Palm vein dynamic capture and deep learning identity recognition terminal

Abstract

The invention provides a palm vein dynamic capture and deep learning identity recognition terminal, which relates to the technical field of biological feature recognition, and comprises a dynamic capture module, a deep learning identity recognition module and an implementation processing and feedback system, the dynamic capture module collects multiple frames of palm vein images and combines a deep convolutional neural network and recurrent neural network hybrid model of the deep learning identity recognition module, abundant space and time sequence features can be extracted, a high-distinction-degree comprehensive feature vector is generated, meanwhile, vein features under the complex background are adaptively learned, and the recognition accuracy is improved. Recognition precision and robustness are remarkably improved, interference such as light changes and skin states is adapted, a palm is supported to move freely within a certain range, the palm does not need to be kept static deliberately, the time of the whole process from dynamic capture to identity recognition is shortened to be within one second through a high-performance embedded processor and an optimization algorithm in combination with a real-time processing and feedback system, and the recognition efficiency is improved. Therefore, the method is suitable for a rapid identity authentication scene.

Description

Technical Field

[0001] The present invention relates to the field of biometric identification technology, and in particular to a palm vein dynamic capture and deep learning identity recognition terminal. Background Art

[0002] According to Chinese publication number "CN105224906B," an intelligent palm vein recognition system for intelligent identification is disclosed. It includes an imaging circuit, an illumination circuit, a recognition module, and a wireless transmission module. The imaging circuit includes a CCD camera, an optical lens, and a filter; the illumination circuit is composed of a near-infrared LED; the recognition module includes a recognition unit, an electronic ranging unit, and a voice unit; and the wireless transmission module transmits video information and control signals to a computer and enables communication with the computer. Through the training and feature matching methods of the present invention, the response speed can be reduced to within 8 seconds, and environmental interference with palm vein recognition can be accurately eliminated, thereby improving the accuracy and efficiency of palm vein recognition.

[0003] The above patent documents and prior art have the following technical problems when used: Problem 1: Traditional palm vein recognition systems rely on single-frame static image acquisition, which provides a single information dimension. Recognition accuracy is easily affected by palm posture deviation or image quality degradation. Furthermore, they are sensitive to environmental conditions, and interference such as light changes and skin stains can increase the false recognition rate, resulting in insufficient robustness. The second problem is that traditional static data collection systems require users to maintain a fixed hand position. If the position is incorrect or there is shaking, the data collection fails, making it inconvenient to use and having a low success rate. In addition, the existing system has a long processing time (2-5 seconds), which cannot meet the high real-time requirements and limits its application in dynamic scenes. Problem three: Traditional palm vein recognition relies on single-frame images, which are easily forged (such as printed images) and lack security. At the same time, manual feature extraction is complex in design and has poor adaptability. It lacks the ability to generalize to low-quality images or diverse scenarios, making it difficult to meet the needs of high security and wide application. Summary of the Invention

[0004] Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a palm vein dynamic capture and deep learning identity recognition terminal, which solves the following problems: 1. To address the problem that palm vein recognition collects only a single piece of information, is easily affected by the environment, and has low recognition accuracy; 2. Address the issues of long recognition time, low success rate, and poor real-time performance during palm vein recognition; 3. Address the issues of palm vein being easily forged, lacking in security and image adaptability.

[0005] Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solutions: a palm vein dynamic capture and deep learning identity recognition terminal, the recognition terminal includes a dynamic capture module, a deep learning identity recognition module and an implementation processing and feedback system, wherein: The dynamic capture module dynamically collects multiple frames of palm vein images of the user's palm through a multi-view near-infrared camera array and a depth sensor, allowing the palm to move freely within a certain range; The deep learning identity recognition module uses a hybrid model of deep convolutional neural networks and recurrent neural networks to extract spatial and temporal features from the multiple frames of palm vein images, generate a comprehensive feature vector, and compare it with the registered features in the database to complete identity recognition; The real-time processing and feedback system processes the multiple frames of palm vein images in real time through an embedded processor, and completes the entire process from dynamic capture to identity recognition within 1 second, while feeding back the recognition results to the user through a display screen and a voice module.

[0006] Preferably, the dynamic capture module includes a multi-view near-infrared camera, a near-infrared light source and a depth sensor. The multi-view near-infrared cameras are arranged in 3-5 positions, covering a viewing angle range of 30°-60° above the palm, with an acquisition frequency of 30-60 frames per second. The wavelength range of the near-infrared light source is 850-950nm, which is used to penetrate the skin to capture vein patterns. The depth sensor measures the distance between the palm and the device in real time, and dynamically adjusts the light source intensity and camera focal length.

[0007] Preferably, the dynamic capture module supports the palm to move freely within the range of 5-15 cm from the device and a left and right movement range of ±10 cm, and automatically screens high-quality frames through a depth sensor and image processing algorithm, and eliminates blurred or blocked frames.

[0008] Preferably, the deep learning identity recognition module further includes the following contents: A feature extraction network adopts a deep convolutional neural network, which takes the multiple frames of palm vein images as input and outputs a vein feature map of each frame; A time series modeling network, which uses a recurrent neural network to perform time series modeling on the multi-frame feature maps to generate a comprehensive feature vector; The classification and matching module compares the comprehensive feature vector with the registered features in the database through a fully connected layer and a Softmax function, and outputs the identity recognition result.

[0009] Preferably, the training data set of the deep learning identity recognition module contains dynamic images of palm veins of 100,000 users, covering different lighting conditions, skin conditions and palm postures, and uses a combination of cross entropy loss and triplet loss to optimize the discrimination of feature vectors.

[0010] Preferably, the real-time processing and feedback system further includes the following: High-performance embedded processor that supports parallel computing and deep learning reasoning; High-speed memory for storing dynamically captured multi-frame images and deep learning model parameters; The display screen and voice module are configured to provide real-time feedback of recognition results.

[0011] Preferably, the real-time processing and feedback system optimizes the computational complexity of the deep learning model through model pruning and quantization technology, so that the processing flow includes dynamic capture 0.2 seconds, image preprocessing 0.1 seconds, feature extraction and matching 0.5 seconds, and the total delay is less than 1 second.

[0012] Preferably, each module of the identification terminal further includes the following steps: Sp1: The user places their palm 5-15 cm above the device, and the dynamic capture module collects 10 frames of palm vein images; Sp2: After preprocessing, the image is input into the deep learning identity recognition module to extract features and generate a comprehensive feature vector; Sp3: The real-time processing and feedback system compares the comprehensive feature vector with the database, outputs the recognition result, and provides feedback through the display screen and voice module.

[0013] Beneficial effects The present invention provides a palm vein dynamic capture and deep learning identity recognition terminal. It has the following beneficial effects: 1. The present invention uses a hybrid model of deep convolutional neural networks and recurrent neural networks, which collect multiple frames of palm vein images through a dynamic capture module and combines it with a deep learning identity recognition module. The present invention can extract rich spatial and temporal features, generate a highly discriminative comprehensive feature vector, and adaptively learn vein features in complex backgrounds, significantly improving recognition accuracy (error rate less than 0.01%) and robustness, and adapting to interference such as light changes and skin conditions.

[0014] 2. The present invention adopts a dynamic capture module to support the free movement of the palm within a certain range without the need to keep it still. Combining the real-time processing and feedback system with a high-performance embedded processor and optimization algorithm, the whole process time from dynamic capture to identity recognition is shortened to less than 1 second, significantly improving the acquisition success rate and authentication efficiency, providing a better user experience and suitable for fast identity authentication scenarios.

[0015] 3. The present invention uses multi-frame time series data collected by a dynamic capture module to increase the difficulty of forgery, combines an end-to-end deep learning framework to replace manual feature extraction, and improves the model's generalization ability for different users and environments through large-scale data set training, making the system more secure and difficult to be deceived by static images or simple forgery methods. At the same time, the algorithm design is simpler and more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a diagram of the identification terminal architecture of the present invention; Figure 2 A line graph showing the quality of multi-frame image acquisition of the dynamic capture module of the present invention; Figure 3 A line graph showing the dimensional changes of the feature graph extracted by the CNN feature of the present invention; Figure 4 A line graph showing hidden state changes for the RNN time series modeling of the present invention; Figure 5 A discrete graph of matching probabilities for feature matching according to the present invention; Figure 6 A line graph of training loss optimized by the loss function of the present invention; Figure 7 It is a line graph of comprehensive performance recognition accuracy and time of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Specific embodiment one: like Figures 1 to 7 As shown, the palm vein dynamic capture and deep learning identity recognition terminal includes a dynamic capture module, a deep learning identity recognition module and an implementation processing and feedback system, wherein: The dynamic capture module uses a multi-view near-infrared camera array and a depth sensor to dynamically capture multiple frames of palm vein images of the user's palm, allowing the palm to move freely within a certain range; The deep learning identity recognition module uses a hybrid model of deep convolutional neural networks and recurrent neural networks to extract spatial and temporal features from multiple frames of palm vein images, generate a comprehensive feature vector, and compare it with the registered features in the database to complete identity recognition; The real-time processing and feedback system uses an embedded processor to process multiple frames of palm vein images in real time, and completes the entire process from dynamic capture to identity recognition within 1 second. At the same time, it feeds back the recognition results to the user through the display screen and voice module.

[0019] The dynamic capture module includes a multi-view near-infrared camera, a near-infrared light source and a depth sensor. There are 3-5 multi-view near-infrared cameras, covering a viewing angle of 30°-60° above the palm, with an acquisition frequency of 30-60 frames per second. The wavelength range of the near-infrared light source is 850-950nm, which is used to penetrate the skin to capture vein patterns. The depth sensor measures the distance between the palm and the device in real time and dynamically adjusts the light source intensity and camera focal length. The dynamic capture module supports free movement of the palm within a range of 5-15cm from the device and a left and right movement range of ±10cm. It automatically screens high-quality frames and eliminates blurred or occluded frames through the depth sensor and image processing algorithm. The present invention uses a deep convolutional neural network and recurrent neural network hybrid model that collects multiple frames of palm vein images through the dynamic capture module and combines it with a deep learning identity recognition module. The present invention can extract rich spatial and temporal features, generate a highly discriminative comprehensive feature vector, and adaptively learn vein features under complex backgrounds, significantly improving recognition accuracy (error recognition rate less than 0.01%) and robustness, and adapting to interference such as light changes and skin conditions.

[0020] The deep learning identity recognition module further includes the following: The feature extraction network uses a deep convolutional neural network, which takes multiple frames of palm vein images as input and outputs a vein feature map for each frame; The time series modeling network uses a recurrent neural network to perform time series modeling on multi-frame feature maps and generate a comprehensive feature vector; The classification and matching module compares the comprehensive feature vector with the registered features in the database through the fully connected layer and the Softmax function, and outputs the identity recognition result.

[0021] The training dataset for the deep learning identity recognition module contains dynamic images of the palm veins of 100,000 users, covering different lighting conditions, skin conditions, and palm postures, and uses a combination of cross-entropy loss and triplet loss to optimize the discriminability of feature vectors.

[0022] The real-time processing and feedback system further includes the following: High-performance embedded processor that supports parallel computing and deep learning reasoning; High-speed memory for storing dynamically captured multi-frame images and deep learning model parameters; The display and voice module are configured to provide real-time feedback on recognition results. A dynamic capture module is used to support free movement of the palm within a certain range without the need to keep it still. The real-time processing and feedback system uses a high-performance embedded processor and optimized algorithm to shorten the entire process from dynamic capture to identity recognition to less than 1 second, significantly improving the collection success rate and authentication efficiency, providing a better user experience and suitable for fast identity authentication scenarios.

[0023] The real-time processing and feedback system optimizes the computational complexity of the deep learning model through model pruning and quantization technology, making the processing flow include dynamic capture in 0.2 seconds, image preprocessing in 0.1 seconds, and feature extraction and matching in 0.5 seconds, with a total delay of less than 1 second. The multi-frame time series data collected by the dynamic capture module increases the difficulty of forgery, and the end-to-end deep learning framework replaces manual feature extraction. The model's generalization ability for different users and environments is improved through large-scale data set training, making the system more secure and difficult to be deceived by static images or simple forgery methods. At the same time, the algorithm design is simpler and more efficient. Specific embodiment two: like Figures 1 to 7 As shown, based on the content in the above specific embodiments, the following contents are further disclosed: Identifying each module of the terminal further includes the following steps: Sp1: Design and assemble a dynamic capture module, using 3-5 multi-view near-infrared cameras (covering a 30°-60° viewing angle range, with a capture frequency of 30-60 frames per second), a near-infrared light source (wavelength 850-950nm), and a depth sensor (ToF, with a resolution of 0.1cm). The module uses an embedded processor to control the coordinated operation of the camera array and light source. This module allows for continuous capture of multiple frames of palm vein images when the palm moves freely within a range of 5-15cm from the device and within a range of ±10cm. The module also uses a depth sensor to measure the distance between the palm and the device in real time, dynamically adjusting the light source intensity and camera focal length to ensure image quality. Sp2: Implements the image screening function of the dynamic capture module. It pre-processes the collected multiple palm vein images using image processing algorithms, including denoising, contrast enhancement, and clarity assessment. It then automatically selects high-quality frames (such as those with clear vein patterns and no obstructions) based on depth sensor data and removes blurry or low-quality frames, ensuring high consistency and reliability of the input data for subsequent feature extraction. Sp3: Build a deep learning identity recognition module. Design a hybrid model based on a deep convolutional neural network (CNN, such as ResNet-50) and a recurrent neural network (RNN, such as LSTM). The CNN extracts spatial features from each frame of the palm vein image to generate a vein feature map. The RNN performs temporal modeling on the multi-frame feature maps to generate a comprehensive feature vector. The comprehensive feature vector is compared with the registered features in the database through a fully connected layer and a softmax function, and the identity recognition result is finally output. Sp4: Train a deep learning identity recognition module using a dataset of dynamic palm vein images from 100,000 users (covering different lighting conditions, skin conditions, and palm postures). Data augmentation techniques (such as rotation, scaling, and noise addition) are used to improve model generalization. A combination of cross-entropy loss and triplet loss is used to optimize the discriminability of feature vectors. After training, model pruning and quantization techniques are used to reduce computational complexity to meet the real-time inference requirements of embedded devices. Sp5: Integrated real-time processing and feedback system, using high-performance embedded processors (such as NVIDIA Jetson Nano) and high-speed memory (such as DDR4) to support parallel computing and deep learning reasoning. Through optimized algorithms, the processing flow is divided into dynamic capture (0.2 seconds), image preprocessing (0.1 seconds) and feature extraction and matching (0.5 seconds), ensuring a total delay of less than 1 second. At the same time, it is equipped with a display and voice module to provide real-time feedback on recognition results (such as "Authentication successful" or "Please try again"); Sp6: Implementing a complete workflow, the user places their palm 5-15cm above the device. The dynamic capture module collects and filters 10 frames of palm vein images. The pre-processed images are input into the deep learning identity recognition module to extract features and generate a comprehensive feature vector. The real-time processing and feedback system compares the feature vector with the database and outputs the recognition result. Finally, the authentication status is fed back to the user through the display and voice module. The entire process is completed within 1 second. Sp7: Verify the application effect of the technical solution, test the convenience and accuracy of users waving their palms to complete identity authentication in financial payment scenarios (such as ATMs or POS machines), verify the high real-time performance of fast passage in access control system scenarios, and confirm the security and reliability of patient identity recognition in medical identity authentication scenarios, ensuring that the system's false recognition rate in different scenarios is less than 0.01% and the collection success rate is higher than 99%, and optimize the device operation experience through user feedback. Specific embodiment three: like Figures 1 to 7 As shown, based on the content in the above specific embodiments, the following contents are further disclosed: The specific contents of the algorithm in the entire recognition terminal are as follows: CNN feature extraction further includes the following:

[0026] in: The feature map of the palm vein image of the tth frame represents the spatial features extracted from the input image, and its dimension is ,in 、 are the height and width of the feature map, is the number of feature channels; No. Frame input palm vein image, dimension is ,in 、 are the image height and width, is the number of input channels (usually 1 for grayscale images; Deep convolutional neural network functions, representing a combination of multiple layers of convolution, pooling, and activation operations; The weight parameter set of CNN, including the convolution kernel weights of each convolution layer; The set of bias parameters of CNN, corresponding to each convolutional layer Input preprocessed palm vein image (such as denoising, contrast enhancement), through multi-layer convolution operations (convolution kernel size such as ), extract local features, apply activation functions (such as ReLU, ) Increase nonlinearity and use pooling layers (such as maximum pooling) to reduce the dimension of the feature map and retain the main features. 5. Repeat convolution, activation and pooling operations until a fixed-dimensional feature map is output. CNN captures the local texture features of palm vein images (such as vein intersections and branches) through multi-layer convolution operations, gradually abstracts them into high-level features, generates spatial feature representations for each frame of the image, accurately extracts the spatial features of vein patterns, improves the uniqueness and accuracy of recognition, and solves the problem that traditional manual feature extraction (such as Gabor filtering) has poor adaptability to complex patterns.

[0027] RNN timing modeling further includes the following:

[0028]

[0029] in: The hidden state at time t represents the RNN's response to the current frame features. and the previous state The dimension of the time series modeling result is (hidden layer dimension); : CNN output Frame feature map, as input to RNN; : The hidden state of the previous moment, initially ; : Recurrent neural network functions (such as LSTM) are used to process time series data; : The weight parameter set of RNN, including the weights from input to hidden layer and from hidden layer to hidden layer; : RNN bias parameter set; : Comprehensive feature vector, representing the final temporal modeling result of multi-frame features, with a dimension of ; : The hidden state of the last frame, is the number of frames (such as 10); FC: fully connected layer function, mapping the hidden state to a feature vector; : Weight parameters of the fully connected layer.

[0030] : Bias parameter of the fully connected layer.

[0031] The feature map output by CNN ( ) input into RNN in sequence, RNN according to the current input and the previous state Update hidden state , calculated using LSTM units as follows: Forget Gate: Input Gate: Unit status update: Output gate: Hidden state: Get the hidden state of the last frame , mapped into a comprehensive feature vector through the fully connected layer , RNN captures the dynamic changes between multiple frames of images (such as the changes in the texture of the palm when it moves slightly) through time series modeling and integrates them into a comprehensive feature vector , enhance the distinguishability of features, use multi-frame time series information to improve the stability and robustness of recognition, solve the problem of misrecognition caused by posture deviation or image quality degradation in traditional single-frame recognition, and increase the difficulty of forgery.

[0032] Feature matching and classification further include the following:

[0033]

[0034] in: : classification score, indicating the matching probability between the input feature vector and the registered feature, with dimension (number of registered users; : Input feature vector With the registered feature vector Euclidean distance Input the comprehensive feature vector of the palm vein image; The feature vector of the rth registered user in the database; : L2 norm, representing the Euclidean distance between vectors; Function, converting distance into probability distribution, ; :: The weight parameters of the layer; : Bias parameters of the layer.

[0035] Calculate input feature vector With all registered feature vectors in the database Euclidean distance , enter the distance layer, converted to matching probability , select the registered user with the highest probability as the recognition result. If the highest probability is lower than the threshold (such as 0.9), it is judged as "unmatched". Classification, determine the matching degree between input features and registered features, realize identity recognition, provide high-precision identity matching, low false recognition rate, solve the problem that traditional matching methods (such as cosine similarity) are sensitive to noise, and support fast comparison of large-scale user databases.

[0036] The loss function further includes the following:

[0037]

[0038]

[0039] in: Total loss function, combining cross entropy loss and triplet loss; : Cross entropy loss, used to optimize classification accuracy; : Triplet loss, used to optimize the discrimination of feature vectors; : Weight coefficient, balancing the two losses (e.g. 0.5); Number of training samples; : No. The true label of each sample (one-hot encoding); : No. The predicted probability of a sample. Anchor feature vector (current sample; Positive sample feature vector (same user); : Negative sample feature vector (different users); Boundary parameters (such as 0.2) ensure the spacing between positive and negative samples.

[0040] During the training phase, the cross entropy loss is calculated , optimize classification accuracy, construct triples (anchor positive samples, negative samples), and calculate triple loss , optimize the discrimination of the feature vector, combine the two losses, and update the model parameters through back propagation The cross entropy loss ensures correct classification, the triplet loss widens the distance between different user features, narrows the distance between the features of the same user, improves the generalization ability of the model, improves the discrimination and generalization ability of the feature vector, solves the problem that traditional training methods (such as single cross entropy loss) are not adaptable to complex backgrounds, and ensures high-precision recognition.

[0041] The entire operation steps are as follows: Data acquisition: Dynamic capture module collects multiple frames of palm vein images (e.g. 10 frames); Feature extraction: CNN for each frame of image Extract feature maps ; Temporal Modeling: RNN for Multi-frame Feature Maps Perform time series modeling and generate comprehensive feature vectors ; Feature Matching: Calculation With registration features The distance is output through Softmax to match the probability ; Training optimization: using loss function Optimize model parameters to improve accuracy and discrimination; Real-time inference: Deploy models on embedded devices to achieve real-time authentication in less than 1 second.

[0042] The algorithm starts with multiple frames of dynamically captured images, extracts the spatial features of each frame through CNN, and generates a comprehensive feature vector by integrating the temporal information through RNN. Classification enables identity recognition. During the training phase, a composite loss function is used to optimize the model to ensure high discriminability and robustness of features. Specific embodiment four: like Figures 1 to 7 As shown, based on the content in the above specific embodiments, the following contents are further disclosed: The following are specific application cases of the entire identification terminal in three scenarios: financial payment, access control system, and medical identity verification. They aim to verify the core technical content of the technical solution. The specific content is as follows: Application Case 1: Financial Payment Scenario - ATM Identity Verification: Scenario Background: At bank ATMs, users need to quickly and securely verify their identities to complete operations such as withdrawals and transfers. Traditional fingerprint or password authentication methods pose forgery risks and are inconvenient to use. This terminal, however, provides greater security and convenience through contactless palm vein recognition. The implementation steps are as follows: Equipment Installation: A palm vein dynamic capture and deep learning identity recognition terminal is installed above the ATM. It is equipped with a multi-view near-infrared camera array (3 cameras, covering a 60° viewing angle), a near-infrared light source (850nm), and a depth sensor to ensure clear images are captured when the palm moves freely within a range of 5-15cm. User operation: The user places their palm on the device without touching the device or holding it in a fixed position. Simply wave or hover naturally for about 1 second, and the system will automatically capture 10 frames of palm vein images. Identity recognition: After the dynamic capture module selects high-quality frames, the images are input into the deep learning identity recognition module. The CNN extracts vein features in each frame, and the RNN generates a comprehensive feature vector, which is compared with the registered features in the bank database. The real-time processing system completes the authentication and feedback within 1 second. Result feedback: The ATM display and voice module will prompt "Authentication successful, please continue" or "Authentication failed, please try again", and the user can complete subsequent financial operations according to the prompts; The entire recognition terminal supports free palm movement and captures multiple frames of images, improving capture success rate and image quality. It adapts to different user palm sizes and postures. Using CNN and RNN to extract spatial and temporal features, it generates highly discriminative feature vectors, ensuring recognition accuracy (false recognition rate <0.01%). The processing time is less than 1 second, meeting the high real-time requirements of financial payment scenarios. At the same time, multi-frame dynamic data increases the difficulty of forgery and improves security. Effectiveness verification: 100 users were tested, with a 99.5% collection success rate. No training was required for proficient operation. Attempts to forge authentication using printed palm vein images were rejected by the system due to the lack of dynamic timing characteristics. The system is significantly more secure than traditional methods, with an average authentication time of 0.8 seconds and a better user experience than traditional fingerprint authentication (2-3 seconds on average). Banks reported that the system reduced the risk of forgery and increased user trust, and plan to expand its use to more ATMs.

[0044] Application Case 2: Access Control System Scenario - Enterprise Fast Pass: Scenario Background: In corporate office buildings or residential communities, access control systems must quickly verify the identities of employees or residents to ensure efficient access. Traditional RFID cards or passwords are prone to loss or theft, while this terminal provides a contactless, efficient solution through palm vein recognition. The implementation steps are as follows: Equipment installation: A palm vein dynamic capture and deep learning identity recognition terminal is installed at the access control entrance. The camera array covers a 60° viewing angle, and the depth sensor is adjusted to capture a range of 5-15cm, ensuring that users can complete authentication while walking. User operation: When an employee or resident approaches the access control system, they naturally extend their palm above the device without stopping or touching it. The system dynamically collects 10 frames of palm vein images as the user passes by; Identity recognition: After the dynamic capture module selects clear frames, the deep learning identity recognition module extracts features and generates a comprehensive feature vector, which is compared with the enterprise or community database. The real-time processing system completes authentication and triggers the access control switch within 1 second; Result feedback: The access control system will give a voice prompt of "Welcome" or "Authentication failed, please try again" and simultaneously record the access log; The entire recognition terminal supports dynamic acquisition, eliminating the need for users to pause or maintain a fixed posture. Multi-frame images improve acquisition success rates, adapting to fast-moving scenarios and learning vein features in complex backgrounds to ensure high-precision recognition. It is highly robust and adaptable to varying lighting and skin conditions, with authentication times of less than one second, meeting the needs of fast-moving traffic during peak hours. Real-time feedback enhances the user experience. Effectiveness verification: A test of 500 employees during peak hours showed an average authentication time of 0.7 seconds, a 50% increase in access efficiency (compared to 1.5 seconds for traditional RFID cards). Tests conducted indoors and outdoors under varying lighting conditions showed a false recognition rate below 0.01% and a 99% acquisition success rate. An attempt to deceive the system using forged images failed due to a lack of dynamic features. The system is more secure than traditional methods. Enterprises report that the system has reduced card loss, improved access efficiency and security, and has been expanded to multiple office building entrances.

[0045] Application Case 3: Medical Identity Verification Scenario - Hospital Patient Identity Confirmation: Scenario Background: Accurately verifying a patient's identity is crucial for hospital registration, medication collection, and surgery to avoid medical errors. Traditional ID or medical insurance card verification methods are prone to loss or fraudulent use, while this terminal provides high security and reliability through palm vein recognition. The implementation steps are as follows: Equipment Installation: Install a palm vein dynamic capture and deep learning identity recognition terminal at the hospital registration window, pharmacy, or operating room entrance. Equipped with a multi-view camera array and depth sensor, it ensures a capture range of 5-15 cm. User operation: The patient places his palm above the device and hovers or moves it slightly for about 1 second. The system dynamically collects 10 frames of palm vein images. Identity recognition: After the dynamic capture module selects high-quality frames, the deep learning identity recognition module extracts features and generates a comprehensive feature vector, which is compared with the hospital's patient database. The real-time processing system completes authentication and outputs the results within 1 second; Result feedback: The display screen and voice prompts "Patient identity confirmation successful, please continue" or "Authentication failed, please contact staff", and the authentication results are synchronized to the hospital information system; The entire identification terminal supports free palm movement, captures multiple frames of images, improves the acquisition success rate, adapts to possible tremors or unstable postures of patients, extracts high-precision features, and adapts to the changes in vein characteristics of patients of different ages and skin types to ensure recognition accuracy. Rapid authentication (less than 1 second) reduces patient waiting time, and multi-frame dynamic data enhances security and prevents identity fraud. Effectiveness verification: A test of 1,000 patients showed an identification accuracy rate of 99.9% and a misidentification rate of less than 0.01%, effectively preventing identity errors. Elderly patients or those with hand tremors could also easily complete authentication, with a collection success rate of 98.5%. Attempts to impersonate using forged images or another person's palm were rejected by the system due to differences in dynamic features and eigenvectors, demonstrating its high security. Hospitals reported that the system significantly reduced the risk of identity theft, improved medical process efficiency and patient trust, and has been rolled out to multiple departments. The above three application cases fully verify that the dynamic capture module achieves non-contact collection with a high success rate in various scenarios and adapts to different user postures and environments; the deep learning identity recognition module ensures the accuracy and reliability of recognition through high-precision feature extraction and robust design; the real-time processing and feedback system achieves rapid authentication in less than 1 second, meeting the real-time requirements of different scenarios while improving security.

[0046] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0047] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Palm vein dynamic capture and deep learning identity recognition terminal, characterized by: The identification terminal includes a dynamic capture module, a deep learning identity recognition module and an implementation processing and feedback system, wherein: The dynamic capture module dynamically collects multiple frames of palm vein images of the user's palm through a multi-view near-infrared camera array and a depth sensor, allowing the palm to move freely within a certain range; The deep learning identity recognition module uses a hybrid model of deep convolutional neural networks and recurrent neural networks to extract spatial and temporal features from the multiple frames of palm vein images, generate a comprehensive feature vector, and compare it with the registered features in the database to complete identity recognition; The real-time processing and feedback system processes the multiple frames of palm vein images in real time through an embedded processor, and completes the entire process from dynamic capture to identity recognition within 1 second, while feeding back the recognition results to the user through a display screen and a voice module.

2. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 1, characterized in that: The dynamic capture module includes a multi-view near-infrared camera, a near-infrared light source and a depth sensor. The multi-view near-infrared cameras are arranged in 3-5 positions, covering a viewing angle range of 30°-60° above the palm, with an acquisition frequency of 30-60 frames per second. The wavelength range of the near-infrared light source is 850-950nm, which is used to penetrate the skin to capture vein patterns. The depth sensor measures the distance between the palm and the device in real time and dynamically adjusts the light source intensity and camera focal length.

3. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 2, characterized in that: The dynamic capture module supports the palm to move freely within the range of 5-15cm from the device and ±10cm left and right, and automatically selects high-quality frames through depth sensors and image processing algorithms, eliminating blurred or blocked frames.

4. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 1, characterized in that: The deep learning identity recognition module further includes the following: A feature extraction network adopts a deep convolutional neural network, which takes the multiple frames of palm vein images as input and outputs a vein feature map of each frame; A time series modeling network, which uses a recurrent neural network to perform time series modeling on the multi-frame feature maps to generate a comprehensive feature vector; The classification and matching module compares the comprehensive feature vector with the registered features in the database through a fully connected layer and a Softmax function, and outputs the identity recognition result.

5. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 1, characterized in that: The training dataset of the deep learning identity recognition module contains dynamic images of palm veins of 100,000 users, covering different lighting conditions, skin conditions and palm postures, and uses a combination of cross entropy loss and triplet loss to optimize the discriminability of feature vectors.

6. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 1, characterized in that: The real-time processing and feedback system further includes the following: High-performance embedded processor that supports parallel computing and deep learning reasoning; High-speed memory for storing dynamically captured multi-frame images and deep learning model parameters; The display screen and voice module are configured to provide real-time feedback of recognition results.

7. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 6, characterized in that: The real-time processing and feedback system optimizes the computational complexity of the deep learning model through model pruning and quantization technology, so that the processing flow includes dynamic capture in 0.2 seconds, image preprocessing in 0.1 seconds, and feature extraction and matching in 0.5 seconds, with a total delay of less than 1 second.

8. The palm vein dynamic capture and deep learning identity recognition terminal according to claim 1, characterized in that: Each module of the identification terminal further includes the following steps: Sp1: The user places their palm 5-15 cm above the device, and the dynamic capture module collects 10 frames of palm vein images; Sp2: After preprocessing, the image is input into the deep learning identity recognition module to extract features and generate a comprehensive feature vector; Sp3: The real-time processing and feedback system compares the comprehensive feature vector with the database, outputs the recognition result, and provides feedback through the display screen and voice module.

Images