Target detection model training method and device, equipment, medium and program product

Abstract

The application provides a target detection model training method and device, equipment, medium and program product, and relates to the field of artificial intelligence. The method comprises the following steps: obtaining an original sample data set; performing M kinds of distortion processing on the original face image of a sample user according to interference factors of an image environment collected by a client to obtain an initial sample data set; obtaining an initial training sample data set from the initial sample data set; performing at least one face local image splicing processing on at least part of sample data of a sample user in the initial training sample data set to obtain a training sample data set; training a target detection model using the training sample set until the model converges to obtain a trained target detection model, and the trained target detection model is used to predict the user identifier of a target user in a bank transaction system according to the face image of the target user. The application improves the identity verification success rate.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, device, medium, and program product for training an object detection model. Background Technology

[0002] When conducting banking transactions through a bank's transaction system, identity verification is often required to ensure that the person conducting the transaction is indeed the account holder.

[0003] Currently, the common method for identity verification is facial recognition. This involves using a neural network model to perform facial recognition on the collected user image and outputting the similarity between the collected user image and a pre-stored reference image of the user to complete the user's identity verification.

[0004] However, when performing facial recognition, there are often situations where the person is identified but the identity verification fails. In other words, the error rate of facial recognition is relatively large, resulting in a low success rate of identity verification and a poor user experience. Summary of the Invention

[0005] This application provides a method, apparatus, device, medium, and program product for training an object detection model to solve the problem of low identity verification success rate.

[0006] Firstly, this application provides a method for training an object detection model, including:

[0007] Obtain the original sample dataset; the original sample dataset includes the original sample data of N sample users, the sample data including: the original face images of the sample users collected by the client of the bank transaction system, and the identifier of the sample users in the bank transaction system;

[0008] Based on the interference factors in the image acquisition environment of the client, the original face images of the sample users are subjected to M types of distortion processing to obtain an initial sample dataset. The initial sample dataset includes at least: face images of N sample users after M types of distortion processing, and the identifier of the sample users in the bank transaction system; the distortion processing includes: illumination distortion processing and / or geometric distortion processing.

[0009] The initial training sample dataset is obtained from the initial sample dataset;

[0010] At least one type of local face image stitching process is performed on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset.

[0011] The target detection model is trained using the training sample set until the model converges, resulting in a trained target detection model. The trained target detection model is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image.

[0012] Optionally, the initial sample dataset is obtained by performing M types of distortion processing on the original face images of the sample users based on interference factors in the image acquisition environment of the client, including:

[0013] Based on the interference factors in the image environment acquired by the client and the image parameters of the original face image of the sample user, M target distortion processing methods are determined from the distortion processing library;

[0014] Based on the determined M types of target distortion processing methods, the original face images of the sample users are subjected to distortion processing to obtain the initial sample dataset.

[0015] Optionally, before performing distortion processing on the original face image of the sample user according to the determined M target distortion processing methods, the method further includes:

[0016] Based on the image environment acquired by at least one client of the bank transaction system, determine the parameter range used by each distortion processing method in the distortion processing library.

[0017] Optionally, before performing local image stitching processing on at least a portion of the sample data from the initial training sample dataset to obtain the training sample dataset, the method further includes:

[0018] Based on the image parameters of the face images in the remaining sample data of the sample users in the initial training sample dataset, determine whether to perform secondary distortion processing on the face image;

[0019] If so, then the secondary distortion processing method is determined based on the M types of distortion processing used on the face image;

[0020] Based on the determined secondary distortion processing method, the face image is subjected to secondary distortion processing.

[0021] Optionally, before performing M types of distortion processing on the original face image of the sample user based on interference factors in the image acquisition environment of the client, the method further includes:

[0022] The value of M is determined based on the maximum number of images required for the partial facial image stitching process.

[0023] Optionally, the method further includes:

[0024] The weight parameters of the trained object detection model are converted to a different format to compress the trained object detection model.

[0025] Secondly, this application provides an authentication method, including:

[0026] Receive transaction requests triggered by the target user through the bank's transaction system client;

[0027] In response to the transaction request, the client is controlled to collect the facial image of the target user;

[0028] Based on the facial image of the target user, a target detection model trained by the method described in any of the first aspects is used to predict the user identifier of the target user in the bank transaction system.

[0029] Based on the user identifier of the target user in the bank's transaction system, determine whether the target user is a legitimate user;

[0030] If the target user is determined to be a legitimate user, then the transaction request is processed;

[0031] The processing result of the transaction request is returned to the client.

[0032] Thirdly, this application provides a target detection model training device, comprising:

[0033] The acquisition module is used to acquire the original sample dataset; the original sample dataset includes the original sample data of N sample users, and the sample data includes: the original face images of the sample users collected by the client of the bank transaction system, and the identifier of the sample users in the bank transaction system;

[0034] The first processing module is used to perform M types of distortion processing on the original face images of the sample users based on the interference factors of the image acquisition environment of the client, to obtain an initial sample dataset. The initial sample dataset includes at least: face images of N sample users after M types of distortion processing, and the identifier of the sample users in the bank transaction system; the distortion processing includes: illumination distortion processing and / or geometric distortion processing.

[0035] The second processing module is used to obtain the initial training sample dataset from the initial sample dataset;

[0036] The third processing module is used to perform at least one type of local face image stitching processing on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset.

[0037] The training module is used to train the target detection model using the training sample set until the model converges, thereby obtaining the trained target detection model. The trained target detection model is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image.

[0038] Fourthly, this application provides an authentication device, comprising:

[0039] The receiving module is used to receive transaction requests triggered by the target user through the bank's transaction system client.

[0040] The control module is used to control the client to collect the facial image of the target user in response to the transaction request;

[0041] The identification module is used to predict the user identifier of the target user in the bank transaction system based on the facial image of the target user and using the target detection model trained by the method described in any one of the first aspects.

[0042] The determination module is used to determine whether the target user is a legitimate user based on the user identifier of the target user in the bank transaction system;

[0043] The processing module is used to process the transaction request after determining that the target user is a legitimate user;

[0044] The output module is used to return the processing result of the transaction request to the client.

[0045] Fifthly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0046] The memory stores computer-executed instructions;

[0047] The processor executes computer execution instructions stored in the memory to implement the target detection model training method or the authentication method as described in any one of the first or second aspects.

[0048] Sixthly, this application provides a computer-readable storage medium, comprising: computer-executable instructions stored in the computer-readable storage medium, wherein the computer-executable instructions, when executed by a processor, are used to implement the target detection model training method or the authentication method as described in any one of the first or second aspects.

[0049] In a seventh aspect, this application provides a computer program product, including a computer program that, when executed by the processor, implements the target detection model training method or the authentication method as described in any one of the first or second aspects.

[0050] Eighthly, this application provides a chip on which a computer program is stored. When the computer program is executed by the chip, it implements the target detection model training method or the authentication method as described in any one of the first or second aspects.

[0051] The target detection model training method, apparatus, device, medium, and program products provided in this application, based on the interference factors in the client's image acquisition environment, set up distortion processing methods corresponding to various interference factors of the client, perform corresponding distortion processing on the original face image acquired by the client, and perform local face image stitching on different face images of the same sample user after distortion processing to obtain a training set including different fine-grained local face features of the sample user under various interference factors. This enables the target detection model to learn the local face features of the sample user under different interference factors, and use the local face features to complete identity verification, thereby improving the success rate of identity verification. At the same time, using the target detection model for identity verification does not require similarity calculation between the acquired image and the reference image, that is, it does not need to consider the complete face information of the target user, avoiding user identity verification problems caused by the presence of interference factors. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] Figure 1 This is a schematic diagram of an application scenario;

[0054] Figure 2 A flowchart illustrating a target detection model training method provided in this application;

[0055] Figure 3 A schematic diagram of the structure of a backbone network convolutional layer is provided in this application;

[0056] Figure 4 A schematic diagram of an activation function image provided in this application;

[0057] Figure 5 A schematic diagram of the derivative function graph of an activation function provided in this application;

[0058] Figure 6 A flowchart illustrating another object detection model training method provided in this application;

[0059] Figure 7 A flowchart illustrating a face recognition method provided in this application;

[0060] Figure 8This is a schematic diagram of the structure of a target detection model training device provided in this application;

[0061] Figure 9 This application provides a schematic diagram of the structure of a face recognition device.

[0062] Figure 10 This is a schematic diagram of the structure of an electronic device provided in this application.

[0063] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0064] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0065] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0066] It should be noted that the methods and apparatus for training object detection models in this application can be used in the field of artificial intelligence, or in any field other than artificial intelligence. The application fields of the methods and apparatus for training object detection models in this application are not limited.

[0067] First, let me explain the terms used in this application:

[0068] Bank self-service equipment: Dedicated equipment provided by banks for users to complete transactions such as deposits, withdrawals, transfers, bill payments, account openings, and information inquiries. It includes cash-based bank self-service equipment (such as ATMs, automatic deposit machines, automatic deposit and withdrawal machines, and foreign exchange machines) and non-cash-based bank self-service equipment (such as inquiry machines, receipt printers, and card issuing machines).

[0069] Figure 1This is a schematic diagram of an application scenario. A bank transaction system includes a client-side application and a server-side application.

[0070] The server-side of the aforementioned banking transaction system, used to process banking transactions based on transaction requests, can be a server or a server cluster. This server-side system can be deployed entirely in a cloud environment, or in a distributed manner, for example, partially in a cloud environment and partially in an edge environment.

[0071] The client of the aforementioned bank transaction system is used to receive user transaction requests and return transaction results. It can be an application (App) version client, such as a mobile banking APP, or a World Wide Web (Web) version client, or a bank self-service device. This application does not limit it.

[0072] When conducting business using a bank's transaction system client, such as a bank's self-service machine, identity verification is frequently required to ensure that the transaction is being conducted by the account holder. Currently, the primary identity verification method used in bank transaction systems is facial recognition to determine if the user is the account holder, which mainly includes the following steps:

[0073] (1) The user's facial image collected using the bank's self-service equipment, and the user's user identifier, are sent to the server of the aforementioned bank transaction system.

[0074] The aforementioned user identifier is used to uniquely identify a user. It can be unique information about the user, such as an identification number or mobile phone number, or a unique ID assigned by the bank's transaction system. It should be noted that when a user uses a bank's self-service device to conduct business, the device has already obtained the user's user identifier (the user must log in to the self-service device to conduct business; the user identifier is obtained after logging in).

[0075] (2) The server of the aforementioned bank transaction system obtains the reference image corresponding to the user identifier from the information database of the bank transaction system based on the user identifier.

[0076] It should be understood that the user identifier, the user's personal information corresponding to the user identifier, and the user's reference image, etc., will be stored in the database of the bank's transaction system in advance. The aforementioned reference image may be an image from the user's identity document, or an image of the user obtained through other means.

[0077] (3) Based on the reference image corresponding to the user identifier and the user face image collected by the bank self-service equipment, a neural network model is used to identify the user face image and output the similarity between the user face image and the reference image.

[0078] (4) Determine whether the similarity exceeds the preset threshold.

[0079] If the number of cases exceeds the limit, it means the user is the actual person, and the user's identity verification is successful, allowing the business to be processed. If the number of cases does not exceed the limit, the data can be collected and verified again. If multiple verifications fail, it means the user is not the actual person, and the business will not be processed.

[0080] However, the above-mentioned facial recognition methods often result in situations where the user is identified but the identity verification fails, indicating a high error rate in facial recognition and a low success rate in identity verification, leading to a poor user experience.

[0081] The inventors of this application discovered through research that the reason why the above-mentioned person is the user but the success rate of identity verification is low is that when the user uses the client of the bank transaction system to collect facial images, the interference of external environmental factors of the client, such as light brightness and collection factors, affects the integrity of the collected user facial information, which in turn makes the similarity of the model output low, resulting in the failure of user identity verification.

[0082] To address the aforementioned technical problems in existing technologies, this application provides a target detection model training method for a bank transaction system client. Based on interference factors in the image environment captured by the client, a pre-defined distortion processing method corresponding to the interference factor is used, and the sample image data is subjected to corresponding distortion processing to simulate the impact of the interference factor on the image. This allows the target detection model to learn the fine-grained features of the sample user's face under different interference factors, thereby enabling the model to identify the user based on a specific fine-grained feature. This eliminates the need to calculate the similarity between the captured image and the reference image, i.e., it eliminates the need to consider complete facial information, thus improving the success rate of identity verification.

[0083] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0084] The execution entity of the following embodiments can be an electronic device such as a computer with processing capabilities, or a model training platform. This platform can be a cloud-based platform, or a distributed platform (e.g., partially deployed in a cloud environment and partially deployed in an edge environment). The embodiments are described below using an electronic device as an example.

[0085] Figure 2 This is a flowchart illustrating a training method for an object detection model provided in this application. Figure 2As shown, the method includes:

[0086] S101. Obtain the original sample dataset.

[0087] The aforementioned original sample dataset includes original sample data from N sample users. This sample data includes: original facial images of the sample users collected by the client of the bank's transaction system, and the identifiers of the sample users in the bank's transaction system (hereinafter referred to as sample user identifiers). The original facial images of the sample users collected by the client of the bank's transaction system can more accurately reflect the environmental conditions when the client was collecting the images.

[0088] It should be understood that the client-side interface of the aforementioned banking system has image capture capabilities.

[0089] Regarding the acquisition method of the aforementioned original sample dataset, one possible implementation is that the electronic device communicates with a database storing the original sample dataset. Another possible implementation is that the electronic device communicates with a database storing the original facial images of sample users collected by the client of the aforementioned bank transaction system, and with a database storing sample user identifiers, respectively, to obtain the original facial images and sample user identifiers of the sample users, thus forming the aforementioned original sample dataset. Yet another possible implementation is that the aforementioned original sample dataset is imported from an external device, such as a USB flash drive.

[0090] Regarding the annotation of the aforementioned original sample dataset, one possible implementation is that the obtained original sample dataset has already annotated the face locations in the original face images of the aforementioned sample users based on the sample user identifiers. That is, the obtained original sample dataset also includes an annotation file for the original face images, which includes the location of the faces in the original face images and the corresponding sample user identifiers. Another possible implementation is that the obtained original sample dataset is not annotated. Instead, the face locations in the original face images of the aforementioned sample users can be annotated based on the obtained sample user identifiers to form an annotation file. The annotation tool can be YOLO_mark, and this application does not limit its scope.

[0091] S102. Based on the interference factors of the image acquisition environment of the above client, perform M types of distortion processing on the original face images of the above sample users to obtain the initial sample dataset.

[0092] The interference factors in the image acquisition environment mentioned above refer to factors that cause the difference between the acquired original face image and the real face image. For example, these may include lighting factors (such as light brightness, light saturation, etc.) and acquisition factors (such as the face being too close or too far from the acquisition device during acquisition, the presence of obstacles, the angle of the face during acquisition, etc.).

[0093] The above distortion processing includes: illumination distortion processing and / or geometric distortion processing.

[0094] The initial sample dataset mentioned above includes at least: M distorted face images of N sample users, and sample user identifiers. It should be understood that the sample user identifiers referred to here are the annotations of the face images.

[0095] Optionally, the value of M can be preset; alternatively, the value of M can be determined based on the maximum number of images required for partial face image stitching processing. The maximum number of images required for partial face image stitching processing is determined by the processing power of the electronic device. Therefore, the maximum number of images required for partial face image stitching processing can be determined based on the processing power of the electronic device, and this maximum number of images can be used as the value of M.

[0096] Optionally, a distortion processing library is preset to address various interference factors in the image acquisition environment of the aforementioned client. Each distortion processing method in this library can be mapped to at least one interference factor, and the interference factors are simulated using the distortion processing methods in this library. The distortion processing library includes P distortion processing methods, where P ≥ M. An example distortion processing library is shown in Table 1:

[0097] Table 1 Distortion Correction Library

[0098]

[0099] It should be understood that the number in the P and M distortion processing methods mentioned above refers to the number of specific distortion processing methods. Applying M distortion processing methods to the original face images of the sample users mentioned above refers to selecting M distortion processing methods from illumination distortion processing and / or geometric distortion processing.

[0100] One possible implementation is to determine M target distortion processing methods from the distortion processing library based on the interference factors of the client-acquired image environment, i.e., M≤P; another possible implementation is to determine M target distortion processing methods from the distortion processing library based on the interference factors of the client-acquired image environment and the image parameters of the original face image of the sample user, i.e., M≤P.

[0101] Based on the M target distortion processing methods determined above, the original face images of the sample users are distorted to obtain the initial sample dataset.

[0102] By determining the value of M based on the interference factors in the image environment acquired by the client, and selecting M target distortion processing methods from the distortion processing library, the interference factors in the image environment acquired by the client can be realistically restored using the M distortion processing methods. Then, the original face image processed by these M distortion methods can be used for model training to improve the accuracy of model training.

[0103] Optionally, based on the interference factors in the image acquisition environment of the client, a reference benchmark is preset for each distortion processing method. This reference benchmark is used to characterize the degree of distortion of a certain parameter of the image. If the image parameters of the original face image of the sample user exceed the reference benchmark of a certain distortion processing method, it indicates that the original face image is distorted and no further distortion processing is needed. If the image parameters of the original face image of the sample user do not exceed the reference benchmark of a certain distortion processing method, it indicates that the original face image is normal and distortion processing is needed.

[0104] One possible implementation involves first identifying the interference factors in the image acquisition environment from the client, then determining the first distortion processing method. Next, based on preset reference standards for each distortion processing method and the parameters of the original face image of the sample user, a second distortion processing method, in addition to the first, is determined. It should be noted that the first and second methods are not quantities; that is, the first or second distortion processing method may include multiple distortion processing methods.

[0105] For example, other candidate distortion processing methods besides the first distortion processing method can be identified first, and then, based on the reference benchmarks of the other candidate distortion processing methods, it can be determined whether the parameters of the original face image exceed the reference benchmarks of the distortion processing method.

[0106] If the distortion is not exceeded, it means that the parameters of the original face image are normal and need to be processed. This distortion processing method is then used as a second distortion processing method. If the distortion is not exceeded, it means that the parameters of the original face image are already distorted and no distortion processing is needed. This process continues until all second distortion processing methods are determined.

[0107] The first distortion processing method, determined based on the interference factors in the image environment acquired by the client, and the second distortion processing method, determined based on the parameters of the original face image, will be used together as the target distortion processing method.

[0108] Another possible implementation involves judging the parameters of the original face image of each sample user one by one based on the reference benchmarks of various distortion processing methods preset according to the interference factors of the image acquisition environment of the client, so as to determine the M target distortion processing methods to be adopted for the original face image.

[0109] Based on the M target distortion processing methods determined above, the original face images of the sample users are distorted to obtain the initial sample dataset.

[0110] Through this step, based on the interference factors in the image acquisition environment of the client, M types of distortion processing were applied to the original face images of the above N sample users. Not only were M types of distortion processing applied to the original face images of each sample user, obtaining face images of each sample user under different interference factors, but these M types of distortion processing also cover the interference factors in the image acquisition environment of the client. The data in the initial sample dataset can be used to characterize the various environmental interference factors when the client acquires images, and face images under the presence of various interference factors can be obtained.

[0111] S103. Obtain the initial training sample dataset from the initial sample dataset mentioned above.

[0112] One possible implementation is to randomly divide the data in the initial sample dataset into a training set and a test set according to a preset ratio, and use the training set as the initial training sample dataset. This ratio could be, for example, 8:1.

[0113] Another possible implementation involves randomly dividing the face images of the same user into different training subsets based on the user's identifier, ensuring that each training subset includes at least one face image of that user. A preset proportion of training subsets is established, and when a preset threshold is reached, M training subsets are obtained. These M training subsets constitute the initial training sample dataset described above.

[0114] S104. Perform at least one type of face local image stitching process on at least a portion of the sample data of the sample users in the above initial training sample dataset to obtain the training sample dataset.

[0115] The aforementioned partial facial image stitching process refers to stitching together different numbers of facial images from the same sample user into a single image of the same size as the original facial image after cropping. It should be noted that during cropping, the facial information must be cropped according to the marked facial position information to ensure that the facial information occupies a preset proportion of the cropped area, for example, 60% of the cropped area. It should also be noted that this application does not limit the number of images stitched together; the specific number is related to the processing power of the electronic device.

[0116] For example, taking the combination of two images as an example, based on the size of the original face image, such as 500*500, two rectangular frames are randomly determined to be cropped, such as one being 500*200 and the other being 500*300. These two rectangular frames are used to crop the face images of the sample user in the initial training sample dataset mentioned above. Before cropping, the proportion of the total number of pixels of the face information in the cropping area to the total number of pixels in the cropping area is calculated. If the proportion is greater than or equal to 60%, cropping is performed. If it is less than 60%, the rectangular frame is moved to obtain a new cropping area. Cropping is performed after the condition is met. The two cropped images are then combined into a new image.

[0117] The aforementioned partial facial image stitching process can also be used to obtain the stitched image through mosaic data augmentation.

[0118] The above-mentioned face partial image stitching process refers to stitching different numbers of face images into a new face image, such as using two images stitched together and four images stitched together.

[0119] Taking two-image stitching and four-image stitching as examples, the following methods may be used to perform at least one type of face local image stitching on at least a portion of the sample data from the initial training dataset:

[0120] One possible implementation is to perform face local image stitching processing on all sample users in the above initial training sample dataset, stitching together four images to obtain the stitched data subset 1; and to perform face local image stitching processing on a preset proportion (e.g., 50%) of sample users in the above initial training sample dataset, stitching together two images to obtain the stitched data subset 2.

[0121] Another possible implementation is to perform face local image stitching processing on all sample users in the above initial training sample dataset, using four images, to obtain the stitched data subset 1; and to perform face local image stitching processing on all sample users in the above initial training sample dataset, using two images, to obtain the stitched data subset 3.

[0122] Another possible implementation is to perform face local image stitching on a portion of the sample users in the initial training sample dataset, stitching together four images to obtain the stitched data subset 4; and to perform face local image stitching on the remaining sample users in the initial training sample dataset, stitching together two images to obtain the stitched data subset 5.

[0123] The training sample dataset consists of a subset of data obtained by stitching together at least one partial facial image, and the aforementioned initial training sample dataset.

[0124] This step can stitch together local facial images of the same sample user to obtain different fine-grained local facial features of the sample user. These local facial features are under various interference factors, making it easier for subsequent models to learn the fine-grained local facial features of the sample user under various interference factors.

[0125] S105. Use the above training sample set to train the object detection model until the model converges, and obtain the trained object detection model.

[0126] The trained object detection model described above is used to predict the user identifier of a target user in a banking transaction system based on the target user's facial image.

[0127] The object detection model mentioned above can be a YOLO series object detection model, such as YOLOv4, or a convolutional neural network model, such as a region-based convolutional neural network (R-CNN) model.

[0128] This application utilizes an object detection model for identity verification, which can learn fine-grained facial features of the target user. Even under the influence of interference factors, it can learn the fine-grained facial features when interference factors are present, and thus identify the target user based on the fine-grained facial features. There is no need to perform similarity calculation between the acquired image and the reference image, that is, there is no need to consider the complete facial information of the target user, thereby improving the success rate of identity verification.

[0129] It should be noted that the methods for training the object detection model using the above training sample set until the model converges can refer to existing technologies, and this application does not limit them here.

[0130] For example, taking the above target detection model as the YOLOv4 model, the model includes three parts: the backbone network, the neck, and the detection head.

[0131] The backbone network of the above model is used to extract facial features at different fine-grained levels. This part uses a Cross-Stage Partial Network (CSPNet). Figure 3 This is a schematic diagram of the structure of a backbone network convolutional layer provided in this application. This part consists of 5 convolutional layers, each processing the input feature map in the same way. The structure of each convolutional layer is as follows: Figure 3As shown, the input feature map is divided into two parts in a 2:3 ratio. 40% of the feature map is directly output through identity mapping, while the remaining part is output through residual structure convolution. Finally, the two outputs are concatenated and combined before being input into the next convolutional layer. Here, w, h, and c represent the width, height, and number of channels of the input feature map, respectively, and A represents the number of residual processes performed in each layer.

[0132] This section utilizes a 2:3 ratio partitioning method to allow for more features to undergo residual processing, thereby increasing the receptive field.

[0133] The neck section of the aforementioned model is used to fuse multi-scale features. This part employs Spatial Pyramid Pooling (SPP) and a Path Aggregation Network (PANet). SPP performs max pooling on the input multi-scale feature map using three grids of different sizes, generating feature vectors of corresponding sizes, which are then concatenated for output, enhancing feature representation. Simultaneously, PANet uses top-down and bottom-up information transmission paths to enhance shallow and deep feature information, achieving effective fusion of features from different layers and maximizing the utilization of output features from each layer. The fused features are then input into the detection head for classification and regression.

[0134] The detection head of the above model is used to output the target classification probability and regression location. An optimized FCOS series detection head can be selected, that is, by abandoning shared weights and using different convolutional groups to complete feature extraction at each layer, thus improving detection capability.

[0135] Optionally, the activation function used in the above three parts can be the function shown in the following formula (1):

[0136]

[0137] The derivative of this function is shown in the following formula (2):

[0138]

[0139] In the above formula, x represents the sum of the weighted values ​​of all inputs to the neuron, and the sum of the preset bias values, i.e., x = ∑ i ω i a i +b, where a i ω is the output of the i-th neuron input to this neuron. i is the weight value corresponding to this input, and b is the preset bias value.

[0140] Figure 4 This is a schematic diagram of an activation function image provided in this application. Figure 5 This is a schematic diagram of the derivative function graph of an activation function provided in this application.

[0141] like Figure 4 and Figure 5 As shown, in the x > 0 part, the activation function is unbounded, and the derivative approaches 1 as x approaches positive infinity. This ensures that the result of the backpropagation multiplication calculation will not approach 0, that is, the gradient vanishing situation will not occur, thus ensuring the weight value update and avoiding the phenomenon of weight value saturation (weight value not being updated).

[0142] In the x≤0 part, the activation function exhibits boundedness and outputs when x is negative, ensuring the integrity of the data distribution and avoiding the situation where there is no data when x is negative; at the same time, the derivative function has an output when x is negative, ensuring the transmission and updating of negative gradient information during backpropagation.

[0143] Using the YOLOv4 model described above, user identity information can be identified by utilizing fine-grained features of the local face, eliminating the need to calculate similarity between the reference image and the captured image, thus improving the success rate of identity verification.

[0144] Optionally, after obtaining the trained object detection model, the weight parameters of the trained object detection model can be formatted to compress the model. For example, the weight parameters can be converted from 32-bit single-precision floating-point format to 16-bit half-precision floating-point format to compress the trained object detection model. This step can reduce the storage space occupied by the object detection model, making it easier to port to mobile devices, such as to bank self-service machines.

[0145] The target detection model training method provided in this application, based on the interference factors in the client-acquired image environment, sets a distortion processing method corresponding to various interference factors of the client, performs corresponding distortion processing on the original face image acquired by the client, and performs local face image stitching on different face images of the same sample user after distortion processing, to obtain a training set including different fine-grained local face features of the sample user under various interference factors. This enables the target detection model to learn the local face features of the sample user under different interference factors, and uses the local face features to complete identity verification, thereby improving the success rate of identity verification. At the same time, using the target detection model for identity verification does not require similarity calculation between the acquired image and the reference image, that is, it does not need to consider the complete face information of the target user, thus avoiding the aforementioned user identity verification situation caused by the existence of interference factors.

[0146] The following explains how to determine the parameter range used for distortion processing.

[0147] Referring to the aforementioned preset distortion processing library and its various distortion processing methods, one possible implementation is to preset the parameter range used by each distortion processing method in the aforementioned distortion processing library; another possible implementation is to determine the parameter range used by each distortion processing method in the distortion processing library based on the image acquisition environment of at least one client of the bank transaction system.

[0148] Referring to Table 1, taking brightness variation as an example, based on the image environment acquired by at least one client of the bank transaction system, the light brightness of the acquired image environment at different time periods is analyzed to determine the range of light brightness variation of the image environment acquired by at least one client. Then, based on the range of light brightness variation, the parameter range used for brightness variation in the distortion processing library is determined.

[0149] One possible implementation is to use the maximum and minimum values ​​of the light intensity variation range of the image environment acquired by at least one client as the parameter range for brightness variation in the distortion processing library.

[0150] Another possible implementation is to multiply the maximum and minimum values ​​of the ambient light brightness variation range of the image acquired by at least one client by a coefficient, and use these as the parameter range for brightness variation in the distortion processing library.

[0151] The range of image rotation parameters in the aforementioned geometric distortion processing can be determined based on the installation location of the acquisition device at least one client and the location of the image acquisition area. This allows us to determine the range of facial angles that the acquisition device can capture, thereby obtaining the parameter range of the image rotation factor in the distortion processing library. For example, the image can be captured from the edge of the image acquisition area, obtaining multiple captured facial angles. It should be understood that the maximum range of facial rotation angles is less than 90 degrees. The aforementioned method can be used to determine how to obtain the parameter range of the image rotation factor in the distortion processing library.

[0152] The range of image scaling parameters in the aforementioned geometric distortion processing can be determined based on the installation location of the acquisition device of at least one client and the location of the image acquisition area. This determines the range of image scaling factors that the acquisition device can acquire for facial information, thus obtaining the parameter range of image scaling factors in the distortion processing library. For example, images can be captured from both the furthest and closest points in the image acquisition area, and the magnification or reduction factor of the two captured facial images can be obtained to determine the range of image scaling factors. The aforementioned method can be used to determine how to obtain the parameter range of image scaling factors in the distortion processing library.

[0153] For example, referring to Table 1, the above distortion processing library can be as shown in Table 2.

[0154] Table 2 Distortion Correction Library

[0155]

[0156] By determining the parameter ranges used for each distortion processing method in the above manner, the distortion processing can more realistically reflect the interference factors of the image acquisition environment of each client, thereby improving the accuracy of the training data.

[0157] The target detection model training method provided in this application utilizes a distortion processing library to pre-set various distortion processing methods and determines the parameter range used for each distortion processing method, thereby improving the efficiency of image distortion processing and thus improving the model training efficiency.

[0158] The following explains how to obtain the training sample dataset using the original sample dataset.

[0159] Figure 6 A flowchart illustrating another object detection model training method provided in this application. Figure 6 As shown, the method includes:

[0160] S201. Obtain the original sample dataset.

[0161] S202. Determine the value of M based on the maximum number of images required for the face local image stitching process.

[0162] The above steps S201 and 202 can be executed in parallel or asynchronously. This application only uses asynchronous execution as an example.

[0163] S203. Based on the interference factors of the image acquisition environment of the above client, perform M types of distortion processing on the original face images of the above sample users to obtain the initial sample dataset.

[0164] S204. Obtain the initial training sample dataset from the initial sample dataset mentioned above.

[0165] Specifically, based on the sample user identifier, the face images of the same sample user are randomly divided into different training subsets, ensuring that each training subset includes at least one face image of that sample user. A preset proportion of training subsets is set, and when a preset threshold is reached, M training subsets are obtained. These M training subsets constitute the initial training sample dataset described above.

[0166] This method ensures that each training subset contains at least one face image of a sample user, avoiding the situation where there are not enough face images corresponding to a sample user when stitching images together later.

[0167] S205. Perform two types of face local image stitching processing on the sample user portion of the above initial training sample dataset to obtain the face image after face local image stitching processing.

[0168] Specifically, from each training subset of the initial training sample dataset, four face images of the sample user are randomly selected and processed into a partial face image by stitching the four images together, resulting in the stitched data subset 1; for 50% of the sample users in each training subset of the initial training sample dataset, two face images of the sample user are randomly selected and processed into a partial face image by stitching the two images together, resulting in the stitched data subset 2.

[0169] S206. Based on the image parameters of the face images in the remaining sample data of the sample users in the initial training sample dataset, determine whether to perform secondary distortion processing on the face image.

[0170] The remaining sample data of the sample users in the aforementioned initial training sample dataset refers to the face local image stitching processing based on the sample user partial data in the aforementioned initial training sample dataset. That is, in each training subset, the remaining face image data of the sample user other than the face image used for face local image stitching.

[0171] Regarding how to determine whether to perform secondary distortion processing, it can be determined whether to perform secondary distortion processing on the face image based on the fact that the image parameters of the face images of the sample users in the initial training sample dataset exceed the reference benchmark of each distortion processing method.

[0172] If yes, it means that the face images of the other sample data of the sample users in the initial training sample dataset are normal images according to the reference benchmark of a certain distortion processing method, and can be subjected to secondary distortion, then proceed to step S207; if no, it means that according to the distortion processing method, the face images of the other sample data of the sample users in the initial training sample dataset are already distorted images, and there is no need to perform secondary distortion.

[0173] S207. Based on the M types of distortion processing used on the face image, determine the secondary distortion processing method.

[0174] For example, based on the above distortion processing library, other distortion processing methods besides the M distortion processing methods used on the face image can be used as secondary distortion processing methods; alternatively, distortion processing methods corresponding to face image parameters that do not exceed the reference benchmark of each distortion processing method can be used as secondary distortion processing methods.

[0175] S208. According to the determined secondary distortion processing method, perform secondary distortion processing on the face image to obtain the face image after secondary distortion processing.

[0176] It should be noted that steps S205 and S206-208 above can be executed in parallel or asynchronously. This application is only an example of asynchronous execution.

[0177] S209. The face image after the above partial face image stitching processing and the face image after secondary distortion processing are used as the training sample set.

[0178] By performing facial local image stitching on the above sample data, different facial local features under the influence of different interference factors can be stitched together onto a single image, thus obtaining different facial local features of the sample user under the influence of different interference factors. Performing secondary distortion processing on the above sample data can apply different interference factors to the same facial image, that is, obtaining a certain local feature of the face under the combined influence of different interference factors.

[0179] Optionally, the above training sample set can be divided into multiple training batches for batch training.

[0180] The target detection model training method provided in this application determines a distortion processing method more specifically for the interference factors in the client's image acquisition environment based on the interference factors in the client's image acquisition environment, and obtains an initial sample dataset including face images after the first distortion processing; the sample data in the initial training sample dataset obtained based on the initial sample dataset is partially processed by face local image stitching and partially processed by secondary distortion processing, so that the influence of various interference factors exists in a face image in different ways, which increases the complexity and diversity of the sample data, making it easier for the target detection model to learn the local fine-grained features of the face under the influence of various interferences, and improving the recognition accuracy of the model.

[0181] The execution entity in the following embodiments can be the server side of the aforementioned bank transaction system.

[0182] Figure 7 This is a flowchart illustrating an authentication method provided in this application. Figure 7 As shown, the method includes;

[0183] S301: Receive a transaction request triggered by the target user through the client of the bank's transaction system.

[0184] The aforementioned transaction request refers to a request to complete a certain business transaction, such as a transfer of funds or an account opening.

[0185] S302. In response to the above transaction request, control the above client to collect the facial image of the target user.

[0186] User authentication is required before processing a transaction request to confirm whether the transaction request was triggered by the target user.

[0187] For example, the acquisition device of the aforementioned client, such as a camera, can be controlled to acquire the facial image of the target user. One or more images can be acquired, and this application does not limit the scope of the acquisition.

[0188] S303. Based on the facial image of the target user, use the target detection model trained by the above method to predict the user identifier of the target user in the bank transaction system.

[0189] S304. Based on the user identifier of the target user in the bank's transaction system, determine whether the target user is a legitimate user.

[0190] It should be understood that when a target user uses the aforementioned client to conduct business, the client has already obtained the target user's user identifier. Therefore, the predicted target user's user identifier can be compared with the target user's user identifier obtained by the client. If they are the same, the target user is determined to be a legitimate user; if they are different, the target user is determined to be an illegitimate user.

[0191] If yes, it means the target user is a legitimate user, and step S305 is executed; if no, it means the user is an illegitimate user, and an authentication failure message is returned.

[0192] S305. Process the above transaction request.

[0193] After confirming that the target user is indeed conducting the transaction, the transaction request for that transaction will be processed.

[0194] S306. Return the processing result of the above transaction request to the above client.

[0195] The processing results of the above transaction requests include various situations such as transaction success, transaction failure, and error. The specific settings can be configured according to actual needs, and this application does not impose any limitations on them.

[0196] Optionally, the target detection model trained by the target detection model training method described in the above embodiments can be compressed and ported to a bank self-service device for use. Then, the above identity verification method applied to the bank self-service device can include the following steps:

[0197] S401. Collect the facial image of the target user based on the transaction request triggered by the target user.

[0198] S402. Based on the facial image of the target user, use the target detection model trained by the above method to predict the user identifier of the target user in the bank transaction system.

[0199] S403. Based on the user identifier of the target user in the bank's transaction system, determine whether the target user is a legitimate user.

[0200] If yes, it means the target user is a legitimate user, and step S405 is executed; if no, it means the user is an illegitimate user, and an authentication failure message is returned.

[0201] S404. Send the aforementioned transaction request to the server of the aforementioned bank transaction system to process the aforementioned transaction request.

[0202] After confirming that the target user is indeed conducting the transaction, the transaction request is sent to the server for processing.

[0203] S405. Receive the processing result of the above transaction request and display it.

[0204] The system receives the processing result of the transaction request returned by the server and displays it on the front-end display interface of the bank's self-service equipment.

[0205] This method can move the user authentication process to bank self-service equipment, reducing the data processing pressure on the bank's transaction system server and improving transaction processing efficiency.

[0206] Optionally, the target detection model can be updated periodically according to actual needs and the training method of the target detection model described in the foregoing embodiments, so as to ensure that it learns the facial features of more users and improves the success rate of identity verification.

[0207] The authentication method provided in this application uses a target detection model trained by the method described in the foregoing embodiments to predict the identity of a face image, obtain a predicted user identifier, and match it with the user identifier in the bank transaction system to complete user authentication. Afterward, the user's transaction request is processed to ensure that the transaction was indeed conducted by the user. This target detection model can predict the user identifier using local fine-grained facial features even under interference factors in the image acquisition environment, and then use the predicted user identifier to complete authentication. Compared with the aforementioned face recognition method for authentication, it does not require considering the complete information of the face in the acquired image, thus improving the success rate of authentication and enhancing the user experience.

[0208] Figure 8 This is a schematic diagram of the structure of a target detection model training device provided in this application. Figure 8 As shown, the device includes;

[0209] The acquisition module 11 is used to acquire the original sample dataset; the original sample dataset includes the original sample data of N sample users, and the sample data includes: the original face images of the sample users collected by the client of the bank transaction system, and the identifier of the sample users in the bank transaction system;

[0210] The first processing module 12 is used to perform M types of distortion processing on the original face images of the sample users according to the interference factors of the image acquisition environment of the client, to obtain an initial sample dataset. The initial sample dataset includes at least: face images of N sample users after M types of distortion processing, and the identifier of the sample users in the bank transaction system; the distortion processing includes: illumination distortion processing and / or geometric distortion processing.

[0211] The second processing module 13 is used to obtain the initial training sample dataset from the initial sample dataset;

[0212] The third processing module 14 is used to perform at least one type of face local image stitching processing on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset.

[0213] Training module 15 is used to train the target detection model using the training sample set until the model converges to obtain the trained target detection model. The trained target detection model is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image.

[0214] One possible implementation is that the first processing module 12 is specifically used to determine M target distortion processing methods from the distortion processing library based on the interference factors of the image acquisition environment of the client and the image parameters of the original face image of the sample user; and to perform distortion processing on the original face image of the sample user according to the determined M target distortion processing methods to obtain the initial sample dataset.

[0215] In this implementation, before the first processing module 12 performs distortion processing on the original face image of the sample user according to the determined M types of target distortion processing methods, the first determining module 16 is used to determine the parameter range of each distortion processing method in the distortion processing library according to the image acquisition environment of at least one client of the bank transaction system.

[0216] In one possible implementation, before the third processing module 14 performs local image stitching on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset, the second determining module 17 is used to determine whether to perform secondary distortion processing on the face image based on the image parameters of the face images in the remaining sample data of the sample users in the initial training sample dataset; if so, then based on the M types of distortion processing used on the face image, a secondary distortion processing method is determined; and the face image is subjected to secondary distortion processing according to the determined secondary distortion processing method.

[0217] In one possible implementation, before the first processing module 12 performs M types of distortion processing on the original face image of the sample user based on the interference factors of the image acquisition environment of the client, the third determining module 18 is used to determine the value of M based on the maximum number of images required for the local face image stitching processing.

[0218] One possible implementation is a conversion module 19, which is used to convert the format of the weight parameters of the trained object detection model in order to compress the trained object detection model.

[0219] The target detection model training device provided in this application can execute the target detection model training method in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0220] Figure 9 A schematic diagram of an authentication device provided in this application. Figure 9 As shown, the device includes;

[0221] Receiving module 21 is used to receive transaction requests triggered by the target user through the client of the bank transaction system;

[0222] Control module 22 is used to control the client to collect the facial image of the target user in response to the transaction request;

[0223] The identification module 23 is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image and the target detection model trained by the target detection model training method described in the foregoing embodiments.

[0224] The determination module 24 is used to determine whether the target user is a legitimate user based on the user identifier of the target user in the bank transaction system;

[0225] Processing module 25 is used to process the transaction request after determining that the target user is a legitimate user;

[0226] Output module 26 is used to return the processing result of the transaction request to the client.

[0227] The authentication device provided in this application can execute the authentication method in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0228] Figure 10 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 10 As shown, the electronic device 500 may include at least one processor 501 and a memory 502, such as a computer, server or other electronic device with processing capabilities.

[0229] Memory 502 is used to store programs. Specifically, the program may include program code, which includes computer operation instructions. Memory 502 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0230] The processor 501 is used to execute computer execution instructions stored in the memory 502 to implement the target detection model training method or the authentication method described in the foregoing method embodiments. The processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0231] The electronic device 500 may also include a communication interface 503, through which it can communicate and interact with external devices. The external device may be, for example, an electronic device such as a computer.

[0232] In practical implementation, if the communication interface 503, memory 502, and processor 501 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0233] Optionally, in a specific implementation, if the communication interface 503, memory 502, and processor 501 are integrated on a single chip, then the communication interface 503, memory 502, and processor 501 can communicate through an internal interface.

[0234] This application also provides a computer-readable storage medium, which may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. Specifically, the computer-readable storage medium stores program instructions, which are used for the target detection model training method or the authentication method in the above embodiments.

[0235] This application also provides a computer program product including executable instructions stored in a readable storage medium. At least one processor of an electronic device 500 can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the electronic device 500 to implement the target detection model training method or the authentication method provided in the various embodiments described above.

[0236] This application also provides a chip on which a computer program is stored. When the computer program is executed by the chip, it implements a target detection model training method or an authentication method provided in various embodiments.

[0237] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0238] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for training an object detection model, characterized in that, The method includes: Obtain the original sample dataset; the original sample dataset includes the original sample data of N sample users, the sample data including: the original face images of the sample users collected by the client of the bank transaction system, and the identifier of the sample users in the bank transaction system; Based on the interference factors in the image acquisition environment of the client, the original face images of the sample users are subjected to M types of distortion processing to obtain an initial sample dataset. The initial sample dataset includes at least: face images of N sample users after M types of distortion processing, and the identifier of the sample users in the bank transaction system; the distortion processing includes: illumination distortion processing and / or geometric distortion processing. The initial training sample dataset is obtained from the initial sample dataset; At least one type of local face image stitching process is performed on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset. The target detection model is trained using the training sample set until the model converges, resulting in a trained target detection model. The trained target detection model is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image. Before performing local image stitching on at least a portion of the sample data from the initial training sample dataset to obtain the training sample dataset, the method further includes: Based on the image parameters of the face images in the remaining sample data of the sample users in the initial training sample dataset, determine whether to perform secondary distortion processing on the face image; If so, then based on the M types of distortion processing applied to the face image, a secondary distortion processing method is determined. The method of determining the secondary distortion processing method includes: based on the distortion processing library, taking other distortion processing methods besides the M types of distortion processing applied to the face image as secondary distortion processing methods; or, taking the distortion processing method corresponding to the face image parameters not exceeding the reference benchmark of each distortion processing method as the secondary distortion processing method. Based on the determined secondary distortion processing method, the face image is subjected to secondary distortion processing.

2. The method according to claim 1, characterized in that, The initial sample dataset is obtained by performing M types of distortion processing on the original face images of the sample users based on interference factors in the image acquisition environment of the client, including: Based on the interference factors in the image environment acquired by the client and the image parameters of the original face image of the sample user, M target distortion processing methods are determined from the distortion processing library; Based on the determined M types of target distortion processing methods, the original face images of the sample users are subjected to distortion processing to obtain the initial sample dataset.

3. The method according to claim 2, characterized in that, Before performing distortion processing on the original face image of the sample user according to the determined M target distortion processing methods, the method further includes: Based on the image environment acquired by at least one client of the bank transaction system, determine the parameter range used by each distortion processing method in the distortion processing library.

4. The method according to any one of claims 1-3, characterized in that, Before performing M types of distortion processing on the original face image of the sample user based on interference factors in the image acquisition environment of the client, the method further includes: The value of M is determined based on the maximum number of images required for the partial facial image stitching process.

5. The method according to any one of claims 1-3, characterized in that, The method further includes: The weight parameters of the trained object detection model are converted to a different format to compress the trained object detection model.

6. An authentication method, characterized in that, The method includes: Receive transaction requests triggered by the target user through the bank's transaction system client; In response to the transaction request, the client is controlled to collect the facial image of the target user; Based on the facial image of the target user, the target detection model trained by the method described in any one of claims 1-5 is used to predict the user identifier of the target user in the bank transaction system. Based on the user identifier of the target user in the bank's transaction system, determine whether the target user is a legitimate user; If the target user is determined to be a legitimate user, then the transaction request is processed; The processing result of the transaction request is returned to the client.

7. A target detection model training device, characterized in that, The apparatus is used to implement the target detection model training method according to any one of claims 1-5, and the apparatus comprises: The acquisition module is used to acquire the original sample dataset; the original sample dataset includes the original sample data of N sample users, and the sample data includes: the original face images of the sample users collected by the client of the bank transaction system, and the identifier of the sample users in the bank transaction system; The first processing module is used to perform M types of distortion processing on the original face images of the sample users based on the interference factors of the image acquisition environment of the client, to obtain an initial sample dataset. The initial sample dataset includes at least: face images of N sample users after M types of distortion processing, and the identifier of the sample users in the bank transaction system; the distortion processing includes: illumination distortion processing and / or geometric distortion processing. The second processing module is used to obtain the initial training sample dataset from the initial sample dataset; The third processing module is used to perform at least one type of local face image stitching processing on at least a portion of the sample data of the sample users in the initial training sample dataset to obtain the training sample dataset. The training module is used to train the target detection model using the training sample set until the model converges, thereby obtaining the trained target detection model. The trained target detection model is used to predict the user identifier of the target user in the bank transaction system based on the target user's face image.

8. An authentication device, characterized in that, The device includes: The receiving module is used to receive transaction requests triggered by the target user through the bank's transaction system client. The control module is used to control the client to collect the facial image of the target user in response to the transaction request; The identification module is used to predict the user identifier of the target user in the bank transaction system based on the facial image of the target user and using the target detection model trained by the method described in any one of claims 1-5. The determination module is used to determine whether the target user is a legitimate user based on the user identifier of the target user in the bank transaction system; The processing module is used to process the transaction request after determining that the target user is a legitimate user; The output module is used to return the processing result of the transaction request to the client.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the target detection model training method as described in any one of claims 1 to 6, or the authentication method.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the target detection model training method as described in any one of claims 1 to 6, or an authentication method.

11. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the target detection model training method as described in any one of claims 1 to 6, or an authentication method.

Images