Intelligent identity verification method and device for self-service terminal

Abstract

The application provides a smart identity verification method and device for a self-service terminal, which can be used in the financial field or other technical fields. The smart identity verification method for the self-service terminal comprises the following steps: obtaining a pair of user ID head portraits and a live user face image captured by a self-service terminal camera; inputting the user ID head portraits and the live user face image into a pre-trained smart identity verification model to obtain a verification result output by the smart identity verification model; wherein the pre-trained smart identity verification model encodes the user ID head portraits and the live user face image by using two twin convolutional neural network modules to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and finally performs calculation on the full-connected feature vector to obtain the verification result. The device is used to execute the above method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an intelligent identity verification method and device for a self-service terminal. BACKGROUND

[0002] With the rapid development of intelligent technology, bank self-service terminals have emerged as the times require and account for a high proportion in user usage. The design concept of self-service terminals is "24-hour service", and users can choose the time to handle business according to their own schedules, which not only effectively solves the problem that net point staff cannot handle business after work, but also effectively alleviates the problem of overcrowding caused by too large a flow of people in bank net points; the working efficiency of self-service terminals is usually higher than that of manual operation, and the time and economic cost of increasing a self-service terminal is much lower than that of hiring a staff, and its simple and intuitive operation provides easy, convenient and thoughtful service for the general public.

[0003] The identity verification process of the self-service terminal: first, place the user's ID card in the certificate recognition place, then aim the lens of the self-service terminal at the user for live shooting, and finally compare and verify the user's identity after comparing the ID card portrait and the image collected by live shooting. The existing identity verification technology of the self-service terminal can achieve a certain accuracy, but there are still great challenges: the first step for users to use the self-service terminal to handle business is identity verification, but the comparison process of the ID card portrait and the live shooting image usually takes a long time, and occasionally the interface will be stuck due to long-time identification, and the net point staff can only restart the business by shutting down and restarting, which not only affects the user experience, but also reduces the working efficiency of the net point staff. SUMMARY

[0004] In view of the problems in the prior art, the embodiments of the present application provide an intelligent identity verification method and device for a self-service terminal, which can at least partially solve the problems in the prior art.

[0005] In one aspect, the embodiments of the present application provide an intelligent identity verification method for a self-service terminal, comprising:

[0006] Obtaining a pair of user ID card portraits and live user face images photographed by a self-service terminal camera;

[0007] Inputting the user ID card portraits and the live user face images into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model;

[0008] The pre-trained intelligent identity verification model encodes the user ID head portrait and the face image of the on-site user by using two twin convolutional neural network modules respectively to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and performs calculation on the full-connected feature vector to obtain a verification result.

[0009] In some embodiments, the convolutional neural network module is a multi-scale width convolution residual network module.

[0010] In some embodiments, the convolutional neural network module encodes the user ID head portrait / on-site user face image to obtain an encoded feature vector, including:

[0011] The user ID head portrait / on-site user face image is sequentially subjected to batch normalization, Mish activation, two-channel convolution and DropBlock operation to obtain a first feature map;

[0012] The first feature vector is sequentially subjected to batch normalization, Mish activation, two-channel convolution and DropBlock operation to obtain a second feature map;

[0013] The second feature map and the user ID head portrait / on-site user face image are pixel-by-pixel added to obtain an encoded feature vector.

[0014] In some embodiments, the two-channel convolution operation is used to extract the features of the image, and the two-channel convolution operation extracts the features of the image, including:

[0015] The input image is convolved by using the first N / 2 convolution kernels and the last N / 2 convolution kernels of the current layer convolutional network respectively to obtain two feature maps, wherein N is the number of convolution kernels of the current layer convolutional network;

[0016] The two feature maps and the input image are spliced to obtain a spliced image;

[0017] The spliced image is convolved by using a convolutional network with a convolution kernel size of 1x1 to obtain an enhanced image.

[0018] In some embodiments, in the two-channel convolution operation, the Mish function is used to activate the neurons.

[0019] In some embodiments, the training process of the intelligent identity verification model is as follows:

[0020] The image acquisition module is configured to acquire pairs of user ID head portraits and live user face images, wherein some of the pairs of user ID head portraits and live user face images are of the same user, and the other pairs of user ID head portraits and live user face images are not of the same user.

[0021] For each pair of the user ID head portrait and the live user face image, the preset Siamese neural network model is trained according to the pair of user ID head portrait and live user face image and a label of the pair of user ID head portrait and live user face image until the pre-trained intelligent identity verification model is obtained.

[0022] In some embodiments, the preset Siamese neural network model uses a binary cross-entropy loss function to calculate a similarity score between the user ID head portrait and the live user face image.

[0023] In another aspect, the present application provides an intelligent identity verification device for a self-service terminal, comprising:

[0024] The acquisition module is configured to acquire a pair of user ID head portrait and live user face image captured by a camera of the self-service terminal;

[0025] The verification module is configured to input the user ID head portrait and the live user face image into the pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model.

[0026] The pre-trained intelligent identity verification model uses two Siamese convolutional neural network modules to respectively encode the user ID head portrait and the live user face image to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and performs calculation on the full-connected feature vector to obtain the verification result.

[0027] In some embodiments, the convolutional neural network module is a multi-scale width convolutional residual network module.

[0028] In some embodiments, the convolutional neural network module encodes the user ID head portrait / live user face image to obtain an encoded feature vector, which includes:

[0029] The user ID head portrait / live user face image is sequentially subjected to batch normalization, Mish activation, two-channel convolution, and DropBlock operation to obtain a first feature map;

[0030] The first feature vector is sequentially subjected to batch normalization, Mish activation, two-channel convolution, and DropBlock operation to obtain a second feature map;

[0031] The second feature map is pixel-by-pixel added to the face image of the user ID head portrait / live user, to obtain an encoded feature vector.

[0032] In some embodiments, the two-branch channel convolution operation is used to extract features of an image, and the two-branch channel convolution operation includes:

[0033] The input image is respectively convolved using the first N / 2 convolution kernels and the last N / 2 convolution kernels of the current layer convolution network, to obtain two feature maps, where N is the number of convolution kernels of the current layer convolution network.

[0034] The two feature maps and the input image are spliced to obtain a spliced image.

[0035] The spliced image is convolved using a convolution network with a convolution kernel size of 1x1 to obtain an enhanced image.

[0036] In some embodiments, in the two-branch channel convolution operation, a Mish function is used to activate neurons.

[0037] In some embodiments, the device further includes a model training module, which is configured to:

[0038] A set of pairs of user ID head portraits and live user face images are obtained, where a part of the pairs of user ID head portraits and live user face images are of the same user, and another part of the pairs of user ID head portraits and live user face images are not of the same user.

[0039] For each pair of user ID head portrait and live user face image, a preset twin neural network model is trained according to the pair of user ID head portrait and live user face image and a label of the pair of user ID head portrait and live user face image, until a pre-trained intelligent identity verification model is obtained.

[0040] In some embodiments, the preset twin neural network model uses a binary cross-entropy loss function to calculate a similarity score between the user ID head portrait and the live user face image.

[0041] Embodiments of the present application also provide an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the program to implement the steps of the intelligent identity verification method for a self-service terminal according to any of the above embodiments.

[0042] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the intelligent identity verification method for a self-service terminal.

[0043] The method and device for intelligent identity verification of a self-service terminal provided by the embodiment of the present application are characterized in that: a pair of user ID head portraits and a face image of a user on site are obtained by using a camera of the self-service terminal; the user ID head portraits and the face image of the user on site are input into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model; and the pre-trained intelligent identity verification model uses two twin convolutional neural network modules to respectively encode the user ID head portraits and the face image of the user on site to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and finally performs calculation on the full-connected feature vector to obtain the verification result. In this way, the intelligent identity verification model is established by using the design idea of the twin neural network, and the gap between the two is estimated from the semantic similarity of the two images, so that the comparison between the user ID head portraits and the face image captured on site can be more ideally realized. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort. In the drawings:

[0045] Figure 1 FIG. 1 is a flowchart of the intelligent identity verification method for a self-service terminal provided by an embodiment of the present application.

[0046] Figure 2 FIG. 2 is a partial flowchart of the intelligent identity verification method for a self-service terminal provided by an embodiment of the present application.

[0047] Figure 3 FIG. 3 is a structure diagram of an improved two-branch channel convolution residual block provided by an embodiment of the present application.

[0048] Figure 4 FIG. 4 is a structure diagram of a two-branch channel convolution module provided by an embodiment of the present application.

[0049] Figure 5 FIG. 5 is a partial flowchart of the intelligent identity verification method for a self-service terminal provided by an embodiment of the present application.

[0050] Figure 6 is a part flow diagram of a smart identity verification method for a self-service terminal provided by an embodiment of the present application.

[0051] Figure 7 is a structural diagram of a twin neural network provided by an embodiment of the present application.

[0052] Figure 8 is a structural diagram of a smart identity verification device for a self-service terminal provided by an embodiment of the present application.

[0053] Figure 9 is a physical structure diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the embodiments of the present application are further described in detail below with reference to the drawings. Herein, the illustrative embodiments of the present application and their descriptions are used to explain the present application, but are not used as limitations of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be arbitrarily sequenced without conflict.

[0055] As to the "first", "second", and the like used herein, they do not particularly mean the order or sequence, nor are used to limit the present application, but are only used to distinguish the elements or operations described by the same technical terms.

[0056] As to the "include", "comprise", "have", "contain", and the like used herein, they are all open terms, that is, they mean to include but not limited to.

[0057] As to the "and / or" used herein, it includes any or all of the ordering of the described things.

[0058] To better understand the present application, the research background of the present application is described in detail as follows.

[0059] The interface freezing of the identity verification of the self-service terminal may be because the current image recognition algorithm cannot quickly determine the similarity between the user's ID photo and the on-site photographed image, thereby causing the system buffering time to be too long and the interface for handling the business to not be loaded.

[0060] In addition, when the user's ID information is input into the self-service terminal system, the ID photo is scanned statically, so that relatively complete and clear facial features can be obtained. However, the self-service terminal on-site photographing is very fast, although it can ensure that the user's front facial contour is photographed, but sometimes the facial features are not clear and the identity verification is difficult.

[0061] To solve the above technical problems, the application provides an intelligent identity verification method for a self-service terminal.

[0062] The execution subject of the intelligent identity verification method for a self-service terminal provided by the application includes but is not limited to a computer.

[0063] Figure 1 is a flowchart of the intelligent identity verification method for a self-service terminal provided by an embodiment of the application, as Figure 1 indicated, the intelligent identity verification method for a self-service terminal provided by the application includes:

[0064] S101, obtaining a pair of user ID head portraits and a live user face image photographed by a self-service terminal camera;

[0065] In step S101, the pair of user ID head portraits and the live user face image photographed by the self-service terminal camera refers to the head portrait on the ID provided by the user and the live face image of the user photographed by the camera when the user handles a business through the self-service terminal. The two images can be photographed by two cameras of the self-service terminal respectively, and after the camera of the self-service terminal photographs a pair of images, the pair of images can be immediately uploaded for intelligent identity verification.

[0066] S102, inputting the user ID head portrait and the live user face image into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model; wherein the pre-trained intelligent identity verification model encodes the user ID head portrait and the live user face image by using two twin convolutional neural network modules to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and performs calculation on the full-connected feature vector to obtain the verification result.

[0067] In step S102, the intelligent identity verification model adopts the design idea of a twin neural network, aiming to compare the similarity between the input two images, and the core idea is to learn from the semantic similarity of the two images to estimate the gap between the two images, so that the comparison between the user ID head portrait and the live face image photographed on site can be more ideally realized.

[0068] The application provides an intelligent identity verification method for a self-service terminal. A pair of user ID head portraits and a live user face image are obtained by using a self-service terminal camera. The user ID head portraits and the live user face image are input into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model. The pre-trained intelligent identity verification model uses two twin convolutional neural network modules to encode the user ID head portraits and the live user face image to obtain two encoded feature vectors, then performs full connection on the two encoded feature vectors to obtain a full-connected feature vector, and performs calculation on the full-connected feature vector to obtain the verification result. In this way, the intelligent identity verification model is established by using the design idea of the twin neural network, and the gap between the two is estimated from the semantic similarity of the two images, so that the comparison between the user ID head portraits and the live face image can be more ideally realized.

[0069] In some embodiments, the convolutional neural network module is a multi-scale width convolution residual network module.

[0070] In some embodiments, when the convolutional neural network module is a multi-scale width convolution residual network module, the convolutional neural network module encodes the user ID head portrait / live user face image to obtain an encoded feature vector, as shown in Figure 2

[0071] S201, sequentially performing batch normalization, Mish activation, two-branch channel convolution, and DropBlock operation on the user ID head portrait / live user face image to obtain a first feature map;

[0072] S202, continuing to sequentially perform batch normalization, Mish activation, two-branch channel convolution, and DropBlock operation on the first feature map to obtain a second feature map;

[0073] S203, pixel-by-pixel adding the second feature map and the user ID head portrait / live user face image to obtain an encoded feature vector.

[0074] Specifically, the multi-scale width convolution residual network uses an improved two-branch channel convolution residual block structure at each layer to extract deep image features, aiming to extract more complete and rich image features, and thus improve the classification accuracy and efficiency of the convolutional neural network module.

[0075] The improved two-branch channel convolution residual block is described in detail as follows:

[0076] As shown in Figure 3 ​As shown, the improved two-part channel convolution residual block in the present application draws on the residual block structure in the residual network and introduces the DropBlock structure. First, batch normalization is used to maintain the distribution of image data, avoiding the phenomenon of covariate shift caused by parameter updates, and the Mish function is used to activate network parameters, and then the two-part channel convolution (see Figure 4 ) is used to extract the features of the image, and the DropBlock operation is used to effectively avoid overfitting. Then, batch normalization, Mish activation, and two-part channel convolution operations are repeatedly performed in sequence. Finally, the input image and the feature map obtained by connecting two convolutions are added pixel by pixel using a residual connection, serving as the output of the improved asymmetric convolution residual block structure. The improved asymmetric convolution residual block structure uses a residual unit direct mapping method to improve the feature extraction capability of the network, and through the improved asymmetric convolution module, the feature extraction capability of the image is strengthened, and the robustness of image flipping and rotation is also enhanced.

[0077] As shown in Figure 5 , in some embodiments, the two-part channel convolution operation is used to extract the features of the image, and the two-part channel convolution operation includes:

[0078] S301, respectively using the first N / 2 convolution kernels and the last N / 2 convolution kernels of the current layer convolutional network to convolve the input image, to obtain two feature maps, wherein N is the number of convolution kernels of the current layer convolutional network;

[0079] S302, splicing the two feature maps and the input image to obtain a spliced image;

[0080] S303, using a convolutional network with a convolution kernel size of 1x1 to perform convolution operation on the spliced image to obtain an enhanced image.

[0081] Specifically, as shown in Figure 4 , the two-part channel convolution module proposed in the present application sets the number of convolution kernels of the current layer network to N and equally divides them into two groups, denoted as group 1 and group 2, i.e. the number of convolution kernels in each group is N / 2. Then, the input image is convolved with N / 2 convolution kernels in group 1 and N / 2 convolution kernels in group 2, and then the two convolution operations are performed respectively to obtain feature maps which are spliced. Finally, the number of channels of the spliced feature map is adjusted to N through convolution, and a branch with a convolution kernel size of 1x1 is introduced from the input end using the short connection method in the residual network to adjust the number of channels and extract the nonlinear features of the input image, enhance the nonlinear characteristics of the network, and then the feature maps obtained by the two branches operating on the input image are aggregated by adding the corresponding position pixels.

[0082] The two-branch channel convolution module equally divides the number of convolution kernels into two groups and acts on the input in parallel, and then splices and fuses the feature maps obtained by each group of convolution kernels and the input image, which helps to extract rich feature information of the input image. The short connection mode introduced can directly supplement the information of the input image through a 1x1 size convolution operation in the case of information loss in forward propagation, to further obtain more complete feature information, thereby enhancing the expression ability of the two-branch channel convolution module in feature extraction.

[0083] In some embodiments, in the two-branch channel convolution operation, a Mish function is used to activate neurons.

[0084] Specifically, compared with other activation functions, the ReLU activation function has a simple derivative and can improve the problem of gradient disappearance to a certain extent, but because a part of data falls into the negative interval, these data will become 0, resulting in that the corresponding weights cannot be updated, thereby appearing "dead" neurons. In order to overcome this shortcoming, the Mish function is selected to activate neurons in the two-branch channel convolution module, and the precision of Mish is better than that of ReLU, and Mish has stability in deep network, and its expression is:

[0085] Mish(x)=x·tanh(ln(1+e x ));

[0086] In the formula, x represents the feature map (also known as feature map) of the previous layer of the network.

[0087] As shown in Figure 6 some embodiments, the training process of the intelligent identity verification model is as follows:

[0088] S401, obtaining pairs of user ID head portraits and on-site user face images in an image set, wherein a part of the pairs of user ID head portraits and on-site user face images are of the same user, and another part of the pairs of user ID head portraits and on-site user face images are not of the same user;

[0089] S402, for each pair of the user ID head portrait and the on-site user face image, training a preset twin neural network model according to the pair of user ID head portrait and on-site user face image and the label of the pair of user ID head portrait and on-site user face image, until a pre-trained intelligent identity verification model is obtained.

[0090] Specifically, the pairs of images in the image set and the labels of the pairs of images are used to train the preset twin neural network model, and after the twin neural network model meets the required recognition accuracy, a trained intelligent identity verification model is obtained.

[0091] In some embodiments, the preset twin neural network model uses a binary cross-entropy loss function to calculate the similarity score between the user ID photo and the live user face image.

[0092] Specifically, the twin neural network uses the similarity score to predict the similarity of two inputs, determines the relative distance between two input images by the method of metric learning, and the loss function is a basic but crucial concept in deep learning, which evaluates the performance of the network model by calculating the gap between the predicted value (label) and the actual data. The smaller the error between the predicted value and the actual value, the smaller the output of the loss function, and the better the performance of the network model. In this application, the binary cross-entropy loss function is used to calculate the similarity score. Therefore, in order to better optimize the network model, it is necessary to train the loss function to minimize and keep convergence. Cross-entropy is an important concept in information theory, which mainly measures the difference between two probability distributions of the same random variable. The expression of the binary cross-entropy loss function is:

[0093]

[0094] Where N represents the number of samples, L represents the loss function, y represents the label of the sample, and p represents the probability of the sample being predicted as a positive class.

[0095] In order to better understand the present application, the intelligent identity verification method for self-service terminals provided by the present application will be described in detail through a specific embodiment.

[0096] The intelligent identity verification method for self-service terminals provided by the embodiments of the present application first inputs the user ID photo and the live face image into two similar sub-networks with the same structure, parameters and weights, i.e. multi-scale width convolution residual network. Then, the difference between the ID photo and the live image is measured by calculating the similarity score. Finally, the identity of the user handling the business on site is verified according to the similarity score.

[0097] The contents involved in the above steps will be described in detail below.

[0098] I. Overall architecture of self-service terminal identity verification

[0099] As Figure 7As shown, the overall architecture of the self-service terminal identity verification proposed in the present application adopts the design idea of Siamese Network, aiming to use the similarity score to judge the consistency of two different inputs. First, the user's ID portrait and the live shot of the front face image are respectively input into the corresponding multi-scale width convolution residual network, and then mapped to a new space to obtain the vector features after coding by the sub-network. It is worth noting that in the Siamese Network, the two parallel coding sub-networks share the encoder parameters. Then, the vector features obtained by each multi-scale width convolution residual network are combined and crossed to help the network model learn the "matching" features of the "paired" input images, and the crossed feature vectors are sent to the feature mapping layer, and then a loss function is designed to measure the similarity of the two inputs, so that the overall architecture of the network can complete the identity verification task.

[0100] II. Sub-network in Siamese Network structure

[0101] The multi-scale width convolution residual network proposed in the present application uses an improved two-channel convolution residual block structure at each layer to extract deep image features, aiming to extract more complete and rich image features, and thus improve the classification accuracy and efficiency of the sub-network.

[0102] (1) Improved two-channel convolution residual block

[0103] As shown in Figure 3 , the present application learns from the residual block structure in the residual network and introduces the DropBlock structure. First, batch normalization is used to maintain the distribution of image data, avoiding the phenomenon of covariate shift caused by parameter updating, and Mish function is used to activate network parameters, and then two-channel convolution (see Figure 4 ) is used to extract image features, and DropBlock operation is used to regularize and effectively avoid overfitting. Then, batch normalization, Mish activation, and two-channel convolution operations are repeatedly performed in sequence. Finally, the input image and the feature map obtained by connecting two convolutions are added pixel by pixel as the output of the improved asymmetric convolution residual block structure. The improved asymmetric convolution residual block structure uses residual units to directly map to improve the feature extraction capability of the network, and through the improved asymmetric convolution module, the feature extraction capability of the image is strengthened, and the robustness of image flipping and rotation is also enhanced.

[0104] (2) Two-channel convolution module

[0105] As shown in Figure 4As shown, the two-part channel convolution module sets the number of convolution kernels of the current layer network as N and divides it into two groups, which are respectively recorded as group 1 and group 2, that is, the number of convolution kernels in each group is N / 2. Then, the input image is convolved in the N / 2 convolution kernels of group 1 and the N / 2 convolution kernels of group 2, and then the feature maps obtained by the two convolution operations are spliced respectively. Finally, the channel number of the spliced feature map is adjusted to N through convolution, and a branch with a convolution kernel size of 1*1 is introduced from the input end in the short connection mode of the residual network to adjust the channel number and extract the nonlinear features of the input image, enhance the nonlinear characteristics of the network, and then the feature information of the feature maps obtained by the two branches of the operation input image in parallel is aggregated by adding the corresponding position pixels.

[0106] The two-part channel convolution module divides the number of convolution kernels into two groups and acts on the input in parallel, and then splices and fuses the feature maps obtained by each group of convolution kernels and the input image, which helps to extract rich feature information of the input image. The introduction of the short connection mode can directly supplement the information of the input image through the 1*1 size convolution operation to further obtain more complete feature information in the case of information loss in forward propagation, thereby enhancing the expression ability of the two-part channel convolution module in feature extraction.

[0107] Compared with other activation functions, the ReLU activation function is simple to derive and can improve the problem of gradient disappearance to some extent, but because a part of data will fall into the negative interval, these data will become 0, resulting in that the corresponding weight cannot be updated, thereby appearing "dead" neurons. In order to overcome this defect, the Mish function is selected to activate the neurons in the two-part channel convolution module, and the precision of Mish is better than that of ReLU, and it has stability in deep network, and its expression is:

[0108] Mish(x)=x·tanh(ln(1+e x ));

[0109] In the formula, x represents the feature map (also called feature map) of the previous layer of the network.

[0110] III. Similarity measurement of ID portrait and on-site photographed image

[0111] The twin neural network uses a similarity score to predict the similarity of two inputs, determines the relative distance between two input images by a method of metric learning, and a loss function is a basic but crucial concept in deep learning, which evaluates the performance of the network model by calculating the gap between the predicted value and the actual data. The smaller the error between the predicted value and the actual value, the smaller the output of the loss function, and the better the performance of the network model. The binary cross-entropy loss function is used to calculate the similarity score in the application. Therefore, in order to better optimize the network model, the loss function needs to be trained to be minimized and keep convergence. Cross-entropy is an important concept in information theory, which mainly measures the difference between two probability distributions of the same random variable. The expression of the binary cross-entropy loss function is:

[0112]

[0113] Wherein, N represents the number of samples, L represents the loss function, y represents the label of the sample, and p represents the probability of the sample being predicted as a positive class.

[0114] The intelligent identity verification method for the self-service terminal provided in the embodiment adopts the design idea of the twin neural network for the overall architecture of the identity verification of the self-service terminal, aims to compare the similarity between the two input images, and the core idea is to learn from the semantic similarity of the two images to estimate the gap between them, so that the comparison between the user's ID photo and the live shooting image can be realized more ideally. The improved asymmetric convolution residual block is used as the subnetwork of the twin network, which aims to extract more rich and complete feature information of the ID photo and the live shooting image in parallel, and further improve the recognition accuracy of the two input images. Although the speed of shooting the front image of the user is fast, sometimes the camera of the automatic terminal directly shoots the user's face before the user is ready, so that the user's face is not clear and the recognition is difficult, and the difficulty of matching the ID photo is increased. The improved asymmetric convolution residual block structure can better solve this difficulty.

[0115] Figure 8 is a structural schematic diagram of the intelligent identity verification device for the self-service terminal provided in an embodiment of the application, as Figure 8 shown, the intelligent identity verification device for the self-service terminal provided in the embodiment of the application comprises:

[0116] The acquisition module 51 is configured to acquire a pair of user ID photos and live user face images shot by the camera of the self-service terminal.

[0117] The verification module 52 is configured to input the user ID photo and the live user face image into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model.

[0118] The pre-trained intelligent identity verification model respectively encodes the user ID head portrait and the live user's face image by using two twin convolutional neural network modules, obtains two encoded feature vectors, then fully connects the two encoded feature vectors to obtain a fully connected feature vector, and calculates the fully connected feature vector to obtain a verification result.

[0119] The intelligent identity verification device for a self-service terminal provided by the embodiments of the present application obtains a pair of user ID head portraits and live user's face images captured by a camera of the self-service terminal; inputs the user ID head portraits and the live user's face images into a pre-trained intelligent identity verification model to obtain a verification result output by the intelligent identity verification model; wherein the pre-trained intelligent identity verification model respectively encodes the user ID head portraits and the live user's face images by using two twin convolutional neural network modules, obtains two encoded feature vectors, then fully connects the two encoded feature vectors to obtain a fully connected feature vector, and calculates the fully connected feature vector to obtain a verification result. In this way, the intelligent identity verification model is established by using the design idea of a twin neural network, and the gap between the two is estimated from the semantic similarity of the two images, so that the comparison between the user ID head portrait and the live face image captured on site can be more ideally realized.

[0120] In some embodiments, the convolutional neural network module is a multi-scale width convolution residual network module.

[0121] In some embodiments, the encoding of the user ID head portrait / live user's face image by the convolutional neural network module to obtain an encoded feature vector includes:

[0122] performing batch normalization, Mish activation, two-channel convolution, and DropBlock operation on the user ID head portrait / live user's face image in sequence to obtain a first feature map;

[0123] continuing to perform batch normalization, Mish activation, two-channel convolution, and DropBlock operation on the first feature vector in sequence to obtain a second feature map;

[0124] performing pixel-by-pixel addition on the second feature map and the user ID head portrait / live user's face image to obtain an encoded feature vector.

[0125] In some embodiments, the two-channel convolution operation is used to extract features of an image, and the two-channel convolution operation includes:

[0126] The input image is convolved by the first N / 2 convolution kernels and the last N / 2 convolution kernels of the current layer convolution network respectively to obtain two feature maps, where N is the number of convolution kernels of the current layer convolution network.

[0127] The two feature maps and the input image are spliced to obtain a spliced image.

[0128] The spliced image is convolved by a convolution network with a convolution kernel size of 1*1 to obtain an enhanced image.

[0129] In some embodiments, in the two-branch convolution operation, a Mish function is used to activate neurons.

[0130] In some embodiments, the apparatus further includes a model training module configured to:

[0131] Obtain pairs of user ID head portraits and live user face images in an image set, where a part of the pairs of user ID head portraits and live user face images are of the same user, and another part of the pairs of user ID head portraits and live user face images are not of the same user.

[0132] For each pair of user ID head portrait and live user face image, train a preset twin neural network model according to the pair of user ID head portrait and live user face image and a label of the pair of user ID head portrait and live user face image until a pre-trained intelligent identity verification model is obtained.

[0133] In some embodiments, the preset twin neural network model uses a binary cross-entropy loss function to calculate a similarity score between the user ID head portrait and the live user face image.

[0134] The embodiments of the apparatus provided by the embodiments of the present application can be specifically used to execute the processing procedures of the above-mentioned method embodiments, and the functions thereof will not be repeated here, and the detailed description can be referred to the above-mentioned method embodiments.

[0135] It should be noted that the intelligent identity verification method and apparatus for self-service terminals provided by the embodiments of the present application can be used in the financial field, and can also be used in any technical field other than the financial field, and the application field of the intelligent identity verification method and apparatus for self-service terminals is not limited by the embodiments of the present application.

[0136] Figure 9 The physical structure diagram of the electronic device provided by an embodiment of the present application is as follows: Figure 9As shown, the electronic device can include a processor 601, a communications interface 602, a memory 603, and a communications bus 604, wherein the processor 601, the communications interface 602, and the memory 603 complete communications with each other through the communications bus 604. The processor 601 can invoke a logical instruction in the memory 603 to execute the method described in any of the above embodiments.

[0137] In addition, the logical instruction in the memory 603 described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or in other words the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0138] The embodiment discloses a computer program product, the computer program product includes a computer program stored on a non-transitory computer readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer, the computer can execute the method provided by the method embodiments.

[0139] The embodiment provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program causes the computer to execute the method provided by the method embodiments.

[0140] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0141] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0142] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0143] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0144] In the description of the specification, the description of the terms "one embodiment", "one specific embodiment", "some embodiments", "for example", "exemplary", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are contained in at least one embodiment or example of the application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0145] The above-described specific embodiments, the purposes, technical solutions and beneficial effects of the application are further described in detail. It should be understood that the above-described is only a specific embodiment of the application and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A smart authentication method for self-service terminals, characterized in that, include: Obtain a pair of user ID card photos and a face image of the user on site, captured by the self-service terminal's camera; The user's ID card photo and the on-site user's facial image are input into a pre-trained intelligent identity verification model to obtain the verification result output by the intelligent identity verification model; The pre-trained intelligent identity verification model uses two twin convolutional neural network modules to encode the user's ID card portrait and the face image of the on-site user, respectively, to obtain two encoded feature vectors. Then, the two encoded feature vectors are fully connected to obtain a fully connected feature vector. The verification result is obtained by calculating the fully connected feature vector. The convolutional neural network module encodes the user's ID card photo / on-site user's facial image to obtain the encoded feature vector, which includes: The user's ID card photo / on-site user's face image are sequentially subjected to batch normalization, Mish activation, binary channel convolution, and DropBlock operations to obtain the first feature map; Continue to perform batch normalization, Mish activation, bichannel convolution and DropBlock operations on the first feature map to obtain the second feature map. The second feature map is added pixel by pixel to the user's ID card photo / on-site user's face image using a residual connection method to obtain the encoded feature vector.

2. The method according to claim 1, characterized in that, The convolutional neural network module is a multi-scale width convolutional residual network module.

3. The method according to claim 1, characterized in that, The bichannel convolution operation is used to extract image features, and the features extracted by the bichannel convolution operation include: The input image is convolved using the first N / 2 convolutional kernels and the last N / 2 convolutional kernels of the current convolutional network layer, respectively, to obtain two feature maps, where N is the number of convolutional kernels in the current convolutional network layer; The two feature maps and the input image are then stitched together to obtain a stitched image. An enhanced image is obtained by performing a convolution operation on the stitched image using a convolutional network with a kernel size of 1×1.

4. The method according to claim 3, characterized in that, In the bichannel convolution operation, the Mish function is used to activate neurons.

5. The method according to any one of claims 1 to 4, characterized in that, The training process of the intelligent identity verification model is as follows: Acquire pairs of user ID card images and on-site user facial images from the image set. Among them, some pairs of user ID card images and on-site user facial images belong to the same user, while other pairs of user ID card images and on-site user facial images do not belong to the same user. For each pair of user ID card photos and on-site user facial images, a preset twin neural network model is trained based on the pair of user ID card photos, on-site user facial images, and the labels of the pair of user ID card photos and on-site user facial images until the pre-trained intelligent identity verification model is obtained.

6. The method according to claim 5, characterized in that, The preset twin neural network model uses a binary cross-entropy loss function to calculate the similarity score between the user's ID card photo and the on-site user's facial image.

7. An intelligent identity verification device for a self-service terminal, characterized in that, include: The acquisition module is used to acquire a pair of user ID card photos and on-site user facial images captured by the self-service terminal's camera; The verification module is used to input the user's ID card photo and the face image of the on-site user into a pre-trained intelligent identity verification model to obtain the verification result output by the intelligent identity verification model. The pre-trained intelligent identity verification model uses two twin convolutional neural network modules to encode the user's ID card portrait and the face image of the on-site user, respectively, to obtain two encoded feature vectors. Then, the two encoded feature vectors are fully connected to obtain a fully connected feature vector. The verification result is obtained by calculating the fully connected feature vector. The convolutional neural network module encodes the user's ID card photo / on-site user's facial image to obtain the encoded feature vector, which includes: The user's ID card photo / on-site user's face image are sequentially subjected to batch normalization, Mish activation, binary channel convolution, and DropBlock operations to obtain the first feature map; Continue to perform batch normalization, Mish activation, bichannel convolution and DropBlock operations on the first feature map to obtain the second feature map. The second feature map is added pixel by pixel to the user's ID card photo / on-site user's face image using a residual connection method to obtain the encoded feature vector.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images