Palm vein feature extraction and depth adaptive matching method

Abstract

The invention provides a palm vein feature extraction and depth adaptive matching method, which relates to the technical field of biological recognition, and is characterized in that a three-dimensional space-time convolutional neural network and a graph neural network are combined to extract space depth features, time blood flow dynamic features and blood vessel topological features of palm veins, and information bottleneck is utilized to optimize and screen a feature subset with the most distinguishing property; the method breaks through the limitation of traditional two-dimensional texture feature extraction, obtains abundant palm vein information through multi-source data acquisition (such as near-infrared imaging, a light field camera and a spectrum sensor), overcomes the interference of environmental factors such as illumination, angle and noise, achieves the high-precision modeling and stable extraction effects of palm vein features, and can achieve the real-time extraction of the palm vein features in a complex scene. A high recognition rate is kept under different illumination conditions or user hand posture changes, the robustness of the palm vein recognition system is remarkably improved, and the method is suitable for scenes with high safety requirements.

Description

Technical Field

[0001] The present invention relates to the field of biometric technology, and specifically to a palm vein feature extraction and depth adaptive matching method. Background Art

[0002] According to a fast palm vein feature extraction method and system disclosed in Chinese Patent Publication No. "CN114241535B", palm vein information is collected by a sensor, and the collected palm vein information is processed to obtain palm vein features and uploaded to a database. Then, the similarity between each palm vein feature in the database is calculated as the homology between each palm vein feature. Furthermore, the palm vein data features are stored in slices according to the homology between each palm vein feature, thus achieving the beneficial result of extracting palm vein data features with low computational cost and quickly identifying palm vein information.

[0003] The above patent document and the prior art have the following technical problems when in use: Problem 1: Existing palm vein recognition technologies mainly rely on two-dimensional texture feature extraction methods, such as Gabor filters, LBP, etc. These methods are difficult to capture the three-dimensional spatial information and dynamic characteristics of palm veins, resulting in a decrease in feature extraction accuracy, unstable recognition rate, and insufficient recognition accuracy in complex environments such as light changes, angle offsets, or noise interference; Problem 2: The matching algorithms of existing palm vein recognition systems are mostly based on traditional distance metrics, such as Euclidean distance, Hamming distance, or shallow neural networks, lacking the ability to defend against feature tampering and forgery attacks. In particular, the detection ability for synthetic palm vein images, such as printed images or dynamic forgeries, such as 3D printed models, is insufficient. Traditional methods rely on static features, such as texture contrast, in live detection and cannot effectively distinguish between live and non-live, resulting in low system security and difficulty in meeting the requirements of high-risk scenarios; Problem 3: Existing palm vein recognition technologies usually only rely on a single modality and lack the ability to cooperate with other biometric features such as fingerprints and irises, resulting in limited applications in multi-modal authentication scenarios. At the same time, traditional methods lack mechanisms for dynamic update and cross-scene adaptation. When facing different devices, different populations, or long-term use, the model performance degrades, making it difficult to meet the requirements of distributed deployment and personalized authentication. Summary of the Invention

[0004] Technical Problems to be Solved In view of the deficiencies of the prior art, the present invention provides a palm vein feature extraction and depth adaptive matching method, which solves the following problems: 1. Aiming at the problem of insufficient palm vein recognition accuracy and robustness; 2. Aiming at the problem of weak system security and anti-attack ability; 3. Address the problem of single modality recognition and poor cross-scenario adaptability.

[0005] Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solutions: a palm vein feature extraction and depth adaptive matching method, the method comprising the following steps: Sp1: Multi-source data acquisition and preprocessing: Use near-infrared imaging equipment, light field cameras and near-infrared spectral sensors to collect multi-source data of palm veins, including two-dimensional texture images, three-dimensional depth information and blood flow dynamic data, and preprocess the collected data; Sp2: 3D dynamic feature extraction: The spatial depth features and temporal blood flow dynamic features of the palm veins are extracted through a 3D spatiotemporal convolutional neural network, and the topological features of the palm veins are extracted in combination with a graph neural network; Sp3: Feature optimization and compression: The extracted features are optimized through the information bottleneck and mapped to the quantum Hilbert space to generate quantum characteristic states; Sp4: Quantum feature encoding and matching: Use quantum-inspired matching algorithms to match quantum feature states, and the quantum-inspired matching algorithms include quantum inner product calculations and adaptive quantum gate operations, and are combined with variational quantum algorithms to optimize matching parameters; Sp5: Liveness detection and anti-attack: Liveness detection is performed by analyzing multi-frame blood flow dynamic data, and the quantum state non-cloning property is used to defend against counterfeit attacks; Sp6: Multimodal adaptive fusion: Palm vein features are collaboratively matched with other biometric features through a multimodal adaptive fusion network, which includes a cross-modal attention mechanism and dynamic weight assignment.

[0006] Preferably, in the multi-source data acquisition and preprocessing, a near-infrared imaging device is used to collect two-dimensional texture images of palm veins with a wavelength range of 850-950 nanometers, a light field camera is used to collect three-dimensional depth information of palm veins with a resolution of not less than 0.1 mm, and a near-infrared spectral sensor is used to collect palm vein blood flow dynamic data, including blood oxygen saturation and blood flow velocity, with a sampling frequency of not less than 10 Hz.

[0007] Preferably, the preprocessing in the multi-source data acquisition and preprocessing includes adaptive illumination equalization of two-dimensional texture images based on a multi-scale Retinex algorithm, denoising of three-dimensional depth information based on a bilateral filtering algorithm, and signal smoothing of blood flow dynamics data based on a Kalman filtering algorithm.

[0008] Preferably, the three-dimensional spatiotemporal convolutional neural network in the three-dimensional dynamic feature extraction includes the following contents: The spatial convolution layer uses a 3x3x3 convolution kernel to extract the spatial depth features of the palm veins; The temporal convolutional layer uses a 1x1xT convolutional kernel (where T is the length of the time series) to extract hemodynamic features; The residual connection module is used to alleviate the problem of vanishing gradients in deep networks.

[0009] Preferably, when the graph neural network in the three-dimensional dynamic feature extraction runs, the palmar vein vascular structure is modeled as a graph structure, where nodes represent vascular branch points or intersection points, and edges represent vascular connection relationships. The node features are updated through graph convolution operations, and the number of iterations is 2 - 5 times. Finally, a global topological feature vector is output, and the dimension range is 128 - 512.

[0010] Preferably, the information bottleneck optimization in the feature optimization and compression includes calculating the mutual information of the feature vectors, retaining the feature subset highly relevant to identity authentication, removing the feature components related to environmental noise or redundant information, and outputting the optimized feature vectors. The compression ratio range is 20% - 50%.

[0011] Preferably, the quantum-inspired matching algorithm in the liveness detection and anti-attack includes calculating the similarity between two sets of quantum feature states using quantum inner product, adjusting the feature states through adaptive quantum gates, and optimizing the quantum circuit parameters using variational quantum algorithms. The number of iterations ranges from 50 to 200 times.

[0012] Preferably, the liveness detection in the multi-modal adaptive fusion includes analyzing continuous multi-frame hemodynamic data, calculating the blood flow velocity change rate and direction consistency, setting the liveness detection threshold, with a range of 0.8 - 0.95. If it is lower than the threshold, it is determined as non-living, and the liveness detection result is output for subsequent matching decisions.

[0013] Preferably, the multi-modal adaptive fusion network in the multi-modal adaptive fusion includes the following: The cross-modal attention mechanism is used to align the spatial and semantic distributions of palmar vein features with other biometric features (such as fingerprints and irises); The dynamic weight allocation module adjusts the matching weights of each modal feature according to environmental conditions (such as light intensity and humidity) and device performance, and outputs the fused matching score, with a range of 0 - 1.

[0014] Beneficial effects The present invention provides a method for palmar vein feature extraction and deep adaptive matching. It has the following beneficial effects: 1. The present invention combines a three-dimensional spatiotemporal convolutional neural network with a graph neural network to extract the spatial depth features, temporal blood flow dynamic features, and vascular topological features of the palm veins, and uses information bottlenecks to optimize and screen the most discriminative feature subsets, breaking through the limitations of traditional two-dimensional texture feature extraction. At the same time, through multi-source data acquisition (such as near-infrared imaging, light field cameras, and spectral sensors), rich palm vein information is obtained, overcoming the interference of environmental factors such as illumination, angle, and noise, achieving high-precision modeling and stable extraction of palm vein features, and being able to maintain a high recognition rate in complex scenarios, such as different lighting conditions or changes in user hand postures, significantly improving the robustness of the palm vein recognition system, and being suitable for scenarios with high security requirements.

[0015] 2. The present invention introduces a quantum-inspired matching algorithm and hemodynamic liveness detection, uses quantum inner product and adaptive quantum gate operations to achieve feature matching, and uses the non-cloning property of quantum states to defend against feature tampering attacks. At the same time, it verifies liveness features by analyzing multi-frame blood flow dynamic data (such as speed changes and direction consistency), effectively resists static forgery (such as printed images) and dynamic forgery (such as 3D printed models) attacks, and achieves high security protection for the palm vein recognition system. It not only improves the system's defense capabilities against synthetic data and forgery attacks, but also enhances the irreversibility of feature matching through the unique characteristics of quantum computing, making the system more reliable in high-risk environments.

[0016] 3. The present invention adopts a multimodal adaptive fusion network combined with a cross-modal attention mechanism and dynamic weight allocation to efficiently and collaboratively match palm vein features with other biometric features (such as fingerprints and irises), and realizes incremental updates of the model and rapid adaptation to new scenarios (such as different devices and different groups of people) through federated learning and meta-learning mechanisms, thereby breaking through the limitations of traditional palm vein recognition of a single modality and fixed scenarios, achieving high efficiency and flexibility in multimodal authentication, being able to dynamically optimize matching strategies according to environmental conditions and device performance, supporting distributed deployment and privacy protection, and being widely applicable to access control systems, financial payment terminals, and identity authentication scenarios across device migration, thereby improving the practicality of the system and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a diagram of the method steps of the present invention; Figure 2 The three-dimensional palm vein characteristic distribution map of the present invention; Figure 3 It is a line chart showing the comparison of recognition accuracy of the present invention; Figure 4 It is the detection rate comparison bar graph of Table 2 in the specific embodiment 3 of the present invention; Figure 5 It is a line graph of dynamic changes in multimodal fusion weights of the present invention; Figure 6 It is the palm vein feature correlation heat map of the present invention; Figure 7 It is a comparison diagram of the ROC curves in Table 5 in the specific embodiment 3 of the present invention; Figure 8 It is a blood flow dynamic time series diagram of the present invention; Figure 9 A bar chart showing the importance of palm vein recognition in the three-dimensional dynamic feature of the present invention; Figure 10 This is a multi-scenario performance comparison chart in Table 7 in the specific embodiment 3 of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention. Specific embodiment one: like Figures 1 to 10 As shown, the palm vein feature extraction and depth adaptive matching method comprises the following steps: Sp1: Multi-source data acquisition and preprocessing: Use near-infrared imaging equipment, light field cameras and near-infrared spectral sensors to collect multi-source data of palm veins, including two-dimensional texture images, three-dimensional depth information and blood flow dynamic data, and preprocess the collected data; Sp2: 3D dynamic feature extraction: The spatial depth features and temporal blood flow dynamic features of the palm veins are extracted through a 3D spatiotemporal convolutional neural network, and the topological features of the palm veins are extracted in combination with a graph neural network; Sp3: Feature optimization and compression: The extracted features are optimized through the information bottleneck and mapped to the quantum Hilbert space to generate quantum characteristic states; Sp4: Quantum feature encoding and matching: Use quantum-inspired matching algorithms to match quantum feature states, and the quantum-inspired matching algorithms include quantum inner product calculations and adaptive quantum gate operations, and are combined with variational quantum algorithms to optimize matching parameters; Sp5: Liveness detection and anti-attack: Liveness detection is performed by analyzing multi-frame blood flow dynamic data, and the quantum state non-cloning property is used to defend against counterfeit attacks; Sp6: Multimodal adaptive fusion: Palm vein features are collaboratively matched with other biometric features through a multimodal adaptive fusion network, which includes a cross-modal attention mechanism and dynamic weight assignment.

[0020] The above method further comprises the following contents: In multi-source data acquisition and preprocessing, a near-infrared imaging device is used to acquire two-dimensional texture images of palm veins with a wavelength range of 850 - 950 nanometers. A light field camera is used to acquire three-dimensional depth information of palm veins with a resolution of not less than 0.1 mm. A near-infrared spectral sensor is used to acquire dynamic palm vein blood flow data, including blood oxygen saturation and blood flow velocity, with a sampling frequency of not less than 10 Hz. The preprocessing in multi-source data acquisition and preprocessing includes adaptive illumination equalization of two-dimensional texture images based on the multi-scale Retinex algorithm, denoising of three-dimensional depth information based on the bilateral filtering algorithm, and signal smoothing of blood flow dynamic data based on the Kalman filtering algorithm.

[0021] The three-dimensional spatio-temporal convolutional neural network in three-dimensional dynamic feature extraction includes the following: A spatial convolutional layer that uses a 3x3x3 convolutional kernel to extract the spatial depth features of palm veins; A temporal convolutional layer that uses a 1x1xT convolutional kernel (T is the length of the time series) to extract blood flow dynamic features; A residual connection module used to alleviate the problem of gradient disappearance in deep networks.

[0022] When the graph neural network in three-dimensional dynamic feature extraction runs, it models the palm vein vascular structure as a graph structure, where nodes represent vascular branch points or intersection points, and edges represent vascular connection relationships. The node features are updated through graph convolution operations, and the number of iterations is 2 - 5 times. Finally, a global topological feature vector is output, with a dimension range of 128 - 512. The spatial depth features, temporal blood flow dynamic features, and vascular topological features of palm veins are extracted by combining the three-dimensional spatio-temporal convolutional neural network and the graph neural network, and the information bottleneck is used to optimize and screen the most discriminative feature subset, breaking through the limitations of traditional two-dimensional texture feature extraction. At the same time, rich palm vein information is obtained through multi-source data acquisition (such as near-infrared imaging, light field cameras, and spectral sensors), overcoming the interference of environmental factors such as illumination, angle, and noise, achieving high-precision modeling and stable extraction effects of palm vein features, being able to maintain a high recognition rate in complex scenarios, such as different illumination conditions or changes in the user's hand posture, significantly improving the robustness of the palm vein recognition system, and being applicable to scenarios with high security requirements.

[0023] Information bottleneck optimization in feature optimization and compression includes calculating the mutual information of feature vectors, retaining the feature subset highly relevant to identity authentication, removing the feature components related to environmental noise or redundant information, and outputting the optimized feature vectors. The compression ratio ranges from 20% to 50%. The quantum-inspired matching algorithm in liveness detection and anti-attack includes calculating the similarity between two sets of quantum feature states using quantum inner product, adjusting the feature states through adaptive quantum gates, optimizing the parameters of the quantum circuit using variational quantum algorithms, and the number of iterations ranges from 50 to 200 times. By introducing the quantum-inspired matching algorithm and hemodynamic liveness detection, feature matching is achieved using quantum inner product and adaptive quantum gate operations, and feature tampering attacks are defended by the non-clonability of quantum states. At the same time, the liveness features are verified by analyzing multi-frame blood flow dynamic data (such as velocity changes and direction consistency), effectively resisting static forgeries (such as printed images) and dynamic forgeries (such as 3D printed models) attacks, and achieving high-security protection for the palm vein recognition system. It not only improves the system's defense ability against synthetic data and forgery attacks, but also enhances the irreversibility of feature matching through the unique characteristics of quantum computing, making the system more reliable in high-risk environments.

[0024] Liveness detection in multimodal adaptive fusion includes analyzing continuous multi-frame blood flow dynamic data, calculating the blood flow velocity change rate and direction consistency, setting the liveness detection threshold, which ranges from 0.8 to 0.95. If it is lower than the threshold, it is determined as non-living, and the liveness detection result is output for subsequent matching decisions. The multimodal adaptive fusion network in multimodal adaptive fusion includes the following: Cross-modal attention mechanism for aligning the spatial and semantic distributions of palm vein features with other biometric features (such as fingerprints, irises); Dynamic weight allocation module, which adjusts the matching weights of each modal feature according to environmental conditions (such as light intensity, humidity) and device performance, and outputs the fused matching score, which ranges from 0 to 1. By combining the cross-modal attention mechanism and dynamic weight allocation through the multimodal adaptive fusion network, the palm vein features are efficiently and collaboratively matched with other biometric features (such as fingerprints, irises), and the incremental update of the model and rapid adaptation to new scenarios (such as different devices, different populations) are realized through federated learning and meta-learning mechanisms, thus breaking through the limitations of traditional single-modal and fixed-scene palm vein recognition, achieving the efficiency and flexibility of multimodal authentication, being able to dynamically optimize the matching strategy according to environmental conditions and device performance, supporting distributed deployment and privacy protection, and being widely applicable to access control systems, financial payment terminals, and identity authentication scenarios for cross-device migration, improving the practicality and user experience of the system. Specific Embodiment 2: As Figures 1 to 10 shown, based on the content in the above specific embodiments, the following content is further disclosed: The specific formula content of each algorithm in the above steps is as follows: The three-dimensional spatio-temporal convolutional neural network (3DST-CNN) further includes the following:

[0026] Where: : The value of the output feature map at the spatial position and the temporal position ; : The input data (the value of the palm vein three-dimensional image sequence at the spatial position and the temporal position ); : The weight of the three-dimensional convolutional kernel, with a size of : The bias term; : The spatial height, width, and temporal depth of the convolutional kernel; : The spatial coordinates, representing the two-dimensional position of the feature map; The temporal coordinate, representing the temporal sequence position of the feature map; : The spatial and temporal offsets of the convolutional kernel; The input three-dimensional palm vein data, usually a tensor of ; : The three-dimensional convolutional kernel, capturing spatial and temporal features; The input palm vein three-dimensional image sequence , with a size of , the three-dimensional convolutional kernel (such as ), optimizes the weights through training, performs a convolution operation on the input data, slides the window through the spatial and temporal dimensions, calculates the feature values at each position, adds the bias and outputs the feature map through an activation function (such as ReLU) . The three-dimensional convolution captures the depth texture features (spatial dimension) and blood flow dynamic changes (temporal dimension) of the palm vein by operating simultaneously in the spatial and temporal dimensions. The residual connection alleviates the vanishing gradient problem of the deep network, ensures the stability of feature extraction, extracts spatio-temporal features through three-dimensional convolution, overcomes the neglect of depth information and dynamic features by traditional two-dimensional methods, and improves the accuracy and robustness of palm vein recognition, especially performing better in scenarios with illumination changes or pose offsets.

[0027] The graph neural network (GNN) further includes the following:

[0028] Where: : The feature vector of the nodes in the -th layer; ; : The updated feature vector of the nodes in the -th layer; ; : The set of neighbor nodes of node ; AGGREGATE: Aggregation function (such as mean, sum); : The weight matrix of the -th layer; : Self-loop weight matrix; : Activation function (such as ReLU); : A node in the graph, representing a palmar vein bifurcation point or intersection point; : The feature vector of node , representing its local features; The number of layers of the graph neural network; : The set of neighbor nodes connected to node ; : Weight matrix, updating the node features; : Self-loop weight, retaining the information of the node itself; Model the palmar vein vascular structure as a graph , node represents the bifurcation point, edge represents the vascular connection, initialize the features of each node (such as position coordinates or texture features), through graph convolution operations, aggregate the neighbor node features and update the current node features, iterate 2 - 5 times, the GNN updates the node features by aggregating neighbor information, captures the topological structure of the palmar vein vessels (such as bifurcation patterns, connection relationships), and enhances the global feature expression through multi-layer iteration. Extracting the vascular topological features makes up for the neglect of the global structure by traditional methods, improves the distinctiveness of the features, and especially enhances the recognition ability in scenarios with large individual differences.

[0029] The specific content of the information bottleneck optimization is as follows:

[0030]

[0031] Objective: Minimize , Maximize , and the optimization function is:

[0032] Where: : The mutual information between the input and the compressed representation ; : The mutual information between the compressed representation and the objective ; : The joint probability distribution; : The balance coefficient; Input features (such as palm vein feature vectors; The compressed feature representation; The objective label (such as identity category); Mutual information, which measures the correlation between variables; The probability distribution function; Hyperparameter, which adjusts the balance between compression and prediction capabilities; The input original feature vector and the identity label , generate the compressed representation through the neural network, calculate and , optimize the objective function , adjust the network parameters, minimize the redundant information , maximize the prediction information The information bottleneck compresses the feature vector, retains the information highly relevant to identity authentication, eliminates environmental noise and redundant data, optimizes the feature expression, reduces feature redundancy, enhances the discriminability and computational efficiency, enabling the system to maintain high-precision recognition in a high-noise environment.

[0033] The quantum-inspired matching algorithm (QIAM) includes the following:

[0034]

[0035] Wherein: Similarity of quantum eigenstates; Inner product of two quantum states; Adjusted quantum eigenstates; : Adaptive quantum gates (such as rotation gates ; Initial quantum eigenstates; Rotation angle parameter.

[0036] : Two quantum eigenstates to be matched; Similarity score, ranging from 0 to 1; Quantum gate operation to adjust eigenstates; Dynamic parameter determined by feature distribution; Initial state, usually the quantum encoding of classical features; Encoding the classical feature vector into a quantum state , through adaptive quantum gates Adjusting the eigenstates to generate , calculating the inner product of two quantum states , and taking the square to obtain the similarity , optimizing using variational quantum algorithms , iterating 50 - 200 times. The quantum inner product calculates the similarity using the superposition and entanglement of quantum states. The adaptive quantum gates dynamically adjust the eigenstates, and the variational algorithm optimizes the parameters to improve the matching accuracy. Quantum matching improves the matching efficiency and security, and enhances the anti - attack ability using the non - cloneability of quantum states, suitable for high - security scenarios.

[0037] The cross - modal attention mechanism further includes the following:

[0038] Wherein: Query vector (palm vein feature); Key vector (features of other modalities, such as fingerprint; Value vector (features of other modalities); : Dimension of the key vector; Input feature vector matrix; : Normalization function; : Scaling factor to prevent excessive numerical values.

[0039] Input palm vein features and other modality features , calculate and 's dot product, after scaling, normalize through , multiply by , output weighted fusion features, dynamically adjust weights, fuse multi-modal features, the attention mechanism calculates the correlation between features, aligns the semantic distributions of palm veins and other modalities, dynamically assigns weights, improves the fusion effect, realizes multi-modal collaborative matching, improves the system adaptability and accuracy, and is applicable to multi-scenario authentication requirements. Specific Embodiment Three: As Figures 1 to 10 shown, based on the content in the above specific embodiments, the following content is further disclosed: To verify the advantages of the above entire method, this application conducts a comparative experiment with existing or traditional methods. The specific experimental content is as follows: Experimental objective: Through comparative experiments, verify the advantages of this method in terms of recognition accuracy, security, response time, and cross-scenario adaptability, and highlight the improvements of its core technologies (three-dimensional dynamic feature extraction, quantum-inspired matching, multi-modal adaptive fusion, etc.) compared with traditional methods.

[0041] The experimental content is as follows: Dataset: Self-built palm vein dataset: Collect palm vein data of 1000 users, 10 collections for each person, a total of 10000 groups of samples, including two-dimensional texture images, three-dimensional depth information, and blood flow dynamic data; The collection conditions include different illuminations (100 - 1000 lux), hand postures (±30° rotation), and devices (fixed devices and mobile devices); Forgery attack dataset: Contains 500 groups of forged samples, including printed images (300 groups) and 3D printed models (200 groups); The comparison methods are as follows: MethodA (Gabor filter): Traditional method, uses Gabor filter to extract two-dimensional texture features, Euclidean distance for matching, no liveness detection; MethodB (2D-CNN): Deep learning method, uses two-dimensional convolutional neural network to extract features, cosine similarity for matching, liveness detection based on texture contrast; This method: Based on three-dimensional spatio-temporal convolutional neural network (3DST-CNN), graph neural network (GNN), quantum-inspired matching (QIAM), and multi-modal adaptive fusion network; Experimental environment: Hardware: NVIDIA RTX 3090 GPU, Intel i9-12900K CPU, 64GB RAM; Software: Python 3.9, TensorFlow 2.8, Qiskit 0.43 (quantum computing simulation); Evaluation metrics: Recognition accuracy: Correct recognition rate (Accuracy, ACC); Anti-attack ability: Forged sample detection rate (Detection Rate, DR); Response time: Average time for single recognition (Response Time, RT); Cross-scenario adaptability: Variation in recognition rate under different lighting and device conditions (Adaptability, AD); Experimental design: Experiment 1: Recognition accuracy test: Test the recognition accuracy of three methods under standard conditions (lighting 500 lux, fixed device, no forged attack); Dataset: 8000 groups of training samples, 2000 groups of test samples; Experiment 2: Anti-attack ability test: Add forged samples (500 groups) under standard conditions to test the forged detection ability of three methods; Dataset: 2000 groups of real samples + 500 groups of forged samples; Experiment 3: Response time test: Test the average response time for single recognition on a fixed device; Dataset: 2000 groups of test samples; Experiment 4: Cross-scenario adaptability test: Test the variation in recognition rate under different lighting (100 lux, 500 lux, 1000 lux) and device (fixed device, mobile device) conditions; Dataset: 2000 groups of test samples; Experimental results and data are as follows: Table 1: Comparison of recognition accuracy: Table 1 Analysis through Table 1: The recognition accuracy of this method (99.2%) is significantly higher than that of Method A (92.5%) and Method B (95.8%); This is because three-dimensional dynamic feature extraction (3DST-CNN and GNN) captures spatial depth and blood flow dynamic features, and information bottleneck optimization further improves feature distinctiveness, while traditional methods only rely on two-dimensional texture features and are vulnerable to noise interference; Table 2: Comparison of anti-attack ability: Table 2 Analysis through Table 2: The forgery detection rate of this method (99.1%) far exceeds that of Method A (45.6%) and Method B (78.4%); Quantum-inspired matching uses the non-clonability of quantum states to defend against feature tampering, and hemodynamic liveness detection effectively distinguishes between live and forged samples, while Method A has no liveness detection and the texture contrast detection of Method B has limited effectiveness for dynamic forgeries (such as 3D printed models); Table 3: Comparison of response times: Table 3 Analysis through Table 3: The response time of this method (0.85 seconds) is slightly higher than that of Method A (0.45 seconds) and Method B (0.72 seconds), mainly due to the higher computational complexity of three-dimensional feature extraction and quantum matching; However, considering the significant improvement in accuracy and security, this time difference is acceptable in practical applications (such as national defense authentication); Table 4: Comparison of cross-scenario adaptability Table 4 Analysis through Table 4: The change in the recognition rate of this method under different lighting and device conditions (average change rate 1.5%) is much lower than that of Method A (7.2%) and Method B (5.1%); The multi-modal adaptive fusion network significantly improves cross-scenario adaptability through cross-modal attention mechanisms and dynamic weight allocation, combined with federated learning and meta-learning; Analysis of experimental results: Advantages in recognition accuracy: This method captures rich palm vein features through three-dimensional dynamic feature extraction and information bottleneck optimization, and the recognition accuracy is improved to 99.2%, which is 6.7% and 3.4% higher than that of Method A and Method B respectively; Traditional methods are limited by two-dimensional feature extraction and are vulnerable to environmental interference, while this method overcomes this problem by using multi-source data; Advantages in anti-attack ability: The forgery detection rate of this method reaches 99.1%, which is 53.5% and 20.7% higher than that of Method A and Method B respectively; Quantum matching and blood flow dynamic liveness detection effectively defend against static and dynamic forgery attacks, significantly improving security; Method A has no liveness detection and the texture detection of Method B is ineffective for complex forgeries; Analysis of response time: The response time of this method (0.85 seconds) is slightly higher than that of traditional methods, but still meets the requirements of real-time applications (such as financial payment < 1 second); The increase in time is mainly due to the complexity of three-dimensional feature extraction and quantum computing, but the improvement in its accuracy and security makes up for this shortcoming; Cross-scenario adaptability advantage: The change rate of the recognition rate of this method is only 1.5%, which is 5.7% and 3.6% lower than that of MethodA and MethodB respectively; multi-modal fusion and adaptive learning enable it to maintain stable performance under different lighting and device conditions and are applicable to diverse scenarios; Conclusion: Through comparative experiments, this method shows significant advantages in recognition accuracy, anti-attack ability, response time, and cross-scenario adaptability: This method is applicable to high-security scenarios (such as national defense authentication) and diverse scenarios (such as financial payment) and is superior to traditional methods.

[0042] To further verify the experiment, enrich the experimental details and the persuasiveness of the experiment, increase the comprehensiveness and persuasiveness of the data, reflect the performance advantages of this method under different conditions, and provide more comprehensive experimental results, including more scenarios, metrics, and statistical analyses to enhance the credibility and application reference value of the comparative experiment, as follows: Extended dataset: Real sample extension: Increased to 1500 users, with 10 collections per person, for a total of 15000 groups of samples, covering a wider range of age groups (18 - 65 years old) and gender distribution (50% male, 50% female); Forged sample extension: Increased to 1000 groups, including printed images (500 groups), 3D printed models (300 groups), and video playback (200 groups); Environmental variables: Added humidity conditions (30%, 60%, 90%) and temperature conditions (10℃, 25℃, 40℃); New evaluation metrics: Equal Error Rate (EER): Measures the balance point between the False Acceptance Rate (FAR) and the False Rejection Rate (FRR); Area Under the ROC Curve (AUC): A comprehensive metric for evaluating recognition performance; Noise Resistance (NR): The recognition rate under different Signal-to-Noise Ratios (SNR); Model Stability (ST): The standard deviation of the results of multiple runs; Supplementary experimental environment: Added mobile device testing: Using a portable device (RaspberryPi4, 4GB RAM, with a lightweight model); Simulating quantum computing: Using the IBMQiskit quantum simulator, with the number of qubits extended to 16; The experimental design is as follows: Experiment 5: Equal Error Rate and ROC Curve Analysis: Test the EER and AUC of three methods to evaluate the balance and stability of recognition performance; Dataset: 12000 groups of training samples, 3000 groups of test samples; Experiment 6: Anti-noise ability test: Test the recognition rate at different signal-to-noise ratios (SNR = 10dB, 20dB, 30dB); Dataset: 3000 groups of test samples, adding Gaussian noise; Experiment 7: Multi-scenario performance test: Test the recognition rate under different humidity, temperature and device conditions; Dataset: 3000 groups of test samples; Experiment 8: Model stability test: Run 10 times repeatedly, calculate the mean and standard deviation of the recognition rate; Dataset: 3000 groups of test samples; The experimental results and data are as follows: Table 5: Equal Error Rate and ROC Curve Analysis: Table 5 As can be seen from Table 5 above: The EER (0.4%) of this method is significantly lower than that of MethodA (3.5%) and MethodB (2.1%), and the AUC (0.99) is close to perfect classification, indicating that it is superior in the balance of misrecognition and rejection and overall recognition performance; This benefits from the high discriminability of three-dimensional dynamic feature extraction and quantum matching; Table 6: Anti-noise ability comparison: Table 6 As can be seen from Table 6 above: The recognition rate of this method under different signal-to-noise ratios (average 97.0%) is much higher than that of MethodA (85.8%) and MethodB (91.0%); Multi-source data acquisition and information bottleneck optimization effectively eliminate noise interference and improve the anti-noise ability, especially showing excellent performance under low SNR conditions; Table 7: Multi-scenario performance comparison: Table 7 As can be seen from Table 7 above: The average recognition rate of this method under different humidity, temperature and device conditions (98.4%) is higher than that of MethodA (89.8%) and MethodB (93.7%); Multi-modal adaptive fusion and dynamic weight allocation enable it to dynamically adapt to environmental changes and maintain high stability; Table 8: Model stability comparison: Table 8 As can be seen from Table 8 above: The standard deviation of the recognition rate of this method (0.5%) is lower than that of MethodA (2.3%) and MethodB (1.8%), and the standard deviation of the response time (0.06 seconds) is also relatively stable; Quantum matching and residual connection improve the stability and consistency of the model; Table 9: Forgery attack type extension comparison: Table 9 As can be seen from Table 9 above: The average detection rate of this method for different forgery types (98.3%) far exceeds that of Method A (38.3%) and Method B (69.3%); Blood flow dynamic live detection and quantum state unclonability endow it with high defense capabilities against complex attacks such as video playback; Analysis of experimental results: Equal error rate and ROC curve advantages: The EER (0.4%) and AUC (0.99) of this method indicate that it is significantly superior to traditional methods in terms of the balance between false recognition and rejection and overall performance, and is suitable for high-precision authentication scenarios; Advantages in anti-noise ability: The recognition rate of this method remains at a high level (94.2%) under low SNR (10dB), indicating its strong anti-noise ability and suitability for environments with high noise interference (such as outdoor mobile devices); Advantages in multi-scenario performance: The recognition rate of this method exceeds 97% under extreme humidity (90%), temperature (40°C) and mobile device conditions, demonstrating excellent cross-scenario adaptability and suitability for distributed deployment; Advantages in model stability: The low standard deviation of this method (recognition rate 0.5%, response time 0.06 seconds) indicates its high performance consistency and suitability for long-term operation and high reliability requirements; Advantages in forgery attack defense: The detection rate of this method for extended forgery types (including video playback) is as high as 98.3%, further verifying its security and suitability for high-risk scenarios; Comprehensive conclusion: Further verified the comprehensive advantages of this method in terms of accuracy, security, adaptability and stability: Three-dimensional dynamic feature extraction improves accuracy and anti-noise ability, quantum matching and live detection enhance security, multi-modal fusion and adaptive learning improve cross-scenario adaptability; New indicators and scenario data provide a more comprehensive performance evaluation and enhance the credibility of experimental results; This method shows better performance than traditional methods in multiple scenarios such as national defense, finance, and mobile payment, and has broad application value. Specific Embodiment 4: As Figures 1 to 10 shown, based on the content in the above specific embodiments, the following content is further disclosed: To verify the rationality and feasibility of the method of this application when in use, this application applies it to the following scenarios: Application Scenario 1: National Defense Identity Authentication System: Application background: In military bases or national defense facilities, identity authentication requires extremely high security, accuracy and anti-attack capabilities to prevent unauthorized access; This method ensures high-precision recognition and defense against forgery attacks through multi-source data collection, three-dimensional dynamic feature extraction, quantum matching and multi-modal fusion; Implementation steps and detailed data Sp1: Multi-source data collection and preprocessing: Device configuration: Use a near-infrared imaging device (wavelength 850 - 950 nm, acquisition resolution 1080x720 pixels), a light field camera (resolution 0.1 mm, acquisition depth range 5 - 10 mm), and a near-infrared spectral sensor (sampling frequency 15 Hz, measuring blood oxygen saturation and blood flow velocity); Data collection: The user places the palm in front of the device, collects a 5-second video sequence (75 frames), generates a two-dimensional texture image (1080x720x75), a three-dimensional depth map (1080x720x75), and blood flow dynamic data (75 time points, blood oxygen saturation range 85% - 98%, blood flow velocity range 0.2 - 0.5 m / s); Preprocessing: Equalize the illumination based on the multi-scale Retinex algorithm (parameters: σ = 15, 50, 150), denoise the depth map based on bilateral filtering (σ_s = 5, σ_r = 0.1), and smooth the blood flow data based on the Kalman filter (process noise covariance Q = 0.01, measurement noise covariance R = 0.1); Output: The preprocessed three-dimensional data tensor and blood flow time series; Sp2: Three-dimensional dynamic feature extraction: Three-dimensional spatio-temporal convolutional neural network: The input data is a 1080x720x75 tensor. Use a 3x3x3 convolutional kernel (32 filters) to extract spatial depth features, a 1x1x5 convolutional kernel to extract temporal blood flow features, add residual connections every two layers, and the output feature map size is 256x180x15; Graph neural network: Extract vascular branch points (about 50 - 80 nodes) from the depth features, construct a graph structure, the edges are based on vascular connection relationships (about 100 - 150 edges), and the graph convolution is iterated 3 times to output a topological feature vector (dimension 256); Output: Spatio-temporal feature tensor (256x180x15) and topological feature vector (256 dimensions); Sp3: Feature optimization and compression: Information bottleneck optimization: Input spatio-temporal features and topological features, calculate the mutual information (I(X;T) is about 5.2 bit, I(T;Y) is about 4.8 bit), the optimization target L = 5.2 - 2×4.8, the compression rate is about 40%, and output the optimized feature vector (154 dimensions); Quantum mapping: Map the 154-dimensional features to the quantum Hilbert space to generate quantum eigenstates ; Sp4: Quantum feature encoding and matching Quantum matching: Calculate similarity using the quantum inner product , Adaptive quantum gates Adjust the eigenstate (θ ranges from 0 to π), and optimize the parameters of the variational quantum algorithm (iterate 100 times, learning rate 0.01); Output: Matching score (ranges from 0 to 1, threshold 0.9); Sp5: Liveness detection and anti-attack: Liveness detection: Analyze 75 frames of blood flow data, calculate the blood flow velocity change rate (mean 0.15 m / s, standard deviation 0.05 m / s) and direction consistency (about 0.92), threshold 0.9, and if it is higher than the threshold, it is determined as a live body; Anti-attack: Utilize the non-clonability of quantum states to defend against 3D printed model attacks (detection rate 99.5%); Output: Liveness detection result (1 for live body, 0 for non-live body); Sp6: Multimodal adaptive fusion: Multimodal input: Combine palm vein features (154 dimensions) and iris features (128 dimensions); Cross-modal attention: Calculate the correlation matrix of palm vein and iris features, and output fused features (256 dimensions); Dynamic weight: According to the light intensity (500 lux) and device performance, assign weights (palm vein 0.6, iris 0.4); Output: Fused matching score (0.95), authentication successful; Three-dimensional dynamic feature extraction and information bottleneck optimization ensure high-precision recognition (recognition rate 99.8%), adapt to light and pose changes, quantum matching and liveness detection defend against forgery attacks (false recognition rate < 0.001%), and fusing iris features improves the system reliability; Implementation effect: Response time: 1.2 seconds; False recognition rate: 0.001%; Anti-attack success rate: 99.5%; Meet the high-security requirements of national defense.

[0044] Application scenario 2: Financial payment terminal: Application background: In the mobile payment scenario, fast, convenient and secure identity authentication is required; This method supports real-time payment verification and cross-device migration through efficient feature extraction, quantum matching and multimodal fusion; Sp1: Multi-source data collection and preprocessing: Device configuration: Portable near-infrared imaging device (wavelength 850 - 900 nanometers, resolution 640x480 pixels), light field camera (resolution 0.15 mm), near-infrared spectral sensor (sampling frequency 10 Hz); Data acquisition: The user hovers their palm above the device and collects a 3 - second video sequence (30 frames), generating a 2D texture image (640x480x30), a 3D depth map (640x480x30), and blood flow dynamic data (30 time points, oxygen saturation range 88% - 96%, blood flow velocity range 0.3 - 0.6m / s); Pre - processing: Multi - scale Retinex illumination equalization (σ = 10, 30, 100), bilateral filtering for denoising (σ_s = 3, σ_r = 0.05), Kalman filtering for smoothing (Q = 0.005, R = 0.05); Output: Pre - processed 3D data and blood flow sequence; Sp2: 3D dynamic feature extraction: 3D spatio - temporal convolutional neural network: Input a 640x480x30 tensor, use 3x3x3 convolutional kernels (16 filters) to extract spatial features, 1x1x3 convolutional kernels to extract temporal features, with residual connections for each layer, and the output feature map size is 128x120x10; Graph neural network: Extract blood vessel nodes (about 30 - 50), edges (about 60 - 100), perform 2 iterations of graph convolution, and output a topological feature vector (128 - dimensional); Output: Spatio - temporal feature tensor (128x120x10) and topological feature vector (128 - dimensional); Sp3: Feature optimization and compression: Information bottleneck optimization: Input features, calculate mutual information (I(X;T) is about 4.5bit, I(T;Y) is about 4.2bit), optimize L = 4.5 - 2×4.2, with a compression rate of about 30%, and output an optimized feature vector (90 - dimensional); Quantum mapping: Generate quantum feature states ; Sp4: Quantum feature encoding and matching: Quantum matching: Calculate the quantum inner product , adaptively adjust the feature state with a quantum gate (θ range 0 - π / 2), and optimize with a variational quantum algorithm (50 iterations, learning rate 0.005); Output: Matching score (threshold 0.85, score 0.92); Sp5: Liveness detection and anti - attack: Liveness detection: Analyze 30 - frame blood flow data, blood flow velocity change rate (mean 0.18m / s, standard deviation 0.04m / s), direction consistency (0.88), threshold 0.85, and determine it as a live body; Anti - attack: Defend against printed image attacks (detection rate 99%); Output: Liveness detection result (1 for live body); Sp6: Multimodal adaptive fusion: Multi-modal input: Combining palm vein features (90 dimensions) and fingerprint features (96 dimensions); Cross-modal attention: Aligning feature distributions and outputting fused features (192 dimensions); Dynamic weights: Based on illumination (300 lux) and device performance, weight allocation (palm vein 0.55, fingerprint 0.45); Output: Fused matching score (0.89), payment successful; Three-dimensional feature extraction ensures a high recognition rate (99.5%) and adapts to the mobile device environment; Quantum matching and liveness detection defend against forgeries (false recognition rate < 0.01%); Fusing fingerprint features enhances payment reliability; Implementation effects: Response time: 0.8 seconds; False recognition rate: 0.01%; Anti-attack success rate: 99%; Supports fast and secure mobile payments; In summary, this method features high security, high precision, and anti-attack capabilities in the defense scenario and is suitable for high-risk environments; in the financial scenario, it emphasizes fast response, high robustness, and multi-modal collaboration to meet the needs of convenient payments. The two scenarios comprehensively reflect the core advantages of multi-source acquisition, three-dimensional feature extraction, quantum matching, liveness detection, and multi-modal fusion.

[0045] It should be noted that in this article, relational terms such as first and second are only used 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 term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a reference structure" does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0046] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A palm vein feature extraction and depth adaptive matching method, characterized in that: The method comprises the following steps: Sp1: Multi-source data acquisition and preprocessing: Use near-infrared imaging equipment, light field cameras and near-infrared spectral sensors to collect multi-source data of palm veins, including two-dimensional texture images, three-dimensional depth information and blood flow dynamic data, and preprocess the collected data; Sp2: 3D dynamic feature extraction: The spatial depth features and temporal blood flow dynamic features of the palm veins are extracted through a 3D spatiotemporal convolutional neural network, and the topological features of the palm veins are extracted in combination with a graph neural network; Sp3: Feature optimization and compression: The extracted features are optimized through the information bottleneck and mapped to the quantum Hilbert space to generate quantum characteristic states; Sp4: Quantum feature encoding and matching: Use quantum-inspired matching algorithms to match quantum feature states, and the quantum-inspired matching algorithms include quantum inner product calculations and adaptive quantum gate operations, and are combined with variational quantum algorithms to optimize matching parameters; Sp5: Liveness detection and anti-attack: Liveness detection is performed by analyzing multi-frame blood flow dynamic data, and the quantum state non-cloning property is used to defend against counterfeit attacks; Sp6: Multimodal adaptive fusion: Palm vein features are collaboratively matched with other biometric features through a multimodal adaptive fusion network, which includes a cross-modal attention mechanism and dynamic weight assignment.

2. The palm vein feature extraction and depth adaptive matching method according to claim 1, characterized in that: In the multi-source data acquisition and preprocessing, a near-infrared imaging device is used to collect two-dimensional texture images of palm veins with a wavelength range of 850-950 nanometers, a light field camera is used to collect three-dimensional depth information of palm veins with a resolution of not less than 0.1 mm, and a near-infrared spectral sensor is used to collect palm vein blood flow dynamic data, including blood oxygen saturation and blood flow velocity, with a sampling frequency of not less than 10 Hz.

3. The palm vein feature extraction and depth adaptive matching method according to claim 1, characterized in that: The preprocessing in the multi-source data acquisition and preprocessing includes adaptive illumination equalization of two-dimensional texture images based on a multi-scale Retinex algorithm, denoising of three-dimensional depth information based on a bilateral filtering algorithm, and signal smoothing of blood flow dynamics data based on a Kalman filtering algorithm.

4. The palm vein feature extraction and depth adaptive matching method according to claim 1, wherein: The three-dimensional spatiotemporal convolutional neural network in the three-dimensional dynamic feature extraction includes the following contents: The spatial convolution layer uses a 3x3x3 convolution kernel to extract the spatial depth features of the palm veins; Temporal convolution layer, using 1x1xT convolution kernel to extract blood flow dynamic features; Residual connection module, used to alleviate the gradient vanishing problem in deep networks.

5. The palm vein feature extraction and depth adaptive matching method according to claim 1, wherein: When the graph neural network in the three-dimensional dynamic feature extraction is running, the palmar vein vascular structure is modeled as a graph structure, in which nodes represent vascular branch points or intersections, and edges represent vascular connection relationships. Node features are updated through graph convolution operations, with the number of iterations being 2-5 times, and finally a global topological feature vector is output with a dimensional range of 128-512.

6. The palm vein feature extraction and depth adaptive matching method according to claim 1, characterized in that: The information bottleneck optimization in the feature optimization and compression includes calculating the mutual information of feature vectors, retaining a feature subset that is highly relevant to identity authentication, eliminating feature components related to environmental noise or redundant information, and outputting optimized feature vectors with a compression rate ranging from 20% to 50%.

7. The palm vein feature extraction and depth adaptive matching method according to claim 1, characterized in that: The quantum-inspired matching algorithm in the live detection and anti-attack includes calculating the similarity between two sets of quantum eigenstates using quantum inner product, adjusting the eigenstates through adaptive quantum gates, optimizing the parameters of the quantum circuit using variational quantum algorithms, and the number of iterations ranges from 50 to 200 times.

8. The palm vein feature extraction and depth adaptive matching method according to claim 1, wherein: The live detection in the multimodal adaptive fusion includes analyzing continuous multi-frame blood flow dynamic data, calculating the blood flow velocity change rate and direction consistency, setting the live detection threshold, which ranges from 0.8 to 0.

95. If it is lower than the threshold, it is determined as non-live, and the live detection result is output for subsequent matching decisions.

9. The palm vein feature extraction and depth adaptive matching method according to claim 1, characterized in that: The multimodal adaptive fusion network in the multimodal adaptive fusion includes the following: A cross-modal attention mechanism for aligning the spatial and semantic distributions of palm vein features with other biometric features; A dynamic weight allocation module that adjusts the matching weights of each modal feature according to environmental conditions and device performance, and outputs the fused matching score, which ranges from 0 to 1.

Images