simulated target identification technology, the model to be migrated which is matched with the category to be migrated and the model performance parameter can be found, then each image sample is identified through the model to be migrated, and a target area with the characteristics of the category to be migrated in each image sample can be identified.
The target area is an area where the image sample is identified by the target detection depth model to meet the category to be migrated, and the prediction performance parameters are the average value of the time consumed by the target detection depth model to identify the image sample and the predicted capacity of the learned model.
In an embodiment, in the step S30, that is, the finding, from the target classification and identification library, the to-be-migrated model matching the to-be-migrated category and the model performance parameter through the simulation target identification technology, and the target area of each image sample identified by the to-be-migrated model include:
s301, finding a target detection depth model set associated with the category to be migrated in the target classification recognition library; the set of target detection depth models includes at least one target detection depth model.
Understandably, in the target classification recognition library, one to-be-migrated category is associated with one target detection depth model set, the target detection depth model set includes target detection depth models of various network structures, and the target detection depth models of different network structures are embodied as models conforming to different performance parameters, because the complexity and the hierarchical structure of the network structure directly affect the performance parameters.
S302, inputting each image sample into each target detection depth model, performing target identification on each image sample through each target detection depth model, acquiring each target detection depth model, identifying the target area corresponding to each image sample, and outputting the prediction performance parameters corresponding to each target area.
Understandably, inputting each image sample into each target detection depth model, wherein each target detection depth model is a trained model used for identifying a target area with characteristics related to the category to be migrated, the network structure of the target detection depth model can be VGG19, Resnet51, SVM, YOLO and the like, the target recognition is to extract features related to the category to be migrated from the input image and recognize a region according to the extracted features, a process of obtaining a target region and counting the time taken to input the image sample to output the target region and the capacity of the model to obtain the predicted performance parameter, wherein for one target detection depth model, the target region and the predicted performance parameter for the image sample are identified.
S303, determining the target detection depth model matched with the model performance parameters according to all the target areas and the prediction performance parameters, and determining the target detection depth model matched with the model performance parameters as a model to be migrated.
Understandably, according to all the target regions and the prediction performance parameters, the model to be migrated which is matched with the model performance parameters may be considered comprehensively, that is, all the target regions and the prediction performance parameters are combined with the model performance parameters, and a winning bid coefficient of each target detection depth model is calculated, that is, the winning bid coefficient of each target detection depth model is determined according to the difference between all the target regions and the prediction performance parameters and the model performance parameters, the winning bid coefficient may measure the difference that each target detection depth model conforms to the model performance parameters, and the target detection depth model which best conforms to the model performance parameters may be identified by the winning bid coefficient and determined as the model to be migrated.
According to the invention, the target detection depth model set associated with the category to be migrated is searched in the target classification recognition library; inputting each image sample into each target detection depth model, performing target identification on each image sample through each target detection depth model, and acquiring a target area, corresponding to each image sample, identified by each target detection depth model and a prediction performance parameter corresponding to each target area; according to the target areas and the predicted performance parameters, the target detection depth model matched with the model performance parameters is determined and determined as the model to be migrated, so that the model to be migrated which is most matched with the model performance parameters is matched by obtaining the target detection depth models related to the category to be migrated and performing target identification output on the target areas and the predicted performance parameters according to the target detection depth models, the cost of manual matching is reduced, the appropriate model to be migrated is matched through two dimensions of the identified target areas and the performance parameters, and the accuracy and the reliability of subsequent annotation identification are improved.
In an embodiment, in the step S303, that is, the determining, according to all the target areas and the predicted performance parameters, the target detection depth model matched with the model performance parameters, and determining the target detection depth model matched with the model performance parameters as the model to be migrated includes:
s3031, performing intersection analysis on all the target regions corresponding to the same image samples to obtain intersection regions corresponding to the image samples, and determining the medium rate of each target detection depth model corresponding to each image sample according to the intersection regions corresponding to each image sample and each target region.
Understandably, performing intersection analysis on all the target regions corresponding to the same image sample, the intersection analysis being an analysis process of taking an intersection of the target regions identified by the target detection depth models each of which has input the same image sample, determining the target region having a largest area ratio of the intersection as the intersection region, the intersection region being the target region corresponding to the area ratio of the intersection of all the target regions corresponding to the same image sample having the largest area ratio of the target region, determining the area ratio of the intersection region corresponding to one image sample to the target region output by one target detection depth model having input the image sample as the winning bid rate of the target detection depth model for the image sample, that is, one image sample corresponds to a plurality of the medium rate ratios corresponding to the target detection depth models one to one, and the area ratio is the ratio of the area of the intersection region to the area of the target region.
S3032, according to the medium rate of the same target detection depth model, determining a target index weight corresponding to the target detection depth model, and meanwhile, according to the predicted performance parameters output by the same target detection depth model, predicting the final predicted performance parameters corresponding to the target detection depth model.
Understandably, an inverse distance weighting interpolation method is applied, the greater the difference value of the inverse distance weighting difference method from a first is, the smaller the weight is given, the distance difference value between the medium rate and a corresponding to the same target detection depth model is given to the weight of each medium rate, the weight of each medium rate can be obtained through a weight function, and the weight function is:
λ(i，j)＝1/(1-1/ln(1-x(i，j)))
wherein λ is(i，j)Weighting, x, to a mid-rate in a jth target detection depth model in a set of target detection depth models for an ith image sample in a set of image samples(i，j)And for the winning rate of the ith image sample in the image sample set in the jth target detection depth model in the target detection depth model set, according to the winning rate corresponding to the same target detection depth model and the weight corresponding to the winning rate, namely multiplying the winning rate corresponding to the same target detection depth model and the weight corresponding to the winning rate respectively, then summing, and then averaging to obtain the target index weight corresponding to the target detection depth model.
And S3033, obtaining a bid-winning coefficient corresponding to each target detection depth model according to the model performance parameters, the target index weights corresponding to each target detection depth model and the final prediction performance parameters.
Understandably, the model performance parameter is divided by the final predicted performance parameter corresponding to each target detection depth model to obtain the performance ratio corresponding to each target detection depth model, the performance ratio is a measure of the difference between the final predicted performance parameter of each target detection depth model and the model performance parameter, the performance ratio corresponding to the same target detection depth model is multiplied by the target index weight to obtain the winning coefficient corresponding to each target detection depth model, and the winning coefficient is a measure of the target detection depth model combined with the target index weight matched with the model performance parameter.
In an embodiment, in the step S3033, that is, obtaining the bid-closing coefficient corresponding to each target detection depth model according to the model performance parameter, the target index weight corresponding to each target detection depth model, and the final predicted performance parameter, includes:
s30331, dividing the model performance parameter with the final prediction performance parameter corresponding to each target detection depth model to obtain a performance ratio corresponding to each target detection depth model.
S30332, multiplying the performance ratio corresponding to the same target detection depth model by the target exponential weight to obtain the winning bid coefficient corresponding to each target detection depth model.
The performance ratio corresponding to each target detection depth model is obtained by dividing the model performance parameter and the final prediction performance parameter corresponding to each target detection depth model; and multiplying the performance ratio corresponding to the same target detection depth model by the target index weight to obtain the bid-winning coefficient corresponding to each target detection depth model, so that a scientific calculation method for the bid-winning coefficient is provided, and the accuracy of selecting the model to be migrated is improved.
S3034, determining the target detection depth model corresponding to the maximum winning coefficient as the model to be migrated.
Understandably, the maximum winning bid coefficient is found out from all the winning bid coefficients, the target detection depth model corresponding to the winning bid coefficient is recorded as the model to be migrated, the parameters of the target detection depth model are determined as the migration parameters of the model to be migrated, and the model to be migrated is the target detection depth model of the region which needs migration learning for identification and the target marking type under the category to be migrated.
The method and the device realize that the intersection area corresponding to the image sample is obtained by performing intersection analysis on all the target areas corresponding to the same image sample, and the bid rate of each target detection depth model corresponding to each image sample is determined according to the intersection area corresponding to each image sample and each target area; determining a target index weight corresponding to each target detection depth model according to all the medium-rate rates; predicting a final predicted performance parameter corresponding to the target detection depth model according to the predicted performance parameter output to the same target detection depth model; obtaining a winning coefficient corresponding to each target detection depth model according to the model performance parameters, the target index weights corresponding to each target detection depth model and the final prediction performance parameters; determining the target detection depth model corresponding to the maximum winning coefficient as the model to be migrated, thus realizing the intersection region through intersection analysis, and identifying the medium-rate of each target detection depth model corresponding to each image sample, and determining the target index weight corresponding to each target detection depth model, thereby calculating the winning bid coefficient corresponding to each target detection depth model, finally matching the model to be migrated, therefore, the method realizes the automatic identification of the intersection region through intersection analysis and the calculation of the target exponential weight, and combining the model performance parameters, the target index weight and the model performance parameters to calculate the winning bid coefficient, and the model to be migrated can be matched scientifically, accurately and automatically, so that the accuracy and reliability of subsequent labeling are improved.
S40, performing target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples, and inputting the image samples, the fine adjustment areas corresponding to the image samples and the target annotation categories into the model to be migrated; the model to be migrated includes migration parameters.
Understandably, the process of fine tuning the target is to perform edge segmentation on an image in a preset range adjacent to an edge of the target region output by the model to be migrated corresponding to the image sample, identify an edge line, and perform an adjustment process of edge reduction on the target region along the edge line.
In an embodiment, in the step S40, the performing the target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples includes:
s401, performing edge segmentation on the preset range adjacent to the target area corresponding to the image sample, and identifying an edge line.
Understandably, the edge segmentation is a process of identifying the edge in the target area, identifying pixel points with the color difference value of the pixel points between the adjacent pixel points larger than the preset color difference value, carrying out closed line segmentation on the identified pixel points, namely segmenting lines which can be enclosed into a closed area, and segmenting edge lines, wherein the edge lines are lines which can be enclosed into a closed area by the pixel points larger than the preset color difference value.
S402, according to the edge line, performing edge reduction on the target area to obtain the fine adjustment area corresponding to the image sample.
Understandably, performing edge reduction on the target area according to the area surrounded by the edge lines, wherein the edge reduction is a process that the target area is reduced to be overlapped with the edge lines and the area surrounded by the edge lines can be covered, and finally, the reduced area is determined as the fine adjustment area, and one image sample corresponds to one fine adjustment area.
The invention realizes the purpose that the edge line is identified by carrying out edge segmentation in the preset range adjacent to the target area corresponding to the image sample; and according to the edge line, performing edge reduction on the target area to obtain the fine adjustment area corresponding to the image sample, so that automatic fine adjustment of the target area is realized, more accurate targets are drawn together, and the accuracy of pre-marking is improved.
In an embodiment, the step S40, namely, the inputting the image sample, the fine-tuning area corresponding to the image sample, and the target annotation category into the model to be migrated includes:
and S403, performing random data enhancement on the image sample to generate a plurality of enhanced images corresponding to the image sample.
Understandably, the random data enhancement includes an enhancement processing procedure of random rotation, random saturation, random contrast and random brightness, and a plurality of enhanced images can be generated by the random data enhancement method, so that samples for training the model to be migrated are increased, and the identification accuracy of the model to be migrated is increased.
S404, associating the enhanced image with the fine tuning area corresponding to the image sample, and inputting the enhanced image into the model to be migrated.
Understandably, the enhanced image and the fine tuning area corresponding to the image sample are associated, and in the case that the enhanced image is obtained through random rotation, the fine tuning area associated with the enhanced image is also associated after corresponding rotation, and the enhanced image and the fine tuning area corresponding to the image sample are associated and input to the model to be migrated.
S405, the target labeling category is used as a category result in the model to be migrated.
Understandably, the model to be migrated represents a model for identifying the target labeling category, and the target labeling category is taken as the category result.
The invention realizes that a plurality of enhanced images corresponding to the image samples are generated by carrying out random data enhancement on the image samples; associating the enhanced image with the fine tuning area corresponding to the image sample, and inputting the enhanced image into the model to be migrated; and the target labeling type is used as a type result in the model to be migrated, so that an enhanced image is generated by random data enhancement, is associated with the fine tuning area and is input into the model to be migrated together, overfitting of the model to be migrated can be prevented, and the accuracy and quality of labeling are improved.
And S50, acquiring a target labeling area corresponding to the image sample marked by the model to be migrated by extracting the target class characteristics of the model to be migrated in a self-adaptive learning manner by using a migration learning technology.
Understandably, the migration learning technology is a machine learning method, which refers to a technology that a pre-trained model is used for training in another task again, the adaptive learning target class feature is a technology that automatically adjusts and learns the relevant implicit features of the target labeling class, namely, the implicit features suitable for the target labeling class are learned in the originally learned features for identifying the to-be-migrated class, the implicit features used for identifying the target labeling class can be continuously learned through extracting the adaptive learning target class feature for identification, so that the target labeling region with the features of the target labeling class can be output, and the target labeling region is a region which is output by the to-be-migrated model and contains the implicit features of the target labeling class.
And S60, determining a loss value according to the target labeling area and the fine adjustment area corresponding to the image sample.
Understandably, the target labeling area and the fine tuning area are input into a loss function in the model to be migrated, and the loss value corresponding to the image sample is obtained through calculation, where the loss function may be set according to a requirement, for example, the loss function is a cross entropy loss function, and the loss function is a logarithm of a difference between the target labeling area and the fine tuning area, which indicates a difference between the target labeling area and the fine tuning area.
And S70, when the loss value does not reach the preset convergence condition, iteratively updating the migration parameters of the model to be migrated until the loss value reaches the preset convergence condition, and recording the converged model to be migrated as a pre-labeled model after training.
Understandably, the convergence condition may be a condition that the loss value is small and does not decrease again after 3000 times of calculation, that is, when the loss value is small and does not decrease again after 3000 times of calculation, stopping training, and recording the model to be migrated after convergence as a pre-labeled model after training; the convergence condition may also be a condition that the loss value is smaller than a set threshold, that is, when the loss value is smaller than the set threshold, the training is stopped, and the model to be migrated after the convergence is recorded as a pre-labeled model after the training is completed, so that when the loss value does not reach the preset convergence condition, the migration parameters of the model to be migrated are continuously adjusted, a migration learning technology is triggered, and a step of obtaining a target labeling area corresponding to the image sample, which is labeled by the model to be migrated, through extraction of a target class characteristic of adaptive learning of the model to be migrated is performed, so that the step can be continuously closed to an accurate result, and the identification accuracy is higher and higher. Therefore, the identification of the pre-labeling can be optimized, and the accuracy and the reliability of the pre-labeling are improved.
Therefore, the method and the device realize the purpose that the target labeling type, the target description corresponding to the target labeling type, the model performance parameters and the image sample set are obtained; crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated; searching a model to be migrated matched with the category to be migrated and the model performance parameter from the target classification recognition library and a target area of each image sample recognized by the model to be migrated by a simulation target recognition technology; performing target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples, and inputting the image samples, the fine adjustment areas corresponding to the image samples and the target labeling types into the model to be migrated; acquiring a target labeling area marked by the model to be migrated by extracting the class characteristics of the model to be migrated in a self-adaptive learning mode by applying a migration learning technology; determining a loss value according to the target labeling area and the fine adjustment area; when the loss value does not reach the preset convergence condition, iteratively updating the migration parameters of the model to be migrated, and recording the converged model to be migrated as a trained pre-labeled model when the loss value reaches the preset convergence condition, so that the class to be migrated is crawled from a target classification recognition library by using a text similarity technology and a crawling technology through a target labeling class, a target description, a model performance parameter and an image sample set, the most matched model to be migrated is found from the target classification recognition library by simulating a target recognition technology, a fine tuning region is obtained by using a target fine tuning technology, the characteristics of the target class characteristics are adaptively learned and a target labeling region is recognized by using a migration learning technology, and the model to be migrated is iteratively updated through the loss value, so that the problems that the model to be migrated is continuously updated through the text similarity technology and the image sample set are solved, The crawling technology, the simulated target recognition technology and the migration learning technology can automatically train the image sample set with zero labeling, and train to obtain the pre-labeling model, so that the manual labeling time and workload are reduced, the labeling efficiency is improved, the investment cost is saved, and the accuracy of pre-labeling is improved.
The certificate pre-labeling method provided by the invention can be applied to the application environment shown in fig. 1, wherein a client (computer device) communicates with a server through a network. The client (computer device) includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, cameras, and portable wearable devices. The server may be implemented as a stand-alone server or as a server cluster consisting of a plurality of servers.
In an embodiment, as shown in fig. 3, a method for pre-labeling a certificate is provided, which mainly includes the following steps S100 to S200:
s100, receiving an image annotation instruction, and acquiring a certificate image in the image annotation instruction.
Understandably, the image annotation instruction is triggered when the collected certificate image needs to be pre-annotated, the image annotation instruction comprises the certificate image, the certificate image is an image containing a certificate, wherein the certificate can be set according to requirements, for example, the certificate is an identity certificate, because the difficulty of photo collection of the identity certificate is high, confidentiality and privacy are involved, the number of collected images of confidentiality or privacy categories is limited, and a certificate pre-annotation model suitable for a certificate annotation area can be trained on a small number of certificate images by using the pre-annotation model training method.
S200, inputting the certificate image into the certificate pre-labeling model trained by the pre-labeling model training method, extracting the learned certificate characteristics through the certificate pre-labeling model, and acquiring a labeling result output by the certificate pre-labeling model according to the certificate characteristics; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
Understandably, only the certificate image needs to be input into the certificate pre-labeling model trained by the pre-labeling model training method, the certificate pre-labeling model is a model which is trained by taking a certificate as the target labeling category, taking the certificate description as the target description, taking preset certificate model performance parameters and a history collected certificate photo as an image sample set, the certificate model performance parameters are related parameters of a preset model applied to the identification certificate, the certificate features are extracted and learned through the certificate pre-labeling model, the certificate features are used as self-adaptive learning target class features in the pre-labeling model training process, therefore, the labeling result identified by the trained certificate pre-labeling model can be obtained, and the region containing the certificate in the certificate image is labeled by the labeling result. The invention realizes that the certificate pre-labeling model is obtained by training in the training mode of the pre-labeling model under the condition of acquiring a small number of certificate photos, can reduce the labeling time of artificial certificate photos and improve the accuracy of pre-labeling.
In an embodiment, a pre-labeled model training device is provided, and the pre-labeled model training device corresponds to the pre-labeled model training method in the embodiment one to one. As shown in fig. 4, the pre-labeling model training device includes an obtaining module 11, a crawling module 12, a recognition module 13, a fine-tuning module 14, a migration module 15, a loss module 16, and a training module 17. The functional modules are explained in detail as follows:
the acquisition module 11 is configured to acquire a target annotation category, a target description corresponding to the target annotation category, a model performance parameter, and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation category;
the crawling module 12 is configured to crawl a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determine the crawled history category as a category to be migrated;
the identification module 13 is configured to search, by using a simulated target identification technology, a to-be-migrated model matched with the to-be-migrated category and the model performance parameter from the target classification identification library, and a target area of each image sample identified by the to-be-migrated model;
a fine-tuning module 14, configured to perform target fine-tuning on all the target regions to obtain fine-tuning regions corresponding to the image samples, and input the image samples, the fine-tuning regions corresponding to the image samples, and the target annotation categories into the model to be migrated; the model to be migrated comprises migration parameters;
the migration module 15 is configured to acquire a target labeling area, corresponding to the image sample, labeled by the model to be migrated through extraction of the class features of the model to be migrated in a self-adaptive learning manner by using a migration learning technology;
a loss module 16, configured to determine a loss value according to the target labeling area and the fine-tuning area corresponding to the image sample;
and the training module 17 is configured to iteratively update the migration parameters of the model to be migrated when the loss value does not reach a preset convergence condition, and record the converged model to be migrated as a pre-labeled model after training until the loss value reaches the preset convergence condition.
For the specific definition of the pre-labeling model training device, reference may be made to the above definition of the pre-labeling model training method, which is not described herein again. The modules in the aforementioned pre-labeling model training device can be implemented in whole or in part by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.
In an embodiment, a certificate pre-labeling device is provided, which corresponds to the certificate pre-labeling method in the above embodiments one to one. As shown in fig. 5, the credential pre-labeling device includes a receiving module 101 and a labeling module 102. The functional modules are explained in detail as follows:
the receiving module 101 is configured to receive an image annotation instruction and acquire a certificate image in the image annotation instruction;
the labeling module 102 is used for inputting the certificate image into the certificate pre-labeling model trained by the pre-labeling model training method, extracting the learned certificate characteristics through the certificate pre-labeling model, and acquiring a labeling result output by the certificate pre-labeling model according to the certificate characteristics; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
For the specific definition of the certificate pre-labeling device, reference may be made to the above definition of the certificate pre-labeling method, which is not described herein again. The modules in the certificate pre-labeling device can be wholly or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.
In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in fig. 6. The computer device includes a processor, a memory, a network interface, and a database connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a pre-labeling model training method, or a certificate pre-labeling method.
In one embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for training the pre-labeling model in the above embodiments when executing the computer program, or implements the method for pre-labeling the certificates in the above embodiments when executing the computer program.
In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which when executed by a processor implements the method for training a pre-annotation model in the above-described embodiments, or which when executed by a processor implements the method for pre-annotation of documents in the above-described embodiments.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, databases, or other media used in embodiments provided herein may include non-volatile and/or volatile memory. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-mentioned functions.
The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present invention, and are intended to be included within the scope of the present invention.

Claims (10)

1. A pre-labeling model training method is characterized by comprising the following steps:
acquiring a target labeling type, target description corresponding to the target labeling type, model performance parameters and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation category;
crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated;
searching a model to be migrated matched with the category to be migrated and the model performance parameter from the target classification recognition library and a target area of each image sample recognized by the model to be migrated by a simulation target recognition technology;
performing target fine adjustment on all the target areas to obtain fine adjustment areas corresponding to the image samples, and inputting the image samples, the fine adjustment areas corresponding to the image samples and the target labeling types into the model to be migrated; the model to be migrated comprises migration parameters;
acquiring a target marking area corresponding to the image sample marked by the model to be migrated by extracting the class characteristics of the model to be migrated in a self-adaptive learning target by applying a migration learning technology;
determining a loss value according to the target labeling area and the fine adjustment area corresponding to the image sample;
and when the loss value does not reach a preset convergence condition, iteratively updating the migration parameters of the model to be migrated until the loss value reaches the preset convergence condition, and recording the converged model to be migrated as a pre-labeled model after training.
2. The method for training the pre-labeled model according to claim 1, wherein the searching the model to be migrated matching the category to be migrated and the model performance parameters from the target classification recognition library by simulating a target recognition technique, and the target area of each image sample recognized by the model to be migrated comprises:
searching a target detection depth model set associated with the category to be migrated in the target classification recognition library; the set of target detection depth models comprises at least one target detection depth model;
inputting each image sample into each target detection depth model, performing target identification on each image sample through each target detection depth model, acquiring each target detection depth model, identifying a target area corresponding to each image sample, and outputting a prediction performance parameter corresponding to each target area;
and determining the target detection depth model matched with the model performance parameters according to all the target areas and the prediction performance parameters, and determining the target detection depth model matched with the model performance parameters as a model to be migrated.
3. The method for training the pre-labeling model according to claim 2, wherein the determining the target detection depth model matching the model performance parameters according to all the target regions and the predicted performance parameters, and determining the target detection depth model matching the model performance parameters as the model to be migrated, comprises:
performing intersection analysis on all the target areas corresponding to the same image samples to obtain intersection areas corresponding to the image samples, and determining the bid rate of each target detection depth model corresponding to each image sample according to the intersection areas corresponding to each image sample and each target area;
determining a target index weight corresponding to the target detection depth model according to the medium rate of the same target detection depth model, and predicting a final prediction performance parameter corresponding to the target detection depth model according to the prediction performance parameter output to the same target detection depth model;
obtaining a winning coefficient corresponding to each target detection depth model according to the model performance parameters, the target index weights corresponding to each target detection depth model and the final prediction performance parameters;
and determining the target detection depth model corresponding to the maximum winning coefficient as the model to be migrated.
4. The method for training the pre-labeled model according to claim 3, wherein the obtaining a bid-closing coefficient corresponding to each of the target detection depth models according to the model performance parameters, the target index weights corresponding to each of the target detection depth models, and the final predicted performance parameters comprises:
dividing the model performance parameters by the final predicted performance parameters corresponding to each of the target detection depth models to obtain performance ratios corresponding to each of the target detection depth models;
and multiplying the performance ratio corresponding to the same target detection depth model by the target exponential weight to obtain the winning bid coefficient corresponding to each target detection depth model.
5. The method for training the pre-labeling model according to claim 1, wherein the inputting the image sample, the fine-tuning region corresponding to the image sample, and the target labeling category into the model to be migrated comprises:
carrying out random data enhancement on the image samples to generate a plurality of enhanced images corresponding to the image samples;
associating the enhanced image with the fine tuning area corresponding to the image sample, and inputting the enhanced image into the model to be migrated;
and taking the target labeling category as a category result in the model to be migrated.
6. A method for pre-labeling a document, comprising:
receiving an image annotation instruction, and acquiring a certificate image in the image annotation instruction;
the certificate image is input into a certificate pre-labeling model trained by the pre-labeling model training method according to any one of claims 1 to 5, the learned certificate features are extracted through the certificate pre-labeling model, and the labeling result output by the certificate pre-labeling model according to the certificate features is obtained; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
7. A pre-labeled model training device, comprising:
the acquisition module is used for acquiring a target labeling type, a target description corresponding to the target labeling type, a model performance parameter and an image sample set; the set of image samples comprises at least one image sample; the image sample corresponds to the target annotation category;
the crawling module is used for crawling a history category similar to the target description in a target classification recognition library by using a text similarity technology, and determining the crawled history category as a category to be migrated;
the identification module is used for searching a model to be migrated which is matched with the category to be migrated and the model performance parameters from the target classification identification library through a simulation target identification technology, and searching a target area of each image sample identified through the model to be migrated;
the fine tuning module is used for performing target fine tuning on all the target areas to obtain fine tuning areas corresponding to the image samples, and inputting the image samples, the fine tuning areas corresponding to the image samples and the target labeling types into the model to be migrated; the model to be migrated comprises migration parameters;
the migration module is used for acquiring a target labeling area corresponding to the image sample marked by the model to be migrated through the extraction of the self-adaptive learning target class characteristics of the model to be migrated by applying a migration learning technology;
the loss module is used for determining a loss value according to the target labeling area and the fine adjustment area corresponding to the image sample;
and the training module is used for iteratively updating the migration parameters of the model to be migrated when the loss value does not reach a preset convergence condition, and recording the converged model to be migrated as a pre-labeled model after training until the loss value reaches the preset convergence condition.
8. A credential pre-labeling device, comprising:
the receiving module is used for receiving the image annotation instruction and acquiring the certificate image in the image annotation instruction;
the marking module is used for inputting the certificate image into the certificate pre-marking model trained by the pre-marking model training method according to any one of claims 1 to 5, extracting the learned certificate characteristics through the certificate pre-marking model and acquiring the marking result output by the certificate pre-marking model according to the certificate characteristics; the certificate pre-labeling model is a model which is trained by taking a certificate as a target labeling category, a certificate description as a target description, preset certificate model performance parameters and a history collected certificate photo as an image sample set; and the marking result marks the certificate-containing area in the certificate image.
9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the pre-labeling model training method of any one of claims 1 to 5 when executing the computer program or the processor implements the document pre-labeling method of claim 6 when executing the computer program.
10. A computer-readable storage medium, in which a computer program is stored, which, when being executed by a processor, carries out the method for training a pre-annotation model according to any one of claims 1 to 5, or which, when being executed by the processor, carries out the method for pre-annotation of documents according to claim 6.
CN202110211382.2A 2021-02-25 2021-02-25 Pre-labeling model training and certificate pre-labeling method, device, equipment and medium Active CN112926654B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110211382.2A CN112926654B (en) 2021-02-25 2021-02-25 Pre-labeling model training and certificate pre-labeling method, device, equipment and medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110211382.2A CN112926654B (en) 2021-02-25 2021-02-25 Pre-labeling model training and certificate pre-labeling method, device, equipment and medium

Publications (2)

Publication Number Publication Date
CN112926654A true CN112926654A (en) 2021-06-08
CN112926654B CN112926654B (en) 2023-08-01

Family

ID=76171813

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110211382.2A Active CN112926654B (en) 2021-02-25 2021-02-25 Pre-labeling model training and certificate pre-labeling method, device, equipment and medium

Country Status (1)

Country Link
CN (1) CN112926654B (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113435463A (en) * 2021-07-27 2021-09-24 平安科技（深圳）有限公司 Object image annotation method, system, equipment and storage medium
CN113469176A (en) * 2021-06-28 2021-10-01 北京有竹居网络技术有限公司 Target detection model training method, target detection method and related equipment thereof
CN113705677A (en) * 2021-08-27 2021-11-26 北京三快在线科技有限公司 Image recognition method and device, electronic equipment and storage medium
CN113806613A (en) * 2021-09-29 2021-12-17 中国平安人寿保险股份有限公司 Training image set generation method and device, computer equipment and storage medium
CN114049499A (en) * 2021-11-18 2022-02-15 南昌黑鲨科技有限公司 Target object detection method, apparatus and storage medium for continuous contour
CN114972883A (en) * 2022-06-17 2022-08-30 平安科技（深圳）有限公司 Target detection sample generation method based on artificial intelligence and related equipment
CN116167010A (en) * 2023-04-25 2023-05-26 南方电网数字电网研究院有限公司 Rapid identification method for abnormal events of power system with intelligent transfer learning capability
WO2024045641A1 (en) * 2022-08-29 2024-03-07 华为技术有限公司 Image annotation method and apparatus

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019200748A1 (en) * 2018-04-17 2019-10-24 平安科技（深圳）有限公司 Transfer learning method, device, computer device, and storage medium
CN110659646A (en) * 2019-08-21 2020-01-07 北京三快在线科技有限公司 Automatic multitask certificate image processing method, device, equipment and readable storage medium
CN111476268A (en) * 2020-03-04 2020-07-31 中国平安人寿保险股份有限公司 Method, device, equipment and medium for training reproduction recognition model and image recognition
WO2021017261A1 (en) * 2019-08-01 2021-02-04 平安科技（深圳）有限公司 Recognition model training method and apparatus, image recognition method and apparatus, and device and medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019200748A1 (en) * 2018-04-17 2019-10-24 平安科技（深圳）有限公司 Transfer learning method, device, computer device, and storage medium
WO2021017261A1 (en) * 2019-08-01 2021-02-04 平安科技（深圳）有限公司 Recognition model training method and apparatus, image recognition method and apparatus, and device and medium
CN110659646A (en) * 2019-08-21 2020-01-07 北京三快在线科技有限公司 Automatic multitask certificate image processing method, device, equipment and readable storage medium
CN111476268A (en) * 2020-03-04 2020-07-31 中国平安人寿保险股份有限公司 Method, device, equipment and medium for training reproduction recognition model and image recognition

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023273570A1 (en) * 2021-06-28 2023-01-05 北京有竹居网络技术有限公司 Target detection model training method and target detection method, and related device therefor
CN113469176A (en) * 2021-06-28 2021-10-01 北京有竹居网络技术有限公司 Target detection model training method, target detection method and related equipment thereof
CN113469176B (en) * 2021-06-28 2023-06-02 北京有竹居网络技术有限公司 Target detection model training method, target detection method and related equipment thereof
CN113435463B (en) * 2021-07-27 2023-06-06 平安科技（深圳）有限公司 Object image labeling method, system, equipment and storage medium
CN113435463A (en) * 2021-07-27 2021-09-24 平安科技（深圳）有限公司 Object image annotation method, system, equipment and storage medium
CN113705677A (en) * 2021-08-27 2021-11-26 北京三快在线科技有限公司 Image recognition method and device, electronic equipment and storage medium
CN113806613A (en) * 2021-09-29 2021-12-17 中国平安人寿保险股份有限公司 Training image set generation method and device, computer equipment and storage medium
CN113806613B (en) * 2021-09-29 2023-07-25 中国平安人寿保险股份有限公司 Training image set generation method, training image set generation device, computer equipment and storage medium
CN114049499A (en) * 2021-11-18 2022-02-15 南昌黑鲨科技有限公司 Target object detection method, apparatus and storage medium for continuous contour
CN114972883A (en) * 2022-06-17 2022-08-30 平安科技（深圳）有限公司 Target detection sample generation method based on artificial intelligence and related equipment
CN114972883B (en) * 2022-06-17 2024-05-10 平安科技（深圳）有限公司 Target detection sample generation method based on artificial intelligence and related equipment
WO2024045641A1 (en) * 2022-08-29 2024-03-07 华为技术有限公司 Image annotation method and apparatus
CN116167010A (en) * 2023-04-25 2023-05-26 南方电网数字电网研究院有限公司 Rapid identification method for abnormal events of power system with intelligent transfer learning capability
CN116167010B (en) * 2023-04-25 2023-12-08 南方电网数字电网研究院有限公司 Rapid identification method for abnormal events of power system with intelligent transfer learning capability

Also Published As

Publication number Publication date
CN112926654B (en) 2023-08-01

Similar Documents

Publication Publication Date Title
CN112926654B (en) Pre-labeling model training and certificate pre-labeling method, device, equipment and medium
US11348249B2 (en) Training method for image semantic segmentation model and server
CN111191791B (en) Picture classification method, device and equipment based on machine learning model
CN112052787B (en) Target detection method and device based on artificial intelligence and electronic equipment
CN111950329A (en) Target detection and model training method and device, computer equipment and storage medium
WO2022042123A1 (en) Image recognition model generation method and apparatus, computer device and storage medium
CN111476268A (en) Method, device, equipment and medium for training reproduction recognition model and image recognition
CN111832581B (en) Lung feature recognition method and device, computer equipment and storage medium
WO2022134805A1 (en) Document classification prediction method and apparatus, and computer device and storage medium
CN114549913B (en) Semantic segmentation method and device, computer equipment and storage medium
CN113204660B (en) Multimedia data processing method, tag identification device and electronic equipment
US11893773B2 (en) Finger vein comparison method, computer equipment, and storage medium
CN113806613B (en) Training image set generation method, training image set generation device, computer equipment and storage medium
CN111898561A (en) Face authentication method, device, equipment and medium
CN111666932A (en) Document auditing method and device, computer equipment and storage medium
CN115063664A (en) Model learning method, training method and system for industrial vision detection
CN115795355A (en) Classification model training method, device and equipment
CN113283388B (en) Training method, device, equipment and storage medium of living body face detection model
CN115374278A (en) Text processing model distillation method, device, computer equipment and medium
CN114547257A (en) Class matching method and device, computer equipment and storage medium
CN115840817A (en) Information clustering processing method and device based on contrast learning and computer equipment
CN113902959A (en) Image recognition method and device, computer equipment and storage medium
CN111091198A (en) Data processing method and device
CN112699762A (en) Food material identification method suitable for embedded equipment
CN111582404A (en) Content classification method and device and readable 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