Face recognition method and equipment based on bank self-service terminal and medium

Abstract

The embodiment of the invention discloses a face recognition method and device based on a bank self-service terminal and a medium, belongs to the technical field of face recognition, and solves the problem that the accuracy of user identity recognition is affected due to insufficient segmentation precision when a face image collected by the bank self-service terminal is processed in the prior art. Acquiring a face image of a customer through a camera device arranged on the bank self-service terminal; inputting the face image into a preset neural network, and performing face image segmentation on the face image through a coding path and a decoding path of the preset neural network to obtain user identity recognition information; acquiring historical service information corresponding to the user according to the user identity identification information, and performing comparison detection on the current service information and the historical service information; and when the detection result is abnormal business, abnormal information is generated, and secondary face recognition is carried out based on the bank self-service terminal so as to check the abnormal information.

Description

Technical Field

[0001] This application relates to the field of facial recognition technology, and in particular to a facial recognition method, device and medium based on a bank self-service terminal. Background Technology

[0002] In the field of financial automation services, facial recognition technology at bank self-service terminals is a core component of user authentication, and its accuracy and efficiency directly impact the security of financial services and user experience. Facial recognition methods based on 2D fully convolutional neural networks face numerous technical challenges when processing 3D facial images captured by bank self-service terminals. This is particularly true given the complex and varied business scenarios at these terminals, including fluctuations in lighting intensity, multi-angle shooting by users, and motion blur, which often result in blurred edges and distorted features in the captured facial images.

[0003] Existing facial recognition methods suffer from insufficient segmentation accuracy when processing facial images captured by bank self-service terminals. This is especially true when complex business scenarios result in blurred images or multi-angle shots, making it difficult to accurately extract facial features and thus affecting the accuracy of user identification. Summary of the Invention

[0004] This application provides a face recognition method, device, and medium based on a bank self-service terminal to solve the following technical problem: existing face recognition methods have insufficient segmentation accuracy when processing face images collected by bank self-service terminals, making it difficult to accurately extract face features, thereby affecting the accuracy of user identification.

[0005] The embodiments of this application adopt the following technical solutions: This application provides a facial recognition method based on a bank self-service terminal. The method includes: acquiring a customer's facial image using a camera device installed on the bank self-service terminal; inputting the facial image into a pre-set neural network to perform facial image segmentation through the encoding and decoding paths of the pre-set neural network to obtain user identification information; obtaining the user's corresponding historical transaction information based on the user identification information, and comparing and detecting the current transaction information with the historical transaction information; if the detection result indicates an abnormal transaction, generating abnormal information, and performing secondary facial recognition based on the bank self-service terminal to view the abnormal information.

[0006] In one implementation of this application, a facial image is input into a preset neural network to perform multi-scale feature extraction on the facial image through the encoding and decoding paths of the preset neural network. Specifically, during multi-scale feature extraction through the encoding path: the facial image is initially processed by the first residual module and the second residual module in the preset neural network to generate a first feature map; the first feature map is downsampled by the third residual module and the second residual module in the preset neural network to obtain a plurality of second feature maps decreasing according to a preset ratio; wherein, decreasing according to the preset ratio includes at least one of decreasing the size according to the preset ratio and decreasing the number of channels according to the preset ratio.

[0007] In one implementation of this application, multiple second feature maps are sequentially obtained in descending order of a preset ratio. Specifically, the extracted second feature maps include at least four scales, and each scale of the feature map includes at least a first path and a second path. The first path consists of two 3×3×3 convolutions with padding of 1, each followed by a group normalization layer and a ReLU layer. The second path consists of one 1×1×1 convolution without padding, followed by a group normalization layer. The features of the first path and the second path are added together to form an identity mapping relationship, and then the feature scale is reduced by max pooling.

[0008] In one implementation of this application, a facial image is input into a pre-set neural network to perform facial image segmentation through the encoding and decoding paths of the pre-set neural network. Specifically, during multi-scale feature extraction through the decoding path: the feature map generated by the encoding path is upsampled by transposed convolution; the feature map obtained from each upsampling is concatenated with the feature map of the corresponding level of the encoding path, and the concatenated feature map is weighted by an attention mechanism; the concatenated image is optimized for feature extraction through a fourth residual module, and the channels are compressed through a 1×1 convolutional layer to output a segmentation result with the same size as the facial image.

[0009] In one implementation of this application, the feature map obtained from each upsampling is concatenated with the feature map of the corresponding level of the encoding path. Specifically, this includes: concatenating the upsampled feature map with the 64×64×64 feature map corresponding to the encoding path to generate a 64×64×128 feature map, which is then processed by the fourth residual module and output as a 64×64×64 feature map; concatenating the upsampled feature map with the 128×128×32 feature map corresponding to the encoding path to generate a 128×128×64 feature map, which is then processed by the fourth residual module and output as a 128×128×32 feature map; and concatenating the upsampled feature map with the 256×256×16 feature map corresponding to the encoding path to generate a 256×256×32 feature map, which is then processed by the fourth residual module and output as a 256×256×16 feature map.

[0010] In one implementation of this application, historical business information corresponding to the user is obtained based on user identification information, and the current business information is compared and detected with the historical business information. Specifically, this includes: obtaining historical business information of the user within a preset time period from the database based on user identification information; wherein the historical business information includes at least one of transaction type, transaction amount, and transaction frequency; determining transaction deviation information between the current business information and the historical business information; wherein the transaction deviation information includes at least one of type overlap rate, amount deviation rate, and transaction spatiotemporal deviation rate; and performing weighted fusion on the transaction deviation information to generate a transaction risk level based on the weighted fusion.

[0011] In one implementation of this application, when the detection result indicates an abnormal business, abnormal information is generated, specifically including: activating a dedicated branch for facial expression features corresponding to the decoding path when the transaction risk level is greater than a preset risk level; increasing feature extraction of key facial expression regions through an attention mechanism; wherein the key facial expression regions include at least one of the eyebrow and eye regions and the lip region; concatenating the facial expression features of adjacent frames based on temporal order to obtain a temporal expression sequence; performing expression change analysis on the temporal expression sequence to obtain the micro-expression change frequency and micro-expression change amplitude corresponding to the key facial expression regions; generating abnormal information when the micro-expression change frequency and / or micro-expression change amplitude are within an abnormal range; wherein the abnormal information includes at least one of the current business information and the current expression feature information.

[0012] In one implementation of this application, secondary facial recognition is performed based on a bank self-service terminal to view abnormal information. Specifically, this includes: receiving an abnormal information query instruction sent by the bank self-service terminal; generating an abnormal code based on the abnormal information, and generating an abnormal information QR code based on the abnormal code; displaying the abnormal information QR code through the bank self-service terminal, and performing secondary facial recognition through the bank self-service terminal after successful scanning verification; comparing the secondary facial recognition result with the facial features at the time of the abnormal transaction, and sending the abnormal information to the user terminal if the comparison result meets the preset comparison conditions.

[0013] This application provides a facial recognition device based on a bank self-service terminal, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: acquire a customer's facial image using a camera device installed on the bank self-service terminal; input the facial image into a preset neural network to perform facial image segmentation through the encoding and decoding paths of the preset neural network to obtain user identification information; obtain the user's corresponding historical business information based on the user identification information, and compare and detect the current business information with the historical business information; if the detection result is an abnormal business, generate abnormal information and perform secondary facial recognition based on the bank self-service terminal to view the abnormal information.

[0014] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: acquire a customer's facial image using a camera device installed on a bank self-service terminal; input the facial image into a preset neural network to perform facial image segmentation through the encoding and decoding paths of the preset neural network to obtain user identification information; obtain the user's corresponding historical business information based on the user identification information, and compare and detect the current business information with the historical business information; if the detection result indicates an abnormal business, generate abnormal information and perform secondary facial recognition based on the bank self-service terminal to view the abnormal information.

[0015] The at least one technical solution adopted in this application embodiment can achieve the following beneficial effects: Firstly, this application embodiment performs image segmentation through the encoding and decoding path of a neural network, accurately extracting key facial features and improving the accuracy of facial image segmentation. Secondly, this application embodiment retrieves historical business data based on user identification information, compares the current business with historical patterns, and can automatically detect abnormal operations that do not conform to user habits, preventing malicious operation of user accounts and reducing the labor costs of manual review by banks. Furthermore, when abnormal business is detected, this application automatically triggers secondary facial recognition and guides the user to view abnormal information through a self-service terminal, improving service response efficiency and reducing regulatory risks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a face recognition method based on a bank self-service terminal is provided for embodiments of this application; Figure 2 This is a schematic diagram of the structure of a face recognition device based on a bank self-service terminal, provided as an embodiment of this application.

[0017] Figure label: 200: Face recognition device based on bank self-service terminal; 201: Processor; 202: Memory. Detailed Implementation

[0018] This application provides a face recognition method, device, and medium based on a bank self-service terminal.

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] The technical solutions proposed in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Figure 1 A flowchart of a face recognition method based on a bank self-service terminal is provided for an embodiment of this application, as follows: Figure 1 As shown, the facial recognition method based on bank self-service terminals includes the following steps: S101. Collect facial images of customers using a camera device installed on a bank's self-service terminal.

[0022] In one implementation of this application, the hardware parameters of the camera device of the bank's self-service terminal are first calibrated. The resolution is set to match the image size input to the subsequent pre-set neural network, such as RAU-Net, and the frame rate is adjusted to ensure real-time acquisition without delay. The self-service terminal guides the customer through a combination of screen display and voice broadcast, prompting the customer to keep their face within the frame using a red rectangular viewfinder, controlling the distance and viewing angle from the terminal. Simultaneously, the camera device acquires a preview image in real time. When a single face occupies 60%-80% of the viewfinder, there is no significant displacement in multiple consecutive frames, and the key facial areas are intact, a acquisition preparation signal is triggered. At this time, the camera device continuously acquires multiple frames of facial images. After each frame is transmitted to the local processing unit, the clarity is determined by the grayscale gradient value, the illumination uniformity is determined by the brightness standard deviation, and the facial integrity is confirmed by edge detection. Finally, the optimal frame is selected from the images that meet the standards as the acquisition result. If none of the frames meet the standards, the acquisition process is restarted.

[0023] Furthermore, the optimally acquired image is scaled to 256×256 pixels using bilinear interpolation, and the pixel values ​​are normalized to the 0-1 range. Then, it is standardized using the mean and variance of the pre-training dataset. The preprocessed image is then encrypted using an encryption algorithm, with metadata such as the acquisition timestamp and terminal number appended, and transmitted to the server via an encrypted channel for subsequent face recognition.

[0024] S102. Input the face image into a preset neural network to perform face image segmentation through the encoding and decoding paths of the preset neural network, and obtain user identification information.

[0025] In one implementation of this application, during multi-scale feature extraction via an encoding path, the facial image is initially processed using the first and second residual modules in the pre-set neural network to generate a first feature map. The first feature map is then downsampled using the third and second residual modules in the pre-set neural network to sequentially obtain multiple second feature maps decreasing according to a preset ratio. This decreasing according to a preset ratio includes at least one of decreasing the size according to a preset ratio and decreasing the number of channels according to a preset ratio.

[0026] Specifically, the pre-built neural network in this embodiment is divided into an encoding path and a decoding path. The encoding path performs feature extraction on the input image at four scales. Each scale of feature extraction includes two paths. One path uses two 3×3×3 convolutions with padding of 1, each followed by a Group Normalization (GN) layer and a ReLU layer. The other path uses one 1×1×1 convolution without padding, followed by a GN layer. The features from the two paths are added to form an identity mapping relationship, and then the feature scale is reduced by a max pooling layer. The encoder ends with a multi-scale feature fusion module.

[0027] Furthermore, the pre-built neural network RAU-Net in this embodiment adopts an encoder-decoder structure. Encoding stage: The input is a 256×256×3 image, encoded using a multi-structure residual structure. Specifically, residual module a and residual module c generate a 256×256×16 feature map. Downsampling stage (repeated three times): After processing by residual module b and residual module c, 128×128×32, 64×64×64, and 32×32×128 feature maps are generated sequentially.

[0028] In one implementation of this application, during multi-scale feature extraction via the decoding path, the feature map generated by the encoding path is upsampled using transposed convolution. The upsampled feature map is then concatenated with the feature map of the corresponding layer in the encoding path, and the concatenated feature map is weighted using an attention mechanism. The concatenated image is then optimized for feature extraction using a fourth residual module, and channels are compressed using a 1×1 convolutional layer to output a segmentation result consistent with the size of the facial image.

[0029] Specifically, in the decoding path of this application embodiment, each scale includes a trilinear interpolation upsampling layer, followed by the same identity mapping as the encoding path. Features of the same scale in the encoding path are concatenated to the decoding path to provide it with the original low-level features. The last layer is a 1×1×1 convolutional layer, and the final output channel number is the number of label categories. Since the input image dimension is 240×240×155, the original features need to be padded before concatenation.

[0030] Furthermore, in the decoding stage, this embodiment uses transposed convolution for upsampling. After each upsampling, it is concatenated with the feature map of the corresponding layer of the encoder. Weighted feature extraction is performed through an attention mechanism, and features are further extracted by the residual module d. Specifically, the decoding process is as follows: The upsampled feature map is concatenated with the encoder's 64×64×64 feature map by the residual module d to generate a 64×64×128 feature map. The 64×64×64 feature map is output by the residual module d. The upsampled feature map is concatenated with the encoder's 128×128×32 feature map to generate a 128×128×64 feature map, and the 128×128×32 feature map is output by the residual module d. The upsampled feature map is concatenated with the encoder's 256×256×16 feature map to generate a 256×256×32 feature map, and the 256×256×16 feature map is output by the residual module d. Finally, the channel is compressed by a 1×1 convolutional layer to output a 256×256×1 feature map.

[0031] Furthermore, the multi-structure residual module in this embodiment includes four residual structures, whose general design follows the residual connection principle: output = main path processing + shortcut path processing. The shortcut path adjusts its dimension through a 1×1 convolution to ensure consistency with the main path output channel, enhancing cross-layer information interaction. This embodiment improves feature reuse efficiency and suppresses the risk of gradient explosion while optimizing segmentation accuracy. The differences between the modules are as follows: Input Block: The main path is 1×1 convolution, Batch Normalization (BN) layer, ReLU function, 3×3 convolution layer, BN, ReLU function, 1×1 convolution layer. The shortcut is a 1×1 convolution layer to achieve channel alignment.

[0032] Downsample Block: The main path is BN, ReLU function, 1×1 convolutional layer with stride=2, BN, ReLU function, 3×3 convolutional layer, BN, ReLU function, 1×1 convolutional layer. The shortcut is a 1×1 convolutional layer with stride=2.

[0033] Bottleneck Block: The main path consists of Batch Normalization (BN), ReLU function, 1×1 convolutional layer, BN, ReLU function, and 3×3 convolutional layer BN. The shortcut is a 1×1 convolutional layer. The key feature is that the output of the main path is added to the input.

[0034] Residual Block: The main path is Batch Normalization (BN), ReLU function, 3×3 convolutional layer, BN, ReLU function, 3×3 convolutional layer. The shortcut is a 1×1 convolutional layer.

[0035] In this embodiment, except for the end of the Bottleneck Block, all convolutional layers are followed by Batch Normalization (BN) and ReLU activation. 1×1 convolutions enforce consistency between input and output channels, and strict dimensionality matching is required before feature summation.

[0036] In one implementation of this application, an embodiment of this application also proposes a fusion loss function to improve network performance and uses identity mapping to prevent gradient explosion or gradient vanishing problems that may occur during backpropagation.

[0037] Specifically, the loss function is a crucial factor affecting the performance of the segmentation model. Based on the cross-entropy loss function, Dice loss is introduced. The cross-entropy loss function is defined as follows: ; in, Representing the i A label of 1 pixel, with a value of 0 or 1. Representing the i The predicted value for each pixel ranges from 0 to 1. During training, the binary cross-entropy loss function may suffer from the vanishing gradient problem. The Dice loss function is introduced to address this issue, and its formula is as follows: ; η This represents the smoothing term, set to 1. The smoothing term prevents the problem of the numerator becoming infinitely large when divided by 0, ensuring training stability. The formula for the fused loss function is as follows: .

[0038] S103. Based on the user's identity information, obtain the user's corresponding historical business information, and compare and detect the current business information with the historical business information.

[0039] In one implementation of this application, historical business information within a preset time period corresponding to the user is retrieved from a database based on user identification information; wherein the historical business information includes at least one of transaction type, transaction amount, and transaction frequency. Transaction deviation information between current business information and historical business information is determined; wherein the transaction deviation information includes at least one of type overlap rate, amount deviation rate, and transaction spatiotemporal deviation rate. The transaction deviation information is weighted and fused to generate a transaction risk level based on the weighted fusion.

[0040] Specifically, in this embodiment, user identification information obtained from face image segmentation in the RAU-Net network is used to initiate a targeted query request to the bank's core business database. The query request carries the user's unique identifier and specifies a preset time period, such as the past 3 months, 6 months, or 12 months. The database is then used to extract the user's historical business information within the preset time period, including at least one of the following: transaction type, transaction amount, and transaction frequency. Simultaneously, the extracted historical business information is structured and converted into feature data in a unified format.

[0041] Furthermore, the current business information, including the type of business being processed, transaction amount, transaction time, and transaction terminal location, is compared with the extracted historical business information dimension by dimension to calculate transaction deviation information. The overlap rate is determined by the proportion of the current business type appearing in the historical preset time period to the total number of transactions. For example, if the current transaction is a cross-border transfer and this type has appeared 5 times in the past 6 months with a total of 100 transactions, the overlap rate is 5%. The amount deviation rate is calculated by |current transaction amount - average amount of similar transactions in the historical preset time period| / average amount of similar transactions in the past. The transaction time-space deviation rate is calculated by combining time deviation and spatial deviation. Time deviation is determined by the overlap between the current transaction time and the historical high-frequency transaction period. For example, if the historical high-frequency period is 9:00-17:00 and the current period is 2:00 AM, the time deviation rate is 100%. Spatial deviation is determined by the proportion of the geographical distance between the current transaction terminal location and the historically commonly used terminal location to the historical maximum transaction distance. Finally, the time deviation rate and spatial deviation rate are merged with a 1:1 weight to obtain the transaction time-space deviation rate, ensuring that the transaction deviation information covers at least one of the type overlap rate, amount deviation rate, and transaction time-space deviation rate.

[0042] Furthermore, the calculated transaction deviation information is weighted and fused. First, the weights of each deviation information are preset according to the bank's risk control rules. Then, the value of each deviation information is multiplied by its corresponding weight, and the sum is obtained to obtain a comprehensive transaction deviation score. Based on this comprehensive score, the transaction risk level is divided, with a preset score of 0-30 for low risk, 31-60 for medium risk, and 61-100 for high risk. At this point, a second facial recognition and manual review are triggered, and the corresponding judgment criteria for the risk level are generated to ensure that the risk level generation process is traceable.

[0043] S104. If the detection result is an abnormal business, generate abnormal information and perform secondary facial recognition based on the bank's self-service terminal to view the abnormal information.

[0044] In one implementation of this application, when the transaction risk level exceeds a preset risk level, a dedicated branch for facial expression features corresponding to the decoding path is activated. An attention mechanism is used to enhance feature extraction of key facial expression regions; these key regions include at least one of the eyebrow / eye region and the lip region. Based on temporal sequence, the facial expression features of adjacent frames are concatenated to obtain a temporal expression sequence. Expression change analysis is performed on the temporal expression sequence to obtain the frequency and amplitude of micro-expression changes corresponding to the key facial expression regions. If the frequency and / or amplitude of micro-expression changes are within an abnormal range, anomaly information is generated; this anomaly information includes at least one of current business information and current expression feature information.

[0045] Specifically, this embodiment compares the fused transaction risk level with a preset risk level. When the transaction risk level is greater than the preset risk level, the function switch of the RAU-Net network decoding path is automatically triggered, and a pre-configured expression feature-specific branch is started. This expression feature-specific branch is based on the transposed convolution and feature concatenation architecture of the RAU-Net decoder, and optimizes the channel dimension and convolution kernel parameters for facial expression feature extraction. During the operation of the expression feature-specific branch, key facial expression regions, including at least the eyebrow and lip regions, are first located from the multi-scale feature map generated by the decoding path. Importance weights for each feature channel are calculated through channel attention, increasing the weights of the feature channels corresponding to the eyebrow and lip regions, for example, to 40%-50%. Simultaneously, a heatmap is generated through spatial attention, highlighting the positions of the eyebrow and lip regions in the feature map to reduce background interference and ensure the accuracy of key expression feature extraction.

[0046] Secondly, this embodiment of the application also utilizes a bank self-service terminal camera device to continuously acquire user facial expression images at a frequency of 10 frames per second. Each frame image is processed by a dedicated branch of RAU-Net to extract a 32-dimensional facial expression feature vector. Based on the chronological order, the facial expression feature vectors corresponding to adjacent frames are sequentially concatenated to form a temporal expression sequence, with each feature vector appended with a timestamp to ensure the temporal continuity and integrity of the sequence. Furthermore, the constructed temporal expression sequence is subjected to frame-by-frame comparative analysis. For key facial expression regions, the difference values ​​of feature vectors between adjacent frames are calculated: the frequency of micro-expression changes is determined by the number of times the feature difference value exceeds a preset threshold per unit time; the amplitude of micro-expression changes is calculated by the difference between the maximum and minimum value of the feature difference value, respectively obtaining the frequency and amplitude of micro-expression changes in the eyebrow and lip regions, forming a multi-dimensional expression change analysis result. The normal range of micro-expression changes is preset. The frequency and amplitude of micro-expression changes obtained from the analysis are compared with the normal range. When the frequency of micro-expression changes exceeds the upper limit of the normal range, or the amplitude of micro-expression changes exceeds the upper limit of the normal range, or both exceed the upper limit, it is judged as an expression abnormality. At this time, abnormal information containing abnormality type, abnormal value and collection timestamp is generated.

[0047] In one implementation of this application, an abnormal information query instruction sent by a bank self-service terminal is received. An abnormal code is generated based on the abnormal information, and an abnormal information QR code is generated based on the abnormal code. The abnormal information QR code is displayed through the bank self-service terminal, and after successful scanning verification, a secondary facial recognition is performed through the bank self-service terminal. The result of the secondary facial recognition is compared with the facial features of the person at the time of the abnormal transaction. If the comparison result meets preset comparison conditions, the abnormal information is sent to the user terminal.

[0048] Specifically, the bank's server monitors information exchange requests from bank self-service terminals in real time. When it receives an abnormal information query command from a self-service terminal, it first verifies the legality of the command, checking whether the terminal number and user identity identifier carried in the command are consistent with the server's registration information. After successful verification, it extracts the core information of the abnormal transaction from the command, such as the abnormal transaction time and transaction number, providing basic data for subsequent abnormal code generation. Secondly, based on the parsed core information of the abnormal transaction, a unique abnormal code is generated according to the bank's preset coding rules. The coding format can be terminal number + abnormal transaction date + 6-digit random sequence to ensure the uniqueness and traceability of the code. The abnormal code and the corresponding abnormal information are integrated into structured data. A QR code generation tool is used to perform URL encoding conversion on this structured data to generate an abnormal information QR code. The QR code contains a unique access path for querying abnormal information and supports automatic redirection to an encrypted query interface after scanning the code.

[0049] Furthermore, the bank server transmits the generated abnormal information QR code data to the bank's self-service terminal that initiated the query request. Upon receiving the data, the terminal invokes its built-in display control module to display the QR code in a designated area of ​​the screen, simultaneously displaying operation instructions. When the user scans the QR code via their mobile terminal, they send a verification request to the server. The server verifies the binding status of the scanning device; upon successful verification, it sends a successful verification signal to the self-service terminal. Upon receiving the successful verification signal from the server, the bank's self-service terminal automatically initiates a secondary face recognition process, using its onboard camera to capture the user's real-time face image. The captured face image is input into a pre-set RAU-Net network, where face image segmentation is performed using the network's encoding and decoding paths. The user's real-time face feature vector is extracted and encrypted to ensure data security during transmission. After receiving the secondary face recognition feature vector transmitted from the self-service terminal, the bank server retrieves the face feature vectors collected during the abnormal transaction from its database and calculates the matching degree between the two sets of feature vectors using a cosine similarity algorithm. The matching score is compared with the preset comparison conditions. If the comparison result meets the conditions, it means that the current user is the same as the user at the time of the abnormal transaction. The server will encrypt and push the abnormal information to the user terminal according to the terminal information reserved by the user. If the comparison result does not meet the conditions, an identity verification failure prompt will be sent to the self-service terminal, requiring a second face recognition or other verification methods to be added.

[0050] Figure 2 This is a schematic diagram of the structure of a face recognition device based on a bank self-service terminal, provided as an embodiment of this application. Figure 2 As shown, a face recognition device 200 based on a bank self-service terminal includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201. The memory 202 stores instructions executable by the at least one processor 201, which, when executed, enable the at least one processor 201 to: acquire a customer's facial image using a camera device installed on the bank self-service terminal; input the facial image into a preset neural network to perform face image segmentation through the encoding and decoding paths of the preset neural network to obtain user identification information; obtain the user's corresponding historical transaction information based on the user identification information, and compare the current transaction information with the historical transaction information for detection; if the detection result indicates an abnormal transaction, generate abnormal information and perform secondary face recognition based on the bank self-service terminal to view the abnormal information.

[0051] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: acquire a customer's facial image using a camera device installed on a bank self-service terminal; input the facial image into a preset neural network to perform facial image segmentation through the encoding and decoding paths of the preset neural network to obtain user identification information; obtain the user's corresponding historical business information based on the user identification information, and compare and detect the current business information with the historical business information; if the detection result indicates an abnormal business, generate abnormal information and perform secondary facial recognition based on the bank self-service terminal to view the abnormal information.

[0052] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0053] The above descriptions are merely embodiments of this application and are not intended to limit the scope of this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions in the embodiments of this application.

Claims

1. A face recognition method based on a bank self-service terminal, characterized in that, The method includes: Customer facial images are captured by cameras installed on bank self-service terminals; The facial image is input into a preset neural network to perform facial image segmentation through the encoding and decoding paths of the preset neural network, thereby obtaining user identification information. Based on the user identification information, obtain the user's corresponding historical business information, and compare and detect the current business information with the historical business information; If the detection result indicates an abnormal transaction, abnormal information is generated, and secondary facial recognition is performed based on the bank's self-service terminal to view the abnormal information.

2. The face recognition method based on the bank self-service terminal according to claim 1, characterized in that, The step of inputting the facial image into a preset neural network to perform multi-scale feature extraction on the facial image through the encoding and decoding paths of the preset neural network specifically includes, when performing multi-scale feature extraction through the encoding path: The facial image is initially processed by the first residual module and the second residual module in the pre-set neural network to generate a first feature map; The first feature map is downsampled by the third residual module and the second residual module in the preset neural network to obtain a plurality of second feature maps decreasing in a preset ratio. The reduction according to a preset ratio includes at least one of the following: a reduction in size according to a preset ratio and a reduction in the number of channels according to a preset ratio. 3.The face recognition method based on a bank self-service terminal of claim 2, wherein, The process of sequentially obtaining multiple second feature maps that decrease according to a preset ratio specifically includes: The extracted second feature map includes at least four scales, and each scale's feature map includes at least the first path and the second path; The first path consists of two 3×3×3 convolutions padded with 1, each followed by a group normalization layer and a ReLU layer. The second path consists of one 1×1×1 convolution without padding, followed by a group normalization layer. The features of the first path and the second path are added together to form an identity mapping relationship, and then feature scaling is performed by max pooling.

4. The face recognition method based on the bank self-service terminal according to claim 1, characterized in that, The step of inputting the facial image into a preset neural network to perform facial image segmentation through the encoding and decoding paths of the preset neural network specifically includes, when performing multi-scale feature extraction through the decoding path: The feature map generated through the encoding path is upsampled by transposed convolution; The feature map obtained from each upsampling is concatenated with the feature map of the corresponding layer of the encoding path, and the concatenated feature map is weighted by an attention mechanism. The fourth residual module optimizes feature extraction of the stitched image and compresses the channels through a 1×1 convolutional layer to output a segmentation result with the same size as the face image.

5. The face recognition method based on a bank self-service terminal according to claim 4, characterized in that, The step of concatenating the feature map obtained from each upsampling with the feature map of the corresponding layer of the encoding path specifically includes: The upsampled feature map is concatenated with the 64×64×64 feature map corresponding to the encoding path to generate a 64×64×128 feature map, which is then processed by the fourth residual module to output a 64×64×64 feature map. The upsampled feature map is concatenated with the 128×128×32 feature map corresponding to the encoding path to generate a 128×128×64 feature map, which is then processed by the fourth residual module to output a 128×128×32 feature map. The upsampled feature map is concatenated with the 256×256×16 feature map corresponding to the encoding path to generate a 256×256×32 feature map. After processing by the fourth residual module, the 256×256×16 feature map is output.

6. The face recognition method based on a bank self-service terminal according to claim 1, characterized in that, The step of obtaining the user's historical business information based on the user identification information and comparing the current business information with the historical business information specifically includes: Based on the user identification information, retrieve the user's historical business information within a preset time period from the database; wherein, the historical business information includes at least one of the following: transaction type, transaction amount, and transaction frequency; The transaction deviation information between the current business information and the historical business information is determined; wherein, the transaction deviation information includes at least one of the following: type overlap rate, amount deviation rate, and transaction time-space deviation rate; The transaction deviation information is weighted and fused to generate a transaction risk level based on the weighted fusion.

7. The face recognition method based on a bank self-service terminal according to claim 6, characterized in that, When the detection result indicates an abnormal service, generating abnormal information specifically includes: If the transaction risk level is greater than the preset risk level, the dedicated branch for facial expression features corresponding to the decoding path will be activated. Feature extraction of key facial expression regions is enhanced through an attention mechanism; wherein, the key facial expression regions include at least one of the eyebrow and eye regions and the lip region; Based on the chronological order, the facial expression features of adjacent frames are concatenated to obtain a temporal expression sequence; The temporal expression sequence was analyzed to obtain the frequency and amplitude of micro-expression changes corresponding to the key facial expression regions. When the frequency and / or amplitude of the micro-expression changes are within an abnormal range, the abnormal information is generated; wherein, the abnormal information includes at least one of the current business information and the current expression feature information.

8. The face recognition method based on a bank self-service terminal according to claim 1, characterized in that, The secondary facial recognition based on the bank's self-service terminal for viewing abnormal information specifically includes: Receive abnormal information query instructions sent by bank self-service terminals; An anomaly code is generated based on the anomaly information, and an anomaly information QR code is generated based on the anomaly code. The abnormal information QR code is displayed on the bank's self-service terminal, and after the QR code verification is successful, a second facial recognition is performed on the bank's self-service terminal. The secondary facial recognition result is compared with the facial features of the person during the abnormal transaction. If the comparison result meets the preset comparison conditions, the abnormal information is sent to the user terminal.

9. A facial recognition device based on a bank self-service terminal, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to perform the method described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-8.

Images