Bank image recognition model training method and device, equipment and storage medium

Abstract

Embodiments of the present application disclose a bank image recognition model training method, device and equipment and a storage medium. The method comprises: receiving first bank image data manually labeled and second bank image data labeled by multiple models; wherein the multiple models are at least two single models that have completed training; inputting training sample sets generated by the first bank image data and the second bank image data in proportion to a bank image recognition model; the bank image recognition model trains by forward propagation before the training sample sets are divided into the first bank image data and the second bank image data; and the bank image recognition model is obtained when the bank image recognition model converges. The method provided by the present application improves the efficiency of data image recognition and has high model recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image recognition, and particularly relates to a bank image recognition model training method and device, equipment and a storage medium. BACKGROUND

[0002] As a kind of archive record certificate, image is widely used in the operation of various industries, such as in bank business. However, image data is large in quantity and various in type, and it is a large workload to sort, audit and archive it, especially when sorting, auditing and archiving it manually, there are many work difficulties.

[0003] Using algorithm to identify image to realize sorting, auditing and archiving is one of the effective ways of practice. However, to achieve high accuracy in image recognition using algorithm, manual annotation is still needed. However, in the big data scenario, manual annotation is costly and inefficient. Therefore, semi-supervised learning is a cost-effective choice. The existing semi-supervised learning is the well-known UDA algorithm. This method belongs to online training, which requires a large amount of GPU resources to improve the recognition effect of the algorithm, and the resource consumption is high. SUMMARY

[0004] The embodiments of the present application propose a bank image recognition model training method, device, equipment and storage medium, which solves the problems of high cost and low efficiency of manual annotation and high resource consumption in existing data image recognition.

[0005] The embodiments of the present application provide a bank image recognition model training method, which comprises:

[0006] Receiving manually annotated first bank image data and multi-model annotated second bank image data; wherein the multi-model is a set of at least two single models completed training;

[0007] The first bank image data and the second bank image data are proportionally generated training sample set and input into the bank image recognition model;

[0008] The bank image recognition model trains the training sample set by forward propagation before separating the first bank image data and the second bank image data;

[0009] When the bank image recognition model converges, the bank image recognition model is obtained.

[0010] Further, the multi-model annotated second bank image data is specifically:

[0011] Inputting unannotated bank image data into at least two single models completed training one by one;

[0012] The completed single model labels the input unlabeled bank image data, and respectively outputs a labeling result;

[0013] The second bank image data labeled by the multiple models is obtained through all the labeling results.

[0014] Further, the second bank image data labeled by the multiple models is obtained through all the labeling results, specifically:

[0015] All the labeling results are normalized to obtain the second bank image data labeled by the multiple models.

[0016] Further, the method further comprises:

[0017] Obtaining a completed single model, specifically:

[0018] Unlabeled bank image data is input into at least two initial models of single models one by one, and the initial model of the single model labels the unlabeled bank image data to obtain a single model labeling data result;

[0019] According to the single model labeling data result and the artificially labeled first bank image data, all single models are trained to obtain a completed single model.

[0020] Further, the training of all single models according to the single model labeling data result and the artificially labeled first bank image data to obtain a completed single model comprises:

[0021] Obtaining labeling difficult samples and labeling non-difficult samples in the initial model labeling data result of the single model;

[0022] The labeling difficult samples are corrected by artificial labeling to obtain a correction result;

[0023] The artificially labeled first bank image data, the correction result and the labeling non-difficult samples are respectively input into the initial model of each single model for training to obtain a completed single model.

[0024] Further, the obtaining of the labeling difficult samples and the labeling non-difficult samples in the initial model labeling data result of the single model comprises:

[0025] Obtaining labeling results respectively output by all single models for the same bank image data;

[0026] The prediction success rate is obtained by the ratio of the number of single models with correct labeling results in all single models to the total number of all single models;

[0027] comparing the predicted success rate with a preset probability threshold value;

[0028] If the predicted success rate is greater than or equal to the preset probability threshold value, the bank image data is labeled as a non-difficult sample; if the predicted success rate is less than the preset probability threshold value, the bank image data is labeled as a difficult sample.

[0029] Further, the method further comprises:

[0030] According to the naming of the bank image data, different image data is determined.

[0031] Further, the method further comprises:

[0032] An initial model of a single model is obtained, specifically:

[0033] At least two different backbone networks are selected;

[0034] The different backbone networks are trained by the artificially labeled first bank image data to obtain the initial model of the single model.

[0035] Further, the different backbone networks are backbone networks that are different in stacking mode, depth, width, module and level.

[0036] Further, the bank image recognition model is trained by forward propagation before the training sample set is divided according to the first bank image data and the second bank image data, comprising:

[0037] The first bank image data is forward propagated to obtain a first loss function; the second bank image data is forward propagated to obtain a second loss function;

[0038] The first loss function is multiplied by a corresponding first proportion to obtain a first multiplication result; the second loss function is multiplied by a corresponding second proportion to obtain a second multiplication result; the sum of the first proportion and the second proportion is 1;

[0039] The first multiplication result and the second multiplication result are added to obtain a loss function of the bank image recognition model.

[0040] The embodiment of the application also provides a bank image recognition model training device, the device comprising:

[0041] A receiving module is configured to receive artificially labeled first bank image data and multi-model labeled second bank image data; wherein the multi-model is a set of at least two trained single models;

[0042] The training sample set generation module is configured to input a training sample set generated by the first bank image data and the second bank image data in proportion into the bank image recognition model.

[0043] The training module is configured to train the bank image recognition model by forward propagation of the training sample set before the training sample set is divided according to the first bank image data and the second bank image data.

[0044] The acquisition module is configured to acquire the bank image recognition model when the training of the bank image recognition model converges.

[0045] The embodiment of the application further provides a bank image recognition model training device, including a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the following steps:

[0046] The first bank image data manually labeled by a person and the second bank image data labeled by multiple models are received, and the multiple models are at least two single models that are trained.

[0047] The first bank image data and the second bank image data are input into the bank image recognition model in proportion to generate a training sample set.

[0048] The bank image recognition model is trained by forward propagation of the training sample set before the training sample set is divided according to the first bank image data and the second bank image data.

[0049] The bank image recognition model is obtained when the training of the bank image recognition model converges.

[0050] The embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the following steps:

[0051] The first bank image data manually labeled by a person and the second bank image data labeled by multiple models are received, and the multiple models are at least two single models that are trained.

[0052] The first bank image data and the second bank image data are input into the bank image recognition model in proportion to generate a training sample set.

[0053] The bank image recognition model is trained by forward propagation of the training sample set before the training sample set is divided according to the first bank image data and the second bank image data.

[0054] The bank image recognition model is obtained when the training of the bank image recognition model converges.

[0055] The training method for the bank image recognition model provided in this invention achieves automatic annotation of image data by combining the differences between individual trained models to form multiple models. Then, combined with manual annotation, the bank image recognition model is trained, which reduces the cost and resource consumption of manual annotation in data image recognition, improves the efficiency of data image recognition, and the model has high recognition accuracy. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] in:

[0058] Figure 1 This is a diagram illustrating the application environment of a training method for a bank image recognition model provided in one embodiment.

[0059] Figure 2 A flowchart illustrating a training method for a bank image recognition model provided in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram illustrating the process of using multi-model annotation for second bank image data, as provided in an embodiment of the present invention.

[0061] Figure 4 A schematic diagram illustrating the process of training a bank image recognition model using separate forward propagation, as provided in an embodiment of the present invention;

[0062] Figure 5 A schematic diagram of the structure of a training device for a bank image recognition model provided in an embodiment of the present invention;

[0063] Figure 6 A schematic diagram of the structure of a training device for a bank image recognition model provided in an embodiment of the present invention;

[0064] Figure 7 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Figure 1 This is a diagram illustrating the application environment of a training method for a bank image recognition model provided in one embodiment. (Refer to...) Figure 1 The training method for this bank image recognition model is applied to a bank image recognition system. This system includes a server 200 and a cluster of terminal devices. The cluster of terminal devices may include one or more terminal devices; in this embodiment, the number of terminal devices is not limited. Figure 1 As shown, multiple terminal devices may specifically include terminal device 1, terminal device 2, terminal device 3, ..., terminal device n; terminal device 1, terminal device 2, terminal device 3, ..., terminal device n are all connected to server 200 through network 300, so that each terminal device can interact with server 200 through network 300. Terminal device 1, terminal device 2, terminal device 3, ..., terminal device n can be: smartphones, tablets, laptops, desktop computers, smart TVs, and other smart terminals. Server 200 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0067] Terminal devices 1, 2, 3, ..., n can be used to collect image data generated during banking operations and send the collected image data to server 200. Server 200 receives manually labeled first bank image data and multi-model labeled second bank image data; wherein the multi-model refers to at least two sets of single models that have completed training; a training sample set is generated from the first and second bank image data in proportion and input into the bank image recognition model; the bank image recognition model trains by forward propagating the training sample set separately for the first and second bank image data; when the bank image recognition model converges, the bank image recognition model is obtained. The manually labeled first bank image data can be labeled on terminal devices 1, 2, 3, ..., n, or on server 200.

[0068] The method provided in this invention uses the differences between trained individual models to form multiple models, thereby achieving automatic annotation of image data. Then, combined with manual annotation, the bank image recognition model is trained, which reduces the cost and resource consumption of manual annotation in data image recognition, improves the efficiency of data image recognition, and achieves high model recognition accuracy.

[0069] like Figure 2 The diagram shown illustrates a training method for a bank image recognition model according to an embodiment of the present invention. This method can be applied to both terminal devices and servers; this embodiment uses server 200 as an example. The method includes:

[0070] Step S101: Receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is a set of at least two trained single models;

[0071] Specifically, in this embodiment, various bank image data are first collected through a terminal device. A portion of the collected bank image data is manually annotated to obtain manually annotated first bank image data; another portion is input into a multi-model consisting of a set of single models that have completed training for annotation to obtain multi-model annotated second bank image data.

[0072] It should be noted that in this application, a single model consists of different backbone networks. At least two different backbone networks can be selected according to actual needs. The backbone networks can be ResNet, SENet, InceptionNet, EfficientNet, ConvNeXt, Vision Transformer, etc.

[0073] It should be noted that, in this embodiment, when annotating the collected bank image data, it is first annotated according to the actual business of the bank image data, that is, according to the business type of the bank image data, such as liability business, loan business, securities investment business, wealth management business, transaction clearing business, payment and settlement business, bank card business, agency business, custody business, and electronic banking business. Then, the bank image data in each business type is annotated according to the materials required in the process. For example, the bank image data collected in the loan business is annotated according to the "Agreement", "Commitment Letter", ID card or valid identity certificate, household registration book, temporary residence permit or valid residence certificate submitted by the loan applicant, retirement certificate, valid marriage certificate, down payment invoice (receipt), academic degree certificate, professional title certificate, proof of owning housing, proof of renting public housing, "Authorization Letter", "Loan Application Form", etc. After annotation, the bank image data is sorted, reviewed and archived according to different business types and the types of materials in the business types to facilitate the management of bank image data.

[0074] Step S102: The first bank image data and the second bank image data are proportionally used to generate a training sample set, which is then input into the bank image recognition model.

[0075] Specifically, the bank image recognition model can be selected from existing models according to the corresponding needs, or it can be constructed according to the actual situation. In this embodiment, a neural network composed of convolutional layers, bn layers and activation layers is used as the bank image recognition model.

[0076] A training sample set is generated from bank image data annotated manually and by multiple models according to business type and material type within that business type. This training sample set includes a training set, a test set, and a validation set. The training set is then input into the bank image recognition model for training. During training, the batch size is set to 32, and each batch consists of first-level bank image data and second-level bank image data annotated by multiple models, in a 1:3 ratio. After training, the model is tested and optimized using the test set, and then further optimized using the validation set.

[0077] Step S103: The bank image recognition model trains by forward propagating the training sample set separately according to the first bank image data and the second bank image data.

[0078] Specifically, the manual annotation results and multi-model annotation results obtained in step S101 are obtained according to the proportion in step S102. Then, the manual annotation results and multi-model annotation results can be forward propagated separately to train, test and optimize the bank image recognition model.

[0079] Step S104: When the bank image recognition model is trained to convergence, the bank image recognition model is obtained.

[0080] Specifically, in this embodiment, when the image is trained to convergence according to step S103, a bank image recognition model is obtained.

[0081] The training method for the bank image recognition model provided in this invention achieves automatic annotation of image data by combining the differences between individual trained models to form multiple models. Then, combined with manual annotation, the bank image recognition model is trained, which reduces the cost and resource consumption of manual annotation in data image recognition, improves the efficiency of data image recognition, and the model has high recognition accuracy.

[0082] In some embodiments, such as Figure 3 The diagram shown is a flowchart illustrating a multi-model annotation process for second bank image data according to an embodiment of the present invention. Specifically:

[0083] Step S201: Input the unlabeled bank image data into at least two trained individual models one by one;

[0084] Step S202: The trained single model annotates the input unannotated bank image data, and the trained single model outputs the annotation results respectively;

[0085] Step S203: Obtain the second bank image data annotated by the multi-model through all the annotation results.

[0086] Specifically, in this embodiment, unannotated bank video data is input into at least two trained individual models. All individual models simultaneously annotate the bank video data for banking business type and material type, with each model outputting an annotation result. Then, the second bank video data annotated by multiple models is obtained from the annotation results of each model.

[0087] In some embodiments, obtaining the second bank image data annotated by the multi-model method using all the annotation results specifically involves:

[0088] All the annotation results are normalized to obtain the second bank image data annotated by the multi-model method.

[0089] In some embodiments, the method further includes: determining different image data based on the naming of the bank image data.

[0090] Specifically, in this embodiment, the annotation results output by all individual models are normalized to obtain the prediction results of the same image data by different individual models, and then the final annotation result of the image data is obtained through processing.

[0091] It should be noted that, in this embodiment, the naming of the bank data is used to determine the input of the same bank image data to different individual models and the prediction results of the same bank image data through different individual models. Then, the annotation results obtained by the same bank image data through different individual models are normalized. In this embodiment, softmax is selected to process the annotation results of different models to obtain the final multi-model annotated second bank image data.

[0092] In some embodiments, the method further includes: obtaining a single model that has completed training, specifically:

[0093] Unlabeled bank image data is input one by one into the initial model of at least two individual models. The initial model of each individual model labels the unlabeled bank image data to obtain the labeling data result of the individual model.

[0094] Based on the individual model annotation data results and the manually annotated first bank image data, all individual models are trained to obtain a single model that has completed training.

[0095] Specifically, in this embodiment, unlabeled bank image data is first input into the initial model of a single model to obtain the labeling data results of the single model containing business type and material type. Then, training samples are generated by manually labeled bank image data with business type and material type and the labeling data results of the single model. The single model is trained, tested and optimized until the error of the labeling results of the initial model of the single model reaches below 20%. At this point, the model training can be stopped and the single model that has been trained is obtained.

[0096] In some embodiments, training all individual models based on the individual model annotation data results and the manually annotated first bank image data to obtain a trained individual model includes:

[0097] Obtain the labeled hard samples and labeled non-hard samples from the initial model annotation data results of a single model;

[0098] The difficult-to-annotate samples were corrected through manual annotation to obtain the correction results;

[0099] The manually labeled first bank image data, the correction results, and the labeled non-difficult samples are respectively input into the initial model of each individual model for training, to obtain a trained individual model.

[0100] Preferably, obtaining the difficult and easy labeled samples from the initial model annotation data results of a single model includes:

[0101] Obtain the annotation results of the same bank image data through all individual models respectively;

[0102] The prediction success rate is obtained by comparing the number of correctly labeled individual models in the annotation results with the total number of all individual models.

[0103] The prediction success rate is compared with a preset probability threshold.

[0104] If the prediction success rate is greater than or equal to the preset probability threshold, the bank image data is labeled as a non-difficult sample; if the prediction success rate is less than the preset probability threshold, the bank image data is labeled as a difficult sample.

[0105] Specifically, in this embodiment, from the single model annotation data results output by the initial model of a single model, difficult-to-annotate samples and easy-to-annotate samples are obtained. A preset probability threshold is set to 0.5. Samples with a prediction success rate greater than or equal to 0.5 are classified as easy-to-annotate samples; samples with a prediction success rate less than 0.5 are classified as difficult-to-annotate samples. Then, all or part of the difficult-to-annotate samples are manually annotated. In this embodiment, 10% of the difficult-to-annotate samples are manually corrected to obtain the corrected results. Then, a training sample set is generated using the manually annotated first bank image data, the corrected results, and the easy-to-annotate samples. This training sample set is input into the initial model of the single model for training, resulting in a trained single model.

[0106] In some embodiments, the method further includes:

[0107] To obtain the initial model for a single model, specifically:

[0108] Choose at least two different backbone networks;

[0109] Preferably, the different backbone networks are backbone networks with different stacking methods, depths, widths, modules, and levels.

[0110] The different backbone networks are trained using the manually labeled first bank image data to obtain the initial model of a single model.

[0111] Specifically, in this embodiment, manually labeled bank image data is divided into training, validation, and test sets in a 3:1:1 ratio and input into different backbone networks. Each backbone network is trained until it converges, resulting in the initial model of a single model. Cross Entropy is set as the training loss function, and the evaluation metric is the average accuracy. Different backbone networks are trained with parameters including a batch size of 32, 90 training epochs, and image enhancements such as color enhancement, rotation, and small-area blurring. The learning rate, image resolution, and other parameters are fine-tuned for the initial model training of each individual model. The accuracy is defined as the ratio of correctly identified images to the total number of correctly identified images.

[0112] In some embodiments, such as Figure 4 The diagram shown is a flowchart illustrating the training process of a bank image recognition model using separate forward propagation, according to an embodiment of the present invention. Step S103 includes:

[0113] Step S1031: Forward propagate the first bank image data to obtain a first loss function; forward propagate the second bank image data to obtain a second loss function;

[0114] Step S1032: The first loss function is multiplied by its corresponding first proportion to obtain a first multiplication result; the second loss function is multiplied by its corresponding second proportion to obtain a second multiplication result; the sum of the first proportion and the second proportion is 1;

[0115] Step S1033: Add the first multiplication result and the second multiplication result to obtain the loss function of the bank image recognition model.

[0116] Specifically, in this embodiment, the first loss function in the manually labeled forward propagation is defined as... The second loss function for forward propagation of the second bank image data is defined as follows: The loss function of the bank image recognition model is then... for:

[0117]

[0118] The first proportion corresponding to the first loss function is The second proportion corresponding to the second loss function is In this embodiment , .

[0119] It should be noted that the first percentage is The second proportion is It can be adjusted according to different types of banking business and can change dynamically; that is, the first proportion is... The second proportion is Adjustments will be made dynamically based on the type of banking business.

[0120] like Figure 5 The diagram shown is a structural schematic of a training device for a bank image recognition model provided in an embodiment of the present invention. The device includes:

[0121] The receiving module 501 is used to receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is at least two sets of individual models that have been trained.

[0122] The training sample set generation module 502 is used to generate a training sample set by proportionally combining the first bank image data and the second bank image data and input it into the bank image recognition model;

[0123] Training module 503 is used for the bank image recognition model to train the training sample set by forward propagation separately according to the first bank image data and the second bank image data;

[0124] The acquisition module 504 is used to acquire the bank image recognition model when the training of the bank image recognition model has converged.

[0125] In some embodiments, the receiving module 501 includes:

[0126] Input model, used to input unlabeled bank image data one by one into at least two trained individual models;

[0127] The annotation module is used by the trained single model to annotate the input unannotated bank image data, and the trained single model outputs the annotation results respectively;

[0128] The second bank image data acquisition module is used to acquire the second bank image data annotated by the multi-model through all the annotation results.

[0129] In some embodiments, the second bank image data acquisition module is further configured to normalize all the annotation results to obtain the second bank image data annotated by the multi-model method.

[0130] In some embodiments, the apparatus further includes:

[0131] The single model acquisition module is used to acquire a single model that has completed training, specifically:

[0132] Unlabeled bank image data is input one by one into the initial model of at least two individual models. The initial model of each individual model labels the unlabeled bank image data to obtain the labeling data result of the individual model.

[0133] Based on the individual model annotation data results and the manually annotated first bank image data, all individual models are trained to obtain a single model that has completed training.

[0134] In some embodiments, the single model acquisition module is further configured to acquire labeled difficult samples and labeled non-difficult samples from the initial model annotation data results of a single model;

[0135] The difficult-to-annotate samples were corrected through manual annotation to obtain the correction results;

[0136] The manually labeled first bank image data, the correction results, and the labeled non-difficult samples are respectively input into the initial model of each individual model for training, to obtain a trained individual model.

[0137] In some embodiments, the single model acquisition module is further configured to acquire the annotation results output by all individual models for the same bank image data;

[0138] The prediction success rate is obtained by comparing the number of correctly labeled individual models in the annotation results with the total number of all individual models.

[0139] The prediction success rate is compared with a preset probability threshold.

[0140] If the prediction success rate is greater than or equal to the preset probability threshold, the bank image data is labeled as a non-difficult sample; if the prediction success rate is less than the preset probability threshold, the bank image data is labeled as a difficult sample.

[0141] In some embodiments, the apparatus further includes:

[0142] The judgment module is used to determine different image data based on the naming of the bank image data.

[0143] In some embodiments, the apparatus further includes:

[0144] The initial model acquisition module for a single model is used to acquire the initial model for a single model, specifically:

[0145] Choose at least two different backbone networks;

[0146] The different backbone networks are trained using the manually labeled first bank image data to obtain the initial model of a single model.

[0147] The different backbone networks are backbone networks with different stacking methods, depths, widths, modules, and levels.

[0148] In some embodiments, the training module 503 includes:

[0149] The first acquisition module is used to forward propagate the first bank image data to obtain a first loss function; and to forward propagate the second bank image data to obtain a second loss function.

[0150] The multiplication module is used to multiply the first loss function with its corresponding first proportion to obtain a first multiplication result; and to multiply the second loss function with its corresponding second proportion to obtain a second multiplication result; the sum of the first proportion and the second proportion is 1;

[0151] The second acquisition module is used to add the first multiplication result and the second multiplication result to obtain the loss function of the bank image recognition model.

[0152] For further details regarding the implementation of the above technical solutions by each module in the training device for the bank image recognition model, please refer to the description in the training method for the bank image recognition model provided above, which will not be repeated here.

[0153] In some embodiments, such asFigure 6 The diagram shown is a structural schematic of a training device for a bank image recognition model according to an embodiment of the present invention. The device includes a memory 601 and a processor 602. The memory 601 stores a computer program. When the computer program is executed by the processor 602, the processor 602 performs the following steps:

[0154] Receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is at least two sets of single models that have been trained.

[0155] The first bank image data and the second bank image data are proportionally used to generate a training sample set, which is then input into the bank image recognition model.

[0156] The bank image recognition model trains by forward propagating the training sample set separately according to the first bank image data and the second bank image data.

[0157] When the bank image recognition model is trained to convergence, the bank image recognition model is obtained.

[0158] For further details regarding the implementation of the above technical solution by the processor 602 in the training device for the bank image recognition model, please refer to the description in the training method for the bank image recognition model provided above, which will not be repeated here.

[0159] The processor 602 can also be called a CPU (Central Processing Unit). The processor 602 may be an integrated circuit chip with signal processing capabilities. The processor 602 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or the processor 602 can be any conventional processor.

[0160] In some embodiments, such as Figure 7The diagram illustrates the structure of a computer-readable storage medium according to an embodiment of the present invention. The storage medium stores a readable computer program 701. This computer program 701 can be stored in the storage medium in the form of a software product, and includes several instructions to cause a computer device (which may be a personal computer, a server machine, or a network device, etc.) or a processor to perform the following steps:

[0161] Receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is at least two sets of single models that have been trained.

[0162] The first bank image data and the second bank image data are proportionally used to generate a training sample set, which is then input into the bank image recognition model.

[0163] The bank image recognition model trains by forward propagating the training sample set separately according to the first bank image data and the second bank image data.

[0164] When the bank image recognition model is trained to convergence, the bank image recognition model is obtained.

[0165] The aforementioned storage media include: USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), and other media that can store program code, or terminal devices such as computers, servers, mobile phones, and tablets.

[0166] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can 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 various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0167] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0168] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A training method for a bank image recognition model, characterized in that, The method includes: Receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is at least two sets of single models that have been trained. The first bank image data and the second bank image data are proportionally used to generate a training sample set, which is then input into the bank image recognition model. The bank image recognition model trains by forward propagating the training sample set separately according to the first bank image data and the second bank image data. When the bank image recognition model is trained to convergence, the bank image recognition model is obtained; Specifically, the second bank image data annotated by the multi-model method includes: Input the unlabeled bank image data one by one into at least two trained individual models; The trained single model annotates the input unannotated bank image data, and the trained single model outputs the annotation results respectively; The second bank image data annotated by the multi-model was obtained by using all the annotation results; The method further includes: To obtain a single model that has completed training, specifically: Unlabeled bank image data is input one by one into the initial model of at least two individual models. The initial model of each individual model labels the unlabeled bank image data to obtain the labeling data result of the individual model. Based on the individual model annotation data results and the manually annotated first bank image data, all individual models are trained to obtain a single model that has completed training.

2. The training method for the bank image recognition model according to claim 1, characterized in that, The process of obtaining the second bank image data annotated by the multi-model method using all the annotation results specifically involves: All the annotation results are normalized to obtain the second bank image data annotated by the multi-model method.

3. The training method for the bank image recognition model according to claim 1, characterized in that, The step of training all individual models based on the labeled data results of the individual models and the manually labeled first bank image data to obtain a single model that has been trained includes: Obtain the labeled hard samples and labeled non-hard samples from the initial model annotation data results of a single model; The difficult-to-annotate samples were corrected through manual annotation to obtain the correction results; The manually labeled first bank image data, the correction results, and the labeled non-difficult samples are respectively input into the initial model of each individual model for training, to obtain a trained individual model.

4. The training method for the bank image recognition model according to claim 3, characterized in that, The process of obtaining the hard-to-label samples and easy-to-label samples from the initial model annotation data results of a single model includes: Obtain the annotation results of the same bank image data through all individual models respectively; The prediction success rate is obtained by comparing the number of correctly labeled individual models in the annotation results with the total number of all individual models. The prediction success rate is compared with a preset probability threshold. If the prediction success rate is greater than or equal to the preset probability threshold, the bank image data is labeled as a non-difficult sample; if the prediction success rate is less than the preset probability threshold, the bank image data is labeled as a difficult sample.

5. The training method for the bank image recognition model according to claim 1, characterized in that, The method further includes: Different image data are identified based on the naming of the bank's image data.

6. The training method for the bank image recognition model according to claim 1, characterized in that, The method further includes: To obtain the initial model for a single model, specifically: Choose at least two different backbone networks; The different backbone networks are trained using the manually labeled first bank image data to obtain the initial model of a single model.

7. The training method for the bank image recognition model according to claim 6, characterized in that, The different backbone networks are backbone networks with different stacking methods, depths, widths, modules, and levels.

8. The training method for the bank image recognition model according to claim 1, characterized in that, The bank image recognition model trains by forward propagating the training sample set separately according to the first bank image data and the second bank image data, including: The first bank image data is forward-propagated to obtain a first loss function; the second bank image data is forward-propagated to obtain a second loss function. The first loss function is multiplied by its corresponding first proportion to obtain a first multiplication result; the second loss function is multiplied by its corresponding second proportion to obtain a second multiplication result; the sum of the first proportion and the second proportion is 1; The loss function of the bank image recognition model is obtained by adding the first multiplication result and the second multiplication result together.

9. A training device for a bank image recognition model, characterized in that, The device includes: The receiving module is used to receive manually labeled first bank image data and multi-model labeled second bank image data; wherein, the multi-model is at least two sets of individual models that have been trained. The training sample set generation module is used to generate a training sample set by proportionally combining the first bank image data and the second bank image data and input it into the bank image recognition model; The training module is used to train the bank image recognition model by forward propagating the training sample set separately according to the first bank image data and the second bank image data. The acquisition module is used to acquire the bank image recognition model when the training of the bank image recognition model has converged. Specifically, the second bank image data annotated by the multi-model method involves: inputting unannotated bank image data one by one into at least two trained individual models; annotating the input unannotated bank image data using the trained individual models; and outputting annotation results using the annotation results. The second bank image data annotated by the multi-model method is then obtained. The device is further configured to: obtain a single model that has completed training, specifically by: inputting unlabeled bank image data one by one into the initial model of at least two single models, wherein the initial model of the single model labels the unlabeled bank image data, and obtains the labeling data result of the single model; and training all the single models based on the labeling data result of the single model and the manually labeled first bank image data, thereby obtaining a single model that has completed training.

10. A training device for a bank image recognition model, comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 8.

Images