Remote Verification Method and System for Terminal Device Access Permission

Abstract

The present invention relates to the field of data processing, and specifically provides a method and system for remotely verifying the access permission of a terminal device, including: performing a biometric extraction operation on the physical sign acquisition image of the terminal device to obtain an initial permission verification feature; decomposing the initial permission verification feature to obtain a plurality of initial verification sub-features; adjusting the feature extraction error of each initial verification sub-feature according to the physical sign acquisition image of the terminal device to obtain a verification adjustment sub-feature; determining a verification adjustment feature based on the plurality of verification adjustment sub-features; improving the feature of the verification adjustment feature according to the physical sign acquisition image of the terminal device to obtain a permission verification improvement feature; determining a target permission verification feature based on the permission verification improvement feature, and performing permission verification based on this. The present invention can obtain a permission verification feature with higher representativeness and accuracy, and the result obtained by performing permission verification based on this permission verification feature has higher precision and reliability.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular, to a method and system for remotely verifying the access permission of a terminal device. Background Art

[0002] In today's digital age, the remote operation and maintenance terminals of government and enterprise customers play a crucial role in various businesses. With the rapid development of information technology, the data processed and tasks undertaken by remote operation and maintenance terminals are becoming increasingly critical, and the security verification of their access permissions has become the core link to ensure the normal operation of government and enterprise services and data security. Traditional methods for verifying the access permissions of terminal devices have exposed many limitations when faced with the increasingly complex network environment and diverse security threats. Early verification methods often relied on simple password or token mechanisms, which were extremely easy to crack or steal and could not provide sufficient security. For example, passwords may be set too simply by users or leaked in an insecure environment, and tokens may also be lost or copied, enabling illegal users to easily obtain access permissions, posing serious risks of data leakage and service interruption to government and enterprise customers. With the progress of technology, biometric recognition technology has gradually been applied to the field of access permission verification. However, existing biometric-based verification methods still have significant deficiencies. On the one hand, the accuracy of biometric extraction needs to be improved. Due to differences in the accuracy of acquisition devices, interference from environmental factors, and the diversity of biometric features themselves, errors are likely to occur during biometric extraction. Summary of the Invention

[0003] In view of this, at least one embodiment of the present invention provides a method and system for remotely verifying the access permission of a terminal device. The technical solution of the present invention is implemented as follows:

[0004] On the one hand, the present invention provides a method for remotely verifying the access permission of a terminal device, the method comprising: performing a biometric extraction operation on a physical sign acquisition image of the terminal device to obtain an initial permission verification feature; decomposing the initial permission verification feature to obtain a plurality of initial verification sub-features; adjusting the feature extraction error of each initial verification sub-feature according to the physical sign acquisition image of the terminal device to obtain a corresponding verification adjustment sub-feature; determining a verification adjustment feature based on the plurality of verification adjustment sub-features; improving the feature of the verification adjustment feature according to the physical sign acquisition image of the terminal device to obtain a permission verification improvement feature; determining a target permission verification feature based on the permission verification improvement feature; and performing identification based on the target permission verification feature to obtain a verification result of the access permission of the terminal device.

[0005] On the other hand, the present invention provides a remote verification system for the access permission of a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, the steps in the above-mentioned method are implemented.

[0006] The remote verification method for the access permission of the terminal device provided by the present invention performs a biometric extraction operation on the body sign acquisition image of the terminal device to obtain an initial permission verification feature. The initial permission verification feature is decomposed according to the body sign parts to obtain a plurality of initial verification sub-features. For each initial verification sub-feature among the plurality of initial verification sub-features, according to the body sign acquisition image of the terminal device, the extraction error of the initial verification sub-feature is determined, and the extraction error is adjusted to obtain a corresponding verification adjustment sub-feature, and a plurality of verification adjustment sub-features are obtained. Then, according to the plurality of verification adjustment sub-features, a verification adjustment feature is obtained. According to the body sign acquisition image of the terminal device, the verification adjustment feature is subjected to feature improvement processing to obtain a permission verification improvement feature. A target permission verification feature is generated according to the permission verification improvement feature.

[0007] In the embodiment of the present invention, after obtaining the initial permission verification feature, the extraction error of the initial permission verification feature is adjusted and the feature is improved to obtain a permission verification improvement feature, and then a target permission verification feature is obtained according to the permission verification improvement feature. The processing process of adjusting the extraction error of the feature is used to adjust the extraction error of the initial permission verification feature for the body sign acquisition image of the terminal device, and the process of feature improvement perfects the missing features of the verification adjustment feature for the body sign acquisition image of the terminal device. Based on this, a permission verification feature with higher representativeness and accuracy can be obtained, and the result obtained by performing permission verification based on this permission verification feature has higher precision and reliability. Description of the Drawings

[0008] The drawings here are incorporated into the specification and constitute a part of this specification. These drawings show embodiments consistent with the present invention and are used together with the specification to explain the technical solutions of the present invention.

[0009] Figure 1 It is a schematic flowchart of the implementation of a remote verification method for the access permission of a terminal device provided by an embodiment of the present invention.

[0010] Figure 2 It is a schematic diagram of the hardware entity of a remote verification system for the access permission of a terminal device provided by an embodiment of the present invention.

[0011] Remote verification system for terminal device access permission - 1000; Processor - 1001; Memory - 1002. Detailed Embodiments

[0012] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further elaborated in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be construed as limitations on the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0013] An embodiment of the present invention provides a method for remote verification of terminal device access rights. This method can be executed by a processor of a remote verification system for terminal device access rights. Among them, the remote verification system for terminal device access rights can refer to devices with data processing capabilities such as servers, laptops, tablets, desktop computers, etc.

[0014] Figure 1 It is a schematic flowchart of the implementation of a method for remote verification of terminal device access rights provided by an embodiment of the present invention. As Figure 1 shown, this method includes the following steps:

[0015] Step S100: Perform a biometric extraction operation on the terminal device physical sign acquisition image to obtain an initial permission verification feature.

[0016] In step S100, the terminal device physical sign acquisition image is obtained by collecting the physical sign information related to the remote operation and maintenance terminal of government and enterprise customers through specific devices and technologies. For example, it may be through a high-definition camera to capture the physical sign parts such as the face and fingerprint of the operation and maintenance personnel, so as to obtain an image containing key information. The quality and accuracy of these images directly affect the effect of subsequent biometric extraction. After obtaining the terminal device physical sign acquisition image, the remote verification system for terminal device access rights starts the biometric extraction operation. This operation aims to accurately extract the biometric information related to permission verification from the collected image. To achieve this goal, a feasible implementation method is to use a pre-calibrated and converged permission verification neural network. This neural network has been trained and optimized with a large amount of data and can effectively identify and extract the biometric features in the image.

[0017] For example, by learning a large number of images containing different facial features, fingerprint features, etc., the permission verification neural network can master how to accurately distinguish the key biometric feature points and feature lines from the image. When the remote verification system for terminal device access rights inputs the terminal device physical sign acquisition image into this pre-calibrated and converged permission verification neural network, the network will analyze and process the image according to the patterns and rules it has learned.

[0018] Specifically, the remote verification system for terminal device access permissions first preprocesses the terminal device physical sign acquisition image. This may include grayscale processing of the image, that is, converting a color image into a grayscale image to simplify subsequent calculations and analyses. At the same time, noise reduction processing of the image may also be performed to remove the noise generated by factors such as the shooting device and environmental interference in the image, and improve the clarity and quality of the image. For example, the Gaussian filtering algorithm is used to perform noise reduction on the image, and by calculating the weighted average of the grayscale values of each pixel point and its surrounding pixel points in the image, the noise interference is removed.

[0019] Next, the remote verification system for terminal device access permissions inputs the preprocessed image into the permission verification neural network. The neural network will perform a series of feature extraction operations on the image. For example, when processing a facial image, the neural network can identify the position and shape information of key facial features such as eyes, nose, and mouth; when processing a fingerprint image, it will extract features such as fingerprint patterns, breakpoints, and bifurcation points. These extracted feature information will constitute an important part of the initial permission verification features.

[0020] Suppose the terminal device physical sign acquisition image is a facial image of an operation and maintenance personnel. The remote verification system for terminal device access permissions first performs grayscale and noise reduction processing on it, and then inputs it into the permission verification neural network. The neural network analyzes the image layer by layer through its internal neurons and connection weights. In the first layer, the edge information in the image can be identified, such as the contour edge of the eyes; in the second layer, more complex features can be further extracted, such as the shape and relative position of the eyes; as the network layers deepen, more advanced and abstract facial features will be continuously extracted, and finally, the initial permission verification features containing the overall facial feature information will be formed. Another example, if the collected is a fingerprint image, the remote verification system for terminal device access permissions also performs preprocessing and then inputs it into the permission verification neural network. The network will extract the detailed features of the fingerprint from the image, such as the position of the breakpoint and the direction of the bifurcation point, and these features combined will constitute the initial permission verification features related to the fingerprint.

[0021] Step S200: Decompose the initial permission verification features to obtain multiple initial verification sub-features.

[0022] In step S200, the initial permission verification features processed by the remote verification system for the access permission of the terminal device are a set of features with permission verification-related information extracted from the physical sign acquisition images of the terminal device. These features may contain various biometric elements. For example, when processing facial physical signs, it may cover feature information of different parts such as eyes, nose, and mouth; when processing fingerprint physical signs, it may include features such as fingerprint patterns, break points, and bifurcation points. The purpose of the remote verification system for the access permission of the terminal device to perform feature decomposition is to decompose these comprehensive features into multiple sub-features that are more independent and targeted according to certain rules.

[0023] Decomposing according to the physical sign part is a feasible and effective way. For example, when the initial permission verification features come from the facial physical sign acquisition image, the remote verification system for the access permission of the terminal device divides according to different regions of the face. The eye region, as an important part of the face, its features are of great significance for identity verification. The remote verification system for the access permission of the terminal device separates the features related to the eyes from the initial permission verification features to form an initial verification sub-feature about the eyes. This may include the shape features of the eyes, such as whether the eyes are round, almond-shaped, etc.; the position features of the eyes, that is, the relative position relationship of the eyes on the face; and the color features of the eyes, etc. By separately extracting and integrating these eye-related features, an independent initial verification sub-feature of the eyes is formed. Similarly, for the nose part, the remote verification system for the access permission of the terminal device extracts feature information such as the shape, length, and width of the nose and combines them into an initial verification sub-feature about the nose. For example, whether the nose is straight or has a sunken bridge of the nose, and the relationship between the length of the bridge of the nose and the overall proportion of the face, etc., these features are all incorporated into the initial verification sub-feature of the nose. The mouth part is no exception. The remote verification system for the access permission of the terminal device pays attention to the shape of the mouth, such as the thickness of the lips, the degree of upward or downward curvature of the corners of the mouth, etc., and organizes these features into an initial verification sub-feature about the mouth.

[0024] When processing the fingerprint feature collection image, the remote verification system for the access permission of the terminal device will also decompose it according to different feature regions of the fingerprint. The trend of the fingerprint lines is an important feature. The remote verification system for the access permission of the terminal device extracts the feature information reflecting the overall trend of the fingerprint lines and forms an initial verification sub-feature regarding the trend of the fingerprint lines. For example, whether the fingerprint lines are clockwise or counterclockwise, or features such as presenting complex spiral shapes will be included. The breakpoint feature of the fingerprint is also one of the key information. The remote verification system for the access permission of the terminal device extracts the features related to the breakpoint position, the number of breakpoints, etc. and constitutes an initial verification sub-feature regarding the breakpoints. For the bifurcation points of the fingerprint, the remote verification system for the access permission of the terminal device analyzes features such as the direction of the bifurcation points and the number of bifurcations and integrates them into an initial verification sub-feature regarding the bifurcation points.

[0025] The implementation method for the remote verification system for the access permission of the terminal device to perform such feature decomposition operations can utilize specific algorithms and models. For example, the convolutional neural network (CNN) technology based on deep learning can be adopted. By designing appropriate convolutional kernels and network structures, the network is enabled to learn how to perform accurate feature decomposition according to the feature differences of the physical sign parts. When the convolutional kernel scans the initial permission verification features, it will extract and classify features at different positions and types according to the set weight parameters. For example, for the feature extraction of the eye part, the convolutional kernel will focus on the pixel information related to the eye shape, color, etc. and separate these features from the overall initial permission verification features through calculation and analysis. Another example is that the remote verification system for the access permission of the terminal device quantifies each feature element in the initial permission verification features and converts them into digital vector forms. Then, through clustering algorithms such as the K-means clustering algorithm, according to the distance metric between the feature vectors, similar feature elements are clustered together. In the case of decomposition according to the physical sign parts, the feature vectors related to the eyes will be clustered into one category to form the initial verification sub-feature of the eyes; the feature vectors related to the nose are clustered into another category to form the initial verification sub-feature of the nose, and so on. In this way, the remote verification system for the access permission of the terminal device decomposes the initial permission verification features into multiple initial verification sub-features with clear physical sign part directions, providing clear and independent data units for subsequent more accurate feature extraction error adjustment and permission verification operations based on these sub-features, which helps to improve the accuracy and reliability of the entire verification process.

[0026] Step S300: According to the physical sign collection image of the terminal device, perform feature extraction error adjustment on each of the initial verification sub-features to obtain corresponding verification adjustment sub-features.

[0027] The initial verification sub-features obtained by the remote verification system for terminal device access rights are obtained by decomposing the initial rights verification features in step S200. Although these initial verification sub-features each represent a part or a characteristic of the physical signs, due to factors such as environmental interference during the image acquisition process, device precision limitations, and limitations of the feature extraction algorithm, there will inevitably be certain extraction errors. The task of the remote verification system for terminal device access rights is to minimize these errors through feature extraction error adjustment.

[0028] Specifically, the remote verification system for terminal device access rights first needs to determine the extraction errors existing in the initial verification sub-features. This requires using the physical sign acquisition images of the terminal device as a reference. For example, when the initial verification sub-feature is a feature related to the eyes part of the face, during the extraction process, due to insufficient image light, some detailed features of the eye shape may not be accurately extracted, resulting in an error in the shape deformation; or when extracting the fingerprint break point feature, due to local blurring of the fingerprint image, the position of the break point has a deviation. To adjust these errors, the remote verification system for terminal device access rights can adopt various implementation methods. A feasible method is to use a deep learning model. By constructing a neural network model specifically for feature extraction error adjustment, this model takes the physical sign acquisition images of the terminal device and the initial verification sub-features as inputs. The neurons and connection weights inside the model are trained and optimized with a large amount of data, and can learn the error patterns that different initial verification sub-features may appear in different situations and the corresponding adjustment methods. For example, for the initial verification sub-feature of the eyes on the face, the remote verification system for terminal device access rights inputs the physical sign acquisition image of the terminal device and this initial verification sub-feature into the error adjustment neural network together. The neural network deeply analyzes the pixel information of the eyes part in the image, compares it with the existing standard eye feature database, and identifies the errors existing in the current initial verification sub-feature. Assuming that it is found that there is a deviation in the eye shape feature, the neural network will adjust this initial verification sub-feature according to the adjustment rules it has learned, such as changing the parameter values of the eye shape to make it closer to the real eye shape, so as to obtain the adjusted verification adjustment sub-feature.

[0029] When processing the initial verification sub-feature of the fingerprint break point, the remote verification system for terminal device access rights also inputs the fingerprint acquisition image and this sub-feature into the error adjustment model. The model will analyze the overall fingerprint pattern and the detailed information around the break point in the image to judge the accuracy of the break point position. If it is found that there is a deviation in the break point position, the model will correct the break point position according to the preset algorithm and training experience to generate a more accurate verification adjustment sub-feature.

[0030] Another implementable implementation method is a statistics-based method. The remote verification system for terminal device access permissions can collect a large number of physical sign acquisition images of the same type of terminal devices and their corresponding accurate initial verification sub-feature data, and establish a statistical model. When adjusting the error of the current initial verification sub-feature, the remote verification system for terminal device access permissions compares and analyzes the sub-feature with the data in the statistical model. By calculating statistical indicators such as the similarity and deviation degree between features, the error range of the initial verification sub-feature is determined. Then, according to a pre-set error adjustment formula, the initial verification sub-feature is adjusted. For example, assuming that for the fingerprint pattern feature extraction error of a certain type, the error adjustment formula obtained from the statistical model is: adjusted feature value = current feature value + (standard feature value - current feature value) × error correction coefficient. Among them, the standard feature value comes from the accurate data in the statistical model, and the error correction coefficient is an empirical value obtained based on a large amount of experimental data. The remote verification system for terminal device access permissions calculates the initial verification sub-feature through this formula to obtain a verified adjustment sub-feature after error adjustment.

[0031] For another example, for the adjustment of facial features, the remote verification system for terminal device access permissions can use the method of geometric transformation. If there is an error in the position relationship of facial features in the initial verification sub-feature, the remote verification system for terminal device access permissions can use geometric transformation algorithms such as affine transformation and perspective transformation to adjust the positions of facial features in the initial verification sub-feature according to the actual positions and proportional relationships of facial features in the physical sign acquisition image of the terminal device, so that they conform to the real facial structure, and then obtain a more accurate verified adjustment sub-feature. Through these implementation methods, the remote verification system for terminal device access permissions can effectively perform feature extraction error adjustment for each initial verification sub-feature based on the physical sign acquisition image of the terminal device, obtain the corresponding verified adjustment sub-feature, provide a higher-quality and more accurate data basis for subsequent determination of the verified adjustment feature and further permission verification operations, thereby improving the reliability and accuracy of the entire remote verification of terminal device access permissions.

[0032] Step S400: Determine the verified adjustment feature based on multiple said verified adjustment sub-features.

[0033] Step S400 requires the remote verification system for terminal device access permissions to determine the verified adjustment feature based on multiple verified adjustment sub-features. This step integrates and refines multiple sub-features after feature extraction error adjustment, aiming to generate a more comprehensive, representative and accurate verification feature, providing key data support for the subsequent permission verification process.

[0034] The multiple verification adjustment sub - features processed by the remote verification system for the access permission of the terminal device are the results obtained after adjusting the feature extraction error of each initial verification sub - feature in step S300. These verification adjustment sub - features depict the physical signs of the terminal device more accurately from different angles, different physical sign parts, or different feature dimensions. For example, when processing facial signs, the eye verification adjustment sub - feature, nose verification adjustment sub - feature, mouth verification adjustment sub - feature, etc. after error adjustment each contain more precise feature information about the corresponding facial parts; when processing fingerprint signs, the verification adjustment sub - features in aspects such as fingerprint line direction, break points, and bifurcation points also more accurately reflect the actual features of the fingerprint. The process by which the remote verification system for the access permission of the terminal device determines the verification adjustment features is a process of organically integrating these scattered but interrelated verification adjustment sub - features. A feasible implementation method is to use the weighted fusion method. The remote verification system for the access permission of the terminal device assigns corresponding weights to each verification adjustment sub - feature according to its importance for the overall permission verification. For example, in facial sign verification, due to its uniqueness and stability, the eye feature has a relatively high importance for identity verification. Therefore, the remote verification system for the access permission of the terminal device can assign a relatively high weight, such as 0.4, to the eye verification adjustment sub - feature; the nose feature is slightly less important than the eye feature, and may be assigned a weight of 0.3; the weight of the mouth feature may be 0.3. The assignment of weights is not fixed, but can be dynamically adjusted and optimized through a large amount of experimental data and actual application scenarios.

[0035] After determining the weights, the remote verification system for the access permission of the terminal device calculates the verification adjustment features according to the following formula: Verification adjustment feature = Eye verification adjustment sub - feature × 0.4 + Nose verification adjustment sub - feature × 0.3 + Mouth verification adjustment sub - feature × 0.3. The calculation here is not a simple numerical addition, but a corresponding weighted operation on the feature vectors or feature data contained in each verification adjustment sub - feature. For example, the eye verification adjustment sub - feature may be a feature vector containing multi - dimensional information such as eye shape, color, and position. The remote verification system for the access permission of the terminal device multiplies each dimension value of the vector by 0.4. Similar weighted operations are also performed on the nose and mouth verification adjustment sub - features, and then the weighted results are combined and integrated to obtain a verification adjustment feature that synthesizes the feature information of multiple facial parts.

[0036] For fingerprint characteristics, the remote verification system for terminal device access permissions also assigns weights to verification adjustment sub-characteristics such as fingerprint line direction, breakpoints, and bifurcation points. Suppose the weight of the fingerprint line direction verification adjustment sub-characteristic is 0.4, the weight of the breakpoint verification adjustment sub-characteristic is 0.3, and the weight of the bifurcation point verification adjustment sub-characteristic is 0.3. The remote verification system for terminal device access permissions integrates and calculates these verification adjustment sub-characteristics according to a similar weighted fusion formula as above to obtain a verification adjustment characteristic reflecting the overall fingerprint characteristics.

[0037] In addition to the weighted fusion method, the remote verification system for terminal device access permissions can also adopt feature splicing and convolutional fusion techniques in deep learning. First, the remote verification system for terminal device access permissions splices multiple verification adjustment sub-characteristics in the feature dimension. For example, the verification adjustment sub-characteristics of eyes, nose, and mouth are connected in a certain order to form a longer feature vector. Then, the spliced feature vector is processed by a convolutional neural network. The convolutional layer in the convolutional neural network uses a specific convolutional kernel to perform a convolution operation on this long feature vector. The convolutional kernel slides on the feature vector, and by calculating the inner product of the convolutional kernel and the local area of the feature vector, more advanced and abstract feature information is extracted. For example, the convolutional kernel can identify features such as the spatial relationship between eyes, nose, and mouth. After multiple convolutional operations, the remote verification system for terminal device access permissions finally obtains a verified and refined verification adjustment characteristic. Through these implementation methods, the remote verification system for terminal device access permissions can effectively integrate and process multiple verification adjustment sub-characteristics to generate a verification adjustment characteristic. This verification adjustment characteristic not only synthesizes the information of multiple sub-characteristics but also further highlights the key features and weakens the secondary features through reasonable weight assignment or deep learning feature fusion methods, thus more comprehensively and accurately reflecting the true characteristics of the terminal device characteristics, providing a solid and reliable data basis for subsequent feature improvement based on this verification adjustment characteristic and finally determining the target permission verification characteristic, and helping to improve the accuracy and reliability of the remote verification of terminal device access permissions.

[0038] Step S500: Based on the image collected from the terminal device characteristics, improve the verification adjustment characteristic to obtain a permission verification improved characteristic.

[0039] The verification adjustment characteristic is obtained by integrating multiple verification adjustment sub-characteristics in step S400. Although these verification adjustment sub-characteristics have undergone feature extraction error adjustment and are more accurate than the initial verification sub-characteristics, due to various complex factors in the characteristic collection process, the verification adjustment characteristic may still have some missing or inaccurate parts. The task of the remote verification system for terminal device access permissions is to make up for these deficiencies through feature improvement operations.

[0040] Specifically, the remote verification system for terminal device access rights clearly verifies the possible missing information of the verification adjustment features. This requires the use of the terminal device's physical sign acquisition images again. For example, in the face physical sign verification scenario, although the verification adjustment features already contain the feature information of the main facial parts such as the eyes, nose, and mouth that have been error-adjusted, due to the angle problem during image acquisition, some detailed features of the facial contour may not be fully reflected in the verification adjustment features; or in fingerprint physical sign verification, the verification adjustment features may miss some tiny features at the edge of the fingerprint.

[0041] To achieve feature improvement, the remote verification system for terminal device access rights can adopt a feature generation network technology based on deep learning. By constructing a dedicated feature generation network, this network takes the terminal device's physical sign acquisition images and verification adjustment features as inputs. This feature generation network is trained with a large amount of data and can learn the mapping relationship between different physical sign acquisition images and the corresponding complete features. For example, for face physical sign verification, the remote verification system for terminal device access rights inputs the face image collected by the terminal device and the existing verification adjustment features into the feature generation network together. The neurons inside the network will conduct in-depth analysis on the facial information in the image, and at the same time, combine the existing information contained in the verification adjustment features to identify the missing detailed features of the facial contour in the current verification adjustment features. Then, the network generates the corresponding missing feature information according to the patterns and rules it has learned, and fuses it with the original verification adjustment features. Suppose the facial contour features in the original verification adjustment features are not complete enough. The feature generation network can generate contour features that conform to the actual situation of the current face through learning a large number of face images and add them to the verification adjustment features, so as to obtain more perfect permission verification and improvement features.

[0042] Another feasible implementation method is based on feature matching and interpolation. The remote verification system for terminal device access rights matches the verification adjustment features with a pre-established physical sign feature library. This physical sign feature library contains a large number of complete physical sign feature samples of different types, different angles, and different states. The remote verification system for terminal device access rights calculates the similarity between the verification adjustment features and each sample in the feature library to find the most matching sample set. For example, in face physical sign verification, find multiple face feature samples that are most similar to the current verification adjustment features in terms of facial structure, expression, etc.

[0043] Then, based on the differences between these matching samples and the verification adjustment features, an interpolation algorithm is used to fill in the missing parts of the verification adjustment features. Suppose the features of the verification adjustment features in a certain area of the face are not clear enough or there are missing parts, while the matching samples have complete and clear features in this area. The remote verification system for terminal device access rights can adopt algorithms such as linear interpolation or spline interpolation. According to the feature values of the matching samples in this area, it calculates and generates supplementary feature values suitable for the current verification adjustment features. For example, for the skin color feature in a certain area of the face, through the interpolation algorithm, based on the skin color information of the matching samples, it generates skin color feature values that conform to the actual situation of the current face and adds them to the verification adjustment features to achieve the improvement of the features.

[0044] Through these implementation manners, the remote verification system for terminal device access rights can effectively perform feature improvement operations on the verification adjustment features based on the terminal device physical sign acquisition images, and obtain the permission verification improvement features. This permission verification improvement feature more comprehensively and accurately reflects the true features of the terminal device physical signs, provides a better and more reliable data basis for subsequent determination of the target permission verification features and the final permission verification, and greatly improves the accuracy and effectiveness of the remote verification of terminal device access rights.

[0045] Step S600: Determine the target permission verification feature based on the permission verification improvement feature.

[0046] The permission verification improvement feature is the result obtained after feature improvement of the verification adjustment feature in step S500. This feature already reflects the information of the terminal device physical signs more comprehensively and accurately. However, to ensure the high precision of the permission verification, it is still necessary to further determine the target permission verification feature.

[0047] A feasible implementation manner is feature selection and fusion. The remote verification system for terminal device access rights analyzes the permission verification improvement feature and judges the importance of each sub-feature in it for the permission verification. For example, in the face physical sign verification, the permission verification improvement feature may include multiple sub-features such as face contour, facial feature proportion, and skin color. The remote verification system for terminal device access rights determines the contribution degree of each sub-feature to the identity verification through a large amount of experimental data and machine learning algorithms. For example, after analysis, it is found that the face contour and facial feature proportion have higher recognition in the identity verification, while the skin color has a relatively small impact on the verification result in some cases.

[0048] Based on this analysis, the remote verification system for terminal device access rights adopts a feature selection algorithm, such as the method based on information gain. Information gain is used to measure the degree of increase in information brought by a feature in a classification task. Its calculation formula is: Information gain = Entropy (original data set) - Entropy (data set divided according to the feature). The calculation formula of entropy is: , where is the probability of the event occurring. By calculating the information gain of each sub - feature, the remote verification system for terminal device access permissions can select sub - features with higher information gain, that is, sub - features that are more important for permission verification. For example, in facial feature verification, sub - features with high information gain such as facial contour and facial feature proportion are selected.

[0049] Then, the remote verification system for terminal device access permissions fuses the selected sub - features. The fusion method can be weighted average fusion. Suppose the facial contour sub - feature F1 and the facial feature proportion sub - feature F2 are selected. The remote verification system for terminal device access permissions assigns weights to them according to their importance in permission verification. For example, the weight of the facial contour sub - feature w1 = 0.6, and the weight of the facial feature proportion sub - feature w2 = 0.4. Through the weighted average formula: target permission verification feature = w1F1+w2F2, these sub - features are fused together to obtain the target permission verification feature. Another implementation is feature transformation based on deep learning. A deep neural network such as an auto - encoder can be used. An auto - encoder consists of an encoder and a decoder. The remote verification system for terminal device access permissions inputs the refined permission verification feature into the encoder part of the auto - encoder. The encoder will compress and transform the feature, extract the most critical and core information, and convert it into a low - dimensional feature representation. This low - dimensional feature representation is the refined target permission verification feature. For example, when processing the refined permission verification feature of facial features, the encoder of the auto - encoder will process the feature containing numerous facial detail information and convert it into a more compact and representative low - dimensional vector, which is the target permission verification feature. In fingerprint feature verification, the auto - encoder will also process the refined permission verification feature and convert the complex fingerprint features into a concise and critical low - dimensional feature representation as the target permission verification feature. In this way, the remote verification system for terminal device access permissions can determine the target permission verification feature based on the refined permission verification feature using a suitable implementation method, providing more accurate and reliable core data support for subsequent terminal device access permission verification based on this feature, thereby effectively improving the accuracy and authority of the verification result.

[0050] Step S700: Identify based on the target permission verification feature to obtain the verification result of the terminal device access permission.

[0051] The target permission verification feature is a highly accurate and representative feature set obtained after multiple previous steps starting from the terminal device feature acquisition image, through a series of operations such as biometric feature extraction, feature decomposition, error adjustment, and feature refinement. This feature set contains key information that can accurately reflect the identity or permissions of the terminal device user.

[0052] Specifically, the remote verification system for terminal device access rights can adopt various implementation methods to identify based on target permission verification features. A feasible method is the matching algorithm based on distance metric. The remote verification system for terminal device access rights compares the target permission verification features with the standard permission verification feature templates pre-stored in the database. For example, in the face recognition scenario, the standard permission verification feature template is a set of facial features of legal users pre-collected and strictly verified. The remote verification system for terminal device access rights calculates the distance between the target permission verification features and each standard template, such as the Euclidean distance formula.

[0053] Suppose the target permission verification feature is the facial feature vector X after current collection and processing, and there are multiple standard facial feature template vectors in the database . The remote verification system for terminal device access rights calculates the Euclidean distance between X and each in turn . If the calculated distance is less than the pre-set threshold, it is considered that the target permission verification feature matches the standard template successfully, that is, the verification passes; otherwise, if all the calculated distances are greater than the threshold, the verification fails. For example, the set threshold is T. When , the remote verification system for terminal device access rights determines that the facial features of the current terminal device user match the k-th standard template and grants access rights; if for all j, , access is refused.

[0054] In the fingerprint recognition scenario, the distance metric algorithm is also used. The target permission verification feature is the fingerprint feature vector after processing, and the standard template is the pre-stored legal fingerprint feature vector. The remote verification system for terminal device access rights judges whether they match by calculating the distance between the two, such as using the Manhattan distance formula: , to determine whether they match. If the distance is within the threshold range, the fingerprint verification passes and access to the terminal device is allowed; otherwise, the verification fails and access is prohibited.

[0055] Another implementation method is the method based on machine learning classifier. The remote verification system for terminal device access rights uses a pre-trained classification model, such as a support vector machine (SVM) classifier. In the training stage, a large number of known permission verification feature data (including legal and illegal samples) are used to train the classifier, so that the classifier learns the patterns and rules of features in different permission states. In the recognition stage, the remote verification system for terminal device access rights inputs the target permission verification features into the trained SVM classifier. The SVM classifier judges its category by calculating the relationship between the target features and the classification hyperplane. For the linearly separable case, the equation of the classification hyperplane is wT w·x + b = 0, where w is the normal vector of the hyperplane, x is the feature vector, and b is the bias term. The classifier determines whether the feature x belongs to the legal permission category by checking which side of the hyperplane it lies on. If it is determined to belong to the legal permission category, the verification passes and the terminal device is granted access permission; if it is determined to belong to the illegal category, the verification fails and access is denied.

[0056] As an implementation, in step S300, based on the physical sign acquisition image of the terminal device, feature extraction error adjustment is performed on each of the initial verification sub-features to obtain corresponding verification adjustment sub-features, including:

[0057] Step S310: Embed the physical sign acquisition image of the terminal device and the initial verification sub-feature to be processed into a feature extraction error adjustment indication framework to obtain an image feature to be adjusted, where the image feature to be adjusted is used to indicate that the image processing network performs feature extraction error adjustment on the initial verification sub-feature to be processed, and the initial verification sub-feature to be processed is any one of the multiple initial verification sub-features;

[0058] Step S320: Load the image feature to be adjusted into the image processing network to obtain an analysis result of the image feature to be adjusted output by the image processing network. The analysis result of the image feature to be adjusted includes the verification adjustment sub-feature corresponding to the initial verification sub-feature to be processed and the extraction error category, and the extraction error category identifies the type of the corresponding extraction feature error.

[0059] Step S310 requires the remote verification system for the terminal device access permission to embed the physical sign acquisition image of the terminal device and the initial verification sub-feature to be processed into a feature extraction error adjustment indication framework to obtain an image feature to be adjusted. The physical sign acquisition image of the terminal device contains rich original information about the user's physical signs, and the initial verification sub-feature to be processed is an information unit initially extracted from these images for a specific physical sign part or feature dimension. The role of the feature extraction error adjustment indication framework is to integrate these two types of information and generate a new feature representation that can guide the subsequent image processing network for error adjustment.

[0060] For example, in the face physical sign verification scenario, the physical sign acquisition image of the terminal device is a clear face image, containing complete information about various facial organs such as eyes, nose, and mouth. The initial verification sub-feature to be processed is assumed to be the feature about the eye part, which may include preliminary extracted information such as the general shape and position of the eyes. The remote verification system for the terminal device access permission inputs this face image and the initial verification sub-feature of the eyes into the feature extraction error adjustment indication framework.

[0061] For example, the feature extraction error adjustment indication framework can be constructed based on the attention mechanism in deep learning. The attention mechanism enables the remote authentication system for terminal device access rights to focus on the key parts when processing complex information, thereby better integrating and analyzing the information. In this framework, the remote authentication system for terminal device access rights first performs feature encoding on the terminal device physical sign acquisition image and the initial verification sub-features to be processed. For the image, a convolutional neural network (CNN) can be used for feature extraction, converting the image into a high-dimensional feature vector representation. For example, through a series of convolutional layers and pooling layer operations, the original image is gradually converted into a feature map containing rich image semantic information. For the initial verification sub-features to be processed, corresponding encoding operations are also performed to convert them into a dimension and format that match the image feature vector. Suppose the feature vector I is obtained after the image is processed by the CNN, and the feature vector F is obtained after the initial verification sub-features to be processed are encoded. In the framework based on the attention mechanism, an attention weight matrix A is calculated, which represents the degree of association between the image features and the initial verification sub-features. A feasible method for calculating the attention weight matrix is through dot product operation and the softmax function. For example, first calculate , where is the i-th element of the image feature vector I, is the j-th element of the initial verification sub-feature vector F to be processed, and n is the dimension of the vector F. This formula calculates the similarity between each element of the image features and the initial verification sub-features and performs normalization processing to obtain the attention weight matrix A.

[0062] Then, according to the attention weight matrix A, the remote authentication system for terminal device access rights performs weighted fusion on the image feature vector I and the initial verification sub-feature vector F to be processed. For example, the image feature G to be adjusted can be calculated by the formula , where m is the dimension of the image feature vector I, represents a certain fusion operation, such as element-wise multiplication or concatenation operation. In this way, the image feature G to be adjusted contains both the overall information of the terminal device physical sign acquisition image and highlights the key information related to the initial verification sub-features to be processed, and can effectively indicate the subsequent image processing network to perform feature extraction error adjustment on the initial verification sub-features. In the fingerprint physical sign verification scenario, the terminal device physical sign acquisition image is a clear fingerprint image, and the initial verification sub-features to be processed are the preliminary extraction features regarding fingerprint breakpoints. The remote authentication system for terminal device access rights will also input the fingerprint image and the breakpoint initial verification sub-features into the feature extraction error adjustment indication framework. Through a similar method based on the attention mechanism, the fingerprint image is feature-encoded, the breakpoint initial verification sub-features are encoded, the attention weight matrix is calculated, and weighted fusion is performed to obtain the image feature to be adjusted that can indicate the fingerprint breakpoint feature extraction error adjustment.

[0063] Step S320 requires the remote authentication system for the terminal device access permission to load the image features to be adjusted into the image processing network to obtain the analysis result of the image features to be adjusted output by the network. This result includes the verification adjustment sub-features corresponding to the initial verification sub-features to be processed and the extraction error category. The image processing network is an intelligent model trained with a large amount of data. It can deeply analyze the input image features to be adjusted, identify the feature extraction errors therein, and generate the corresponding adjusted verification adjustment sub-features.

[0064] Continuing with the example of facial feature verification, the remote authentication system for the terminal device access permission inputs the image features to be adjusted obtained in step S310 into the image processing network. This image processing network can be a deep convolutional neural network, which uses a large amount of image data containing different facial features and corresponding accurate feature annotations during the training process. The network internally includes multiple convolutional layers, pooling layers, and fully connected layers.

[0065] When the image features to be adjusted are input into the network, the convolutional layer performs local feature extraction on the features, capturing various detail information in the image through different convolutional kernels. For example, some convolutional kernels may be specifically used to extract the contour features of the eyes, and others may be used to extract the texture features of the eyes. The pooling layer then downsamples the features extracted by the convolutional layer, reducing the data volume while retaining the key features. The fully connected layer integrates and classifies the features extracted by the previous layers. When processing the image features to be adjusted regarding the eyes, the image processing network analyzes the difference between the image features to be adjusted and the pre-learned standard eye features. If it is found that there is an error in the extraction of the eye shape, for example, the actual eye shape should be almond-shaped, but the shape extracted by the initial verification sub-feature is closer to circular, the network will adjust the eye shape feature according to the adjustment rules it has learned. At the same time, the network will also identify that this error belongs to the shape extraction error category. Finally, the network outputs the adjusted eye verification adjustment sub-features and the extraction error category.

[0066] For example, the training process of the image processing network can use the backpropagation algorithm. During the training phase, a large number of facial images with accurate labels and the corresponding initial verification sub-features are input into the network, and the network outputs the predicted verification adjustment sub-features and error categories. By calculating the loss function between the prediction result and the true label, such as the mean squared error loss function. Then, use the backpropagation algorithm to update the weight parameters of the network according to the gradient of the loss function, enabling the network to gradually learn accurate feature extraction and error adjustment capabilities.

[0067] The image processing network can also adopt a recurrent neural network (RNN) or its variant, such as a long short-term memory network (LSTM), to process the image features to be adjusted. For some physiological sign information with sequential features, such as the direction of fingerprint patterns, the RNN or LSTM can better capture the temporal or sequential relationships between features. For example, when processing the image features to be adjusted for the direction of fingerprint patterns, the LSTM can, through its internal memory units and gating mechanisms, remember the starting direction, turning information, etc. of the patterns, so as to more accurately analyze and adjust the errors in the extraction of pattern direction features, and output more accurate verification adjustment sub-features and accurate extraction error categories.

[0068] Through steps S310 - S320, the remote verification system for the access permission of the terminal device can utilize the feature extraction error adjustment indication framework and the image processing network, and for each initial verification sub-feature, in combination with the physiological sign acquisition image of the terminal device, effectively identify and adjust the errors generated in the feature extraction process, and obtain the corresponding verification adjustment sub-features and extraction error categories. This not only improves the accuracy of the verification sub-features, but also provides more detailed error information for the subsequent verification process, which helps to further improve the reliability and accuracy of the entire remote verification method for the access permission of the terminal device. Whether in different scenarios such as facial physiological sign verification or fingerprint physiological sign verification, this implementation method can, through precise error adjustment, provide high-quality data support for permission verification, and ensure the security of the remote operation and maintenance of the terminal by government and enterprise customers and the effectiveness of access control.

[0069] As an implementation method, step S500 of improving the features of the verification adjustment features according to the physiological sign acquisition image of the terminal device to obtain the improved features for permission verification includes:

[0070] Step S510: Embedding the physiological sign acquisition image of the terminal device and the verification adjustment features into the feature improvement indication framework to obtain the image features to be improved, where the image features to be improved are used to indicate that the image processing network improves the verification adjustment features;

[0071] Step S520: Loading the image features to be improved into the image processing network to obtain the analysis result of the image features to be improved output by the image processing network for the image features to be improved, and the analysis result of the image features to be improved includes the improved features for permission verification.

[0072] In step S510, the remote authentication system that requests access permission for the terminal device embeds the terminal device's physical sign acquisition image and the authentication adjustment feature into the feature improvement indication framework, thereby obtaining the image feature to be improved. The terminal device's physical sign acquisition image is an intuitive record of the physical signs of the terminal device user, containing rich details and overall features; the authentication adjustment feature is a relatively more accurate feature representation obtained after adjusting the initial authentication sub-feature through the previous steps, but there may still be some missing information or inaccuracies. The purpose of the feature improvement indication framework is to generate a new feature representation that can guide the subsequent image processing network to improve features.

[0073] Taking facial physical sign authentication as an example, the terminal device's physical sign acquisition image is a clear face image that covers the complete information of all parts of the face, such as skin texture, subtle changes in facial contours, etc. After the previous error adjustment, the authentication adjustment feature already contains relatively accurate key facial features, such as the shape and position of eyes, nose, mouth, etc., but may be insufficient in capturing the details of facial expressions. The remote authentication system for the terminal device's access permission inputs this face image and the existing facial authentication adjustment feature into the feature improvement indication framework.

[0074] For example, the feature improvement indication framework can be constructed based on a deep learning fusion mechanism. First, the remote authentication system for the terminal device's access permission performs feature encoding on the terminal device's physical sign acquisition image and the authentication adjustment feature. For the terminal device's physical sign acquisition image, a convolutional neural network (CNN) is used for feature extraction to convert the image into a high-dimensional feature vector representation. For example, through a series of convolutional layers and pooling layer operations, the original face image is gradually converted into a feature map containing rich facial semantic information. Suppose the image feature vector I is obtained after processing. For the authentication adjustment feature, corresponding encoding is also performed to make it match the image feature vector in dimension and format, obtaining the feature vector V.

[0075] Next, in the feature improvement indication framework, the image feature vector I and the authentication adjustment feature vector V are combined through a fusion operation. A feasible fusion method is weighted fusion based on the attention mechanism. The remote authentication system for the terminal device's access permission calculates an attention weight matrix A to determine the relative importance of the image feature and the authentication adjustment feature in the fusion process. One method for calculating the attention weight matrix is to use the dot product operation and the softmax function.

[0076] In step S520, the remote authentication system for the access permission of the terminal device loads the image features to be improved into the image processing network, so as to obtain the analysis result of the image features to be improved output by the network. This result includes the improved features for permission authentication. The image processing network is an intelligent model trained with a large amount of data, with powerful data analysis and feature generation capabilities. It can deeply analyze the input image features to be improved, mine the potential information therein, and generate more complete and accurate improved features for permission authentication.

[0077] Still taking the facial feature authentication as an example, the remote authentication system for the access permission of the terminal device inputs the image features to be improved obtained in step S510 into the image processing network. This image processing network can be a deep convolutional neural network, which uses a large amount of image data containing different facial features and corresponding accurate feature annotations during the training process. The network internally includes multiple convolutional layers, pooling layers, and fully connected layers.

[0078] After the image features to be improved are input into the network, the convolutional layer extracts local features of the features. Different convolutional kernels capture various detailed information in the image features to be improved. For example, some convolutional kernels may focus on extracting micro-expression features related to facial expressions, such as the fine wrinkles at the corners of the eyes, the slight upward or downward curvature of the mouth, etc.; some other convolutional kernels may be used to extract facial skin texture features, which may not have been fully reflected in the previous verified adjustment features. The pooling layer downsamples the features extracted by the convolutional layer, retaining the key features while reducing the data volume. The fully connected layer integrates and comprehensively analyzes the features extracted by the previous layers. When processing the facial image features to be improved, the image processing network deeply analyzes the input features according to the facial feature patterns and rules learned during the training process. If it is found that there are deficiencies in the facial expression details of the verified adjustment features, the network will generate corresponding facial expression detail features according to the image information contained in the image features to be improved, and fuse them with the original verified adjustment features. For example, the facial expression features in the original verified adjustment features are relatively blurred, and the image features to be improved, through the fusion with the body sign acquisition images of the terminal device, contain richer expression information. The image processing network analyzes this information, generates more accurate facial expression features, and integrates them with other existing facial features, and finally outputs the improved features for permission authentication containing more complete facial features.

[0079] For example, the training process of an image processing network can use the backpropagation algorithm. During the training phase, a large number of facial images with accurate labels and corresponding verification adjustment features are input into the network, and the network outputs the predicted permission verification improvement features. By calculating the loss function between the prediction result and the true label, such as the mean squared error loss function. Then, the backpropagation algorithm is used to update the weight parameters of the network according to the gradient of the loss function, so that the network can gradually learn the accurate feature improvement ability. The image processing network can also adopt the architecture of a generative adversarial network (GAN) to achieve feature improvement. In this architecture, there is a generator and a discriminator. The role of the generator is to generate permission verification improvement features based on the input features of the image to be improved, while the discriminator is responsible for judging whether the generated features are real and accurate. During the training process, the generator and the discriminator compete with each other and continuously optimize. The generator tries to generate results closer to the real permission verification improvement features to deceive the discriminator; the discriminator continuously improves its judgment ability to accurately distinguish between real features and generated features. Through this adversarial training, the generator can learn how to generate high-quality permission verification improvement features based on the features of the image to be improved.

[0080] Through steps S510 - S520, the remote verification system for the access permission of the terminal device effectively combines the information of the terminal device's physical sign acquisition image with the verification adjustment features by using the feature improvement indication framework and the image processing network, comprehensively improves the features of the verification adjustment features, and obtains the permission verification improvement features.

[0081] As an implementation, step S600 of determining the target permission verification features based on the permission verification improvement features includes:

[0082] Step S610: Decompose the permission verification improvement features to obtain multiple permission verification improvement sub-features;

[0083] Step S620: Adjust the feature extraction error of each permission verification improvement sub-feature according to the terminal device's physical sign acquisition image to obtain the corresponding permission verification improvement adjustment sub-features;

[0084] Step S630: Determine the target permission verification features based on multiple permission verification improvement adjustment sub-features.

[0085] Step S610 requires the remote verification system for the access permission of the terminal device to decompose the permission verification improvement features to obtain multiple permission verification improvement sub-features. The permission verification improvement features are a relatively complete and accurate feature set obtained after multiple previous steps of processing. However, in order to analyze and utilize these features more carefully, the remote verification system for the access permission of the terminal device needs to disassemble them into multiple more targeted sub-features.

[0086] Taking facial feature verification as an example, the permission verification perfecting feature may be a vector containing comprehensive information about the overall facial features, covering multiple aspects such as facial contour, facial feature characteristics, skin color, etc. The remote verification system for terminal device access permissions decomposes features based on different facial regions and feature dimensions. For the facial contour, features such as face shape (round, square, oval, etc.) and the curvature of facial lines are extracted to form the permission verification perfecting sub-features for the facial contour; for the facial features, detailed feature extraction is performed on the eyes, nose, mouth, etc. respectively. Taking the eyes as an example, the shape of the eyes (single eyelid, double eyelid, phoenix eye, etc.), the size of the eyes, and the distance between the eyes are obtained as the permission verification perfecting sub-features for the eye part; for the nose, sub-features such as the height of the nose bridge, the shape of the nose tip, and the width of the nasal wings are extracted; for the mouth, sub-features such as the thickness of the lips and the shape of the mouth corners are separated. In addition, features such as the hue and brightness of the skin color are also extracted as part of the permission verification perfecting sub-features. For example, the remote verification system for terminal device access permissions can use a deep learning-based feature extraction network for feature decomposition. For example, a convolutional neural network (CNN) model can be constructed, and the structure of the model is designed to be able to extract features for different feature regions and dimensions. For the permission verification perfecting feature corresponding to the facial image, through the design of different convolutional kernels, some convolutional kernels are specifically used to extract facial contour features, and other convolutional kernels focus on the extraction of facial feature characteristics. When these convolutional kernels perform convolutional operations on the permission verification perfecting feature, they will extract the corresponding feature information according to their weight parameters and separate it into different permission verification perfecting sub-features. For the permission verification perfecting feature of the fingerprint image, similarly, through carefully designed convolutional kernels, different types of features such as fingerprint patterns, breakpoints, and bifurcation points can be extracted respectively to achieve feature decomposition.

[0087] Step S620 requires the remote verification system for terminal device access permissions to adjust the feature extraction error for each permission verification perfecting sub-feature based on the terminal device physical sign acquisition image to obtain the corresponding permission verification perfecting adjusted sub-feature. Although the permission verification perfecting sub-feature has gone through the previous error adjustment and feature perfecting steps, due to the influence of various factors such as the acquisition environment and device accuracy, there may still be certain extraction errors. Therefore, the remote verification system for terminal device access permissions needs to finely adjust these sub-features again based on the terminal device physical sign acquisition image.

[0088] Continuing with the example of facial feature verification, for the sub-feature of eye shape permission verification improvement, assume that the eye shape extracted through the previous steps is "round", but in the actual terminal device feature collection image, it is found that the actual shape of the eyes is closer to "almond-shaped", with a certain shape extraction error. The remote verification system for terminal device access permission inputs the terminal device feature collection image and the sub-feature of eye shape permission verification improvement into a dedicated error adjustment model. This error adjustment model can be a neural network based on deep learning, which has learned a large number of accurate features of eye shapes and feasible error patterns in different facial images during the training process. The model will analyze the actual shape information of the eyes in the image, compare it with the current sub-feature of eye shape permission verification improvement, and adjust the sub-feature of eye shape permission verification improvement according to the learned adjustment rules, correcting it to "almond-shaped" which is closer to the actual situation, thus obtaining the adjusted sub-feature of eye shape permission verification improvement.

[0089] For example, the error adjustment model can adopt the Residual Network (ResNet) structure. By introducing skip connections, the Residual Network can effectively solve the problem of gradient vanishing in deep learning, enabling the model to learn more complex error adjustment patterns. During the training process, a large number of terminal device feature collection images with accurate labels and the corresponding sub-features of permission verification improvement are input into the Residual Network, and the network outputs the adjusted sub-feature of permission verification improvement. By calculating the loss function between the prediction result and the true label, such as the mean squared error loss function. Then, the backpropagation algorithm is used to update the weight parameters of the network according to the gradient of the loss function, enabling the network to learn accurate error adjustment capabilities.

[0090] Step S630 requires the remote verification system for terminal device access permission to determine the target permission verification feature based on multiple adjusted sub-features of permission verification improvement. This step is to integrate the multiple sub-features after error adjustment to form a comprehensive target permission verification feature that best represents the actual situation of the terminal device features.

[0091] In facial feature verification, the remote verification system for terminal device access rights fuses the refined sub-features for permission verification such as the adjusted facial contour, eyes, nose, mouth, skin color, etc. A feasible fusion method is weighted fusion. The remote verification system for terminal device access rights assigns corresponding weights to each sub-feature according to its importance for facial identity verification. For example, the facial contour feature has a high distinctiveness for overall facial recognition and may be assigned a weight of 0.3; due to its uniqueness and stability, the eye feature may have a weight of 0.3; the weights of the nose and mouth features are 0.15 respectively; and the weight of the skin color feature is 0.1. Through the weighted fusion formula: Target permission verification feature = Adjusted facial contour permission verification refined sub-feature × 0.3 + Eye permission verification refined sub-feature × 0.3 + Nose permission verification refined sub-feature × 0.15 + Mouth permission verification refined sub-feature × 0.15 + Skin color permission verification refined sub-feature × 0.1, each sub-feature is integrated to obtain the final facial target permission verification feature.

[0092] In addition to weighted fusion, the remote verification system for terminal device access rights can also adopt the feature concatenation and convolution fusion technology in deep learning. First, multiple refined sub-features for permission verification are concatenated in the feature dimension. For example, the refined sub-features for permission verification of each part of the face are connected together in a certain order to form a longer feature vector. Then, the concatenated feature vector is processed by a convolutional neural network. The convolutional layer in the convolutional neural network will use a specific convolutional kernel to perform a convolutional operation on this long feature vector to extract higher-level and more abstract feature information. After multiple convolutional operations, the remote verification system for terminal device access rights finally obtains a target permission verification feature that has been fused and refined.

[0093] As an implementation manner, step S100, performing a biometric extraction operation on the terminal device feature collection image to obtain an initial permission verification feature, includes:

[0094] Step S110: Performing a biometric extraction operation on the terminal device feature collection image through a pre-calibrated and converged permission verification neural network to obtain an initial permission verification feature.

[0095] When the remote authentication system for terminal device access rights receives the physical sign acquisition image of the terminal device, it inputs it into the pre-calibrated and converged rights authentication neural network, which can be, for example, a convolutional neural network or any other feasible neural network. Taking the facial image as an example, first, the remote authentication system for terminal device access rights preprocesses the physical sign acquisition image of the terminal device. This may include the normalization operation of the image, that is, mapping the pixel values of the image to a specific range, such as [0, 1], to eliminate the differences caused by factors such as brightness and contrast between different images. Then, the normalized image is input into the rights authentication neural network. Each neuron and layer inside the network will gradually analyze and extract features from the image. For example, in the first convolutional layer of the network, multiple different convolutional kernels are used to scan the image. These convolutional kernels can be regarded as small filters, and each convolutional kernel has its own specific weight parameters. When the convolutional kernel slides on the image, it performs a dot product operation with the local area of the image, thereby extracting low-level features such as edges and lines in the image. For example, a vertical convolutional kernel can detect the vertical edge information in the image.

[0096] As the network depth increases, subsequent layers will further extract higher-level and more abstract features based on the low-level features extracted by the previous layers. For example, after multiple convolutional and pooling operations, the network can extract the feature information of facial organs such as eyes, nose, and mouth. These feature information will be gradually integrated and processed, and finally form the initial rights authentication features. For example, the initial rights authentication features may be a set containing the feature vectors of each key part of the face, and these feature vectors can accurately represent the biometric information of the face.

[0097] As an implementation manner, before the step S110 of performing biometric feature extraction operation on the physical sign acquisition image of the terminal device through the pre-calibrated and converged rights authentication neural network to obtain the initial rights authentication features, the method further includes the training process of the network:

[0098] Step S101: Obtain the initial rights authentication neural network, where the initial rights authentication neural network is obtained by calibrating the pre-debugged rights authentication neural network according to the physical sign acquisition image samples and the first-order prior rights authentication features corresponding to the physical sign acquisition image samples. The physical sign acquisition image samples correspond to second-order prior rights authentication features and first-order prior rights authentication features. The second-order prior rights authentication features are the image features with the first accuracy corresponding to the physical sign acquisition image samples, and the first-order prior rights authentication features are the image features with the second accuracy corresponding to the physical sign acquisition image samples.

[0099] In step S101, the physical sign acquisition image sample corresponds to a second-order prior permission verification feature and a first-order prior permission verification feature. The second-order prior permission verification feature is the image feature of the first precision corresponding to the physical sign acquisition image sample, and the first-order prior permission verification feature is the image feature of the second precision corresponding to the physical sign acquisition image sample. Among them, the first precision is weak precision, and the second precision is strong precision. The specific division threshold of the precision can be set according to the actual situation.

[0100] Taking facial physical sign acquisition as an example, the remote verification system for the access permission of the terminal device collects a large number of facial images of different individuals as physical sign acquisition image samples. These samples cover various different facial features, such as different face shapes (round, square, oval), facial feature proportions (big eyes, small eyes, high nose bridge, flat nose bridge, etc.), and skin color differences, etc. For each facial image sample, through a combination of specific implementation methods and manual annotation, its second-order prior permission verification feature and first-order prior permission verification feature are determined. The second-order prior permission verification feature may be a relatively broad and basic facial feature description. For example, the general contour range of the face in the image, the general shape category of the eyes (such as a rough judgment of single eyelid or double eyelid). The precision of such features is relatively low and can only provide some basic image information pointers. The first-order prior permission verification feature is more accurate and detailed. For example, the precise facial contour coordinate data, the specific shape parameters of the eyes (such as the angle of the eye corner, the length of the eye fissure, etc.). These features can more accurately depict the real features of the face.

[0101] When obtaining the initial permission verification neural network, the remote verification system for the access permission of the terminal device uses the pre-debugged permission verification neural network. This pre-debugged neural network has undergone preliminary design and some basic training and has a certain feature extraction ability, but it still needs to be further optimized to adapt to specific physical sign acquisition image samples and verification requirements. The remote verification system for the access permission of the terminal device inputs the physical sign acquisition image sample and its corresponding first-order prior permission verification feature into the pre-debugged permission verification neural network for calibration. During the calibration process, by adjusting the weight parameters and bias parameters of the network, the network can better extract the feature information that matches the first-order prior permission verification feature from the physical sign acquisition image sample. For example, for facial image samples, the network gradually learns how to accurately identify key features such as the precise shape of the eyes and the specific height of the nose from the image during the calibration process. By continuously adjusting the weights and biases, the features output by the network are made closer to the first-order prior permission verification feature. For example, the remote verification system for the access permission of the terminal device can use the stochastic gradient descent (SGD) algorithm to adjust the weight parameters and bias parameters of the network.

[0102] Step S102: Based on the network weight parameters and bias parameters of the initial permission verification neural network, generate a quasi-tuning extraction network, where the network composition architecture of the quasi-tuning extraction network is the same as that of the pre-debugged permission verification neural network, and the network weight parameters and bias parameters of the quasi-tuning extraction network are the same as those of the initial permission verification neural network;

[0103] Step S103: Load the physical sign acquisition image sample into the quasi-tuning extraction network to obtain the first predicted feature output by the quasi-tuning extraction network; and determine the first training error value based on the first predicted feature and the first-order prior permission verification feature;

[0104] Step S104: Load the physical sign acquisition image sample into the initial permission verification neural network to obtain the second predicted feature output by the initial permission verification neural network; and determine the second training error value based on the second predicted feature, the first predicted feature, the second-order prior permission verification feature, and the first-order prior permission verification feature;

[0105] Step S105: Determine the comprehensive training error value based on the first training error value and the second training error value;

[0106] Step S106: Iterate the network weight parameters and bias parameters of the quasi-tuning extraction network according to the comprehensive training error value until the preset iteration stop condition is met, and obtain a pre-tuned and converged extraction neural network, where the pre-tuned and converged extraction neural network is the pre-tuned and converged permission verification neural network.

[0107] Step S102 requires the remote verification system for terminal device access permissions to generate a quasi-tuning extraction network based on the network weight parameters and bias parameters of the initial permission verification neural network. The network composition architecture of the quasi-tuning extraction network is the same as that of the pre-debugged permission verification neural network, and its network weight parameters and bias parameters are also the same as those of the initial permission verification neural network. The purpose of this step is to provide a copy network for subsequent network training and optimization, so as to perform various adjustments and tests without affecting the original initial permission verification neural network.

[0108] Taking the neural network for face recognition as an example, the initial permission verification neural network already has a certain feature extraction ability after being calibrated in step S101. Suppose this neural network has multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract local features in the face image, such as edges and textures in different directions; the pooling layer downsamples the features extracted by the convolutional layer to reduce the data volume and retain the main features; the fully connected layer integrates and classifies the features extracted by the previous layers. The remote verification system for terminal device access rights precisely replicates the structure of this neural network, including the number of neurons in each layer, connection methods, etc. At the same time, the weight parameters (i.e., the weight values connected to each neuron) and bias parameters (the bias values of each neuron) of the initial permission verification neural network are copied unchanged into the proposed calibrated extraction network. In this way, the proposed calibrated extraction network has exactly the same feature extraction ability and behavioral performance as the initial permission verification neural network in the initial state.

[0109] Step S103 requires the remote verification system for terminal device access rights to load the physical sign acquisition image sample into the proposed calibrated extraction network to obtain the first predicted feature output by the proposed calibrated extraction network; and determine the first training error value based on the first predicted feature and the first-order prior permission verification feature. This step is to conduct a preliminary test on the proposed calibrated extraction network to evaluate the gap between its feature extraction ability for the physical sign acquisition image sample and the real features in the current state.

[0110] Continuing with the face recognition example, the remote verification system for terminal device access rights sequentially inputs a large number of face physical sign acquisition image samples into the proposed calibrated extraction network. The network processes the images according to its internal structure and parameters, extracts various feature information from the input face images, and finally outputs the first predicted feature. For example, the first predicted feature may be a set of feature vectors containing the features of each part of the face, and these vectors attempt to describe information such as the shape of the face and the proportion of facial features. Then, the remote verification system for terminal device access rights compares these first predicted features with the corresponding first-order prior permission verification features. The first-order prior permission verification features are accurately labeled information representing the real features of the face. By calculating the difference between the two, the remote verification system for terminal device access rights can determine the first training error value.

[0111] Step S104 requires the remote verification system for terminal device access rights to load the physical sign acquisition image sample into the initial permission verification neural network to obtain the second predicted feature output by the initial permission verification neural network; and determine the second training error value based on the second predicted feature, the first predicted feature, the second-order prior permission verification feature, and the first-order prior permission verification feature. This step more comprehensively evaluates the performance and error situation of the network by performing feature extraction on the same sample on the initial permission verification neural network and combining more hierarchical prior feature information.

[0112] Taking facial recognition as an example again, the remote verification system for terminal device access rights inputs the same facial feature collection image sample into the initial permission verification neural network. After processing, this network outputs a second predicted feature, which is also a set of information collections describing facial features. At this time, the remote verification system for terminal device access rights comprehensively considers multiple factors to determine the second training error value. The second-order prior permission verification feature, as a relatively broad and basic description of facial features, together with the first-order prior permission verification feature (exact feature), is compared and analyzed with the second predicted feature and the first predicted feature.

[0113] The process of calculating the second training error value is relatively complex, and it needs to consider the differential relationships between multiple features. A possible calculation method is to comprehensively consider the weighted differences between different features. For example, for different parts of facial features (such as eyes, nose, mouth, etc.), different weights are assigned according to their importance for identity recognition. Suppose the weight of the eye feature is , the weight of the nose feature is , and the weight of the mouth feature is . For each sample, the error of the eye part is calculated respectively, where is the eye feature in the first-order prior permission verification feature, is the eye feature in the first predicted feature, is the eye feature in the second predicted feature; the error of the nose part ; the error of the mouth part . Then, the second training error value . In this way, the prior features at different levels and the predicted features of the two networks are comprehensively considered, more comprehensively reflecting the error situation of the initial permission verification neural network in feature extraction.

[0114] Step S105 requires the remote verification system for terminal device access rights to determine the comprehensive training error value based on the first training error value and the second training error value. This step integrates the error values obtained in the previous two steps to obtain a single error index that can comprehensively reflect the current training state and performance of the network, providing a unified basis for subsequent parameter iteration. Multiple methods can be used to determine the comprehensive training error value. One method is to perform a weighted sum of the first training error value and the second training error value.

[0115] In step S106, the remote verification system for the access permission of the terminal device iteratively approximates and tunes the network weight parameters and bias parameters of the extraction network based on the comprehensive training error value until the preset iteration stop condition is met, obtaining a pre-tuned and converged extraction neural network, which is the pre-tuned and converged permission verification neural network. This step is the core part of network training. By continuously adjusting the network parameters, the feature extraction ability of the network is continuously optimized until a stable and accurate state is finally achieved.

[0116] The remote verification system for the access permission of the terminal device uses an optimization algorithm to iteratively approximate and tune the parameters of the extraction network. One optimization algorithm is Stochastic Gradient Descent (SGD) and its variants, such as Adagrad, Adadelta, Adam, etc. Taking the Stochastic Gradient Descent algorithm as an example, in each iteration, the remote verification system for the access permission of the terminal device calculates the gradient according to the comprehensive training error value.

[0117] During the iteration process, the remote verification system for the access permission of the terminal device continuously repeats the process of calculating the comprehensive training error value, gradient, and updating the parameters. After each update, it checks whether the preset iteration stop condition is met. The preset iteration stop condition can be in various forms. For example, the comprehensive training error value is less than a specific threshold, such as 0.01. When the comprehensive training error value drops below this threshold, the remote verification system for the access permission of the terminal device considers that the network has reached a sufficiently accurate state and stops the iteration. Or, an upper limit on the number of iterations can also be set, such as a maximum of 1000 iterations. When this iteration number limit is reached, regardless of whether the comprehensive training error value reaches the threshold, the iteration stops.

[0118] Through a series of operations in steps S102 - S106, the remote verification system for the access permission of the terminal device comprehensively optimizes and tunes the initial permission verification neural network. From generating the approximate and tunable extraction network, to separately evaluating the errors of the two networks, to determining the comprehensive training error value and iteratively optimizing the parameters, finally obtaining the pre-tuned and converged permission verification neural network.

[0119] As an implementation, before obtaining the initial permission verification neural network in step S101, the method further includes:

[0120] Step S10: Perform a biometric extraction operation on the target sign acquisition image sample to obtain an initial sample permission verification feature; wherein, the target sign acquisition image sample is any one of the multiple sign acquisition image samples;

[0121] Step S20: Perform a feature adjustment operation on the initial sample permission verification feature to obtain a debug transition image feature;

[0122] Step S30: Determine the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample according to the debugging transition image feature; the second-order prior permission verification feature corresponding to the target physical sign acquisition image sample is the initial sample permission verification feature or the debugging transition image feature.

[0123] Step S10 requires the remote verification system for terminal device access permission to perform biometric extraction operations on the target physical sign acquisition image sample to obtain the initial sample permission verification feature. The target physical sign acquisition image sample is a representative sample selected from numerous physical sign acquisition images for subsequent analysis and processing. The remote verification system for terminal device access permission uses specialized biometric extraction techniques to process these samples.

[0124] Taking facial physical sign acquisition as an example, the target physical sign acquisition image sample obtained by the remote verification system for terminal device access permission may be a clear face photo. To extract the biometric features therein, the remote verification system for terminal device access permission can adopt the convolutional neural network (CNN) technology based on deep learning. First, preprocess the image, such as normalization processing, mapping the pixel values of the image to a specific range, such as [0, 1], to eliminate differences in brightness, contrast, etc. among different images.

[0125] Next, input the normalized image into the CNN. The CNN contains multiple convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, different convolutional kernels scan the image to extract local features. For example, some convolutional kernels are specifically used to extract the edge features of the face. By performing dot product operations with local regions of the image, the contours of parts such as eyes, nose, and mouth can be detected. After being processed by multiple convolutional layers, the image is gradually converted into a feature map containing rich semantic information. The pooling layer downsamples the feature map, reducing the data volume while retaining the key features. Finally, the fully connected layer integrates these features and outputs the initial sample permission verification feature. These features may be a vector containing information about the key parts of the face, such as the shape of the eyes and the position of the nose.

[0126] Step S20 requires the remote verification system for terminal device access permission to perform feature adjustment operations on the initial sample permission verification feature to obtain the debugging transition image feature. This step aims to optimize and improve the initially extracted features so that they can more accurately reflect the true situation of the target physical signs.

[0127] Still taking facial signs as an example, the remote verification system for terminal device access rights further processes the initial sample permission verification feature containing facial key part information. First, the initial sample permission verification feature is decomposed into multiple more targeted initial sample verification sub-features. For example, for the eye feature, it can be decomposed into initial sample verification sub-features such as the shape, size, and color of the eyes; for the nose, it can be decomposed into sub-features such as the height of the nasal bridge and the shape of the nasal tip.

[0128] Then, according to the target sign to collect image samples, feature extraction error adjustment is performed on each initial sample verification sub-feature. The remote verification system for terminal device access rights can use a dedicated error adjustment model, which learns a large amount of facial image data to understand the possible error patterns of different facial features during the extraction process. For example, if the extraction of the eye shape in the initial sample verification sub-feature is inaccurate, the error adjustment model will adjust the initial sample verification sub-feature of the eye shape according to the actual shape information of the eyes in the target sign collection image sample and the learned adjustment rules to obtain a more accurate sample verification adjustment sub-feature.

[0129] Based on multiple sample verification adjustment sub-features, the remote verification system for terminal device access rights determines the sample verification adjustment feature. For example, the sample verification adjustment sub-features of the adjusted eyes, nose, mouth, etc. are integrated to form a comprehensive sample verification adjustment feature. Then, according to the target sign to collect image samples, feature improvement is performed on the sample verification adjustment feature. The remote verification system for terminal device access rights can utilize generative adversarial network (GAN) technology. GAN consists of a generator and a discriminator. The generator generates more perfect features based on the information of the sample verification adjustment feature and the target sign collection image sample, and the discriminator determines whether the generated features are real and accurate. Through the adversarial training of the two, the features generated by the generator continuously approach the real facial features, thereby obtaining the sample permission verification perfect feature. The sample verification adjustment feature and the sample permission verification perfect feature together constitute the debugging transition image feature.

[0130] Step S30 requires the remote verification system for terminal device access rights to determine the first-order prior permission verification feature corresponding to the target sign collection image sample based on the debugging transition image feature. At the same time, the second-order prior permission verification feature corresponding to the target sign collection image sample is any one of the initial sample permission verification feature or the debugging transition image feature.

[0131] Continuing with the example of facial signs, the remote verification system for terminal device access rights further analyzes the improved features of sample rights verification. First, the improved features of sample rights verification are decomposed into multiple sub-features of sample rights verification. For example, the improved features of sample rights verification for the face are decomposed into more detailed sub-elements of features such as eyes, nose, mouth, etc.

[0132] Then, according to the target sign, image samples are collected, and the feature extraction error adjustment is performed on each sub-feature of sample rights verification improvement. The remote verification system for terminal device access rights uses the error adjustment model again, combines the detailed information of the image samples collected by the target sign, and precisely adjusts each sub-feature to obtain the corresponding adjusted sub-feature of sample rights verification improvement.

[0133] Finally, based on multiple adjusted sub-features of sample rights verification improvement, the first-order prior rights verification feature corresponding to the image sample of the target sign is determined. For example, the adjusted sub-features of sample rights verification improvement for parts such as eyes, nose, and mouth are comprehensively considered and integrated to form a first-order prior rights verification feature that can accurately represent the true features of the face. The second-order prior rights verification feature can be the initial sample rights verification feature, which is relatively broad and basic and can provide a general feature range; or it can be the debug transition image feature, which has been optimized and improved to better reflect the actual situation of facial features.

[0134] Through steps S10 - S30, starting from the image sample of the target sign, after a series of operations such as biometric feature extraction, feature adjustment, and determination of prior rights verification features, accurate and reliable prior knowledge is provided for the subsequent acquisition of the initial rights verification neural network. These prior rights verification features, as an important basis for network training, can help the initial rights verification neural network learn faster and more accurately how to extract effective biometric features from the image of the terminal device sign, thereby improving the accuracy and stability of the entire remote verification method for terminal device access rights, ensuring that only legitimate users can obtain access rights to the terminal device, and effectively protecting the security of the remote operation and maintenance terminals of government and enterprise customers.

[0135] As an implementation manner, the debug transition image feature includes the sample verification adjustment feature and the improved feature of sample rights verification; based on this, in step S20, the feature adjustment operation is performed on the initial sample rights verification feature to obtain the debug transition image feature, including:

[0136] Step S21: Decompose the initial sample rights verification feature to obtain multiple initial sample verification sub-features;

[0137] Step S22: Collect image samples according to the target physical signs, and perform feature extraction error adjustment on each of the initial sample verification sub-features to obtain corresponding sample verification adjusted sub-features;

[0138] Step S23: Determine the sample verification adjusted feature according to multiple sample verification adjusted sub-features;

[0139] Step S24: Collect image samples according to the target physical signs, and perform feature improvement on the sample verification adjusted feature to obtain a sample permission verification improved feature.

[0140] Step S21 requires the remote verification system for terminal device access permission to perform feature decomposition on the initial sample permission verification feature to obtain multiple initial sample verification sub-features. The initial sample permission verification feature is extracted by the remote verification system for terminal device access permission from the image samples of target physical signs in Step S10. It is a comprehensive feature set that contains multi-faceted information of the target physical signs. In order to process and optimize these features more meticulously, the remote verification system for terminal device access permission needs to decompose it into multiple initial sample verification sub-features with more pertinence.

[0141] Taking facial physical signs as an example, assume that the initial sample permission verification feature is a vector containing overall facial feature information, which may comprehensively include information such as facial contour, facial feature characteristics, and skin color. The remote verification system for terminal device access permission performs feature decomposition based on different parts and feature dimensions of the face. For the facial contour, features such as face shape (e.g., round, square, oval, etc.) and the curvature of facial lines will be extracted to form the initial sample verification sub-features regarding the facial contour. For example, through specific algorithms and models, analyze the pixel distribution of the contour line in the facial image, calculate the curvature change of the contour, and thus determine the specific type of face shape. This process can utilize edge detection algorithms, such as the Canny edge detection algorithm, to first extract the edge information of the face, and then determine the initial sample verification sub-features related to the face shape through the fitting and analysis of the edge curve.

[0142] For example, a remote verification system for terminal device access permissions can utilize a feature extraction network based on deep learning for feature decomposition. For example, a convolutional neural network (CNN) model can be constructed, and its structure is designed to be able to extract features for different feature regions and dimensions. For the initial sample permission verification features corresponding to facial images, through the design of different convolutional kernels, some convolutional kernels are specifically used to extract facial contour features, while others focus on the extraction of facial feature features. When these convolutional kernels perform convolutional operations on the initial sample permission verification features, they will extract corresponding feature information according to their weight parameters and separate it into different initial sample verification sub-features. For the initial sample permission verification features of fingerprint images, different types of features such as fingerprint lines, breakpoints, and bifurcation points can also be extracted through carefully designed convolutional kernels to achieve feature decomposition.

[0143] Step S22 requires the remote verification system for terminal device access permissions to adjust the feature extraction error of each initial sample verification sub-feature based on the target physical sign acquisition image sample to obtain the corresponding sample verification adjustment sub-feature. Although biometric feature extraction has been performed in step S10, due to factors such as the image acquisition environment, device accuracy, and limitations of the extraction algorithm, there may be certain extraction errors in the initial sample verification sub-features. The remote verification system for terminal device access permissions needs to correct and optimize these sub-features based on the complete information contained in the target physical sign acquisition image sample.

[0144] Still taking facial physical sign verification as an example, for the initial sample verification sub-feature of the eye shape, assuming that the eye shape extracted through the previous steps is "round", but in the actual target physical sign acquisition image sample, it is found that the real shape of the eye is closer to "almond-shaped", there is a certain shape extraction error. The remote verification system for terminal device access permissions inputs the target physical sign acquisition image sample and the initial sample verification sub-feature of this eye shape into a dedicated error adjustment model. This error adjustment model can be a neural network based on deep learning, which has learned a large number of accurate features of eye shapes and feasible error patterns in different facial images during the training process.

[0145] The model analyzes the actual shape information of the eyes in the target physical sign acquisition image sample and compares it with the current initial sample verification sub-feature. By calculating the difference between the two, for example, the Euclidean distance formula can be used to measure the distance between two shape feature vectors. According to this distance and the learned adjustment rules, the initial sample verification sub-feature of the eye shape is adjusted and corrected to the "almond shape" closer to the real situation, so as to obtain the adjusted eye shape sample verification adjustment sub-feature. For example, the error adjustment model can adopt the Residual Network (ResNet) structure. By introducing skip connections, the Residual Network can effectively solve the problem of gradient disappearance in deep learning, enabling the model to learn more complex error adjustment patterns. During the training process, a large number of target physical sign acquisition image samples with accurate labels and the corresponding initial sample verification sub-features are input into the Residual Network, and the network outputs the adjusted sample verification adjustment sub-features. By calculating the loss function between the prediction result and the real label, such as the mean squared error loss function. Then, the backpropagation algorithm is used to update the weight parameters of the network according to the gradient of the loss function, enabling the network to learn accurate error adjustment capabilities.

[0146] In step S23, the remote authentication system for the access permission of the terminal device verifies the adjusted sub-features based on multiple samples to determine the sample verification adjustment feature. This step integrates multiple sub-features after error adjustment to form a comprehensive sample verification adjustment feature that can more accurately reflect a certain aspect of the target physical sign. In facial physical sign verification, the remote authentication system for the access permission of the terminal device fuses multiple sample verification adjustment sub-features such as the adjusted eyes, nose, mouth, and facial contour. A feasible fusion method is weighted fusion. The remote authentication system for the access permission of the terminal device assigns corresponding weights to each sub-feature according to its importance in describing the facial features. For example, the facial contour feature plays an important role in describing the overall facial structure and may be assigned a weight of 0.3; due to its uniqueness and key role in facial recognition, the eye feature may have a weight of 0.3; the weights of the nose and mouth features are 0.15 respectively; and the skin color feature has a weight of 0.1. Through the weighted fusion formula: sample verification adjustment feature = facial contour sample verification adjustment sub-feature × 0.3 + eye sample verification adjustment sub-feature × 0.3 + nose sample verification adjustment sub-feature × 0.15 + mouth sample verification adjustment sub-feature × 0.15 + skin color sample verification adjustment sub-feature × 0.1, each sub-feature is integrated to obtain the final facial sample verification adjustment feature. This sample verification adjustment feature combines the adjusted feature information of multiple facial parts and more comprehensively and accurately describes the facial feature situation. In addition to weighted fusion, the remote authentication system for the access permission of the terminal device can also adopt feature splicing and convolutional fusion techniques in deep learning. First, multiple sample verification adjustment sub-features are spliced in the feature dimension. For example, the sample verification adjustment sub-features of each facial part are connected in a certain order to form a longer feature vector. Then, the spliced feature vector is processed by a convolutional neural network. The convolutional layer in the convolutional neural network uses a specific convolutional kernel to perform a convolutional operation on this long feature vector to extract more advanced and abstract feature information. After multiple convolutional operations, the remote authentication system for the access permission of the terminal device finally obtains a sample verification adjustment feature that has been fused and refined.

[0147] In step S24, the remote authentication system for the access permission of the terminal device acquires an image sample based on the target physical sign and refines the sample verification adjustment feature to obtain a sample permission verification refinement feature. Although the sample verification adjustment feature has undergone error adjustment and preliminary fusion, in order to make it more comprehensively and accurately reflect the true features of the target physical sign, the remote authentication system for the access permission of the terminal device still needs to further refine it based on the image sample acquired for the target physical sign.

[0148] The remote verification system for terminal device access permissions can utilize Generative Adversarial Network (GAN) technology to improve the features of sample verification adjustment features. A generative adversarial network consists of a generator and a discriminator. The role of the generator is to generate more perfect features based on the sample verification adjustment features and the information of the target physical sign acquisition image samples. The discriminator is responsible for judging whether the generated features are real and accurate.

[0149] In facial physical sign verification, the generator will generate more realistic facial features based on the sample verification adjustment features of the face and other information in the target physical sign acquisition image samples, such as facial expressions, lighting and other factors. For example, the generator can generate facial features with corresponding expression details based on the basic information of the facial contour and facial features in the sample verification adjustment features, combined with the facial expression information in the target physical sign acquisition image samples. The discriminator will judge the generated features to see if they match the real facial features. If the discriminator judges that the generated features are not real, the generator will adjust according to the feedback of the discriminator and continuously improve the generated features. Through the repeated adversarial training of the generator and the discriminator, the generator can finally generate more accurate and perfect facial sample permission verification improvement features.

[0150] In addition, the remote verification system for terminal device access permissions can also utilize reinforcement learning algorithms to improve the features of sample verification adjustment features. Reinforcement learning algorithms interact with the environment through an agent and optimize their own behavior according to the reward signals feedback by the environment. In the process of feature improvement, the agent can be regarded as a module that operates on the sample verification adjustment features, and the environment is the target physical sign acquisition image samples. The agent generates a tentative feature based on the sample verification adjustment features, and then obtains a reward signal according to the feedback provided by the target physical sign acquisition image samples (such as the similarity between the generated features and the real features). The agent adjusts its operation strategy according to the reward signal and continuously generates sample permission verification improvement features that are closer to the real features.

[0151] As an implementation manner, step S30, determining the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample according to the debug transition image feature, includes:

[0152] Step S31: Decompose the sample permission verification improvement feature to obtain multiple sample permission verification improvement sub-features;

[0153] Step S32: According to the target physical sign acquisition image sample, perform feature extraction error adjustment on each sample permission verification improvement sub-feature to obtain the corresponding sample permission verification improvement adjustment sub-feature;

[0154] Step S33: Based on the multiple sample permission verification and improvement of the regulatory sub-features, determine the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample; the second-order prior permission verification feature corresponding to the target physical sign acquisition image sample is any one of the following features: the initial sample permission verification feature, the sample verification adjustment feature, and the sample permission verification improvement feature.

[0155] Step S31 requires the remote verification system with terminal device access permission to perform feature decomposition on the sample permission verification improvement feature to obtain multiple sample permission verification improvement sub-features. The sample permission verification improvement feature is a comprehensive feature set obtained after feature adjustment and improvement in the previous steps, which comprehensively reflects the relevant information of the target physical sign. However, in order to analyze and utilize these features more deeply, the remote verification system with terminal device access permission disassembles it into multiple more targeted and refined sub-features.

[0156] For example, the remote verification system with terminal device access permission can use a deep convolutional neural network (DCNN) for feature decomposition. DCNN has powerful feature extraction capabilities. By designing different convolutional kernels and network hierarchical structures, it can accurately extract features for different feature regions and dimensions. For example, for the sample permission verification improvement feature corresponding to a facial image, the remote verification system with terminal device access permission can construct a DCNN with a multi-branch structure. Different branches are responsible for extracting features of different parts of the face, and each branch contains multiple convolutional layers and pooling layers. By carefully designing the parameters and sizes of the convolutional kernels, some convolutional kernels are specifically used to extract the detailed features of the facial contour, while others focus on the microscopic feature extraction of the facial features. When these convolutional kernels perform convolutional operations on the sample permission verification improvement feature, they will extract the corresponding feature information according to their weight parameters and separate it into different sample permission verification improvement sub-features. For the sample permission verification improvement feature of a fingerprint image, similarly, through a customized DCNN structure, different convolutional kernels can be used to extract different types of detailed features such as fingerprint patterns, breakpoints, and bifurcation points respectively to achieve accurate feature decomposition.

[0157] Step S32 requires the remote verification system for terminal device access permissions to collect image samples based on the target physical signs, perform feature extraction error adjustment on each sub-feature of the sample permission verification refinement, and obtain the corresponding adjusted sub-feature of the sample permission verification refinement. Although the sub-features of the sample permission verification refinement have already undergone the previous feature adjustment and refinement steps, due to various uncertain factors during the image acquisition process (such as light changes, device precision limitations) and the inherent limitations of the feature extraction algorithm, there may still be certain extraction errors. The remote verification system for terminal device access permissions needs to perform fine error adjustment on these sub-features based on the complete information contained in the image samples collected for the target physical signs.

[0158] Continuing with the example of facial physical sign verification, for the sub-feature of the sample permission verification refinement of the eye shape, assume that there are deviations in some details between the eye shape parameters extracted through the previous steps and the actual situation in the target physical sign acquisition image sample. For example, there is a certain error between the extracted angle parameter of the phoenix eye and the actual angle. The remote verification system for terminal device access permissions inputs the target physical sign acquisition image sample and the sub-feature of the sample permission verification refinement of this eye shape into a dedicated error adjustment model. This error adjustment model can be a residual network (ResNet) based on deep learning. ResNet can effectively learn complex error adjustment patterns by introducing skip connections and solve the problem of gradient disappearance in deep learning.

[0159] The model will first perform in-depth analysis on the eye region in the target physical sign acquisition image sample to extract the true eye shape features. Then, it will compare them with the current sub-feature of the sample permission verification refinement. By calculating the difference between the two, for example, the mean squared error (MSE) formula can be used to measure the degree of difference. According to the calculated difference and the adjustment rules learned by the model during training, the sub-feature of the sample permission verification refinement of the eye shape is adjusted. For example, if it is found that the angle of the phoenix eye in the current sub-feature is larger than the true angle, the model will make a corresponding reduction adjustment to the angle parameter according to the adjustment rules, thereby obtaining the adjusted sub-feature of the sample permission verification refinement of the eye shape.

[0160] Step S33 requires the remote verification system for terminal device access permissions to determine the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample based on multiple adjusted sub-features of the sample permission verification refinement. This step is to organically integrate the multiple sub-features after error adjustment to form a first-order prior permission verification feature that can most accurately and comprehensively reflect the true situation of the target physical signs.

[0161] In facial feature verification, the remote verification system for terminal device access rights fuses multiple sample rights verification and refined adjustment sub-features such as the adjusted eyes, nose, mouth, and facial contour. A feasible and effective fusion method is weighted fusion. The remote verification system for terminal device access rights assigns corresponding weights to each sub-feature according to its importance in describing the overall facial features. For example, the facial contour feature plays a key role in determining the overall structure and identity of the face and may be assigned a weight of 0.3; due to its uniqueness and high recognition rate in facial recognition, the eye feature may have a weight of 0.3; although the nose and mouth features have a slightly lower recognition rate compared to the eyes and facial contour, they also make important contributions to the overall facial features, with weights of 0.15 respectively; other detail features (such as skin color, facial texture, etc.) have a weight of 0.1. Through the weighted fusion formula: first-order prior rights verification feature = facial contour sample rights verification and refined adjustment sub-feature × 0.3 + eye sample rights verification and refined adjustment sub-feature × 0.3 + nose sample rights verification and refined adjustment sub-feature × 0.15 + mouth sample rights verification and refined adjustment sub-feature × 0.15 + other detail sample rights verification and refined adjustment sub-feature × 0.1, each sub-feature is integrated to obtain the final facial first-order prior rights verification feature. This first-order prior rights verification feature synthesizes the feature information of multiple facial parts after fine adjustment and can highly accurately represent the true feature situation of the face.

[0162] In addition, the remote verification system for terminal device access rights can also use feature concatenation and convolutional fusion techniques in deep learning to determine the first-order prior rights verification feature. First, multiple sample rights verification and refined adjustment sub-features are concatenated in the feature dimension. For example, the sample rights verification and refined adjustment sub-features of each part of the face are connected together in a certain logical order to form a longer and richer feature vector. Then, the concatenated feature vector is processed by a convolutional neural network. The convolutional layer in the convolutional neural network will use carefully designed convolutional kernels to perform convolutional operations on this long feature vector. These convolutional kernels can capture the associations and interactions between different sub-features and extract higher-level and more abstract feature information. After multiple layers of convolutional operations, the remote verification system for terminal device access rights finally obtains a first-order prior rights verification feature that has been deeply fused and refined.

[0163] Meanwhile, the second-order prior permission verification feature corresponding to the target physical sign acquisition image sample can be selected from the initial sample permission verification feature, the sample verification adjustment feature, and the sample permission verification improvement feature. The initial sample permission verification feature is relatively broad and basic, which can provide a rough feature range and serve as a second-order prior permission verification feature to provide a basic reference framework for subsequent network training. The sample verification adjustment feature has undergone certain error adjustment and is relatively more accurate. Selecting it as the second-order prior permission verification feature can balance the accuracy and generalization of the feature to a certain extent. The sample permission verification improvement feature is a feature that has undergone multiple adjustments and improvements and is the most accurate, but may be slightly weaker in generalization. The remote verification system for the terminal device access permission can flexibly select an appropriate feature as the second-order prior permission verification feature according to specific verification requirements and data characteristics.

[0164] As an implementation manner, the initial permission verification feature includes a plurality of verification feature elements, and the plurality of verification feature elements include a first verification feature element. Based on this, in step S110, the biometric feature extraction operation is performed on the physical sign acquisition image of the terminal device through a pre-calibrated and converged permission verification neural network to obtain the initial permission verification feature, including:

[0165] Step S111: Vectorize the physical sign acquisition image of the terminal device to obtain an image matrix representation;

[0166] Step S112: Map the image matrix representation to a high-dimensional space to obtain an image tensor representation;

[0167] Step S113: Perform a feature weight focusing operation on the image tensor representation to obtain a focused image tensor representation;

[0168] Step S114: Perform position extraction on the focused image tensor representation to obtain a position extraction result;

[0169] Step S115: Obtain the position code corresponding to the position extraction result in a preset feature library, where the position code corresponding to the position extraction result in the preset feature library is the first position code;

[0170] Step S116: Perform a reduction mapping on the first position code to obtain the first verification feature element of the initial permission verification feature.

[0171] Step S111 requires the remote authentication system for the terminal device access permission to vectorize the terminal device's physical sign acquisition image to obtain an image matrix representation. The terminal device's physical sign acquisition image is the original data containing physical sign information. The vectorization operation aims to convert it into a matrix form that is easy for the remote authentication system for the terminal device access permission to process and analyze. Taking the facial physical sign acquisition image as an example, assume that a color image with a resolution of m×n is captured. A color image usually consists of three color channels (such as the RGB channels), and each pixel has a value ranging from 0 to 255 on each channel, representing the intensity of the pixel on that color channel. The remote authentication system for the terminal device access permission arranges the RGB values of each pixel of this image in a certain order to form a three-dimensional matrix. For example, taking the R value, G value, and B value of the pixel in the first row and first column of the image as the first element group of the matrix, the RGB values of the second column of pixels as the second element group, and so on, until the RGB values of the pixels in the last row and last column of the image, finally obtaining an image matrix with a size of m×n×3. This matrix completely preserves the color and spatial information of the image, providing the basic data structure for subsequent processing.

[0172] Step S112 requires the remote authentication system for the terminal device access permission to map the image matrix representation to a high-dimensional space to obtain an image tensor representation. Mapping the image matrix to a high-dimensional space is to enable the remote authentication system for the terminal device access permission to capture more complex and abstract feature information in the image. In deep learning, a tensor is a general data structure for representing high-dimensional data. Taking the m×n×3 matrix of the facial image as an example, the remote authentication system for the terminal device access permission uses some layer structures in the neural network, such as fully connected layers or convolutional layers, to perform linear and non-linear transformations on the image matrix. For example, through the convolution operation of the convolutional layer, the image matrix is convolved with multiple convolutional kernels. A convolutional kernel is a small matrix that slides on the image matrix and performs a dot product operation with a local area of the image matrix to generate a new feature matrix. Assume that the size of the convolutional kernel is k×k and it slides on the image matrix with a stride of s. For each sliding position, calculate the dot product of the convolutional kernel and the local area of the image matrix to obtain a new value, and these new values form a new matrix. After the convolution operations of multiple convolutional kernels, multiple such new matrices will be obtained. Combining these new matrices in a certain way forms a higher-dimensional tensor. For example, after a series of convolution operations, it may be possible to obtain an image tensor with a size of m'×n'×c from the original m×n×3 image matrix, where m' and n' are the spatial dimension sizes after the convolution operation, and c is the number of convolutional kernels, that is, the number of newly generated feature channels.

[0173] Step S113 requires the remote authentication system for terminal device access rights to perform feature weight focusing operations (such as performing attention) on the image tensor representation to obtain a focused image tensor representation. The purpose of the feature weight focusing operation is to enable the remote authentication system for terminal device access rights to pay more attention to the important parts related to biometric features in the image tensor, ignore some unimportant information, and thus improve the accuracy and efficiency of feature extraction. Taking the facial image tensor as an example, the remote authentication system for terminal device access rights can adopt an attention mechanism. The attention mechanism determines which parts need to be focused on by calculating the importance weights of each element in the image tensor. For example, using an attention network that analyzes the image tensor and calculates the importance scores of each element for representing facial biometric features. The importance scores can be obtained by calculating the similarity between the feature vectors at different positions in the image tensor and a learnable query vector. Suppose the image tensor is T and the query vector is q. For each element in the image tensor, calculate the similarity score (where represents the transpose of the query vector q). Then, normalize these scores through the softmax function to obtain the weight of each element, where N is the total number of elements in the image tensor. These weights represent the degree of importance of each element in feature extraction.

[0174] Next, the remote authentication system for terminal device access rights multiplies each element in the image tensor by its corresponding weight to achieve focusing on important features. For example, the element of the focused image tensor. After such an operation, the parts related to the key facial biometric features in the image tensor are enhanced, while some unimportant background information or noise information is weakened, thus obtaining a focused image tensor representation.

[0175] Step S114 requires the remote authentication system for terminal device access rights to perform position extraction on the focused image tensor representation to obtain a position extraction result. The position extraction operation aims to determine the key position information related to biometric features from the focused image tensor. Taking the focused image tensor of a facial image as an example, the remote authentication system for terminal device access rights can use some algorithms or models based on spatial positions to extract the key positions. For example, through an object detection algorithm, such as the Faster R-CNN algorithm based on a convolutional neural network.

[0176] Step S115 requires the remote authentication system for the terminal device access permission to obtain the location code corresponding to the location extraction result in the preset feature library, where the location code corresponding to the location extraction result in the preset feature library is the first location code. The preset feature library is a database pre-constructed by the remote authentication system for the terminal device access permission, which contains various biometric location information, and the location code is a digital representation of this location information. Taking facial features as an example, the preset feature library may contain the standard location information of key facial features (such as eyes, nose, mouth) with different facial expressions and different individuals. The remote authentication system for the terminal device access permission matches the location information of the key facial features extracted in step S114 with the preset feature library. For example, for the extracted eye location information, the remote authentication system for the terminal device access permission searches in the preset feature library for the eye location template that best matches it. Assuming that the eye location templates in the preset feature library are represented by coordinate ranges and relative proportional relationships, the remote authentication system for the terminal device access permission calculates the similarity between the extracted eye location and these templates. Metrics such as Euclidean distance or cosine similarity can be used to calculate the similarity.

[0177] After finding the most matching template, the remote authentication system for the terminal device access permission obtains the location code corresponding to this template as the first location code. This location code is a digitally processed feature representation, which can be a vector or a specific numerical sequence, used to uniquely identify this location information. In the processing of fingerprint features, the preset feature library contains the standard location information of key features (such as breakpoints, bifurcation points) of different types of fingerprints (such as loop fingerprints, whorl fingerprints). The remote authentication system for the terminal device access permission matches the extracted fingerprint key feature locations with the preset feature library, calculates the similarity, finds the most matching template and obtains the location code corresponding to it as the first location code. The implementation of this operation mainly depends on database query and matching algorithms, and through rapid and accurate matching of the extracted location information and the templates in the preset feature library, the corresponding location code is obtained.

[0178] Step S116 requires the remote authentication system for the terminal device access permission to perform an inverse mapping on the first location code to obtain the first verification feature element of the initial permission verification feature. The inverse mapping operation converts the location code back into information with actual biometric significance, serving as the first verification feature element of the initial permission verification feature. Taking facial features as an example, assume that the first location code is a digitized vector that represents the position information of the eyes in the facial image. The remote authentication system for the terminal device access permission uses a pre-trained inverse mapping model that has learned the mapping relationship between the location code and the actual facial features. This model can be a neural network that takes the first location code as input and, through a series of neuron calculations and non-linear transformations, outputs a facial feature information with actual significance, such as the specific shape and size of the eyes, as the first verification feature element of the initial permission verification feature. For example, the model may determine whether the eyes are single eyelids or double eyelids and the approximate size range of the eyes based on the location code.

[0179] As an implementation, the initial permission verification feature includes multiple verification feature elements, and the multiple verification feature elements further include verification feature elements other than the first one. The verification feature elements other than the first one are any verification feature element other than the first verification feature element. Based on this, in step S110, the biometric feature extraction operation is performed on the physical sign acquisition image of the terminal device through the permission verification neural network that has been pre-tuned to converge to obtain the initial permission verification feature, including:

[0180] Step S117: Map the location code corresponding to the previous verification feature element to a high-dimensional space to obtain a tensor representation of the location code;

[0181] Step S118: Perform a feature weight focusing operation on the tensor representation of the location code to obtain a focused tensor representation of the location code;

[0182] Step S119: Obtain the location code corresponding to the focused tensor representation of the location code in the preset feature library, where the location code corresponding to the focused tensor representation of the location code in the preset feature library is the second location code;

[0183] Step S1110: Perform an inverse mapping on the second location code to obtain the verification feature element other than the first one of the initial permission verification feature.

[0184] In step S117, the remote verification system for the access permission of the terminal device obtains the position code corresponding to the previous verification feature element. The position code is a coding method for digitally representing position information, which records the position information of the feature in a specific space. In practical applications, for example, when processing the physical sign collection image of the terminal device, different physical sign parts in the image have different spatial positions, and the position code is used to identify these positions. For example, assume that the physical sign collection image of the terminal device is an image containing a human face, and different parts such as eyes, nose, and mouth have their own positions in the image, and the position code is a digital code used to accurately describe the positions of these parts.

[0185] Next, the remote verification system for the access permission of the terminal device maps the position code to a high-dimensional space. Mapping to a high-dimensional space is to enable richer expression and processing of position information in a higher-dimensional vector space. Usually, mathematical methods such as linear transformation can be used to achieve this mapping. For example, assume that the position code is a one-dimensional vector [x], and the remote verification system for the access permission of the terminal device can map it to a high-dimensional space through a matrix multiplication. For example, by using an m×1 matrix W, through calculation Y = W×[x] (where Y is the high-dimensional vector after mapping), the tensor representation of the position code is obtained. Here, m represents the dimension of the high-dimensional space, and m is usually greater than 1. For example, m = 10, then after mapping, the original one-dimensional position code becomes a ten-dimensional vector, so that more information can be carried in the high-dimensional space.

[0186] In step S118, the remote verification system for the access permission of the terminal device performs a feature weight focusing operation on the obtained location encoding tensor representation, and this operation is similar to performing an attention mechanism. The core idea of the attention mechanism is to enable the remote verification system for the access permission of the terminal device to automatically focus on the important parts in the data. In this step, the remote verification system for the access permission of the terminal device weights each dimension in the location encoding tensor representation according to different feature weights. For example, assume that the location encoding tensor representation is a ten-dimensional vector [y1, y2, y3, y4, y5, y6, y7, y8, y9, y10], and the remote verification system for the access permission of the terminal device assigns a weight value to each dimension, such as [w1, w2, w3, w4, w5, w6, w7, w8, w9, w10], and these weight values represent the importance of each dimension in the current task. Through weighted calculation, that is, the new vector element zi = wi × yi (i = 1 to 10), a new vector [z1, z2, z3, z4, z5, z6, z7, z8, z9, z10] is obtained, and this new vector is the focused location encoding tensor representation after the feature weight focusing operation. The purpose of doing this is to highlight the more important information in the location encoding tensor representation, so that subsequent processing can pay more attention to the key features.

[0187] In step S119, the remote verification system for the access permission of the terminal device matches the focused position encoding tensor representation obtained through the feature weight focusing operation with a preset feature library. The preset feature library is a pre-established database containing various features and their corresponding position encodings. For example, in the scenario of processing the physical sign collection image of the terminal device, the preset feature library may contain the standard position encoding information of different face features, fingerprint features, etc. at different positions. The remote verification system for the access permission of the terminal device searches in the preset feature library for the position encoding that best matches the focused position encoding tensor representation, and this found position encoding is the second position encoding. The specific search process can use a similarity calculation method, such as cosine similarity. Suppose the focused position encoding tensor representation is vector A, and each position encoding vector in the preset feature library is Bi (i = 1 to n, where n is the number of position encoding vectors in the preset feature library). By calculating the cosine similarity Sim(A, Bi) = (A·Bi) / (||A||×||Bi||) (where A·Bi is the dot product of vectors, and ||A|| and ||Bi|| are the norms of vectors A and Bi respectively), the Bi with the highest similarity is found, and this Bi is the second position encoding. For example, when processing a face image, the focused position encoding tensor representation may correspond to the position information of a certain local feature in the image. By searching in the preset feature library, the standard position encoding that best matches it is found, and this standard position encoding is the second position encoding, which can more accurately reflect the position of this local feature in the overall feature system.

[0188] In step S1110, the remote verification system for the access permission of the terminal device performs an inverse mapping on the found second position encoding. The inverse mapping is an operation opposite to the previous mapping to the high-dimensional space, and its purpose is to restore the position encoding processed in the high-dimensional space to an appropriate dimension to obtain the verification feature element. For example, if the one-dimensional position encoding was previously mapped to the high-dimensional space through matrix multiplication W×[x], then the inverse mapping may be calculated through where is the inverse matrix of matrix W, Y' is the second position encoding vector, and [x'] is the restored vector). After the inverse mapping, the obtained result is the verification feature element other than the first of the initial permission verification feature. This verification feature element further enriches the content of the initial permission verification feature and provides more feature information for more accurate permission verification in the subsequent process. For example, when processing a face image for permission verification, the verification feature element obtained through the inverse mapping of the second position encoding may be a detailed feature description of a specific area of the face, such as the texture feature of the cheek. These features, together with other verification feature elements, constitute a more comprehensive and accurate initial permission verification feature, which helps to improve the accuracy and reliability of permission verification.

[0189] Figure 2 This is a schematic diagram of the hardware entity of a remote authentication system for terminal device access rights provided by an embodiment of the present invention. As Figure 2 shown, the hardware entity of the remote authentication system 1000 for terminal device access rights includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any of the above embodiments.

[0190] The memory 1002 stores a computer program that can run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001, and can also cache data to be processed or already processed by the processor 1001 and each module in the remote authentication system 1000 for terminal device access rights (for example, image data, audio data, voice communication data, and video communication data), and can be implemented by flash memory (FLASH) or random access memory (Random Access Memory, RAM).

[0191] When the processor 1001 executes the program, it implements the steps of the remote authentication method for terminal device access rights in any of the above items. The processor 1001 generally controls the overall operation of the remote authentication system 1000 for terminal device access rights.

[0192] As described above, it is only the embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, and all should be covered within the protection scope of the present invention.

Claims

1. A remote verification method for the access permission of a terminal device, characterized in that, The method includes: Performing a biometric extraction operation on the physical sign acquisition image of the terminal device to obtain an initial permission verification feature; Decomposing the initial permission verification feature to obtain a plurality of initial verification sub-features; Embedding the physical sign acquisition image of the terminal device and the to-be-processed initial verification sub-feature into a feature extraction error adjustment indication framework to obtain a to-be-adjusted image feature, where the to-be-adjusted image feature is used to indicate that an image processing network performs extraction feature error adjustment on the to-be-processed initial verification sub-feature, and the to-be-processed initial verification sub-feature is any one of the plurality of initial verification sub-features; Loading the to-be-adjusted image feature into the image processing network to obtain an analysis result of the to-be-adjusted image feature output by the image processing network, where the analysis result of the to-be-adjusted image feature includes a verification adjustment sub-feature corresponding to the to-be-processed initial verification sub-feature and an extraction error category, and the extraction error category identifies the type corresponding to the extraction feature error; Determining a verification adjustment feature based on the plurality of verification adjustment sub-features; Embedding the physical sign acquisition image of the terminal device and the verification adjustment feature into a feature improvement indication framework to obtain a to-be-improved image feature, where the to-be-improved image feature is used to indicate that the image processing network performs feature improvement on the verification adjustment feature; Loading the to-be-improved image feature into the image processing network to obtain an analysis result of the to-be-improved image feature output by the image processing network, where the analysis result of the to-be-improved image feature includes the permission verification improvement feature; Determining a target permission verification feature based on the permission verification improvement feature; Performing identification based on the target permission verification feature to obtain a verification result of the access permission of the terminal device.

2. The method according to claim 1, wherein The determining the target permission verification feature based on the permission verification improvement feature includes: Decomposing the permission verification improvement feature to obtain a plurality of permission verification improvement sub-features; Performing extraction feature error adjustment on each of the permission verification improvement sub-features according to the physical sign acquisition image of the terminal device to obtain corresponding permission verification improvement adjustment sub-features; Determining the target permission verification feature based on the plurality of permission verification improvement adjustment sub-features.

3. The method according to claim 1, characterized in that, The performing a biometric extraction operation on the physical sign acquisition image of the terminal device to obtain an initial permission verification feature includes: Performing a biometric extraction operation on the physical sign acquisition image of the terminal device through a pre-calibrated and converged permission verification neural network to obtain an initial permission verification feature.

4. The method according to claim 1, wherein Before the performing a biometric extraction operation on the physical sign acquisition image of the terminal device through a pre-calibrated and converged permission verification neural network to obtain an initial permission verification feature, the method further includes: Obtain an initial permission verification neural network, where the initial permission verification neural network is obtained by calibrating a pre-debugged permission verification neural network based on a physical sign acquisition image sample and a first-order prior permission verification feature corresponding to the physical sign acquisition image sample. The physical sign acquisition image sample corresponds to a second-order prior permission verification feature and a first-order prior permission verification feature. The second-order prior permission verification feature is an image feature with a first precision corresponding to the physical sign acquisition image sample, and the first-order prior permission verification feature is an image feature with a second precision corresponding to the physical sign acquisition image sample; Generate a proposed calibration extraction network based on the network weight parameters and bias parameters of the initial permission verification neural network, where the network composition architecture of the proposed calibration extraction network is the same as that of the pre-debugged permission verification neural network, and the network weight parameters and bias parameters of the proposed calibration extraction network are the same as those of the initial permission verification neural network; Load the physical sign acquisition image sample into the proposed calibration extraction network to obtain a first predicted feature output by the proposed calibration extraction network; and determine a first training error value based on the first predicted feature and the first-order prior permission verification feature; Load the physical sign acquisition image sample into the initial permission verification neural network to obtain a second predicted feature output by the initial permission verification neural network; and determine a second training error value based on the second predicted feature, the first predicted feature, the second-order prior permission verification feature, and the first-order prior permission verification feature; Determine a comprehensive training error value based on the first training error value and the second training error value; Iterate the network weight parameters and bias parameters of the proposed calibration extraction network according to the comprehensive training error value until a preset iteration stop condition is met, and obtain a pre-calibrated and converged extraction neural network, where the pre-calibrated and converged extraction neural network is the pre-calibrated and converged permission verification neural network.

5. The method according to claim 4, wherein Before the obtaining of the initial permission verification neural network, the method further includes: Perform a biometric extraction operation on a target physical sign acquisition image sample to obtain an initial sample permission verification feature; where the target physical sign acquisition image sample is any one of the multiple physical sign acquisition image samples; Perform a feature adjustment operation on the initial sample permission verification feature to obtain a debug transition image feature; Determine a first-order prior permission verification feature corresponding to the target physical sign acquisition image sample based on the debug transition image feature; the second-order prior permission verification feature corresponding to the target physical sign acquisition image sample is the initial sample permission verification feature or the debug transition image feature.

6. The method according to claim 5, characterized in that The debug transition image feature includes a sample verification adjustment feature and a sample permission verification improvement feature; The performing of the feature adjustment operation on the initial sample permission verification feature to obtain a debug transition image feature includes: Perform a feature decomposition on the initial sample permission verification feature to obtain a plurality of initial sample verification sub-features; Collect image samples according to the target physical signs, and perform feature extraction error adjustment on each of the initial sample verification sub-features to obtain corresponding sample verification adjustment sub-features; Determine a sample verification adjustment feature based on multiple sample verification adjustment sub-features; Collect image samples according to the target physical signs, and perform feature improvement on the sample verification adjustment feature to obtain a sample permission verification improvement feature.

7. The method according to claim 6, wherein The determining the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample according to the debug transition image feature includes: Perform feature decomposition on the sample permission verification improvement feature to obtain multiple sample permission verification improvement sub-features; Collect image samples according to the target physical signs, and perform feature extraction error adjustment on each of the sample permission verification improvement sub-features to obtain corresponding sample permission verification improvement adjustment sub-features; Determine the first-order prior permission verification feature corresponding to the target physical sign acquisition image sample based on multiple sample permission verification improvement adjustment sub-features; the second-order prior permission verification feature corresponding to the target physical sign acquisition image sample is any one of the following features: the initial sample permission verification feature, the sample verification adjustment feature, the sample permission verification improvement feature.

8. The method according to claim 3, characterized in that, The initial permission verification feature includes multiple verification feature elements, and the multiple verification feature elements include a first verification feature element; the obtaining of the initial permission verification feature by performing a biometric extraction operation on the terminal device physical sign acquisition image by a pre-calibrated and converged permission verification neural network includes: Vectorize the terminal device physical sign acquisition image to obtain an image matrix representation; Map the image matrix representation to a high-dimensional space to obtain an image tensor representation; Perform a feature weight focusing operation on the image tensor representation to obtain a focused image tensor representation; Perform position extraction on the focused image tensor representation to obtain a position extraction result; Obtain the position encoding corresponding to the position extraction result in a preset feature library, where the position encoding corresponding to the position extraction result in the preset feature library is the first position encoding; Perform a reduction mapping on the first position encoding to obtain the first verification feature element of the initial permission verification feature; The multiple verification feature elements further include verification feature elements other than the first one, and the verification feature elements other than the first one are any one verification feature element other than the first verification feature element; the obtaining of the initial permission verification feature by performing a biometric extraction operation on the terminal device physical sign acquisition image by a pre-calibrated and converged permission verification neural network includes: Map the position encoding corresponding to the previous verification feature element to a high-dimensional space to obtain a position encoding tensor representation; Perform a feature weight focusing operation on the position encoding tensor representation to obtain a focused position encoding tensor representation; Obtain the position encoding corresponding to the focused position encoding tensor representation in a preset feature library, where the position encoding corresponding to the focused position encoding tensor representation in the preset feature library is the second position encoding; Perform a reduction mapping on the second position encoding to obtain the verification feature element other than the first one of the initial permission verification feature.

9. A remote verification system for access rights of a terminal device, including a memory and a processor, where the memory stores a computer program that can run on the processor, characterized in that, When the processor executes the program, it implements the steps in the method according to any one of claims 1 to 8.

Images