A jewelry inventory management authentication method and system based on multimodal data recognition
By identifying and processing noise data and optimizing the multimodal identification and authentication algorithm, the accuracy and adaptability of the multimodal identification and authentication algorithm in the noisy environment in jewelry inventory management is solved, and the accuracy and security of authentication are improved.
Patent Information
- Application Number
- CN202411068464.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-06
AI Technical Summary
The existing multimodal identification authentication algorithm is susceptible to noise data interference when processing multimodal live scan information in jewelry inventory management, resulting in insufficient authentication accuracy and robustness, and failing to effectively adapt to changes in biometric and behavioral patterns. .
By obtaining the multimodal live scan abnormality detection information of the initial multimodal identification authentication algorithm, identifying and processing a priori comment of the safety authentication elements carrying noise, using the training information of past jewelry inventory managers to generate prediction information, and debug the algorithm. Optimize the algorithm to improve its performance and adaptability in noisy environments.
It improves the accuracy and robustness of multimodal identification and authentication, improves the security and efficiency of jewelry inventory management, and ensures the accuracy and reliability of identity authentication.
Smart Images

Figure CN118587780B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a jewelry inventory management authentication method and system based on multimodal data recognition. Background Art
[0002] In jewelry inventory management, multimodal identification and authentication technology is an important means of security assurance. However, existing multimodal identification and authentication algorithms often face interference from noisy data when processing multimodal live scan information, resulting in insufficient authentication accuracy and robustness. In particular, when there is noise in the prior annotations of security authentication elements, the performance of the algorithm will be further affected. In addition, since biometrics and behavioral patterns may change in actual operations, the algorithm needs to have a certain degree of adaptability and dynamic adjustment capabilities. However, existing technical solutions often ignore this point, resulting in unsatisfactory authentication results. Summary of the invention
[0003] The present invention provides a jewelry inventory management authentication method and system based on multimodal data recognition, which can solve or partially solve the technical problems involved in the above-mentioned background technology.
[0004] An embodiment of the present invention provides a jewelry inventory management authentication method based on multimodal data recognition, which is applied to a management authentication system. The method includes:
[0005] Acquire multimodal liveness scanning anomaly detection information of an initial multimodal recognition and authentication algorithm, wherein the multimodal liveness scanning anomaly detection information includes X groups of security authentication element prior annotations, each group of security authentication element prior annotations includes a jewelry inventory manager's prior feature variables for Y authentication elements, and there is at least one group of security authentication element prior annotations carrying noise in the X groups of security authentication element prior annotations, and there is at least one prior feature variable of the authentication element abnormal in the prior feature variables of the Y authentication elements included in the prior feature variables of the security authentication element carrying noise, and X and Y are both positive integers;
[0006] For any past jewelry inventory manager, if the security authentication factor prior annotation corresponding to the past jewelry inventory manager in the multi-mode liveness scanning anomaly detection information carries noise, then based on the security authentication factor training information and the security authentication factor prior annotation of the past jewelry inventory manager, the security authentication factor prediction information of the past jewelry inventory manager is obtained; wherein the security authentication factor training information of the past jewelry inventory manager includes: the identified prediction result of the feature variable of the past jewelry inventory manager for at least one authentication factor;
[0007] According to the security authentication factor prediction information of the past jewelry inventory managers, the initial multimodal recognition and authentication algorithm is debugged to obtain a debugged multimodal recognition and authentication algorithm; the debugged multimodal recognition and authentication algorithm is used to mine the multimodal living body information to be authenticated that contains the target security authentication factor, and the multimodal living body information to be authenticated is used to perform authority security authentication on the target jewelry inventory manager, and there is a corresponding relationship between the target jewelry inventory manager and the multimodal living body information to be authenticated.
[0008] In one implementation, obtaining the prediction information of the security authentication elements of the past jewelry inventory managers based on the security authentication element training information and the security authentication element prior annotations of the past jewelry inventory managers includes:
[0009] For each of the Y authentication factors, if the prior feature variable of the authentication factor contained in the prior annotation of the security authentication factor of the past jewelry inventory manager is abnormal, obtaining the feature variable prediction result of the authentication factor from the security authentication factor training information of the past jewelry inventory manager as the current feature variable of the authentication factor;
[0010] If the prior characteristic variables of the authentication factor included in the prior annotations of the security authentication factor of the past jewelry inventory manager are not abnormal, determining the prior characteristic variables of the authentication factor as the current characteristic variables of the authentication factor;
[0011] Based on the current characteristic variables of the Y authentication factors, the security authentication factor prediction information of the past jewelry inventory managers is obtained.
[0012] In one implementation, the method further includes:
[0013] Generate a liveness scanning feature relationship network corresponding to the multi-mode liveness scanning anomaly detection information, wherein the liveness scanning feature relationship network includes X*Y relationship network nodes, and each relationship network node is used to indicate whether a priori feature variable of an authentication factor of a jewelry inventory manager is abnormal.
[0014] In one implementation, the step of obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm includes:
[0015] Obtain at least two groups of reference prior annotations, each group of reference prior annotations includes at least one prior characteristic variable of a jewelry inventory manager for at least one authentication factor; wherein the authentication factors included in different groups of reference prior annotations are different, and different groups of reference prior annotations correspond to different jewelry inventory managers;
[0016] The at least two groups of reference prior annotations are clustered based on each jewelry inventory manager to obtain the multi-mode liveness scanning anomaly detection information.
[0017] In one implementation, the initial multimodal recognition and authentication algorithm is debugged based on the security authentication factor prediction information of the past jewelry inventory managers to obtain a debugged multimodal recognition and authentication algorithm, including:
[0018] Obtaining multimodal liveness prediction information corresponding to the past jewelry inventory manager based on the past jewelry inventory manager and the security authentication factor prediction information of the past jewelry inventory manager through an initial multimodal recognition and authentication algorithm;
[0019] Based on the past jewelry inventory manager and the multimodal liveness prediction information, determining an algorithm training error of an initial multimodal recognition and authentication algorithm, wherein the algorithm training error is used to characterize a scanning offset weight of the past jewelry inventory manager and the multimodal liveness prediction information;
[0020] Based on the algorithm training error, the algorithm parameters of the initial multimodal recognition and authentication algorithm are optimized to obtain the debugged multimodal recognition and authentication algorithm.
[0021] In one implementation, the algorithm training error includes an authentication factor mining error and a joint training error, and the joint training error includes a first training error, a second training error, a third training error, and a fourth training error;
[0022] The step of determining the algorithm training error of the initial multimodal recognition and authentication algorithm based on the past jewelry inventory management personnel and the multimodal liveness prediction information includes:
[0023] Determine the authentication factor mining error based on the prior feature variables of the prior authentication factor in the prior annotation of the security authentication factor and the feature variable prediction result of the prior authentication factor in the security authentication factor training information; wherein the authentication factor mining error is used to characterize the mining result confidence score of the security authentication factor training information, and the prior authentication factor is the authentication factor corresponding to the prior feature variable without abnormality in the prior annotation of the security authentication factor;
[0024] Based on the multimodal living body prediction information and the past jewelry inventory management personnel, determining the first training error, the first training error being used to characterize a real-time offset error between the multimodal living body prediction information and the past jewelry inventory management personnel;
[0025] The second training error and the third training error are determined based on the security authentication factor prediction information, the discrete evaluation quality of the security authentication factor prediction information, the average feature of the security authentication factor prediction information, and the average feature of the security authentication factor training information; wherein the second training error is used to characterize the current feature variable of the prior authentication factor in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information, and the third training error is used to characterize the current feature variable of the remaining authentication factors other than the prior authentication factor among the Y types of authentication factors included in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information;
[0026] According to the depth residual feature map, and the distribution parameters and average characteristics of the thermodynamic relationship spectrum matched by the depth residual feature map, the fourth training error is determined, and the fourth training error is used to characterize the noise proportion of the depth residual feature map compared to the thermodynamic relationship spectrum; wherein the depth residual feature map is the deep liveness characterization vector of the past jewelry inventory manager obtained by the fully connected branch of the initial multimodal recognition and authentication algorithm;
[0027] Based on the first training error, the second training error, the third training error and the fourth training error, determining the joint training error, wherein the joint training error is used to characterize the matching weight between the multimodal liveness prediction information generated based on the security authentication element prediction information and the past jewelry inventory manager;
[0028] The algorithm training error is determined based on the authentication factor mining error and the joint training error.
[0029] In one implementation, based on the continuous correction of the past jewelry inventory manager's security authentication factor training information along with the debugging of the initial multimodal recognition and authentication algorithm, the discrete evaluation quality and average characteristics of the security authentication factor prediction information are also in a continuous correction state.
[0030] In one implementation, the initial multimodal identification and authentication algorithm includes a first data encoding and decoding branch and a second data encoding and decoding branch;
[0031] The first data encoding and decoding branch is used to generate a depth residual feature map based on the past jewelry inventory management personnel and the security authentication factor prediction information of the past jewelry inventory management personnel, and the average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the depth residual feature map are determined by the algorithm parameters of the first data encoding and decoding branch, the past jewelry inventory management personnel and the security authentication factor prediction information of the past jewelry inventory management personnel;
[0032] The second data encoding and decoding branch is used to obtain the multimodal liveness prediction information based on the depth residual feature map and the security authentication factor prediction information of the past jewelry inventory managers. The average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the multimodal liveness prediction information are determined by the algorithm parameters of the second data encoding and decoding branch, the depth residual feature map and the security authentication factor prediction information of the past jewelry inventory managers.
[0033] In one implementation, the past security authentication element training information of jewelry inventory managers is obtained by a security authentication element embedding algorithm, and the method further includes:
[0034] On the basis that the number of persons in the multi-modal liveness scanning abnormality detection information is greater than the set number, the security authentication factor embedding algorithm adopts a general first long short-term memory algorithm, and in the debugging process of the initial multi-modal recognition and authentication algorithm, the algorithm parameters of the first long short-term memory algorithm are optimized by using the difference between the security authentication factor prior annotations of the past jewelry inventory managers and the security authentication factor training information of the past jewelry inventory managers;
[0035] Alternatively, on the basis that the number of persons in the multi-modal liveness scanning anomaly detection information is not greater than the set number, the security authentication factor embedding algorithm adopts a non-universal second long short-term memory algorithm, and the second long short-term memory algorithm is an AI algorithm with a security authentication factor mining function obtained by pre-debugging using multi-modal liveness scanning detection information, and the multi-modal liveness scanning detection information includes scanning detection information of multiple dimensions and corresponding liveness feature variables.
[0036] In one implementation, the second long short-term memory algorithm includes a residual connection branch and a soft threshold branch;
[0037] During the debugging process of the initial multimodal recognition and authentication algorithm, the algorithm parameters of the soft threshold branch are optimized by utilizing the difference between the prior annotations of the security authentication factors of the past jewelry inventory managers and the training information of the security authentication factors of the past jewelry inventory managers, while the algorithm parameters of the residual connection branch are maintained unchanged.
[0038] An embodiment of the present invention provides a management authentication system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.
[0039] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.
[0040] The embodiment of the present invention obtains multi-modal liveness scanning anomaly detection information, identifies and processes the prior annotations of security authentication elements with noise, generates prediction information using the training information and prior annotations of past jewelry inventory managers, and debugs the initial multi-modal recognition and authentication algorithm, ultimately obtaining a more accurate and robust multi-modal recognition and authentication algorithm. This design improves the performance of multi-modal recognition and authentication, and brings higher security and efficiency to jewelry inventory management. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 The present invention provides a flowchart of a jewelry inventory management authentication method based on multimodal data recognition.
[0042] Figure 2 A schematic diagram of the structure of a management authentication system provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0044] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.
[0045] Figure 1 A jewelry inventory management authentication method based on multimodal data recognition is shown, which is applied to a management authentication system. The method includes the following steps 110 to 130.
[0046] In the field of jewelry inventory management, it is crucial to ensure the accuracy and security of the identity of jewelry inventory managers. Traditional single-modal authentication methods (such as fingerprint recognition or password authentication) can no longer meet current security requirements. Therefore, the embodiment of the present invention adopts a management authentication system that combines multiple biometrics and behavioral patterns for identity authentication, which becomes a safer and more reliable option. The following is a specific application scenario example, which details how to implement and manage the identity authentication of jewelry inventory managers through a management authentication system.
[0047] Step 110: Obtain multi-modal liveness scanning anomaly detection information of an initial multi-modal recognition and authentication algorithm.
[0048] The multi-mode liveness scanning anomaly detection information includes X groups of security authentication factor prior annotations, each group of security authentication factor prior annotations includes a jewelry inventory manager's prior feature variables for Y authentication factors, and there is at least one group of security authentication factor prior annotations carrying noise in the X groups of security authentication factor prior annotations, and there is at least one prior feature variable of the authentication factor abnormal among the prior feature variables of the Y authentication factors included in the prior annotations of the security authentication factors carrying noise, and X and Y are both positive integers.
[0049] In the embodiment of the present invention, the first step of managing the authentication system is to obtain the multi-modal liveness scanning anomaly detection information of the initial algorithm. The multi-modal liveness scanning anomaly detection information includes X groups of security authentication element prior annotations, each group of annotations is for a jewelry inventory manager, and covers Y different authentication elements. These authentication elements may include fingerprints, facial features, iris scanning, voice recognition, and specific behavior patterns.
[0050] The data in the security authentication factor prior annotations are obtained through pre-training and scanning. Each set of security authentication factor prior annotations records in detail the characteristic variables of each jewelry inventory manager for various authentication factors. However, in actual applications, due to various internal and external factors (such as equipment errors, environmental interference, human operation errors, etc.), these security authentication factor prior annotations may contain noisy data. Noisy security authentication factor prior annotations mean that at least one of the characteristic variables of the authentication factor is abnormal, which may affect the accuracy and reliability of the authentication algorithm.
[0051] Step 120: For any past jewelry inventory manager, if the corresponding security authentication factor prior annotation of the past jewelry inventory manager in the multi-mode liveness scanning anomaly detection information carries noise, then based on the security authentication factor training information and security authentication factor prior annotation of the past jewelry inventory manager, obtain the security authentication factor prediction information of the past jewelry inventory manager.
[0052] The security authentication element training information of past jewelry inventory managers includes: the identified prediction results of the characteristic variables of the past jewelry inventory managers for at least one authentication element.
[0053] In the management authentication system, in order to improve the robustness and accuracy of the algorithm, it is necessary to process the noisy prior annotations. For any past jewelry inventory manager, if the corresponding security authentication factor prior annotation in the multi-mode liveness scanning anomaly detection information is noisy, the management authentication system will use the manager's security authentication factor training information to generate prediction information.
[0054] Security authentication element training information is data acquired in real time through a variety of sensors and devices, including the predicted results of the manager's characteristic variables for at least one authentication element. These data are obtained through continuous scanning and monitoring, and can reflect the manager's biometric characteristics and behavior patterns in actual operations.
[0055] Based on these training information and noisy prior annotations, the management authentication system generates security authentication factor prediction information of the jewelry inventory manager through advanced machine learning algorithms (such as deep learning, support vector machine, etc.). The security authentication factor prediction information is an estimate of the real biometrics and behavior patterns of the jewelry inventory manager, which is used for subsequent algorithm debugging and optimization.
[0056] Step 130: Debug the initial multimodal recognition and authentication algorithm based on the past security authentication factor prediction information of the jewelry inventory manager to obtain a debugged multimodal recognition and authentication algorithm.
[0057] Among them, the debugged multimodal recognition and authentication algorithm is used to mine multimodal living information to be authenticated that contains target security authentication elements, and the multimodal living information to be authenticated is used to perform authority security authentication on the target jewelry inventory manager. There is a corresponding relationship between the target jewelry inventory manager and the multimodal living information to be authenticated.
[0058] In the embodiment of the present invention, after obtaining the security authentication factor prediction information of the past jewelry inventory managers, the management authentication system uses the security authentication factor prediction information to debug the initial multimodal recognition authentication algorithm. The purpose of debugging is to optimize the algorithm so that it can better handle complex situations in actual scenarios, especially those data with noise.
[0059] During the debugging process, the management authentication system compares the security authentication factor prediction information with the prior annotations to analyze the differences and consistency between the two. In this way, the algorithm can learn which feature variables are stable and reliable and which ones may be affected by noise. Based on these learning results, the algorithm will make corresponding adjustments to improve the robustness to noisy data and the accuracy to normal data.
[0060] After debugging, the multimodal recognition and authentication algorithm can more effectively mine the multimodal liveness information to be authenticated that contains the target security authentication elements. The authenticated multimodal liveness information is used to perform authority security authentication on the target jewelry inventory managers, ensuring that only strictly verified managers can access and operate the jewelry inventory.
[0061] In an exemplary application scenario, a jewelry company has a complete management authentication system for managing its inventory. The system needs to perform strict identity authentication on each inventory manager to ensure the safety of jewelry. The management authentication system first obtains 100 sets of prior annotations of security authentication elements for jewelry inventory managers. Each set of annotations contains five authentication elements: fingerprint, facial features, iris scan, voice recognition, and behavior patterns. However, in actual applications, it was found that 10 sets of data carried noise, mainly manifested as abnormal feature variables of fingerprint and iris scan. For these 10 sets of noisy data, the system uses past security authentication element training information to generate prediction information. The training information is obtained through continuous monitoring and scanning, including the biometrics and behavior patterns of managers in daily operations. The system uses a deep learning algorithm to generate new feature variable prediction results based on the training information and the noisy prior annotations. The system compares these predictions with the prior annotations and finds that the noise mainly comes from the sensor error of fingerprint and iris scans. Based on this finding, the algorithm is adjusted accordingly, reducing the weight of fingerprint and iris scans in the authentication process, while increasing the weight of facial features and behavior patterns. After debugging, the multimodal recognition and authentication algorithm was used to authenticate the target jewelry inventory manager. The system verifies the manager's identity by mining the multimodal living information to be authenticated that contains the target security authentication elements, such as facial features, voice recognition, and behavior patterns. Only when all authentication elements are matched and verified by the algorithm can the manager obtain the permission to access and operate the jewelry inventory. In this way, not only the accuracy and security of identity authentication are improved, but also the data with noise can be effectively processed, ensuring the stability and reliability of the system in practical applications.
[0062] On the basis of the above, step 110 specifically involves obtaining the multimodal liveness scanning anomaly detection information of the initial multimodal recognition and authentication algorithm. This step is the basis of the entire authentication system and provides key data support for subsequent security authentication. First, the multimodal liveness scanning anomaly detection information contains X groups of security authentication element prior annotations, each group of annotations is for a specific jewelry inventory manager and covers Y different authentication elements. These authentication elements may include fingerprints, facial features, iris scanning, voice recognition, and specific behavior patterns, which together constitute the basis of multimodal recognition. Secondly, each group of security authentication element prior annotations records in detail the characteristic variables of the manager for various authentication elements. These characteristic variables are obtained through pre-training and scanning, and they constitute an important basis for the identity authentication of managers. However, in actual applications, due to the influence of various internal and external factors, these prior annotations may contain noisy data. Then, the management authentication system needs to pay attention to the noisy security authentication element prior annotations. At least one of these annotations contains an abnormality in the characteristic variable of one authentication element, which may be caused by equipment errors, environmental interference, human operation errors, etc. These noise data will affect the accuracy and reliability of the authentication algorithm, so special processing is required. Finally, the management authentication system emphasizes the importance of X and Y as positive integers in this step. X represents the number of groups of prior annotations, that is, the number of jewelry inventory managers; Y represents the number of authentication elements corresponding to each manager. These two parameters together define the scale and complexity of multi-modal liveness scanning anomaly detection information, and also provide a basis for subsequent algorithm debugging and optimization. It can be seen that step 110 provides key data support for subsequent security authentication by obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition authentication algorithm. This step not only involves a large amount of biometric and behavioral pattern data, but also needs to process possible noise data to ensure the accuracy and reliability of the entire authentication system.
[0063] Next, step 120 focuses on processing the noise data in the multi-mode liveness scanning anomaly detection information, and generates security authentication factor prediction information for jewelry inventory managers based on past security authentication factor training information and prior annotations of these managers.
[0064] First, for any past jewelry inventory manager, if their corresponding security authentication element prior annotations in the multi-modal liveness scanning anomaly detection information are noisy, then the management authentication system needs to perform special processing on these noisy data. This is because the noise data may affect the accuracy and reliability of the authentication algorithm, thereby affecting the performance of the entire authentication system. Secondly, in order to process these noisy data, the management authentication system uses the security authentication element training information of past jewelry inventory managers. These training information are data obtained in real time through a variety of sensors and devices. They contain the prediction results of the characteristic variables of past jewelry inventory managers for at least one authentication element obtained by identification. These prediction results reflect the biometrics and behavior patterns of managers in actual operations and are an important basis for generating prediction information. Then, the management authentication system combines the security authentication element training information and prior annotations of past jewelry inventory managers, and generates the security authentication element prediction information of these managers through advanced machine learning algorithms (such as deep learning, support vector machines, etc.). This prediction information is an estimate of the real biometrics and behavior patterns of managers. It takes into account the influence of noise data and reflects the identity characteristics of managers as accurately as possible. Finally, the obtained security authentication element prediction information will be used in the subsequent multi-modal recognition and authentication process. By comparing and analyzing with the prior annotations, the management authentication system can further optimize the authentication algorithm and improve its ability to process noise data, thereby ensuring the accuracy and reliability of the entire authentication system. Therefore, step 120 processes the noise data in the multi-modal liveness scanning anomaly detection information and generates prediction information based on the past jewelry inventory management personnel's security authentication element training information and prior annotations, providing key data support and algorithm optimization direction for subsequent multi-modal recognition authentication.
[0065] Finally, step 130 focuses on debugging and optimizing the initial multimodal recognition and authentication algorithm using the prediction information of security authentication elements of past jewelry inventory managers. This process aims to improve the accuracy and robustness of the algorithm in practical applications, and ensure that the security authentication of the target jewelry inventory manager can be carried out efficiently and reliably. First of all, it should be clear that the prediction information of security authentication elements of past jewelry inventory managers is obtained by in-depth analysis of their training information and prior annotations. This information not only reflects the real biometrics and behavior patterns of managers, but also takes into account the influence of noise data, so it has a very high reference value. Secondly, when debugging the initial multimodal recognition and authentication algorithm, the management authentication system closely combines these prediction information with the actual authentication process. Through comparison and analysis, the advantages and disadvantages of the algorithm in processing different authentication elements can be found, so as to make targeted adjustments and optimizations. Then, a key goal in the debugging process is to improve the algorithm's ability to handle noise data. In practical applications, noise data is inevitable. Therefore, the management authentication system needs to ensure that the algorithm can still maintain high accuracy and stability under noise interference. Through continuous debugging and optimization, the management authentication system can achieve this goal and make the algorithm more robust. Finally, the debugged multimodal recognition and authentication algorithm will be used to mine the multimodal living body information to be authenticated that contains the target security authentication elements. This information will serve as an important basis for the authorization security authentication of the target jewelry inventory management personnel. Since the algorithm has been deeply debugged and optimized, the management authentication system can ensure the accuracy and reliability of the authentication process. At the same time, since the algorithm can process noisy data, the management authentication system can also reduce misjudgments and missed judgments in practical applications, further improving the overall performance of the authentication system.
[0066] It is worth noting that in the multimodal identification and authentication system, each jewelry inventory manager will have a set of corresponding prior feature variables. These variables are obtained through pre-training and scanning, and constitute an important basis for manager identity authentication. Taking fingerprints and facial features as an example, the fingerprint texture features of a manager may be represented by a vector containing 200 floating-point numbers, such as [0.56, 0.78, 0.23, ..., 0.45], where each floating-point number represents the texture intensity at a specific location on the fingerprint. Similarly, his facial geometric features may also be represented by a vector containing 50 floating-point numbers, such as [0.34, 0.67, 0.55, ..., 0.29], where each floating-point number represents the geometric feature value at a specific location on the face.
[0067] However, in actual applications, due to the influence of various internal and external factors, there may be noisy data in the prior feature variables. To deal with these problems, the system uses the past training information of security authentication elements of jewelry inventory managers to generate feature variable prediction results. These prediction results are obtained by analyzing and mining the training information through advanced machine learning algorithms, and they reflect the biometrics and behavioral patterns of managers in actual operations.
[0068] Taking facial geometry as an example, for a specific manager, the predicted result of his facial geometry may be a new vector containing 50 floating point numbers, such as [1.23, 0.87, 0.95, ..., 0.32]. This prediction result vector is consistent with the prior feature variable vector in dimension, but may be numerically different because it takes into account more actual operation data and the influence of noise factors. By comparing and analyzing the prior feature variables and the feature variable prediction results, the system can more accurately determine the identity of the manager and improve the accuracy and robustness of authentication.
[0069] Therefore, the prior feature variables provide the basic information for manager identity authentication, which is relatively stable and reliable; while the feature variable prediction results take into account the impact of noise data and reflect the real biometrics and behavior patterns of managers as accurately as possible. In the multimodal recognition and authentication system, combining these two concepts can build a more accurate and robust identity authentication system, providing strong security for jewelry inventory management.
[0070] In summary, the embodiment of the present invention first obtains the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm, and identifies the prior annotations of the security authentication elements that may carry noise, thereby enhancing the system's ability to handle noisy data and improving the accuracy and robustness of authentication. Secondly, the security authentication element training information and prior annotations of past jewelry inventory managers are used to generate their security authentication element prediction information, which not only optimizes the authentication algorithm, but also enables the algorithm to better adapt to the changes in biometrics and behavioral patterns in actual operations, further improving the accuracy and reliability of authentication. Finally, by debugging the initial multi-modal recognition and authentication algorithm, a debugged algorithm is obtained, which can more effectively mine the multi-modal liveness information to be authenticated that contains the target security authentication elements, providing jewelry inventory managers with more powerful authority security authentication guarantees.
[0071] In some possible embodiments, the security authentication factor prediction information of the past jewelry inventory manager is obtained based on the security authentication factor training information and the security authentication factor prior annotations of the past jewelry inventory manager, including: for each of the Y authentication factors, if the prior feature variables of the authentication factor contained in the prior annotations of the security authentication factors of the past jewelry inventory manager are abnormal, obtaining the feature variable prediction results of the authentication factor from the security authentication factor training information of the past jewelry inventory manager as the current feature variables of the authentication factor; if the prior feature variables of the authentication factor contained in the prior annotations of the security authentication factors of the past jewelry inventory manager are not abnormal, determining the prior feature variables of the authentication factor as the current feature variables of the authentication factor; and obtaining the security authentication factor prediction information of the past jewelry inventory manager based on the current feature variables of the Y authentication factors.
[0072] It can be understood that this embodiment generates accurate and reliable security authentication factor prediction information for each past jewelry inventory manager by comprehensively considering prior knowledge and actual training data, thereby improving the performance and robustness of the entire multimodal recognition and authentication system.
[0073] First, the above process is carried out independently for each past jewelry inventory manager. For each jewelry inventory manager, the management authentication system will examine the Y authentication factors contained in the prior annotations of their security authentication factors. These Y authentication factors may cover fingerprints, facial features, irises, voices, and specific behavioral patterns, which together form the basis of manager identity authentication.
[0074] Next, for each of the Y authentication factors, the management authentication system will conduct a detailed check. Specifically, if the prior characteristic variable of a certain authentication factor included in the prior annotation of the security authentication factor of the jewelry inventory manager in the past is marked as abnormal, then this means that there may be noise or errors in the prior annotation of this authentication factor. In this case, the management authentication system will turn to the security authentication factor training information of the manager and obtain the characteristic variable prediction result of this authentication factor. This prediction result is obtained by analyzing and mining the training information through advanced machine learning algorithms, and it reflects the real characteristics of this authentication factor in the actual operation of the manager. Therefore, the management authentication system uses this characteristic variable prediction result as the current characteristic variable of this authentication factor for the subsequent generation of security authentication factor prediction information.
[0075] However, if the prior feature variables of a certain authentication factor included in the prior annotations of the security authentication factors of the jewelry inventory manager in the past are marked as non-abnormal, then this means that this authentication factor is relatively accurate and reliable in the prior annotations. In this case, the management authentication system does not need to check the feature variable prediction results in the training information. Instead, the management authentication system will directly determine the prior feature variables of this authentication factor as its current feature variables, and also use them for the generation of subsequent security authentication factor prediction information.
[0076] In this way, the management authentication system can determine the current feature variables of Y authentication factors for past jewelry inventory managers. These current feature variables take into account both the basic information in the prior annotations and the noise and error factors in actual operations. Therefore, they can more accurately reflect the real biometrics and behavior patterns of managers.
[0077] Finally, based on the current feature variables of these Y authentication factors, the management authentication system can generate prediction information of security authentication factors of past jewelry inventory managers through certain algorithms and models. This prediction information is a comprehensive result, which covers the characteristic information of managers in multiple aspects such as fingerprints, facial features, irises, voices, and behavioral patterns. By comparing and analyzing the information in the prior annotations, the system can more accurately determine the identity of the manager, thereby improving the accuracy and robustness of authentication.
[0078] This design generates accurate and reliable security authentication factor prediction information by meticulously checking and processing Y authentication factors of past jewelry inventory managers. This prediction information not only takes into account the basic information in the prior annotations, but also fully considers the noise and error factors in actual operations. Therefore, it can significantly improve the performance and robustness of the multimodal recognition and authentication system, bringing higher security and efficiency to jewelry inventory management.
[0079] In practical applications, this embodiment can be widely used in various scenarios that require identity authentication and authority management. For example, in the field of jewelry inventory management, it can be used to authenticate the identity and authority of managers entering and leaving the warehouse, ensuring that only personnel with corresponding authority can enter the warehouse to operate. At the same time, it can also be used to monitor and record the operating behavior of managers so that alarms and processing can be carried out in time when abnormal situations occur. In addition, this technical solution can also be customized and expanded according to specific application scenarios and requirements to meet the identity authentication and authority management requirements in different fields.
[0080] In addition, this embodiment proposes a new method for predicting security authentication factors by combining prior knowledge and actual training data. This method not only takes into account the basic information in the prior annotations, but also fully considers the noise and error factors in actual operations. Therefore, it can more accurately reflect the real biometrics and behavior patterns of managers and improve the accuracy and robustness of authentication. At the same time, this technical solution also uses advanced machine learning algorithms and models for feature extraction and predictive analysis, further improving the accuracy and efficiency of prediction.
[0081] This design, by comprehensively considering prior knowledge and actual training data to generate accurate and reliable security authentication factor prediction information, provides strong technical support and guarantee for identity authentication and authority management in jewelry inventory management and other fields.
[0082] In other optional embodiments, the method further includes: generating a live scan feature relationship network corresponding to the multi-mode live scan anomaly detection information, the live scan feature relationship network comprising X*Y relationship network nodes, each relationship network node being used to indicate whether a prior feature variable of an authentication factor of a jewelry inventory manager is abnormal.
[0083] Based on this embodiment, the method not only covers the basic live scanning anomaly detection, but also further introduces a generation mechanism of the live scanning feature relationship network. The implementation of this mechanism aims to build a comprehensive, detailed and dynamic feature monitoring network to achieve efficient identification and response to various potential abnormal behaviors in the identity authentication process of jewelry inventory managers.
[0084] First, the live scan feature relationship network is a highly integrated and complex information structure. Its core purpose is to convert multimodal live scan data into a relationship network that is easy to analyze and can quickly locate anomalies. This network consists of X*Y relationship network nodes, where X represents different types of authentication elements (such as fingerprints, facial features, iris information, etc.), and Y represents the number of feature variables that can be subdivided under each authentication element (for example, facial features may include multiple variables such as eye shape, nose size, mouth width, etc.). Through this design, each relationship network node can accurately correspond to a specific feature variable in the identity verification process of jewelry inventory managers, thereby achieving refined and comprehensive anomaly detection.
[0085] Secondly, the construction of the liveness scanning feature relationship network is not static, but a dynamic updating and learning process. Every time a liveness scan is performed, the system will collect and analyze the biometric data of the jewelry inventory manager in real time. These data will then be converted into a series of feature variable values and matched and updated with the corresponding nodes in the relationship network. If a feature variable value deviates from the normal distribution range, that is, it is considered abnormal, then the corresponding relationship network node will be marked as abnormal, thus triggering further abnormality handling procedures.
[0086] Then, in order to improve the accuracy of anomaly detection, the live scan feature network also incorporates advanced machine learning algorithms. These algorithms can automatically learn and adjust the normal feature range of each network node based on historical data, so that the network can adaptively cope with the individual differences of different jewelry inventory managers and physiological changes that may occur over time. At the same time, through continuous learning and optimization, the network can gradually identify and filter out false alarms caused by environmental factors (such as light, angle, etc.), further improving the accuracy of anomaly detection.
[0087] In addition, the liveness scanning feature network also has a strong ability to identify abnormal patterns. In practical applications, a single abnormal feature is often difficult to directly determine as fraud, because there may be multiple reasons for the temporary deviation of a feature variable. However, when multiple network nodes are abnormal at the same time, or certain specific abnormal patterns (such as the combination of abnormalities of certain feature variables) appear frequently in the network, the system can perform a higher level of risk assessment based on these abnormal patterns, thereby more accurately identifying potential fraud or security threats.
[0088] Finally, it is worth mentioning that the liveness scanning feature network is not only a detection tool, but also a management and decision support system. By real-time monitoring of the status of the network, the jewelry inventory management department can intuitively understand the health of the current identity verification process and promptly discover and resolve potential security risks. At the same time, based on the data analysis provided by the network, the management department can also formulate more scientific and effective security strategies, continuously optimize the identity verification process, and ensure the safety and efficient management of jewelry inventory.
[0089] It can be seen that the introduction of the live scanning feature relationship network has brought a new perspective and solution to identity authentication anomaly detection in the field of jewelry inventory management. By building a dynamic, intelligent and comprehensive feature monitoring network, this method not only significantly improves the accuracy and efficiency of anomaly detection, but also provides powerful decision-making support for jewelry inventory management departments, helping them to achieve a safer and more efficient operation model in the context of the new era.
[0090] In a preferred embodiment, the method of obtaining multi-modal liveness scanning anomaly detection information of an initial multi-modal recognition and authentication algorithm includes: obtaining at least two groups of reference prior annotations, each group of reference prior annotations including at least one jewelry inventory manager's prior feature variable for at least one authentication factor; wherein different groups of reference prior annotations contain different authentication factors, and different groups of reference prior annotations correspond to different jewelry inventory managers; and clustering the at least two groups of reference prior annotations based on each jewelry inventory manager to obtain the multi-modal liveness scanning anomaly detection information.
[0091] It should be noted that obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm is a comprehensive task involving multi-dimensional data processing and complex algorithm application. Its core goal is to build an efficient and accurate multi-modal liveness scanning anomaly detection system by deeply mining and analyzing the multi-dimensional feature data generated by jewelry inventory managers during the liveness scanning process. The technical solution of this embodiment particularly emphasizes the key role of clustering processing with reference to prior annotations.
[0092] First, the management authentication system needs to obtain at least two sets of reference prior annotations. These annotations are the prior feature variables provided by jewelry inventory managers for at least one authentication factor when performing live scanning. These feature variables may cover a variety of biometric information such as fingerprints, facial features, iris information, etc., which together constitute the unique identification of the jewelry inventory manager. It is worth noting that the authentication elements contained in different groups of reference prior annotations are different, which means that the management authentication system has comprehensively collected the identity characteristics of jewelry inventory managers from multiple different dimensions. At the same time, the jewelry inventory managers corresponding to each set of reference prior annotations are also different, which ensures the diversity and breadth of the data and provides a solid foundation for subsequent data analysis.
[0093] Next, the management certification system clusters these at least two groups of reference prior annotations, taking each jewelry inventory manager as a benchmark. Clustering is an unsupervised machine learning algorithm that can group data points with similar characteristics into one category, thereby helping the management certification system discover hidden patterns and structures in the data. In this step, the management certification system uses clustering algorithms to deeply analyze and organize the feature variables in the reference prior annotations, clustering those annotations with similar characteristics together to form different clusters.
[0094] The clustering process is an iterative and optimized process. First, the algorithm randomly selects several data points as the initial cluster centers. Then, it calculates the distance between each data point and these cluster centers and assigns each data point to the nearest cluster center. Next, the algorithm recalculates the center of each cluster to make it the mean or median of all the points in the cluster. This process is repeated until the cluster center no longer changes significantly or a predetermined number of iterations is reached.
[0095] Through clustering, the management authentication system can more clearly see the feature differences and commonalities of different jewelry inventory managers during live scanning. Those annotations that are clustered together often represent jewelry inventory managers with similar identity characteristics. Those annotations far away from the cluster center may represent potential anomalies or fraudulent behavior. Therefore, clustering not only helps the management authentication system understand the internal structure of the data, but also provides strong support for subsequent anomaly detection.
[0096] Based on the clustering process, the management authentication system further obtains multi-mode liveness scan anomaly detection information. This information is obtained by in-depth mining and analysis of the clustering results. It reveals which feature variables are within the normal range and which feature variables may be abnormal. For example, if the management authentication system finds that the liveness scan results of a jewelry inventory manager are significantly different from the normal clustering clusters, then the management authentication system can consider that this scan result may be abnormal and needs further verification and review.
[0097] It can be seen that obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm is a complex and delicate process. In this process, clustering processing with reference to prior annotations plays a vital role. It not only helps the management and authentication system understand the inherent structure and pattern of the data, but also provides strong support for subsequent anomaly detection. By deeply mining and analyzing the results of clustering processing, the management and authentication system can more accurately identify potential anomalies and fraudulent behaviors, thereby ensuring the safety and efficient management of jewelry inventory.
[0098] Under some other optional design ideas, the initial multimodal recognition and authentication algorithm is debugged based on the security authentication factor prediction information of the past jewelry inventory managers to obtain the debugged multimodal recognition and authentication algorithm, including: obtaining the multimodal liveness prediction information corresponding to the past jewelry inventory managers based on the past jewelry inventory managers and the security authentication factor prediction information of the past jewelry inventory managers through the initial multimodal recognition and authentication algorithm; determining the algorithm training error of the initial multimodal recognition and authentication algorithm based on the past jewelry inventory managers and the multimodal liveness prediction information, and the algorithm training error is used to characterize the scanning offset weight of the past jewelry inventory managers and the multimodal liveness prediction information; based on the algorithm training error, optimizing the algorithm parameters of the initial multimodal recognition and authentication algorithm to obtain the debugged multimodal recognition and authentication algorithm.
[0099] Under this design idea, the initial multimodal recognition and authentication algorithm is debugged according to the past prediction information of security authentication factors of jewelry inventory managers to obtain an optimized and more accurate implementation method of the multimodal recognition and authentication algorithm. The core of this method is to use past data and prediction information to evaluate and optimize the performance of the algorithm, so as to ensure its accuracy and reliability in practical applications.
[0100] First, the management authentication system needs to obtain the corresponding multimodal live prediction information of these managers through the initial multimodal recognition and authentication algorithm based on the past jewelry inventory managers and their security authentication factor prediction information. This step is the basis for algorithm debugging. It allows the management authentication system to simulate the performance of the algorithm in actual applications and generate a series of prediction data that can be used for subsequent analysis. These multimodal live prediction information includes various biometric data, such as fingerprints, facial features, iris information, etc., which are all key elements that the algorithm needs to analyze and compare during the recognition process.
[0101] Secondly, based on past jewelry inventory managers and their multimodal live prediction information, the management authentication system can determine the algorithm training error of the initial multimodal recognition authentication algorithm. This error is an important indicator that is used to characterize the scan offset weight between past jewelry inventory managers and multimodal live prediction information. In other words, it reflects the accuracy and stability of the algorithm in the prediction process. If the error is large, it means that the algorithm has deviations or deficiencies in some aspects and needs further optimization and adjustment.
[0102] To illustrate this process more specifically, the management authentication system can introduce a mathematical formula to calculate the algorithm training error. The management authentication system has n past jewelry inventory managers. For each manager i, the management authentication system has its actual security authentication factor x_i and the algorithm-predicted multimodal living information y_i. Then, the algorithm training error E can be expressed as: E=Σ(w_i*|x_i-y_i|) / n.
[0103] Among them, w_i represents the scan offset weight, which reflects the importance or representativeness of each manager in the overall data set. |x_i-y_i| represents the absolute difference between the actual value and the predicted value, that is, the magnitude of the error. By calculating the sum of the weighted errors of all managers and dividing it by the total number of managers n, the management certification system can obtain an average error value as a measure of the algorithm training error.
[0104] Then, based on the calculated algorithm training error, the management authentication system can optimize the algorithm parameters of the initial multimodal recognition authentication algorithm. This step is the key to algorithm debugging, which aims to reduce the training error and improve the accuracy and performance of the algorithm by adjusting the parameters and configuration of the algorithm. The optimization process may involve multiple aspects, such as adjusting the feature extraction algorithm, improving the classifier design, optimizing the decision boundary, etc. The specific optimization method can be selected and customized according to the specific implementation and application scenario of the algorithm.
[0105] During the optimization process, the management authentication system can use various machine learning algorithms and techniques, such as gradient descent, random forest, support vector machine, etc. These algorithms can help the management authentication system automatically adjust and optimize the parameters of the algorithm to find the best configuration solution. At the same time, the management authentication system can also use cross-validation, regularization and other technologies to avoid overfitting and underfitting problems, ensuring that the optimized algorithm has good generalization ability and stability.
[0106] Finally, after continuous optimization and adjustment, the management authentication system can obtain a fully debugged multimodal recognition and authentication algorithm. This algorithm has been significantly improved in accuracy and performance, and can better cope with various complex and changing application scenarios. In practical applications, it can help the jewelry inventory management department to more accurately and efficiently identify and verify the identity of jewelry inventory managers, and improve the security and management efficiency of jewelry inventory.
[0107] With this design, the management authentication system can obtain a fully debugged and optimized, more accurate and reliable multimodal recognition and authentication algorithm. This algorithm will play an important role in areas such as jewelry inventory management, providing more advanced and effective technical support for identity recognition and verification.
[0108] In some other preferred embodiments, the algorithm training error includes an authentication factor mining error and a joint training error, and the joint training error includes a first training error, a second training error, a third training error, and a fourth training error; the algorithm training error of the initial multimodal recognition authentication algorithm is determined based on the past jewelry inventory management personnel and the multimodal living body prediction information, including: determining the authentication factor mining error based on the prior feature variables of the prior authentication factor in the prior annotation of the security authentication factor and the feature variable prediction results of the prior authentication factor in the security authentication factor training information; wherein, the authentication factor mining error is used A mining result confidence score characterizing the training information of the security authentication factor is provided, wherein the prior authentication factor is an authentication factor corresponding to a prior feature variable without abnormality in the prior annotation of the security authentication factor; based on the multimodal liveness prediction information and the past jewelry inventory management personnel, the first training error is determined, and the first training error is used to characterize the real-time offset error between the multimodal liveness prediction information and the past jewelry inventory management personnel; based on the security authentication factor prediction information, the discrete evaluation quality of the security authentication factor prediction information, the average feature of the security authentication factor prediction information and the average feature of the security authentication factor training information, the The second training error and the third training error; wherein the second training error is used to characterize the current feature variable of the prior authentication factor in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information, and the third training error is used to characterize the current feature variable of the remaining authentication factors other than the prior authentication factor in the Y types of authentication factors included in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information; the fourth training error is determined according to the depth residual feature map, and the distribution parameters and average characteristics of the thermodynamic relationship spectrum matched by the depth residual feature map, the fourth The training error is used to characterize the noise proportion of the deep residual feature map compared to the thermal relationship spectrum; wherein the deep residual feature map is the deep liveness characterization vector of the past jewelry inventory manager obtained by the fully connected branch of the initial multimodal recognition and authentication algorithm; based on the first training error, the second training error, the third training error and the fourth training error, the joint training error is determined, and the joint training error is used to characterize the matching weight between the multimodal liveness prediction information generated based on the security authentication element prediction information and the past jewelry inventory manager; based on the authentication element mining error and the joint training error, the algorithm training error is determined.
[0109] In this embodiment, the core of obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm lies in the in-depth analysis and accurate calculation of the algorithm training error to ensure that the management authentication system can effectively evaluate and optimize the performance of the initial multi-modal recognition and authentication algorithm.
[0110] First, it is necessary to clarify the composition of the algorithm training error. In an embodiment of the present invention, the algorithm training error is decomposed into two main parts: authentication factor mining error and joint training error. The authentication factor mining error mainly focuses on the confidence score of the mining result of the security authentication factor training information, while the joint training error is further refined into the first training error, the second training error, the third training error and the fourth training error, which are used to characterize different training deviations and weights respectively.
[0111] For the determination of the authentication factor mining error, the management authentication system mainly relies on the prior feature variables of the prior authentication factors in the prior annotations of the security authentication factors, and the prediction results of the feature variables of the prior authentication factors in the security authentication factor training information. By comparing the prior feature variables and the prediction results, the management authentication system can evaluate the confidence score of the mining results of the security authentication factor training information, that is, the authentication factor mining error. This error indicator is crucial for the management authentication system to understand the performance of the algorithm in mining and identifying security authentication factors.
[0112] Next, the management certification system focuses on the determination of the joint training error. The joint training error is composed of four different training errors, which reflect the training bias and weight of the algorithm from different perspectives.
[0113] The first training error is mainly used to characterize the real-time offset error between the multimodal live prediction information and the past jewelry inventory managers. By calculating the difference between the prediction information and the actual manager, the management certification system can evaluate the accuracy of the algorithm in real-time recognition.
[0114] The second training error and the third training error focus more on the scan offset weight of the characteristic variables in the security authentication element prediction information compared to the prediction information. Among them, the second training error is aimed at the current characteristic variables of the prior authentication element, and the third training error is aimed at the current characteristic variables of the remaining authentication elements except the prior authentication element. Through these two error indicators, the management authentication system can gain a deep understanding of the algorithm's performance in processing and identifying different authentication elements.
[0115] Finally, the fourth training error is determined based on the distribution parameters and average characteristics of the deep residual feature map and its matching thermal relationship spectrum. This error indicator is mainly used to characterize the noise ratio of the deep residual feature map compared to the thermal relationship spectrum, thereby helping the management and authentication system to evaluate the performance of the algorithm in feature extraction and noise suppression.
[0116] After determining the four training errors, the management authentication system can further calculate the joint training error. The joint training error is a comprehensive indicator that combines the information of the four training errors and is used to characterize the matching weights of the multimodal live prediction information generated based on the security authentication factor prediction information and the past jewelry inventory managers. This indicator is of great significance for the management authentication system to comprehensively evaluate the training performance and optimization direction of the algorithm.
[0117] Finally, based on the authentication factor mining error and the joint training error, the management authentication system can determine the algorithm training error. This comprehensive error indicator not only reflects the performance of the algorithm in mining and identifying security authentication factors, but also comprehensively considers the performance of the algorithm in real-time recognition, feature processing, noise suppression and other aspects. Through in-depth analysis and optimization of this error indicator, the management authentication system can continuously improve the performance and accuracy of the initial multimodal recognition authentication algorithm, thereby providing more solid technical support for its application in areas such as jewelry inventory management.
[0118] Before we delve into these training errors, it should be noted that each training error carries different dimensions of performance evaluation for the initial multimodal recognition and authentication algorithm.
[0119] First, the first training error is mainly used to measure the real-time deviation between the multimodal live prediction information and the past jewelry inventory manager. In other words, it focuses on the difference between the results predicted by the algorithm and the actual manager status. The management certification system can formalize it into a formula: first training error = function (multimodal live prediction information, past jewelry inventory manager status). This function calculates the difference between the two and gives a value, which represents the deviation between the prediction result and the actual status. A lower first training error means that the algorithm's prediction result is closer to reality and has higher real-time and accuracy.
[0120] Next, we turn to the second training error. This error focuses on the relationship between the current characteristic variables of the prior authentication elements in the security authentication element prediction information and the overall scan offset weight of the prediction information. In other words, it examines the stability and consistency of the characteristic variables of the algorithm when processing the prior authentication elements. The management authentication system can express it as: second training error = function (current characteristic variables of the prior authentication elements, scan offset weights of the security authentication element prediction information). This function evaluates the offset of the characteristic variables of the prior authentication elements in the prediction information, and thus gives a value that reflects its stability. A lower second training error indicates that the algorithm has better stability and consistency when processing prior authentication elements.
[0121] Then, let's look at the third training error. Similar to the second training error, the third training error also focuses on the characteristic variables in the security authentication factor prediction information, but it targets the remaining authentication factors other than the prior authentication factors. The purpose of this error is to evaluate the performance of the algorithm when processing non-prior authentication factors. The management authentication system can formalize it as: third training error = function (current characteristic variables of the remaining authentication factors, scanning offset weights of the security authentication factor prediction information). This function calculates the offset of the characteristic variables of the remaining authentication factors in the prediction information and gives a numerical value to reflect the performance of the algorithm when processing these factors. A lower third training error means that the algorithm also has excellent performance when processing non-prior authentication factors.
[0122] Finally, let's discuss the fourth training error. This error is determined based on the distribution parameters and average features of the deep residual feature map and the thermodynamic relationship spectrum it matches. It focuses on the noise ratio of the deep residual feature map compared to the thermodynamic relationship spectrum, that is, the noise suppression ability of the algorithm during the feature extraction process. It can be expressed as: Fourth training error = function (depth residual feature map, distribution parameters and average features of the thermodynamic relationship spectrum). This function evaluates the noise level in the deep residual feature map and compares it with the thermodynamic relationship spectrum to give a value reflecting the noise ratio. A lower fourth training error indicates that the algorithm has better noise suppression ability during feature extraction and can generate clearer and more accurate feature representations.
[0123] It can be seen that the first training error, the second training error, the third training error and the fourth training error have comprehensively evaluated the performance of the initial multimodal recognition and authentication algorithm from different perspectives. Together, they constitute a complete system of algorithm training errors, which helps to deeply understand the performance of the algorithm in practical applications and guide targeted optimization. By continuously reducing these training errors, the performance and accuracy of the algorithm can be continuously improved, providing more solid technical support for its application in areas such as jewelry inventory management.
[0124] In the next step, based on the continuous correction of the past jewelry inventory manager's security authentication factor training information along with the debugging of the initial multimodal recognition and authentication algorithm, the discrete evaluation quality and average characteristics of the security authentication factor prediction information are also in a state of continuous correction.
[0125] In this limited state, the debugging of the initial multimodal recognition and authentication algorithm is an iterative and optimization process. In this process, the algorithm will continuously receive the security authentication element training information of the past jewelry inventory managers, and self-correct and improve based on this information. This correction not only involves the algorithm's recognition and judgment of security authentication elements, but also includes the adjustment and optimization of the discrete evaluation quality and average characteristics of the security authentication element prediction information.
[0126] Specifically, the discrete evaluation quality of the security certification factor prediction information is an important indicator for measuring the stability and consistency of the prediction information. During the debugging process of the algorithm, the management certification system will pay close attention to the changes in this indicator and make corresponding adjustments to the algorithm based on its feedback. For example, if the discrete evaluation quality shows large fluctuations or inconsistencies, the management certification system may adjust certain parameters in the algorithm or optimize the feature extraction method to improve the stability and accuracy of the prediction information.
[0127] At the same time, the average characteristics of the security certification factor prediction information are also an important focus during the algorithm debugging process. The average characteristics represent the overall level and trend of the prediction information, which can help the management certification system understand the performance of the algorithm when processing large amounts of data. During the debugging process, the management certification system will continue to monitor the changes in the average characteristics and fine-tune the algorithm based on its feedback. For example, if the average characteristics show a trend that deviates from the normal range, the management certification system may adjust certain strategies in the algorithm or add some new constraints to ensure that the prediction information is more in line with the actual situation.
[0128] It is worth noting that the continuous revision of the security authentication factor training information provides valuable data support for algorithm debugging. This training information contains the real data and characteristics of past jewelry inventory managers, and is an important basis for the algorithm to learn and optimize itself. By continuously receiving and analyzing this training information, the algorithm can gradually discover the laws and patterns therein, and adjust its parameters and strategies accordingly to improve its ability to identify and judge security authentication factors.
[0129] In this process, the management and certification system also needs to pay attention to balancing the complexity and performance of the algorithm. On the one hand, the management and certification system hopes that the algorithm can capture and utilize the useful features in the training information as much as possible to improve the accuracy and stability of the prediction; on the other hand, the management and certification system also needs to avoid problems such as low computational efficiency and overfitting caused by overly complex algorithms. Therefore, during the debugging process, the management and certification system needs to continuously simplify and optimize the algorithm to find a balance point that is both accurate and efficient.
[0130] In addition, in order to ensure the smooth progress of algorithm debugging, the management and certification system also needs to establish a complete evaluation system. This evaluation system should include comprehensive collection and analysis of various training information, regular testing and evaluation of algorithm performance, and timely feedback and adjustment of debugging results. Through this evaluation system, the management and certification system can more accurately understand the performance and existing problems of the algorithm, and make targeted optimization and improvements accordingly.
[0131] It can be seen that the debugging process of the initial multimodal recognition and authentication algorithm is a complex and meticulous task. In this process, the management authentication system needs to pay close attention to the changes in the discrete evaluation quality and average characteristics of the security authentication factor prediction information, and adjust and optimize the algorithm accordingly based on its feedback. At the same time, the management authentication system also needs to make full use of the security authentication factor training information of past jewelry inventory managers to provide strong data support for algorithm debugging. Through continuous iteration and optimization, the management authentication system can gradually improve the performance and accuracy of the algorithm, providing a more solid technical foundation for its application in fields such as jewelry inventory management.
[0132] In some further optional embodiments, the initial multimodal recognition and authentication algorithm includes a first data encoding and decoding branch and a second data encoding and decoding branch; the first data encoding and decoding branch is used to generate a depth residual feature map based on the past jewelry inventory managers and the security authentication factor prediction information of the past jewelry inventory managers, and the average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the depth residual feature map are determined by the algorithm parameters of the first data encoding and decoding branch, the past jewelry inventory managers and the security authentication factor prediction information of the past jewelry inventory managers; the second data encoding and decoding branch is used to obtain the multimodal living body prediction information based on the depth residual feature map and the security authentication factor prediction information of the past jewelry inventory managers, and the average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the multimodal living body prediction information are determined by the algorithm parameters of the second data encoding and decoding branch, the depth residual feature map and the security authentication factor prediction information of the past jewelry inventory managers.
[0133] In this embodiment, the initial multimodal recognition and authentication algorithm specially designs two key data encoding and decoding branches: a first data encoding and decoding branch and a second data encoding and decoding branch, which each undertakes different tasks and work together to generate accurate multimodal living body prediction information.
[0134] First, let's focus on the first data encoding and decoding branch. The main responsibility of this branch is to generate deep residual feature maps based on the information of past jewelry inventory managers and their security authentication factor prediction information. This is a complex and delicate process that requires the algorithm to be able to deeply explore and utilize the key features in this data. These feature maps not only reflect the complex relationship between jewelry inventory managers and security authentication factors, but also further describe the nuances of this relationship through the average features and distribution parameters of the thermodynamic relationship spectrum. It is worth noting that these average features and distribution parameters are not fixed, but are jointly determined by the algorithm parameters of the first data encoding and decoding branch, the information of past jewelry inventory managers, and their security authentication factor prediction information. This dynamism ensures the accuracy and adaptability of the feature map, enabling it to better reflect the actual situation.
[0135] Next, we turn to the second data encoding and decoding branch. The task of this branch is to obtain multimodal live prediction information based on the deep residual feature map generated by the first data encoding and decoding branch and the prediction information of security authentication factors of past jewelry inventory managers. This is a higher-level task that requires the algorithm to integrate and utilize information from different sources to generate comprehensive and accurate prediction results. Similar to the first data encoding and decoding branch, the complexity and diversity of multimodal live prediction information are further described by the average characteristics and distribution parameters of the thermodynamic relationship spectrum. However, these feature parameters are jointly determined by the algorithm parameters of the second data encoding and decoding branch, the deep residual feature map, and the prediction information of security authentication factors of past jewelry inventory managers. This multi-dimensional information fusion ensures the accuracy and reliability of the prediction results.
[0136] In detail, the first data encoding and decoding branch generates deep residual feature maps rich in key features by deeply mining and utilizing the information of past jewelry inventory managers and their security authentication factor prediction information. These feature maps not only provide valuable input information for the second data encoding and decoding branch, but also lay a solid foundation for the accuracy and performance of the entire algorithm. After receiving these deep residual feature maps, the second data encoding and decoding branch further integrates and utilizes the security authentication factor prediction information of past jewelry inventory managers. Through complex algorithm processing and calculation, it finally generates comprehensive and accurate multimodal live prediction information. This information not only reflects the complex relationship between jewelry inventory managers and security authentication factors, but also provides strong technical support and guarantee for jewelry inventory management.
[0137] In addition, it is also necessary to note the subtlety of the design of the two data encoding and decoding branches. They each undertake different tasks, but they are interdependent and supportive of each other. The first data encoding and decoding branch provides key input information to the second data encoding and decoding branch, and the second data encoding and decoding branch further processes and utilizes this information to generate the final prediction result. This division of labor and cooperation design not only improves the efficiency and performance of the algorithm, but also enhances its adaptability and robustness.
[0138] In this way, by introducing the design of two data encoding and decoding branches, the initial multimodal recognition and authentication algorithm can not only more accurately identify and judge the security authentication factors of jewelry inventory managers, but also generate more comprehensive and accurate prediction information.
[0139] In a preferred design idea, the security authentication factor training information of past jewelry inventory managers is obtained by a security authentication factor embedding algorithm, and the method also includes: on the basis that the number of personnel in the multi-modal liveness scanning anomaly detection information is greater than a set number, the security authentication factor embedding algorithm adopts a general first long short-term memory algorithm, and in the debugging process of the initial multi-modal recognition and authentication algorithm, the difference between the security authentication factor prior annotations of the past jewelry inventory managers and the security authentication factor training information of the past jewelry inventory managers is used to optimize the algorithm parameters of the first long short-term memory algorithm; or, on the basis that the number of personnel in the multi-modal liveness scanning anomaly detection information is not greater than a set number, the security authentication factor embedding algorithm adopts a non-general second long short-term memory algorithm, and the second long short-term memory algorithm is an AI algorithm with a security authentication factor mining function obtained by pre-debugging using multi-modal liveness scanning detection information, and the multi-modal liveness scanning detection information includes scanning detection information of multiple dimensions and corresponding liveness feature variables.
[0140] Combined with the above content, the past security authentication factor training information of jewelry inventory managers is obtained through the security authentication factor embedding algorithm. The core task of this algorithm is to extract key features related to security authentication from a large amount of data, thereby providing strong support for the subsequent identification and authentication process. Under this design idea, the selection and application of the algorithm is not static, but will be flexibly adjusted according to actual conditions.
[0141] Specifically, when the number of people in the multi-mode liveness scanning anomaly detection information is greater than the set number, the security authentication factor embedding algorithm will choose to use the general first long short-term memory algorithm. The long short-term memory algorithm is a deep learning algorithm that performs well in sequence data processing. It can effectively capture and utilize the temporal features in the data. Here, the first long short-term memory algorithm is used to process the security authentication factor training information of past jewelry inventory managers. By learning and analyzing this data, the algorithm can extract key features related to security authentication and provide support for subsequent identification and authentication processes.
[0142] However, simply using a general long short-term memory algorithm is not enough to meet all needs. In practical applications, the debugging and optimization of the algorithm also need to be considered. Therefore, under this design idea, the difference between the prior annotations and training information of the security authentication elements of past jewelry inventory managers is also used to optimize the algorithm parameters of the first long short-term memory algorithm. This step is crucial because it can ensure that the algorithm has higher accuracy and stability in practical applications. Through continuous debugging and optimization, the first long short-term memory algorithm can be made more suitable for processing the security authentication element training information of jewelry inventory managers, thereby improving the efficiency and accuracy of the entire security authentication process.
[0143] On the other hand, when the number of people in the multi-modal liveness scanning anomaly detection information is not greater than the set number, the security authentication factor embedding algorithm will choose to use the non-universal second long short-term memory algorithm. Unlike the universal first long short-term memory algorithm, the second long short-term memory algorithm is an AI algorithm with security authentication factor mining function obtained by pre-debugging multi-modal liveness scanning detection information. This algorithm not only has the advantages of the long short-term memory algorithm, but also has been specially optimized and debugged for the characteristics of multi-modal liveness scanning detection information. By utilizing scanning detection information in multiple dimensions and corresponding liveness feature variables, the second long short-term memory algorithm can more accurately mine key features related to security authentication and provide strong support for subsequent identification and authentication processes.
[0144] In practical applications, the performance of the second LSTM algorithm is also very outstanding. It can effectively process multi-mode liveness scanning detection information in various complex scenarios and extract valuable security authentication elements from it. Whether in the case of a small number of personnel or in the case of complex scanning detection information, the second LSTM algorithm can maintain high accuracy and stability. This provides strong technical support and guarantee for jewelry inventory management.
[0145] In this way, by flexibly selecting and applying different long-short-term memory algorithms and combining the security authentication element training information of past jewelry inventory managers and multi-mode liveness scanning anomaly detection information, the security authentication process is fully optimized. It not only improves the accuracy and efficiency of security authentication, but also enhances the adaptability of the algorithm to different scenarios.
[0146] In some other examples, the second long short-term memory algorithm includes a residual connection branch and a soft threshold branch; during the debugging process of the initial multimodal recognition and authentication algorithm, the algorithm parameters of the soft threshold branch are optimized by utilizing the difference between the prior annotations of the security authentication factors of the past jewelry inventory managers and the training information of the security authentication factors of the past jewelry inventory managers, and the algorithm parameters of the residual connection branch are maintained unchanged.
[0147] In this example, the second LSTM algorithm specially designs a residual connection branch and a soft threshold branch, each of which undertakes different tasks and works together to achieve effective processing of security authentication factor training information of past jewelry inventory managers.
[0148] First, the residual connection branch plays an important role in the second long short-term memory algorithm. It is responsible for maintaining the stability and continuity of the algorithm during processing. Residual connection, as the name implies, maintains the difference between input and output by introducing residual terms, thereby ensuring that the algorithm does not lose important features when processing complex information. In this example, the algorithm parameters of the residual connection branch remain unchanged during the debugging process of the initial multimodal recognition and authentication algorithm. This means that no matter how the input security authentication element training information changes, the residual connection branch can maintain its stable processing performance, providing strong support for the overall stability and accuracy of the algorithm.
[0149] Next, the task of the soft threshold branch is to optimize the algorithm parameters based on the difference between the prior annotations of the security authentication factors of past jewelry inventory managers and the training information. Soft threshold, as a common mathematical tool, is cleverly applied here in the algorithm optimization process. By adjusting the parameters of the soft threshold, the algorithm can adapt to different input information more flexibly and extract more accurate security authentication factor features. During the debugging process of the initial multimodal recognition and authentication algorithm, the algorithm parameters of the soft threshold branch will be continuously optimized and adjusted to ensure that the algorithm can achieve the best performance when processing the security authentication factor training information of past jewelry inventory managers.
[0150] In detail, during the debugging process of the initial multimodal recognition and authentication algorithm, the residual connection branch maintained its stable processing performance, providing a solid foundation for the overall stability and accuracy of the algorithm. The soft threshold branch flexibly adjusted and optimized according to the difference in input information to extract more accurate security authentication factor features. The collaborative work of these two branches enables the second long short-term memory algorithm to perform well when processing the security authentication factor training information of past jewelry inventory managers.
[0151] In addition, it is also necessary to note the optimization strategy in this example. During the debugging process of the initial multimodal recognition and authentication algorithm, the optimization strategy mainly focused on optimizing the algorithm parameters of the soft threshold branch. This is because the soft threshold branch has greater flexibility and adaptability, and can adapt to different input information by adjusting parameters. The residual connection branch, on the other hand, maintains the algorithm parameters unchanged due to its stable processing performance. This targeted optimization strategy not only improves the efficiency of the algorithm, but also enhances its adaptability and robustness.
[0152] Finally, it is important to emphasize the technological innovation and progress demonstrated by this example. By introducing the design of residual connection branches and soft threshold branches, the second long short-term memory algorithm can not only more accurately process and identify the security authentication element training information of past jewelry inventory managers, but also flexibly adjust and optimize according to the differences in input information. This provides strong technical support and guarantee for jewelry inventory management, and also brings new opportunities and challenges to the development of related fields.
[0153] In some independent embodiments, the method also includes: based on the debugged multimodal recognition and authentication algorithm, mining the multimodal living information to be authenticated that contains the target security authentication elements corresponding to the target jewelry inventory manager; and performing authority security authentication on the target jewelry inventory manager based on the multimodal living information to be authenticated.
[0154] In this embodiment, it involves how to mine the security authentication elements corresponding to the target jewelry inventory management personnel based on the multimodal recognition and authentication algorithm, and perform the authority security authentication process accordingly. This embodiment not only improves the security of jewelry inventory management, but also ensures the accuracy and efficiency of authority authentication.
[0155] First, the core of this embodiment lies in the application of a multimodal recognition and authentication algorithm. Multimodal recognition, as the name implies, is to integrate a variety of different biometric features or behavioral patterns for identity recognition. Compared with traditional single-modal recognition, multimodal recognition has higher accuracy and robustness because it can integrate multiple sources of information to more comprehensively characterize a person's identity characteristics. In this embodiment, the management authentication system uses the multimodal recognition and authentication algorithm that has been debugged to mine the multimodal living information to be authenticated corresponding to the target jewelry inventory manager that contains the target security authentication elements.
[0156] Specifically, the mining process first requires the collection of multimodal information on the target jewelry inventory manager. This information may include multiple biometric features such as facial images, fingerprints, irises, and voiceprints, as well as non-biological features such as behavioral patterns and operating habits. The collected multimodal information will be input into the multimodal recognition and authentication algorithm, which will conduct a comprehensive analysis and processing of this information to extract key features related to the identity of the target jewelry inventory manager.
[0157] In this process, the algorithm will pay special attention to those features related to the target security authentication elements. Target security authentication elements refer to those features that can uniquely identify the identity of a jewelry inventory manager, such as specific facial contours, fingerprint patterns, iris textures, etc. By mining these features, the algorithm can construct a multimodal living body information library to be authenticated that contains the target security authentication elements. This information library will serve as the basis for subsequent permission security authentication.
[0158] After completing the mining of the multimodal living body information to be authenticated, the next step is to perform authority security authentication on the target jewelry inventory management personnel based on this information. Authority security authentication is a crucial part of jewelry inventory management, which ensures that only personnel with corresponding authority can access and operate the jewelry inventory. In this embodiment, the management authentication system uses the mined multimodal living body information to be authenticated to build an efficient authority security authentication system.
[0159] Specifically, when the target jewelry inventory manager attempts to access the jewelry inventory, the system will first ask him to provide multimodal live information. This information may include biometric information such as facial images and fingerprints, as well as non-biometric information such as behavioral patterns. The provided information will be input into the permission security authentication system, and the system will compare and match this information with the previously mined multimodal live information database to be authenticated.
[0160] During the comparison and matching process, the system will comprehensively consider multiple information sources and feature elements to ensure the accuracy and robustness of authentication. For example, the system will check whether the provided facial image matches the facial contour in the multimodal living body information database to be authenticated; it will verify whether the provided fingerprint is consistent with the fingerprint pattern in the multimodal living body information database to be authenticated; it will also analyze whether the provided behavior pattern is consistent with the behavior habits in the multimodal living body information database to be authenticated. Only when all of this information is matched successfully will the system consider the target jewelry inventory manager to be legitimate and grant him or her corresponding access and operation permissions.
[0161] The advantage of this embodiment is that it fully utilizes the advantages of multimodal recognition, and performs identity recognition and authorization authentication by integrating multiple different biometric features and behavior patterns. This comprehensive recognition method not only improves the accuracy of authentication, but also enhances the robustness and anti-attack capabilities of the system. Because even if a certain biometric feature or behavior pattern is forged or imitated, the system can still identify the true identity through other features or patterns.
[0162] In addition, in actual applications, the management authentication system can select appropriate biometrics and behavior patterns for identification and authentication based on specific needs and scenarios. For example, in scenarios with higher security requirements, the management authentication system can select more biometrics for identification and authentication; while in scenarios with higher convenience requirements, the management authentication system can select fewer biometrics for identification and authentication. At the same time, with the continuous development and advancement of technology, the management authentication system can also incorporate more biometrics and behavior patterns into this technical solution to further improve the accuracy and efficiency of identification and authentication.
[0163] This design uses a multimodal recognition and authentication algorithm to mine the security authentication elements corresponding to the target jewelry inventory management personnel, and performs authority security authentication based on this. It not only improves the security of jewelry inventory management, but also ensures the accuracy and efficiency of authority authentication. In future research and practice, the management authentication system can further explore and optimize the application and implementation methods of this technical solution to promote the continuous development and progress in the field of jewelry inventory management. For example, the management authentication system can try to introduce more deep learning technologies and machine learning algorithms to improve the performance and accuracy of the multimodal recognition and authentication algorithm; at the same time, the management authentication system can also combine actual application scenarios and needs to carry out more refined design and optimization of the technical solution to meet the security authentication needs in different scenarios.
[0164] When deeply discussing the process of security authentication of the target jewelry inventory management personnel based on the multimodal live information to be authenticated, it can be regarded as a multi-dimensional matching and verification task, the core of which is to accurately compare the multimodal live information collected in real time with the data in the pre-stored multimodal live information database to be authenticated, so as to determine whether the person currently trying to access the jewelry inventory has legal authority. In this process, the numerical feature vector, as a key data representation form, carries the abstraction of multimodal information in mathematical space, providing a computable and comparable basis for the algorithm.
[0165] First, multimodal liveness information covers a variety of biometric features and behavioral patterns. Taking facial images and fingerprints as examples, facial images can be converted into a set of numerical feature vectors that describe key facial features, such as eye position, nose shape, mouth size, etc., through facial recognition algorithms. These features can form a vector, such as [x1, y1, z1, ...], where x1, y1, z1, etc. represent different facial feature parameters. Similarly, fingerprint information can also be extracted through fingerprint recognition algorithms to extract fingerprint lines, nodes, bifurcations and other features, and converted into numerical vectors, such as [a1, b1, c1, ...], where a1, b1, c1, etc. represent different feature parameters of fingerprints.
[0166] When the target jewelry inventory manager attempts to access the jewelry inventory, the system will ask him to provide multimodal liveness information. This information is then converted into a numerical feature vector and compared with the data in the multimodal liveness information database to be authenticated. The multimodal liveness information database to be authenticated is a pre-built database that stores the numerical feature vectors of the multimodal liveness information of all legitimate jewelry inventory managers. For example, the facial feature vector of legitimate person A may be [xA, yA, zA, ...], and the fingerprint feature vector may be [aA, bA, cA, ...].
[0167] When performing the comparison, the system calculates the similarity between the numerical feature vector of the multimodal liveness information provided by the person currently trying to access and the numerical feature vector of each legitimate person in the multimodal liveness information database to be authenticated. Taking the facial feature vector as an example, the system can use the Euclidean distance to calculate the similarity between the two vectors. The calculation formula of the Euclidean distance is √[(x1-xA)^2+(y1-yA)^2+(z1-zA)^2+...], where (x1, y1, z1, ...) is the facial feature vector of the person currently trying to access, and (xA, yA, zA, ...) is the facial feature vector of the legitimate person A. The smaller the Euclidean distance, the more similar the two vectors are. In addition to the Euclidean distance, the system can also use cosine similarity to calculate the similarity between two vectors.
[0168] In actual applications, in order to improve the accuracy and robustness of permission security authentication, the system will combine a variety of different similarity calculation methods and thresholds for comprehensive judgment. For example, the system can set the Euclidean distance threshold of the facial feature vector to T1 and the Euclidean distance threshold of the fingerprint feature vector to T2. Only when the Euclidean distance between the facial feature vector provided by the person currently trying to access and the facial feature vector of a legitimate person is less than T1, and the Euclidean distance between the fingerprint feature vectors is less than T2, will the system consider the person currently trying to access to be legitimate and grant him the corresponding access and operation permissions.
[0169] In addition, in order to further improve the accuracy and robustness of permission security authentication, the system can also pre-process and quality control the collected multi-modal living information. For example, for facial images, the system can perform illumination normalization, size normalization and other processing to improve the accuracy of facial recognition; for fingerprint information, the system can perform denoising, enhancement and other processing to improve the accuracy of fingerprint recognition.
[0170] In addition to traditional mathematical methods, machine learning algorithms and deep learning algorithms are also widely used in feature extraction and classification of multimodal living information. These algorithms can automatically extract the most useful features through learning and training, and build more accurate and efficient classification models. For example, convolutional neural networks (CNNs) can be used for feature extraction and classification of facial images. By training a large number of facial image data sets, CNNs can automatically learn the key features in facial images and convert them into efficient numerical feature vectors. Similarly, recurrent neural networks (RNNs) can be used for feature extraction and classification of behavioral patterns. By capturing key information in time series data, RNNs can effectively identify different behavioral patterns and convert them into numerical feature vectors.
[0171] In this way, by combining a variety of different biometrics and behavioral patterns, as well as advanced machine learning algorithms and deep learning algorithms, a more accurate and efficient permission security authentication system can be built. For example, when the target jewelry inventory manager attempts to access the jewelry inventory, the system can require him to provide a variety of multimodal live information such as facial images, fingerprint information, and operating habits. Then, the system can use CNN to extract and classify facial images, use traditional fingerprint recognition algorithms to extract and classify fingerprint information, and use RNN to extract and classify operating habits. Finally, the system compares and verifies these numerical feature vectors with the data in the multimodal live information database to be authenticated, so as to determine whether the person currently trying to access is a legitimate person, and grant him corresponding access and operating permissions accordingly.
[0172] In summary, the authorization security authentication of the target jewelry inventory management personnel based on the multimodal living body information to be authenticated involves multiple links such as the collection, conversion, comparison and verification of various types of biometrics and behavior patterns. In this process, the numerical feature vector, as a key data representation form, carries the abstraction of multimodal information in mathematical space, providing a computable and comparable basis for the algorithm. By combining a variety of different similarity calculation methods and thresholds for comprehensive judgment, and using advanced machine learning algorithms and deep learning algorithms for feature extraction and classification, a more accurate and efficient authorization security authentication system can be constructed, providing more reliable and secure protection for jewelry inventory management.
[0173] Further, Figure 2 FIG. 2 is a schematic diagram of a management authentication system 200 provided in an embodiment of the present invention. Figure 2 The management authentication system 200 shown includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method in the embodiment of the present invention.
[0174] Alternatively, if Figure 2 As shown, the management authentication system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention.
[0175] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .
[0176] Alternatively, if Figure 2 As shown, the management authentication system 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices.
[0177] Optionally, the management authentication system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present invention, which will not be described in detail here for the sake of brevity.
[0178] It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capability.
[0179] It is understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.
[0180] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.
[0181] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.
[0182] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), including several instructions for a terminal (which can be a mobile phone, computer, server, or network equipment, etc.) to execute the methods described in each embodiment of the present invention. The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are only illustrative and not restrictive. Under the enlightenment of the present invention, those skilled in the art can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which belong to the protection of the present invention.
Claims
1. A jewelry inventory management authentication method based on multimodal data recognition, characterized in that: The method is applied to a management authentication system, and the method comprises: Acquire multimodal liveness scanning anomaly detection information of an initial multimodal recognition and authentication algorithm, wherein the multimodal liveness scanning anomaly detection information includes X groups of security authentication element prior annotations, each group of security authentication element prior annotations includes a jewelry inventory manager's prior feature variables for Y authentication elements, and there is at least one group of security authentication element prior annotations carrying noise in the X groups of security authentication element prior annotations, and there is at least one prior feature variable of the authentication element abnormal in the prior feature variables of the Y authentication elements included in the prior feature variables of the security authentication element carrying noise, and X and Y are both positive integers; For any past jewelry inventory manager, if the security authentication factor prior annotation corresponding to the past jewelry inventory manager in the multi-mode liveness scanning anomaly detection information carries noise, then based on the security authentication factor training information and the security authentication factor prior annotation of the past jewelry inventory manager, the security authentication factor prediction information of the past jewelry inventory manager is obtained; wherein the security authentication factor training information of the past jewelry inventory manager includes: the identified prediction result of the feature variable of the past jewelry inventory manager for at least one authentication factor; According to the security authentication factor prediction information of the past jewelry inventory managers, the initial multimodal recognition and authentication algorithm is debugged to obtain a debugged multimodal recognition and authentication algorithm; the debugged multimodal recognition and authentication algorithm is used to mine the multimodal living body information to be authenticated that contains the target security authentication factor, and the multimodal living body information to be authenticated is used to perform authority security authentication on the target jewelry inventory manager, and there is a corresponding relationship between the target jewelry inventory manager and the multimodal living body information to be authenticated.
2. The method according to claim 1, characterized in that The method of obtaining the prediction information of the safety authentication elements of the past jewelry inventory managers based on the safety authentication element training information and the safety authentication element prior annotations of the past jewelry inventory managers includes: For each of the Y authentication factors, if the prior feature variable of the authentication factor contained in the prior annotation of the security authentication factor of the past jewelry inventory manager is abnormal, obtaining the feature variable prediction result of the authentication factor from the security authentication factor training information of the past jewelry inventory manager as the current feature variable of the authentication factor; If the prior characteristic variables of the authentication factor included in the prior annotations of the security authentication factor of the past jewelry inventory manager are not abnormal, determining the prior characteristic variables of the authentication factor as the current characteristic variables of the authentication factor; Based on the current characteristic variables of the Y authentication factors, the security authentication factor prediction information of the past jewelry inventory managers is obtained.
3. The method according to claim 2, characterized in that The method further comprises: Generate a liveness scanning feature relationship network corresponding to the multi-mode liveness scanning anomaly detection information, wherein the liveness scanning feature relationship network includes X*Y relationship network nodes, and each relationship network node is used to indicate whether a priori feature variable of an authentication factor of a jewelry inventory manager is abnormal.
4. The method according to claim 1, characterized in that: The step of obtaining the multi-modal liveness scanning anomaly detection information of the initial multi-modal recognition and authentication algorithm includes: Obtain at least two groups of reference prior annotations, each group of reference prior annotations includes at least one prior characteristic variable of a jewelry inventory manager for at least one authentication factor; wherein the authentication factors included in different groups of reference prior annotations are different, and different groups of reference prior annotations correspond to different jewelry inventory managers; The at least two groups of reference prior annotations are clustered based on each jewelry inventory manager to obtain the multi-mode liveness scanning anomaly detection information.
5. The method according to claim 1, characterized in that The method of debugging the initial multimodal recognition and authentication algorithm based on the past security authentication factor prediction information of the jewelry inventory manager to obtain a debugged multimodal recognition and authentication algorithm includes: Obtaining multimodal liveness prediction information corresponding to the past jewelry inventory manager based on the past jewelry inventory manager and the security authentication factor prediction information of the past jewelry inventory manager through an initial multimodal recognition and authentication algorithm; Based on the past jewelry inventory manager and the multimodal liveness prediction information, determining an algorithm training error of an initial multimodal recognition and authentication algorithm, wherein the algorithm training error is used to characterize a scanning offset weight of the past jewelry inventory manager and the multimodal liveness prediction information; Based on the algorithm training error, the algorithm parameters of the initial multimodal recognition and authentication algorithm are optimized to obtain the debugged multimodal recognition and authentication algorithm.
6. The method according to claim 5, characterized in that The algorithm training error includes an authentication factor mining error and a joint training error, and the joint training error includes a first training error, a second training error, a third training error and a fourth training error; The step of determining the algorithm training error of the initial multimodal recognition and authentication algorithm based on the past jewelry inventory management personnel and the multimodal liveness prediction information includes: Determine the authentication factor mining error based on the prior feature variables of the prior authentication factor in the prior annotation of the security authentication factor and the feature variable prediction result of the prior authentication factor in the security authentication factor training information; wherein the authentication factor mining error is used to characterize the mining result confidence score of the security authentication factor training information, and the prior authentication factor is the authentication factor corresponding to the prior feature variable without abnormality in the prior annotation of the security authentication factor; Based on the multimodal living body prediction information and the past jewelry inventory management personnel, determining the first training error, the first training error being used to characterize a real-time offset error between the multimodal living body prediction information and the past jewelry inventory management personnel; The second training error and the third training error are determined based on the security authentication factor prediction information, the discrete evaluation quality of the security authentication factor prediction information, the average feature of the security authentication factor prediction information, and the average feature of the security authentication factor training information; wherein the second training error is used to characterize the current feature variable of the prior authentication factor in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information, and the third training error is used to characterize the current feature variable of the remaining authentication factors other than the prior authentication factor among the Y types of authentication factors included in the security authentication factor prediction information compared to the scanning offset weight of the security authentication factor prediction information; According to the depth residual feature map, and the distribution parameters and average characteristics of the thermodynamic relationship spectrum matched by the depth residual feature map, the fourth training error is determined, and the fourth training error is used to characterize the noise proportion of the depth residual feature map compared to the thermodynamic relationship spectrum; wherein the depth residual feature map is the deep liveness characterization vector of the past jewelry inventory manager obtained by the fully connected branch of the initial multimodal recognition and authentication algorithm; Based on the first training error, the second training error, the third training error and the fourth training error, determining the joint training error, wherein the joint training error is used to characterize the matching weight between the multimodal liveness prediction information generated based on the security authentication element prediction information and the past jewelry inventory manager; The algorithm training error is determined based on the authentication factor mining error and the joint training error.
7. The method according to claim 6, characterized in that Based on the continuous correction of the past jewelry inventory manager's security authentication factor training information along with the debugging of the initial multimodal recognition and authentication algorithm, the discrete evaluation quality and average characteristics of the security authentication factor prediction information are also in a state of continuous correction.
8. The method according to claim 5, characterized in that The initial multimodal identification and authentication algorithm includes a first data encoding and decoding branch and a second data encoding and decoding branch; The first data encoding and decoding branch is used to generate a depth residual feature map based on the past jewelry inventory management personnel and the security authentication factor prediction information of the past jewelry inventory management personnel, and the average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the depth residual feature map are determined by the algorithm parameters of the first data encoding and decoding branch, the past jewelry inventory management personnel and the security authentication factor prediction information of the past jewelry inventory management personnel; The second data encoding and decoding branch is used to obtain the multimodal liveness prediction information based on the depth residual feature map and the past security authentication factor prediction information of the jewelry inventory manager. The average characteristics and distribution parameters of the thermodynamic relationship spectrum matched by the multimodal liveness prediction information are determined by the algorithm parameters of the second data encoding and decoding branch, the depth residual feature map and the past security authentication factor prediction information of the jewelry inventory manager.
9. The method according to claim 1, characterized in that: The security authentication factor training information of the past jewelry inventory managers is obtained by a security authentication factor embedding algorithm. The method further includes: on the basis that the number of personnel in the multi-mode liveness scanning anomaly detection information is greater than a set number, the security authentication factor embedding algorithm adopts a general first long short-term memory algorithm, and in the debugging process of the initial multi-modal recognition and authentication algorithm, the algorithm parameters of the first long short-term memory algorithm are optimized by using the difference between the security authentication factor prior annotations of the past jewelry inventory managers and the security authentication factor training information of the past jewelry inventory managers; or, on the basis that the number of personnel in the multi-mode liveness scanning anomaly detection information is not greater than a set number, the security authentication factor embedding algorithm adopts a non-general second long short-term memory algorithm, the second long short-term memory algorithm is an AI algorithm with a security authentication factor mining function obtained by pre-debugging using multi-mode liveness scanning detection information, and the multi-mode liveness scanning detection information includes scanning detection information of multiple dimensions and corresponding liveness feature variables; The second long short-term memory algorithm includes a residual connection branch and a soft threshold branch; during the debugging process of the initial multimodal recognition and authentication algorithm, the algorithm parameters of the soft threshold branch are optimized by utilizing the difference between the prior annotations of the security authentication factors of the past jewelry inventory managers and the training information of the security authentication factors of the past jewelry inventory managers, and the algorithm parameters of the residual connection branch are maintained unchanged.
10. A management authentication system, characterized in that: The method comprises a memory and at least one processor; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.
