Multi-level permission access control identification method and system based on biometric feature weight distribution

Abstract

The present application relates to the field of computer vision and biometric recognition, and particularly relates to a multi-level permission access control recognition method and system based on biometric feature weight distribution. The method comprises the steps of: acquiring a target continuous image frame, extracting macrostructure and microtexture features; determining a weight reduction ratio based on actual displacement between frames and a limit displacement tolerance, and stripping a target weight amount from a microtexture reference weight; extracting corneal reflection spot coordinates, calculating relative displacement in combination with actual eye size, inputting a mapping network to obtain a predicted compensation feature, and assigning the above weight amount to obtain a semantic compensation feature; superimposing and fusing the macroscopic, microtexture after weight reduction, and semantic compensation feature into a comprehensive authentication feature, matching a database, and verifying permission to drive the lock to be opened and upload records. Through dynamic weight distribution and eye spot compensation, motion blur interference is effectively overcome, and the recognition accuracy and access control security in a dynamic scene are improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and biometric recognition, and in particular to a multi-level access control recognition method and system based on biometric weight allocation. Background Technology

[0002] In biometric scenarios such as smart glass door lock control and rapid facial verification in office areas, it is often necessary to guide the target to complete identity authentication and access control while they are approaching rapidly without cooperation. To improve recognition stability in complex environments, existing multimodal visual fusion technologies mainly rely on the parallel extraction of multi-dimensional features.

[0003] For example, patent application CN121259886A discloses an intelligent face recognition and early warning system for monitoring. This scheme mainly relies on static multidimensional feature extraction from images captured by image acquisition devices, and then performs similarity-weighted matching by combining gait features and body temperature indicators. To address image quality degradation caused by lighting or posture, this scheme employs a static quality assessment model. After the feature extraction process is completed, a post-processing score is performed, and corresponding static comparison weights are assigned to each modality feature based on the score results.

[0004] However, the aforementioned patent applications have significant underlying data processing defects when dealing with non-cooperative, rapid approach scenarios, namely, severe time lag in feature weight adjustment, which makes it unable to handle transient motion blur.

[0005] In real-world scenarios involving rapid approach and door opening, the image is highly susceptible to non-uniform motion blur due to the combined constraints of equipment exposure time and target movement speed. As the target rapidly approaches the lens, low-frequency macroscopic facial contour features become clearer due to image magnification, but high-frequency microscopic iris details and periorbital texture features suffer severe semantic dimensionality reduction and information loss due to pixel movement. The quality assessment model used in this patent application performs a static evaluation only after all features have been extracted. This approach cannot respond in real-time to the degradation rate of the image within a microsecond-level window. Lacking the ability to transiently adjust feature weights within a millisecond-level exposure window, those already blurred and damaged microscopic high-frequency features still carry significant erroneous noise in the final similarity calculation, easily leading to incorrect rejection or misidentification by the system.

[0006] Therefore, how to overcome the fuzzy feature interference caused by the quality scoring lag in existing feature matching schemes under high-speed motion conditions, and achieve dynamic and real-time feature weight reduction and accurate compensation, has become a technical problem that urgently needs to be solved in the field of remote control of smart glass door locks. Summary of the Invention

[0007] To address the problem that existing feature matching schemes are prone to misidentification due to blurred or damaged features interfering with the comparison results caused by quality scoring delays under high-speed motion conditions, this invention provides a multi-level access control identification method and system based on biometric feature weight allocation.

[0008] In a first aspect, the present invention provides a multi-level access control identification method based on biometric weight allocation, which adopts the following technical solution: S1: Obtain consecutive image frames containing the target object, and extract the macroscopic structural features and microscopic texture features of the current frame; S2: Determine the feature weighting ratio based on the actual motion displacement between consecutive image frames and the pre-calibrated limit displacement tolerance. Based on the feature weighting ratio, extract the target weight amount from the baseline weight of the micro-texture features to obtain the micro-texture features after weight reduction. S3: Extract the coordinate data of the corneal reflective spot in the current frame, calculate the relative displacement feature based on the measured eye size, input the relative displacement feature into the pre-trained mapping network to output the corresponding prediction compensation feature, and assign the target weight to the prediction compensation feature to obtain the semantic compensation feature. S4: The macroscopic structural features, the microscopic texture features after weight reduction, and the semantic compensation features are superimposed and fused at the element level to generate a comprehensive authentication feature. The comprehensive authentication feature is matched with the base database features. After the match is successful, the multi-level door lock permissions corresponding to the target object are retrieved and verified. Based on the legal permission level, the control command is output to drive the electric glass door lock to open, and a door lock record containing the operator's identity and unlocking time is generated and uploaded to the remote control management system.

[0009] This invention first measures the degree of motion blur directly through physical displacement, and automatically calculates the feature weight reduction ratio in real time when people move quickly. It also accurately extracts the target weight from the baseline weight of micro-texture features, effectively preventing damaged and blurred feature data from participating in the final comparison and causing misjudgment.

[0010] Furthermore, this invention extracts the coordinate data of corneal reflective spots that maintain physical stability during movement, combines this with measured eye dimensions to convert them into relative displacement features, uses a mapping network to calculate predictive compensation features, and directly assigns the previously stripped target weights to these predictive compensation features to obtain semantic compensation features. This processing method effectively achieves accurate replacement of feature energy. Finally, macroscopic structural features, weighted microscopic texture features, and semantic compensation features are element-wise superimposed and fused, ensuring that the final comprehensive authentication features used for comparison not only eliminate fuzzy noise but also complete lost information, significantly improving the recognition accuracy in high-speed traffic scenarios.

[0011] Preferably, the mapping network is a fully connected neural network, and its pre-training process includes: Acquire clear images and simulated motion-blurred images of the same target object; The baseline semantic features of the clear image and the incomplete semantic features of the simulated motion-blurred image are extracted using a feature extraction network, and the feature difference between the two is calculated as the true value of the high-frequency residual lost due to blurring. The relative displacement features extracted from the simulated motion-blurred image are input into the mapping network to obtain the prediction compensation features; With the goal of reducing the mean square error between the predicted compensation features and the true value of the high-frequency residuals, the network parameters of the mapping network are adjusted using the backpropagation algorithm.

[0012] This invention clarifies the pre-training process of the mapping network. This training process directly uses the feature difference between the baseline semantic features of the clear image and the incomplete semantic features of the simulated motion-blurred image as the sole learning benchmark. By calculating the mean square error between the true value of the high-frequency residual and the predicted compensation features output by the network, the fully connected neural network is forced to accurately fit the feature data actually lost due to motion blur. This training method ensures that the compensation features generated by the mapping network have extremely high fidelity and prevents the network from generating false features that interfere with the final base database matching.

[0013] Preferably, the fully connected neural network includes an input layer, at least two hidden layers, and an output layer; The hidden layer is used to perform linear dimensionality increase and nonlinear mapping transformation on the input data and connect activation functions. The output layer outputs the predicted compensation feature and its number of neurons is consistent with the number of dimensions of the baseline semantic feature.

[0014] Preferably, the extraction of macroscopic structural features and microscopic texture features of the current frame includes: The current frame is decomposed in the frequency domain using the median filtering algorithm and the Laplacian operator to separate the low-frequency macroscopic structural components that characterize the facial contours and the high-frequency microscopic texture components that characterize the periorbital texture. Feature extraction is performed on the two separated components to obtain the macroscopic structural features and microscopic texture features.

[0015] This invention utilizes a median filtering algorithm and the Laplacian operator to perform frequency domain decomposition on the current frame, achieving physical isolation of features across different frequency bands from a visual level. This separation process enables the system to isolate low-frequency macroscopic structural components, which are less susceptible to blurring, from high-frequency microscopic texture components, which are highly susceptible to blurring. This provides a priori conditions for subsequent precise weight reduction and compensation targeting only the microscopic parts, preventing cross-contamination of global facial features by local blurring.

[0016] Preferably, the step of stripping the target weight amount from the baseline weight of the microtexture feature according to the feature reduction ratio to obtain the weighted microtexture feature includes: The ratio of the actual displacement to the limit displacement tolerance is calculated as the weighting ratio coefficient. The target weight is calculated by multiplying the baseline weight of the microtexture feature by the weight reduction ratio coefficient. The target weight is subtracted from the baseline weight to obtain the reduced feature weight, which is then applied to the microtexture feature.

[0017] This invention calculates a weighting ratio coefficient by measuring the ratio of actual displacement to the tolerance limit displacement, and then multiplies this coefficient by the baseline weight of the micro-texture features to obtain the target weight amount. This calculation step establishes a precise linear adjustment closed loop in engineering: the faster the movement speed and the greater the actual displacement, the more target weight amount the system deducts proportionally. This allows the system to adaptively control the intensity of feature repair, maintaining the authenticity of system features at various movement speeds.

[0018] Preferably, the method for obtaining the actual motion displacement based on the consecutive image frames is as follows: Multi-level Gaussian pyramid downsampling is performed on two adjacent frames to establish the scale space of the image; The initial displacement value is calculated from the top of the pyramid and then passed down and corrected layer by layer. At the original layer level, the transient pixel offset, which characterizes the actual motion displacement, is calculated and output using a local optical flow statistical algorithm.

[0019] This invention employs a multi-layer Gaussian pyramid downsampling process. At the top layer of the pyramid, the actual large-span displacement is proportionally compressed into a small displacement for initial value calculation. This value is then passed down and corrected layer by layer, ultimately outputting the transient pixel offset at the original layer level. This combined computational process completely solves the computational defect of conventional local optical flow statistical algorithms, which easily loses tracking targets when facing non-cooperative high-speed targets. It ensures that the output actual motion displacement has extremely high numerical accuracy, providing reliable physical data support for the subsequent accurate calculation of feature weighting ratios.

[0020] Preferably, extracting the coordinate data of the corneal reflective spot in the current frame includes: In the high-frequency image region of the current frame, an adaptive threshold segmentation algorithm is applied to extract the corneal reflective spot region, and the geometric center of the spot region is calculated as the coordinate data.

[0021] Preferably, the calculation of relative displacement features based on measured eye size includes: Detect the actual pixel interpupillary distance of the human eye in the current frame; The ratio of the displacement represented by the coordinate data to the actual pixel interpupillary distance is used as the relative displacement feature.

[0022] Preferably, the step of element-level superposition and fusion of macroscopic structural features, weighted microscopic texture features, and semantic compensation features to generate comprehensive authentication features includes: After ensuring that the macroscopic structural features, the weighted microscopic texture features, and the semantic compensation features are on the same dimension, the feature vectors corresponding to the three are added element by element to generate a fused feature vector, and the fused feature vector is normalized to the second norm to generate the comprehensive authentication feature.

[0023] Secondly, this invention provides a multi-level access control system based on biometric weight allocation, employing the following technical solution: A multi-level access control system based on biometric weight allocation includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned multi-level access control method based on biometric weight allocation.

[0024] By adopting the above technical solution, the multi-level access control identification method based on biometric weight allocation is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0025] The present invention has the following technical effects: This invention abandons the traditional static scoring method after feature extraction. By directly converting the actual physical motion displacement into a dynamic weighting basis for micro-texture features, the system can remove interference noise caused by damaged data in real time and accurately.

[0026] Furthermore, this invention precisely and equally assigns the reduced target weight to the predictive compensation features, generating semantic compensation features, and finally completes feature repair through element-level superposition and fusion. This effectively ensures that even within a millisecond-level blurred exposure window, the system can still combine comprehensive authentication features with standardized dimensions, rich information, and free of dirty data. This fundamentally improves the interception accuracy of multi-level access control systems under complex dynamic conditions and provides highly reliable underlying data support for cloud-based intelligent control and management systems, achieving seamless integration of physical authentication and networked access control. Attached Figure Description

[0027] Figure 1 This is a flowchart of the multi-level access control identification method based on biometric weight allocation provided in the embodiments of the present invention; Figure 2This is a diagram illustrating the effect of matching results changing with actual motion displacement, provided in an embodiment of the present invention. Figure 3 The diagram shows the effect of the existing method matching result changing with the actual motion displacement provided in the embodiment of the present invention. Detailed Implementation

[0028] The access control and authentication system of this invention relies on a specific physical hardware carrier and network architecture for collaborative operation. This system is integrated into a smart control device for a glass door lock with biometric recognition and networked remote management functions. Its basic hardware includes an authentication terminal body installed on the side of the door. The front recognition area of ​​the authentication terminal body is equipped with a high-definition image acquisition device for capturing visible light images of the target object. Status indicator lights are fixedly installed on the surface of the authentication terminal body, and an interactive display screen supporting high information density data presentation is provided in the center to intuitively display the physical authentication parameter flow and real-time device operation curves. The internal structure integrates a processor for image processing and feature modeling calculations, as well as a memory for storing program instructions and base database data.

[0029] In addition, the system also includes an access control module and a network communication module that are communicatively connected to the processor. The access control module is used to directly drive external access devices, namely the glass door electric / magnetic lock bodies. The network communication module establishes a real-time bidirectional data link with the remote visual management host computer, used to receive remote door opening and closing control commands and multi-level door lock permission configuration policies issued by the remote visual management host computer, and to upload detailed door lock opening records and data flow information to the remote visual management host computer in real time. Simultaneously, the system is equipped with an abnormal vibration / forced tampering monitoring sensor and a local audible and visual alarm, serving as the underlying physical security auxiliary execution terminal for linking the alarm notification service module of the management host computer.

[0030] Example 1: This invention discloses a multi-level access control identification method based on biometric weight allocation, referring to... Figure 1 This includes steps S1-S4: S1: Obtain consecutive image frames containing the target object, and extract the macroscopic structural features and microscopic texture features of the current frame.

[0031] First, during system operation, the aforementioned image acquisition device acquires continuous visible light image frames of the target object as it approaches the authentication terminal.

[0032] Then, to achieve frequency band feature separation and extraction, the processor uses a median filtering algorithm and a Laplacian operator to perform frequency domain decomposition on the acquired current frame. Specifically, the median filter is used as a low-pass filter to process the current frame, separating macroscopic low-frequency structural components representing facial contours and face shapes; the Laplacian operator is used as a high-frequency filter to process the current frame, separating microscopic high-frequency texture components representing iris details and periocular textures. The effect of this operation is that it achieves physical isolation of frequency band features from the visual level, providing prior conditions for subsequent differentiated weighting of different quality features, and effectively avoiding cross-contamination of global features by motion blur.

[0033] Next, the processor first performs size scaling and pixel grayscale normalization on the macroscopic structure components and microscopic texture components, respectively, to convert them into a numerical matrix format that conforms to the network input standard. The processor then separately inputs the processed macroscopic structure components and microscopic texture components into the ResNet-18 feature extraction network. This network contains multiple residual blocks composed of skip connections, performing multi-layer convolutional forward propagation operations. During this process, the processor truncates the feature tensor output from the third stage of the network and uses global average pooling to reduce the dimensionality and flatten the feature tensor. Finally, the network outputs a fixed 512-dimensional vector for each of the macroscopic structure components and microscopic texture components, denoted as the macroscopic structure feature and the microscopic texture feature, respectively.

[0034] It should be added that the ResNet-18 feature extraction network was pre-trained on an offline server before the device left the factory. The specific training process is as follows: (1) Construct a training dataset to obtain a large number of clear face and periorbital images of known identities. For any identity in the dataset, select two different images of that identity as a positive sample pair; select an image of that identity and an image of another different identity as a negative sample pair.

[0035] (2) The above sample images are preprocessed by scaling and normalizing pixel gray values ​​in sequence, and then input into the initial state of the ResNet-18 network to extract the corresponding 512-dimensional semantic feature vector.

[0036] (3) The error is calculated using the triplet loss function. The calculation logic of this loss function is as follows: calculate the Euclidean distance between the two semantic feature vectors in the positive sample pair and the Euclidean distance between the two semantic feature vectors in the negative sample pair; with the optimization objective of minimizing the distance between the positive sample pairs and making the distance between the negative sample pairs greater than the distance between the positive sample pairs plus a preset numerical boundary, the current loss value is calculated.

[0037] (4) Calculate the gradient using the backpropagation algorithm and continuously update the weight parameters of each convolutional layer within the ResNet-18 network until the loss value output by the triplet loss function no longer decreases and tends to converge. After training is completed, solidify the network parameters and store them in the device's memory.

[0038] By employing a residual block skip connection structure within the network, the gradient vanishing problem, which is prone to occur during training and feature extraction in deep networks, is effectively avoided, ensuring the computational stability of feature extraction. By truncating the third stage of the network and using global average pooling to output a fixed 512-dimensional vector, sufficient local feature information is preserved while redundant spatial location information is eliminated. This significantly compresses the massive image data into a simplified one-dimensional array, greatly reducing the hardware load on subsequent feature fusion and computation. Furthermore, rigorous training using the triplet loss function in the offline stage forces the network to bring vectors of the same identity closer together and push vectors of different identities apart in mathematical space. This ensures that the 512-dimensional feature vector output by the network has extremely high identity discriminative power, providing accurate data for the final identity matching.

[0039] While extracting the aforementioned features, the processor uses a feature detection operator within the target object region of the current frame to identify and extract a set of feature points with high gradient changes, which serve as physical anchor points for subsequent cross-frame motion tracking.

[0040] S2: Determine the feature weighting ratio based on the actual motion displacement between the consecutive image frames and the pre-calibrated limit displacement tolerance, and strip the target weight amount from the base weight of the micro-texture feature according to the feature weighting ratio to obtain the micro-texture feature after weight reduction.

[0041] Since high-speed approach of the target can cause high-frequency micro-textures to become blurred and lose their functionality due to convolution, this step aims to dynamically reduce the weight of damaged micro-texture features through physical displacement calculations. The specific operation is as follows: First, the system performs L-level Gaussian pyramid downsampling on two adjacent frames to establish a scale space. The iterative process for specific displacement calculation includes: For ease of description, the previous frame will be referred to as the reference frame, and the current frame will be referred to as the frame to be matched.

[0042] (1) At the top of the pyramid, the processor maps each feature point in the feature point set to the coordinate system of that scale. For each feature point, the processor compares the gray-level distribution of the top-level image of the reference frame with that of the top-level image of the frame to be matched in the neighborhood of that point, initializes the displacement estimate as a zero vector, and solves the initial residual displacement corresponding to each feature point through the Lucas-Kanade constraint equation. The purpose of this operation is to take advantage of the characteristic that high-frequency noise in the image is filtered out at extremely low resolution, to provide a reliable initial search starting point for high-speed displacement with a large span, and to avoid the computational dilemma of traditional optical flow methods easily getting trapped in local optima when dealing with non-cooperative targets.

[0043] (2) The processor multiplies the displacement output calculated for each feature point at the previous level by a scale factor, which serves as the initial displacement estimate for that feature point at the current level. For each feature point, the processor resamples and translates its neighboring pixels in the matching frame image at the current level based on its corresponding initial displacement estimate, so that the local region is visually initially aligned with the corresponding neighboring region in the reference frame image. The processor then performs optical flow calculation again in the aligned local region to solve for the small positional residual of each feature point relative to the reference frame. The processor adds the initial displacement estimate of each feature point to the small positional residual to obtain the final displacement output of that feature point at the current level. In this embodiment, the scale factor is preferably set to 2.

[0044] (3) Repeat the above process until backtracking to the bottom of the pyramid, and finally output the transient pixel offset vector of each feature point in the feature point set. The processor summarizes the transient pixel offset vectors corresponding to all feature points in the feature point set, and calculates the average value of the magnitude of all vectors to obtain the actual motion displacement of the current frame.

[0045] The above method utilizes a pyramid structure to progressively shrink the image layer by layer. At the smallest top layer, the actual large-span displacement is proportionally compressed into a tiny displacement that the optical flow algorithm can handle. Then, through reverse mapping, the image is restored layer by layer, and calculation errors are corrected. This combined processing completely solves the problem of traditional optical flow methods being unable to keep up with high-speed motion. It can not only capture large-scale motion shifts but also ensure that the final output pixel displacement data is highly accurate, thus providing reliable data support for the subsequent accurate subtraction of the weights of damaged textures.

[0046] Then, the system retrieves the pre-calibrated limit displacement tolerance, which represents the maximum pixel offset allowed when the system's recognition rate is within acceptable limits.

[0047] Next, the processor removes the actual moving parts with a limit displacement tolerance to obtain a weight reduction ratio. The original baseline weight of the microtexture features is multiplied by this weight reduction ratio to calculate the target weight amount to be stripped.

[0048] Finally, the target weight is subtracted from the baseline weight to obtain the reduced feature weight, which is then applied to the microtexture features.

[0049] This stage directly uses the actual physical displacement values ​​measured at the front end as a clear basis for feature weight adjustment, clearly describing the amount of detail loss caused by motion in the current image. This allows for real-time and precise reduction of the impact of damaged micro-texture features on the final recognition and comparison, based on the actual speed of movement of the person. This effectively prevents misjudgments caused by the device using blurred or invalid feature data for authentication.

[0050] The above embodiments involve limit displacement tolerance. The calibration process of the limit displacement tolerance is described below: During the factory preset stage, a standard test model is carried on a constant speed motion slide rail, and multiple sets of continuous image frames are collected at different moving speeds to test the system matching degree. When the system matching pass rate falls below the preset safety threshold, such as 99%, the average maximum pixel offset of feature points between two adjacent frames is recorded and fixed as the limit displacement tolerance of the device.

[0051] S3: Extract the coordinate data of the corneal reflective spot in the current frame, calculate the relative displacement feature based on the measured eye size, input the relative displacement feature into a pre-trained mapping network to output the corresponding prediction compensation feature, and assign the target weight to the prediction compensation feature to obtain the semantic compensation feature.

[0052] Because S2 reduces the weight of micro-texture features affected by motion blur, the total amount of feature information used by the system for final identity comparison is lost. To fill this gap in recognition information, the system must find an alternative feature source that remains stable even under high-speed motion. Considering that the geometric properties of the corneal reflective spot are highly resistant to motion blur, the system aims to extract the physical displacement of this spot and use it as the base data for calculating compensation information. At the same time, in order to eliminate the absolute scale error caused by different users' distances from the lens, the system must introduce the inherent physiological scale of the human body, transforming absolute physical motion into a relative displacement ratio unaffected by shooting distance.

[0053] Based on the above logic, the specific execution operation is as follows: First, to eliminate external high-light interference such as eyeglass reflections and skin sweat, the processor first locates the contour of the human eye based on macroscopic structural components, and extracts the region of interest (ROI) containing the eyeball. Within this ROI, the processor acquires microscopic texture components as high-frequency image regions. The processor then processes these high-frequency image regions using an adaptive threshold segmentation algorithm to extract the geometric center of the corneal reflection spot to obtain coordinate data. The difference between the coordinate data of the corneal reflection spot in two adjacent frames is calculated to obtain the absolute displacement. Simultaneously, the processor detects the actual interpupillary distance (IPD) of the human eye in the current frame, divides the absolute displacement by the actual IPD, and outputs the relative displacement feature.

[0054] By first locating the eye region and then extracting the highlight points, false highlight points caused by complex ambient lighting can be accurately filtered out, ensuring that the selected physical reference object is unique and highly reliable. At the same time, by using the actual pixel interpupillary distance as the denominator for division, the error of image size magnification or reduction caused by differences in user height and distance from the lens is effectively eliminated, ensuring that the final calculated displacement feature only reflects the pure motion amplitude, providing standardized input data for subsequent network calculations that is not affected by shooting distance.

[0055] Then, the system runs a pre-trained fully connected neural network as a mapping network. The input layer of this mapping network receives the aforementioned 2D relative displacement features; the hidden layers employ a three-layer structure: the first hidden layer increases the 2D data to 128 dimensions, the second expands it to 256 dimensions, and the third hidden layer further extracts feature associations through a nonlinear mapping transformation; each hidden layer is followed by a ReLU activation function to introduce nonlinear correction. The number of neurons in the output layer is strictly set to 512, thus outputting a 512-dimensional vector with the exact same dimension as the feature extraction network, defined as the prediction compensation feature. This feature represents the feature residuals lost due to motion displacement.

[0056] The training process for this mapping network is as follows: (1) The processor acquires a clear image and a simulated motion-blurred image of the same target object and separates the corresponding micro-texture components.

[0057] (2) Using a feature extraction network, forward propagation is performed on the two micro-texture components respectively to extract the 512-dimensional baseline semantic features corresponding to the clear image and the 512-dimensional incomplete semantic features corresponding to the simulated motion-blurred image. The baseline semantic features and the incomplete semantic features are subtracted element by element to calculate the numerical difference vector between the two, which is defined as the true value of the high-frequency residual.

[0058] (3) Input the relative displacement features corresponding to the blurred image into the mapping network, and use the mean square error between the output result and the above high-frequency residual true value as the loss function. With the goal of minimizing the loss function, continuously adjust the connection weights of each layer through the backpropagation algorithm until the loss value output by the loss function converges, and complete the training.

[0059] Instead of letting the network generate features aimlessly, the training process directly uses the true numerical difference between the clear image and the blurred image as the sole benchmark for network learning. Through the strict constraints of the mean squared error loss function, the mapping network is forced to accurately fit the part of data actually lost due to motion blur, thereby ensuring that the predicted compensation features generated by the network have extremely high realism and effectively preventing false features from interfering with the recognition results.

[0060] Finally, the processor multiplies the target weight by each element of the 512-dimensional predicted compensation feature to obtain the semantic compensation feature. The target weight represents the proportion of damage to the original texture, and by using it as a coefficient in the predicted compensation feature, adaptive adjustment of the compensation intensity is achieved.

[0061] S4: The macroscopic structural features, the microscopic texture features after weight reduction, and the semantic compensation features are superimposed and fused at the element level to generate a comprehensive authentication feature. The comprehensive authentication feature is matched with the base database features. After the match is successful, the multi-level door lock permissions corresponding to the target object are retrieved and verified. Based on the legal permission level, the control command is output to drive the electric glass door lock to open, and a door lock record containing the operator's identity and unlocking time is generated and uploaded to the remote control management system.

[0062] First, the processor acquires macroscopic structural features, microscopic texture features with reduced weights, and semantic compensation features, all of which are in the same 512-dimensional range. Then, it performs element-wise addition on the feature vectors corresponding to these three features to generate an initial fused feature vector. Subsequently, it performs second-norm normalization on the initial fused feature vector to uniformly scale its vector magnitude to 1, thereby generating a comprehensive authentication feature that maintains the 512-dimensional range.

[0063] Subsequently, the processor retrieves the pre-registered base database feature vectors from memory, which have also undergone second-norm normalization, and calls the industry-standard cosine similarity algorithm to calculate the spatial morphological similarity between the two. Since the magnitudes of both the comprehensive authentication feature vector and the base database feature vector involved in the comparison have been constrained to 1, the cosine similarity calculation is mathematically equivalent to the single-dot product operation of the two vectors. This design completely eliminates the conventional and complex division and square root operations, minimizing the hardware overhead of high-frequency transient comparisons and perfectly meeting the real-time requirements of high-information-density industrial terminals for rapid passage through millisecond-level exposure windows.

[0064] When the calculated cosine similarity is greater than or equal to the preset matching threshold, the processor determines that the target object's identity verification has passed. Simultaneously, it immediately extracts the identifier code of this legitimate identity, retrieves the corresponding hierarchical role-based access control policy matrix, and initiates the following multi-level spatiotemporal Boolean dynamic verification of permissions at the local hardware level: When the processor verifies that the identity code belongs to the super administrator role, the system assigns it the highest access priority. At this time, the Boolean logic is always true, which not only directly allows the person to open the current glass door lock at any time without regard to time window restrictions, but also activates the dedicated high-priority channel of the network communication module, allowing the terminal to receive and execute remote door opening and closing unauthorized control commands issued by the remote visual management host computer.

[0065] When the processor verifies that the identification code belongs to an internal employee role, the system automatically retrieves the set of authorized physical lock nodes for that person, as well as the preset normal working day access window matrix, such as 08:00-20:00 on weekdays. The processor reads local RTC real-time clock data for comparison. Only when both the current timestamp and the spatial node are within the valid window width will the spatiotemporal Boolean check return a true value, and then issue an access command; if it is outside the working period, it is determined to be unauthorized, and the lock remains locked.

[0066] When the processor verifies that the identity code belongs to a temporary external visitor, the system initiates a lifecycle single-use restriction verification. This permission must meet a specific dynamic time interval. Once the preset single-use validity period or access period is exceeded, its permission policy matrix automatically becomes invalid and returns a false value.

[0067] After the multi-level permission Boolean verification is passed, the processor outputs control commands to the access control module, driving the connected glass door electric lock or magnetic lock to perform the physical release action to unlock; at the same time, the status indicator light on the surface of the authentication terminal body lights up the release mark, and the physical authentication dot product value stream and permission verification passed mark are displayed on the interactive display screen in real time.

[0068] Meanwhile, the processor, through the network communication module, encrypts and uploads structured data packets containing door lock number, user identity, unlocking time, and door lock status to the remote visual management host computer in real time, so that the door lock record query module on the system side can synchronize and retain them. If the system detects a forced signal on the door through the local abnormal vibration / forced removal monitoring sensor in an unauthenticated state, the processor will immediately activate the local audible and visual alarm and push a warning command to the alarm notification module of the host computer.

[0069] To demonstrate the effectiveness of the solution, relevant experiments were conducted. Below are the images obtained from the experiments: Figure 2This image illustrates the effect of the matching results of this invention changing with actual motion displacement. The horizontal dotted line represents the preset matching threshold for safe passage, and the solid line formed by the concatenation of solid square markers represents the dynamic trajectory of the cosine similarity after feature extraction and fusion under different motion displacements. The image shows that the solid line trajectory of this invention exhibits extremely strong anti-interference capabilities. Throughout the entire testing period, the solid line maintains a significant safe distance from the horizontal dotted line, firmly locking within the high confidence range. More importantly, even when facing large-span actual motion displacements, the solid line trajectory can still be quickly reconstructed to the similarity peak using compensation logic and maintain translational stability.

[0070] Figure 3 The image shows the effect of existing matching methods changing with actual motion displacement. The horizontal dotted line represents the preset matching threshold, and the dashed line formed by connected hollow circles represents the decay trajectory of cosine similarity with increasing motion displacement when using the traditional, unpatched existing feature matching scheme. The image shows that as the actual motion displacement increases, the dashed line representing the existing technology exhibits a continuous and rapid downward trend. When the actual motion displacement expands to a certain range, the dashed line completely breaks through the horizontal dotted line, and even more severe numerical collapse occurs in subsequent large displacement ranges.

[0071] By comparison Figure 2 and Figure 3 As can be seen, this invention can accurately remove the weights of damaged features that have been blurred and contaminated, while simultaneously using predictive compensation features generated by physical laws to fill in the missing information in equal amounts. This energy replacement method ensures that the system can still output comprehensive authentication features with standard dimensions and purity when facing complex motion conditions, fundamentally overcoming the problem of identity misidentification in high-speed access control scenarios.

[0072] This invention also discloses a multi-level access control system based on biometric weight allocation, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the multi-level access control method based on biometric weight allocation according to this invention.

[0073] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A multi-level access control identification method based on biometric feature weight allocation, characterized in that, include: S1: Obtain consecutive image frames containing the target object, and extract the macroscopic structural features and microscopic texture features of the current frame; S2: Determine the feature weighting ratio based on the actual motion displacement between consecutive image frames and the pre-calibrated limit displacement tolerance. Based on the feature weighting ratio, extract the target weight amount from the baseline weight of the micro-texture features to obtain the micro-texture features after weight reduction. S3: Extract the coordinate data of the corneal reflective spot in the current frame, calculate the relative displacement feature based on the measured eye size, input the relative displacement feature into the pre-trained mapping network to output the corresponding prediction compensation feature, and assign the target weight to the prediction compensation feature to obtain the semantic compensation feature. S4: The macroscopic structural features, the microscopic texture features after weight reduction, and the semantic compensation features are superimposed and fused at the element level to generate a comprehensive authentication feature. The comprehensive authentication feature is matched with the base database features. After the match is successful, the multi-level door lock permissions corresponding to the target object are retrieved and verified. Based on the legal permission level, the control command is output to drive the electric glass door lock to open, and a door lock record containing the operator's identity and unlocking time is generated and uploaded to the remote control management system.

2. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The mapping network is a fully connected neural network, and its pre-training process includes: Acquire clear images and simulated motion-blurred images of the same target object; The baseline semantic features of the clear image and the incomplete semantic features of the simulated motion-blurred image are extracted using a feature extraction network, and the feature difference between the two is calculated as the true value of the high-frequency residual lost due to blurring. The relative displacement features extracted from the simulated motion-blurred image are input into the mapping network to obtain the prediction compensation features; With the goal of reducing the mean square error between the predicted compensation features and the true value of the high-frequency residuals, the network parameters of the mapping network are adjusted using the backpropagation algorithm.

3. The multi-level access control identification method based on biometric weight allocation according to claim 2, characterized in that, The fully connected neural network includes an input layer, at least two hidden layers, and an output layer; The hidden layer is used to perform linear dimensionality increase and nonlinear mapping transformation on the input data and connect activation functions. The output layer outputs the predicted compensation feature and its number of neurons is consistent with the number of dimensions of the baseline semantic feature.

4. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The extraction of macroscopic structural features and microscopic texture features of the current frame includes: The current frame is decomposed in the frequency domain using the median filtering algorithm and the Laplacian operator to separate the low-frequency macroscopic structural components that characterize the facial contours and the high-frequency microscopic texture components that characterize the periorbital texture. Feature extraction is performed on the two separated components to obtain the macroscopic structural features and microscopic texture features.

5. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The step of removing the target weight from the baseline weight of the micro-texture features according to the feature reduction ratio to obtain the weighted micro-texture features includes: The ratio of the actual displacement to the limit displacement tolerance is calculated as the weighting ratio coefficient. The target weight is calculated by multiplying the baseline weight of the microtexture feature by the weight reduction ratio coefficient. The target weight is subtracted from the baseline weight to obtain the reduced feature weight, which is then applied to the microtexture feature.

6. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The specific method for obtaining the actual motion displacement between consecutive image frames is as follows: Multi-level Gaussian pyramid downsampling is performed on two adjacent frames to establish the scale space of the image; The initial displacement value is calculated from the top of the pyramid and then passed down and corrected layer by layer. At the original layer level, the transient pixel offset, which characterizes the actual motion displacement, is calculated and output using a local optical flow statistical algorithm.

7. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The step of extracting the coordinate data of the corneal reflective spot in the current frame includes: In the high-frequency image region of the current frame, an adaptive threshold segmentation algorithm is applied to extract the corneal reflective spot region, and the geometric center of the spot region is calculated as the coordinate data.

8. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The calculation of relative displacement features based on measured eye size includes: Detect the actual pixel interpupillary distance of the human eye in the current frame; The ratio of the displacement represented by the coordinate data to the actual pixel interpupillary distance is used as the relative displacement feature.

9. The multi-level access control identification method based on biometric weight allocation according to claim 1, characterized in that, The process of generating a comprehensive authentication feature by element-level superposition and fusion of macroscopic structural features, weighted microscopic texture features, and semantic compensation features includes: After ensuring that the macroscopic structural features, the weighted microscopic texture features, and the semantic compensation features are on the same dimension, the feature vectors corresponding to the three are added element by element to generate a fused feature vector, and the fused feature vector is normalized to the second norm to generate the comprehensive authentication feature.

10. A multi-level access control system based on biometric weight allocation, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the multi-level access control identification method based on biometric weight allocation according to any one of claims 1-9.

Images