A method for collaborative calling of character recognition information based on artificial intelligence
By performing homologous and cross-source similarity calculations in a multi-camera, multi-terminal environment, constructing a locally consistent relational domain and performing bidirectional confirmation, the problem of erroneous candidates being carried into subsequent stages is solved, achieving stable and traceable relational face recognition results and improving the reliability and consistency of judgment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUTURE MAN (XIAMEN) ARTIFICIAL INTELLIGENCE CO LTD
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-31
AI Technical Summary
In environments with multiple cameras, multiple terminals, and multiple data sources, existing technologies for facial recognition are prone to carrying erroneous candidates into subsequent stages. Furthermore, there is a lack of systematic utilization of the relationship between shared identity determination, source-specific determination, and source attribute determination, making it difficult to form stable, traceable, and reusable relational conclusions in the determination results.
By performing similarity calculations on upstream information sources from the same source dual-path and cross-source information acquisition channels, a local consistency relation domain is constructed for the same source and a local consistency relation domain for the cross-source. Two-way confirmation is performed to establish a formal same-source confirmation relation domain and a formal cross-source confirmation relation domain. The optimal identity category is selected as the exclusive identity determination, and a consensus synthesis is performed to generate a consensus level and a formal consensus order. Multi-level screening and progressive call expansion are then carried out.
It improves the reliability and anti-interference capability of subsequent collaborative judgment, forms a stable judgment basis for continued progressive calls, transforms scattered identification results into traceable relational call results, and enhances the consistency of subsequent link verification and continuous calls.
Smart Images
Figure CN122490152A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information collaborative retrieval technology, and in particular to a method for collaborative retrieval of person recognition information based on artificial intelligence. Background Technology
[0002] In real-world environments where multiple cameras, terminals, and data sources coexist, facial recognition and related technologies are becoming increasingly widespread. Furthermore, in processing personal information, cross-source recognition, candidate ranking, identity determination, and result fusion are being introduced to improve recognition efficiency while ensuring privacy. A common approach in existing solutions is to first extract personal recognition information from upstream information acquisition channels and then compare the similarity of candidate objects.
[0003] From the perspective of technological development trends, although existing related technologies can calculate similarity and ranking, they often only look at a single path or a single ranking result. This makes it easy to regard a candidate that happens to be ranked first as a reliable result. Once the quality of upstream information fluctuates, the source is different, or there is significant local interference, it is easy to carry erroneous candidates into subsequent stages. In addition, there is a lack of systematic utilization of the relationship between shared identity determination, source-specific determination, and source attribute determination. As a result, although the determination result may appear to have a score, it is difficult to explain whether it is formed by joint support or by a single path that has boosted it. Therefore, it is difficult to form stable, traceable, and reusable relational conclusions. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a collaborative method for invoking person recognition information based on artificial intelligence, which solves the problem that existing technologies easily carry erroneous candidates into subsequent stages. In addition, it lacks a systematic utilization of the relationship between shared identity determination, source-specific determination, and source attribute determination, making it difficult to form stable, traceable, and reusable relational conclusions in the determination results.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for collaboratively accessing person recognition information based on artificial intelligence, comprising, For the upstream information sources of the character recognition information acquisition path, the recognition information and candidate recognition information of the same upstream information source are calculated for the same source dual-path similarity. At the same time, for the cross-source candidate recognition information of different upstream information sources, the cross-source similarity under different paths is calculated, and the same source initial local consistency relation domain and cross-source local consistency relation domain are constructed respectively. Bidirectional confirmation is performed on the initial local consensus relation domains of the same source and the local consensus relation domains of the cross-source. The joint sort value is calculated and the formal same-source confirmation relation domains and the formal cross-source confirmation relation domains are formed according to the corresponding joint sort value. Construct a shared identity original score vector, select the optimal identity category as the exclusive identity determination, and establish a determination collaboration record for the three determination paths; The consensus synthesis process is executed, the maximum joint identity probability is calculated, and a consensus level, consensus identity, and formal consensus order are generated for each candidate to form a formal consensus sequence. Based on the formal consensus sequence, candidate objects are screened in multiple layers and grouped into several identity relationship clusters according to consensus identity. The most stable main relationship cluster is selected and sent to the final link to complete the progressive call expansion and form a relational call conclusion as the output result.
[0007] As a preferred embodiment of the AI-based collaborative invocation method for character recognition information according to the present invention, wherein: the construction of the same-source initial local consistency relation domain and the cross-source local consistency relation domain respectively includes; A similarity ranking sequence is constructed for the similarity of two paths with the same origin. Candidate identification information is used as candidate objects, and a candidate ranking is constructed based on the similarity ranking of each candidate object. At the same time, the difference in the ranking values of the candidate rankings in the two paths is considered. A joint ranking value with the same origin is defined, and the mean of all joint ranking values with the same origin is calculated to define the initial local consistent relation domain. A ranking sequence is constructed for cross-source similarity. For each candidate ranking, the ranking values of different pathways and whether the ranking sequences of different pathways for cross-source candidate identification information are consistent are considered simultaneously. The cross-source joint ranking value is calculated, and the mean of all calculated cross-source joint ranking values is calculated to generate a cross-source locally consistent relation domain.
[0008] As a preferred embodiment of the artificial intelligence-based collaborative invocation method for character recognition information described in this invention, the step of bidirectionally confirming the initial local consistency relation domain and the cross-source local consistency relation domain includes, based on the initial local consistency relation domain and the cross-source local consistency relation domain, taking the candidate recognition information of the same upstream information source as the query object. For each query object, reverse local relation domains and reverse cross-source relation domains are constructed. The similarity, ranking, and joint ranking values of the two paths are recalculated based on the query object. The mean of the corresponding joint ranking values is calculated to construct reverse initial local consistent relation domains and reverse cross-source initial local consistent relation domains respectively.
[0009] As a preferred embodiment of the collaborative invocation method for character recognition information based on artificial intelligence described in this invention, the formation of the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain includes: determining the same-source mutual selection mark for each candidate recognition information of the same upstream information source; if the character recognition information falls into the reverse initial local consistency relation domain of the candidate recognition information, then the same-source mutual selection mark is performed to form the formal same-source confirmation relation domain. Simultaneously, if the person identification information falls into the reverse cross-source initial local consistency relation domain of the candidate identification information, cross-source mutual selection labeling is performed; the same-source bridging support number is calculated, and the bridging support mean is calculated for all candidates in the cross-source initial relation domain; For candidate identification information from cross-source upstream information sources, multi-level information filtering is performed to retain candidate identification information that has completed cross-source mutual selection marking, while also retaining candidate identification information whose same-source bridging support number is not lower than the average bridging support. Based on the filtering results, a formal cross-source confirmation ranking value is defined by the same-source bridging support number and the cross-source joint ranking value, and a formal cross-source confirmation relation domain is formed.
[0010] As a preferred embodiment of the AI-based collaborative retrieval method for person recognition information described in this invention, the step of selecting the optimal identity category as the exclusive identity determination includes: Based on the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, a unified candidate set is formed, and the ranking values of the candidate objects in the unified candidate set in the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain are recorded. The sum of the ranking values of the candidate objects in the same-source confirmation relation domain and the ranking values in the cross-source confirmation relation domain is used as the initial cooperative ranking value of the candidate objects. For each candidate object in the unified candidate set, the shared identity determination path, the source-specific identity determination path, and the source attribute determination path are respectively called up. The shared identity determination path obtains the original score of the identity category to which the candidate object belongs through the upstream information source and outputs the shared identity original score vector. The source-specific identity determination pathway selects a determination model based on the actual source category of the object to be confirmed. It calculates the original score vector through normalized probability to obtain the source-specific identity determination probability vector, and selects the maximum determination probability vector as the optimal identity category, which is used as the exclusive identity determination. The maximum probability corresponding to the corresponding category is recorded.
[0011] As a preferred embodiment of the artificial intelligence-based collaborative retrieval method for person recognition information described in this invention, the establishment of a collaborative determination record for three determination paths includes: The original score vector of shared identity is transformed into a shared identity determination probability vector, and the original score vector of source-specific identity determination is transformed into a source-specific identity determination probability vector. The original score vector of the source attribute is converted into the source determination probability, and the optimal category of shared identity, the optimal category of source-specific identity, and the optimal category of source attribute are determined respectively, thereby forming a determination collaboration record corresponding to each candidate object.
[0012] As a preferred embodiment of the AI-based collaborative invocation method for person recognition information described in this invention, the calculation of the maximum joint identity probability includes: for each candidate object in the unified candidate set, calculating the joint original score of the shared identity determination and the source-specific identity determination, converting the joint original score into a joint identity probability, and selecting the maximum determination probability vector as the optimal category of joint identity; for each candidate object in the unified candidate set, calculating the consistency of the joint identity determination with the shared identity determination using the cross-entropy algorithm.
[0013] As a preferred embodiment of the AI-based collaborative invocation method for person recognition information described in this invention, the step of generating a consensus level, consensus identity, and formal consensus order for each candidate object to form a formal consensus sequence includes: Determine whether the optimal category of shared identity is consistent with the optimal category of source-specific identity, whether the optimal category of joint identity is consistent with the optimal category of shared identity, and whether the optimal category of source attribute is consistent with the actual source category of the current object to be confirmed. Based on this, generate three consistency flags and a total consistency count. Candidates are divided into four categories—strong consensus, medium consensus, weak consensus, and veto—based on the total number of consensuses. The optimal category of joint identity is used as the consensus identity of non-veto candidates. The candidates are then sorted in order of consensus level, total number of consensuses, consensus quantity, maximum joint identity probability, and initial coordination order to form a formal consensus sequence.
[0014] As a preferred embodiment of the AI-based collaborative invocation method for character recognition information described in this invention, wherein: the step of selecting the most stable master-relationship cluster and sending it into the final link includes, Based on the formal consensus sequence, a first-level screening is performed on all candidate objects, and the non-rejected candidate objects constitute the first-level callable candidate set. Based on consensus identity, candidate objects in the first-level callable candidate set are merged to construct identity relationship clusters, thereby obtaining the second-level callable relationship cluster set and the second-level callable candidate set. In the second-level callable relationship cluster set, the main relationship cluster is selected by comparing the number of members, and the candidate objects in the main relationship cluster form the third-level callable candidate set. The main representative candidate is selected by comparing the order in the formal consensus sequence, thus forming the progressive link expansion call result.
[0015] As a preferred embodiment of the AI-based collaborative invocation method for character recognition information described in this invention, the step of completing the progressive invocation expansion and forming a relational invocation conclusion as the output result includes: combining the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain to determine the source relation category for each candidate object; and combining consensus identity, formal consensus sequence order, link affiliation category, source relation category, consensus level, consistency quantity, and initial collaborative order to form a candidate-level relational conclusion table, which together constitutes the relational invocation conclusion output result.
[0016] The beneficial effects of this invention are as follows: By continuously connecting the same-source relation domain and the cross-source relation domain, the originally scattered identification results can be transformed into a stable set of relations that can be progressively invoked, thereby improving the reliability and anti-interference ability of subsequent collaborative judgments. By solidifying the three-way judgment results into a unified judgment collaboration record, the results of the preceding relation screening and the subsequent identity judgment results can be closed and connected at the same object level, forming a stable judgment basis that can be progressively invoked. By synchronously outputting the main relation cluster, secondary relation cluster, exclusion cluster, and candidate-level and relation cluster-level conclusions, the identification results can be upgraded from single-point judgments to traceable relational invocation results, thereby enhancing the consistency of subsequent link verification and continuous invocation. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of the AI-based collaborative invocation method for human recognition information in Example 1.
[0019] Figure 2 This is a formal cross-source confirmation flowchart of the AI-based collaborative invocation method for character recognition information in Example 1.
[0020] Figure 3 This is a flowchart of the relationship determination process for the AI-based collaborative invocation method for human recognition information in Example 1. Detailed Implementation
[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0024] Example 1, referring to Figure 1 , Figure 2 and Figure 3 This is the first embodiment of the present invention, which provides a method for collaboratively invoking person recognition information based on artificial intelligence, including the following steps: S1, for the upstream information source of the person recognition information acquisition path, the same source dual-path similarity calculation is performed for the recognition information and candidate recognition information of the same upstream information source. At the same time, for the cross-source candidate recognition information of different upstream information sources, the cross-source similarity under different paths is calculated. The candidate recognition information is used as the candidate object, and each candidate object is ranked in sequence. The same source initial local consistency relation domain and the cross-source local consistency relation domain are constructed respectively. S1.1 Rank each candidate in sequence, construct an initial locally consistent relation domain and a cross-source locally consistent relation domain, including constructing a similarity ranking sequence for the similarity of the same source dual pathways, ranking each candidate, considering the difference in ranking values of the candidates in the two pathways, defining the same source joint ranking value, and calculating the mean of all same source joint ranking values to define the initial locally consistent relation domain (retaining candidates whose joint ranking value is not higher than the overall average level). A ranking sequence is constructed for cross-source similarity. For each candidate ranking, the ranking values of different pathways and whether the ranking sequences of different pathways for cross-source candidate identification information are consistent are considered simultaneously. The cross-source joint ranking value is calculated, and the mean of all calculated cross-source joint ranking values is calculated to generate a cross-source locally consistent relation domain (retaining cross-source candidates whose cross-source joint ranking value is not higher than the overall average level). Specifically, the calculation of similarity between two originating pathways can include the similarity between the current identified information object and the candidate identified information of the information acquisition pathway under the first information acquisition pathway, and the similarity between the current identified information object and the candidate identified information of the information acquisition pathway under the second information acquisition pathway, which can be specifically expressed as:
[0025]
[0026] in, This indicates that the current object in the first path is the same as the first homologous candidate. similarity, This indicates that the current object and the second homologous candidate are in the second path. similarity, and These represent the first and second currently identified information objects, respectively. 1 and 2 represent the candidate indices, while 1 and 2 represent the indices of the current identified information object. For calculating the joint ranking value of homologous sources, based on the similarity ranking sequence under different information acquisition pathways, for example, the joint ranking value of homologous sources under two different information acquisition pathways can be expressed as:
[0027] in, Indicates the value of the joint sorting of the same source. This indicates the similarity ranking under the first information acquisition path. This indicates the similarity ranking under the second information acquisition path; By linking the same-source dual-path ranking results with the cross-source ranking results within a unified relational domain, we can first compress randomly similar candidates within the same-source range (referring to the range of identification information and candidate identification information from the same upstream information source), and then use cross-source consistency to further filter out false candidates that only hold true under a single source. This allows subsequent calls to objects to simultaneously possess local stability and cross-source continuity. By constraining the top positions and ranking deviations through joint ranking values, candidates with high similarity in a single path but inconsistency in the dual paths can be eliminated in advance. By directly limiting the relational domain boundary with the statistical results of all joint ranking values, we can avoid the expansion distortion caused by fixed boundaries. Through the continuous connection between the same-source relational domain and the cross-source relational domain, the originally scattered identification results can be transformed into a stable set of relations that can be progressively invoked, improving the reliability and anti-interference capability of subsequent collaborative judgments.
[0028] S2, perform bidirectional confirmation on the initial local consensus relation domain and the cross-source local consensus relation domain, calculate the joint sort value, and form the formal local consensus relation domain and the formal cross-source confirmation relation domain based on the corresponding joint sort value; S2.1, perform bidirectional confirmation of the initial local consistency relation domain and the cross-source local consistency relation domain, including, based on the initial local consistency relation domain and the cross-source local consistency relation domain, taking the candidate identification information of the same upstream information source as the query object, performing the construction of the reverse local relation domain and the reverse cross-source relation domain respectively, and recalculating the similarity, ranking and joint ranking value of the two paths through the query object, and calculating the mean according to the corresponding joint ranking value, and constructing the reverse initial local consistency relation domain and the reverse cross-source initial local consistency relation domain respectively (retaining candidates whose corresponding joint ranking value is not higher than the overall average level). Specifically, constructing a reverse local relation domain first requires building a reverse homology candidate set for the query object, which can be represented as:
[0029] in, Represents the reverse homology candidate set. Indicates the identification information object, Indicates the j-th homologous candidate. Represents a set of candidates with the same origin; The same-source mutual selection tag can be marked according to the person recognition information to determine whether it falls into the reverse initial local consistency relation domain of the candidate recognition information. If it falls into the domain, the same-source mutual selection tag is set to 1; otherwise, it is set to 0. By continuously combining forward screening and reverse mutual selection confirmation of the initial relation domain, the accidental candidates formed by a single top ranking can be transformed into stable candidates that can be bidirectionally reverted, significantly improving the closure of the relation domain. By continuing to use the formal confirmation results of the same source as the support for cross-source bridging, and then superimposing cross-source mutual selection filtering, the cross-source candidates can not only match the current identification results, but also be compatible with the already stable same source relations, thus forming a progressive constraint for cross-source confirmation.
[0030] S2.2, forming the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, including: for each candidate identification information from the same upstream information source, performing same-source mutual selection marking; if the person identification information falls into the reverse initial local consistency relation domain of the candidate identification information, then performing same-source mutual selection marking to form the formal same-source confirmation relation domain; simultaneously, if the person identification information falls into the reverse cross-source initial local consistency relation domain of the candidate identification information, then performing cross-source mutual selection marking; for all candidates in the cross-source initial relation domain, calculating the bridging support mean, expressed as:
[0031]
[0032] in, This indicates that bridging supports the mean. This represents the number of candidates in the reverse cross-source initial locally consistent relation domain. Indicates the number of source bridge support. This represents candidate identification information from cross-source upstream information sources. This indicates a formal homology confirmation relation domain; Specifically, to construct the reverse cross-origin relationship domain, we first need to build the corresponding reverse cross-origin candidate set based on the cross-origin query object, which can be represented as:
[0033] in, Represents the reverse cross-source candidate set; For candidate identification information from cross-source upstream information sources, multi-layer information filtering is performed to retain candidate identification information that has completed cross-source mutual selection marking, while also retaining candidate identification information with a bridging support number not lower than the average value. Based on the retained candidate identification information, a formal cross-source confirmation ranking value is defined through the same-source bridging support number and the cross-source joint ranking value, and a formal cross-source confirmation relation domain is formed. Specifically, cross-source confirmation sorting values are no longer combined into a single weighted number, but instead adopt a fixed sorting rule. They can be sorted from largest to smallest based on the same-source bridging support number, and then sorted from smallest to largest based on the cross-source joint sorting value.
[0034] By using a sequential joint determination of bridging support numbers and cross-source joint ranking values, we can simultaneously ensure that cross-source candidates possess both group support strength and their own ranking stability, avoiding the dominance of a single highly similar candidate. Ultimately, this makes the formally confirmed relation domain more resistant to spurious matching diffusion and more reliable for subsequent collaborative invocation. S3. Analyze the collaborative ranking values of the same-source ranking and cross-source ranking, construct the original score vector of the shared identity, select the optimal identity category as the exclusive identity determination, and establish the collaborative determination record of the three determination paths.
[0035] S3.1 Select the optimal identity category as the exclusive identity determination, record the maximum probability corresponding to the category, including forming a unified candidate set based on the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, and recording the ranking value of the candidate objects in the unified candidate set in the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain. The sum of the ranking value of the candidate object in the same-source confirmation relation domain and the ranking value in the cross-source confirmation relation domain is used as the initial collaborative ranking value of the candidate object. It should be noted that the ranking value of the candidate object in the same-source relation domain and the ranking value of the candidate object in the cross-source relation domain are based on the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, respectively, in order to transform the stability of the cross-source relation and the same-source relation into a numerical position that can be directly used in the calculation. The stability of the candidate object in the same-source confirmation structure and the stability of the cross-source confirmation structure can be considered at the same time. For each candidate object in the unified candidate set, the shared identity determination path, the source-specific identity determination path, and the source attribute determination path are respectively introduced. The shared identity determination path obtains the original score of the identity category to which the determination path belongs by the determination path through the upstream information source, and constructs and outputs the shared identity original score vector. The source-specific identity determination pathway selects a determination model based on the actual source category of the object to be confirmed. It calculates the original score vector through normalized probability to obtain the source-specific identity determination probability vector, and selects the maximum determination probability vector as the optimal identity category, which is used as the exclusive identity determination. It also records the maximum probability corresponding to the corresponding category. Specifically, for the calculation of the co-location ranking value, based on the same-source ranking and cross-source ranking respectively, for each candidate object, the sequence values of the corresponding formal same-source confirmation relation domain and formal cross-source confirmation relation domain are used as the definition values of the same-source ranking and cross-source ranking, and the sum of the two definition values is used as the initial co-location ranking value; Furthermore, regarding the original score of the identity category to which the candidate object belongs, the original score vector of shared identity determination can be converted into a shared identity determination probability vector. Then, the optimal identity category given by the shared identity determination path can be determined, and the shared identity determination probability corresponding to the optimal identity category can be recorded as a basis. Regarding the shared identity determination probability vector, it should be noted that in the shared identity determination, the determination model will output a set of original scores corresponding to different identity categories for each candidate object in the unified candidate set. The original score vector represents the determination strength of each identity category on the current candidate object. Specifically, each identity category score in the original score vector is first subjected to exponential mapping, which further amplifies the categories with larger original scores while maintaining the relative relationships between categories. Then, all exponential mapping results corresponding to the candidate object are summed, and the exponential mapping result corresponding to each category is divided by the summation value to obtain the normalized probability value corresponding to each identity category. After the above processing, the sum of the probability values of all identity categories is 1, thus forming a shared identity determination probability vector, which represents the probability distribution of each identity category on the current candidate object. After obtaining the shared identity determination probability vector, the probability values of each identity category in the probability vector are compared, and the identity category with the highest probability value is selected as the optimal identity category given by the shared identity determination path. At the same time, the highest probability value is recorded as the determination confidence of the candidate object under the shared identity determination path, which is used for subsequent joint analysis and consensus determination with the source-specific identity determination and source attribute determination results.
[0036] S3.2, establish a collaborative decision record for the three decision paths, including transforming the shared identity original score vector into a shared identity decision probability vector, transforming the source-specific identity original score vector into a source-specific identity decision probability vector, and transforming the source attribute original score vector into a source decision probability, and determining the optimal category for shared identity, the optimal category for source-specific identity, and the optimal category for source attribute respectively, thereby forming a collaborative decision record for each candidate object, represented as follows:
[0037]
[0038]
[0039] in, and These represent the probabilities of determining the source attribute of the first and second information acquisition pathways, respectively. and These represent the original scores for the first and second information acquisition pathways, respectively. Let r represent the optimal category of the source attribute, r represent the set of information acquisition pathways, c represent the candidate object index of the unified candidate set, and m represent the set of information acquisition pathways.
[0040] By forming a unified candidate set and injecting the two types of confirmation strengths into subsequent judgments using a collaborative ranking value, subsequent identity determinations can no longer be performed independently of the preceding relational structure. Through the continuous linkage of shared identity determination, source-specific identity determination, and source attribute determination, cross-source commonality, current source specificity, and source legitimacy constraints can be retained simultaneously on the same candidate, thereby avoiding the amplification of local biases caused by single-path determinations. By solidifying the three-way determination results into a unified determination collaboration record, the preceding relational screening results and subsequent identity determination results can be closed and connected at the same object level, forming a stable determination basis that can be further progressively invoked.
[0041] S4, execute consensus synthesis, calculate the maximum joint identity probability and the consistency of joint identity determination with shared identity determination, generate consensus level, consensus identity and formal consensus order for each candidate object, and form a formal consensus sequence; S4.1 Calculate the maximum joint identity probability and the consistency of the joint identity determination with the shared identity determination. This includes calculating the joint raw score of the shared identity determination and the source-specific identity determination for each candidate object in the unified candidate set, converting the joint raw score into a joint identity probability, obtaining the optimal joint identity category, and recording it as the maximum joint identity probability, expressed as:
[0042]
[0043]
[0044] in, This represents the combined original score. and These represent the original scores for shared identity determination and source-specific identity determination, respectively. L represents the total number of identities, k represents the identity index, and h represents the non-k identity index. Represents the probability of joint identity. Represents the probability of the maximum joint identity. This represents the joint raw score of identity index k; For each candidate object, the consistency quantity between the joint identity determination and the shared identity determination is calculated using the cross-entropy algorithm. The consistency quantity is calculated by determining the degree of difference between the shared identity probability distribution and the joint identity probability distribution for each candidate object in the unified candidate set, and is expressed as:
[0045] in, Represents a consistency quantity. This represents the probability of shared identity determination for the k-th identity index; By first combining the shared identity determination and the source-specific identity determination into a joint identity result, and then using this joint result to measure its consistency with the shared identity determination, the originally separate commonality determination and source-specificity determination can be converged into the same determination benchmark, thereby simultaneously suppressing single-path bias and source local distortion.
[0046] S4.2, generate consensus level, consensus identity and formal consensus order for each candidate object to form a formal consensus sequence, including determining whether the optimal category of shared identity is consistent with the optimal category of source-specific identity, whether the optimal category of joint identity is consistent with the optimal category of shared identity, and whether the optimal category of source attribute is consistent with the actual source category of the current object to be confirmed, and generate three consistency flags and the total consistency number accordingly. Candidates are divided into four categories—strong consensus, medium consensus, weak consensus, and veto—based on the total number of consensuses. The category with the best joint identity is used as the consensus identity of non-veto candidates. The candidates are then sorted in order of consensus level, total number of consensuses, consensus quantity, maximum joint identity probability, and initial coordination order to form a formal consensus sequence. Specifically, for the generation of three consistency markers, values are assigned based on whether the optimal category of shared identity is consistent with the optimal category of source-specific identity, whether the optimal category of joint identity is consistent with the optimal category of shared identity, and whether the optimal category of source attribute is consistent with the actual source category of the current object to be confirmed. The marker can be 1 when consistent and 0 when inconsistent, and the sum of the marker values is used as the total consistency value. Regarding the classification of candidate objects into four categories—strong consensus, medium consensus, weak consensus, and veto—it should be noted that this classification can be based on three consensus markers and their summed total consensus count. Since the three consensus markers include 1 or 0, it can be defined that when the total consensus count is equal to 3, the candidate object is classified as strong consensus; when the total consensus count is equal to 2, the candidate object is classified as medium consensus; when the total consensus count is equal to 1, the candidate object is classified as weak consensus; and when the total consensus count is equal to 0, the candidate object is classified as veto. By applying the three consistency markers, total consistency count, consistency quantity, maximum joint identity probability, and initial collaborative order to the generation of the formal consensus sequence in a fixed order, the candidate order can simultaneously possess identity consistency, source legitimacy, judgment stability, and continuity of prior relationships. This closes the prior relationship confirmation result and the subsequent judgment result into a highly reliable consensus link that can be progressively invoked.
[0047] S5, based on the formal consensus sequence, performs multi-level screening of candidate objects to obtain a candidate set that can enter the basic link, and groups the candidate objects into several identity relationship clusters according to consensus identity, screens out the relationship clusters that can enter the second layer link from the identity relationship clusters, selects the most stable main relationship cluster and sends it into the final link, completes the progressive call expansion and constitutes the relational call conclusion as the output result. S5.1, Select the most stable master relationship cluster and send it into the final link, including, based on the formal consensus sequence, first-level screening of all candidate objects, and forming the first-level callable candidate set of non-veto candidate objects; Based on consensus identity, candidate objects in the first-level callable candidate set are merged to construct identity relation clusters. The number of members, optimal priority, total priority, and average consistency of each identity relation cluster are calculated. Then, based on the rule that the number of members is not less than the median of the number of members in all identity relation clusters and the optimal priority is not higher than the median of the optimal priority of all identity relation clusters, the second-level callable relation cluster set is obtained, and thus the second-level callable candidate set is obtained. When selecting the specific set of callable relation clusters for the second layer, the median value of the number of members in all relation clusters is first calculated, and then the median value of the optimal order of all relation clusters is calculated. After that, only relation clusters that meet both of the following conditions are retained: First, the number of members in the relation cluster is not less than the median value of the number of members in all relation clusters; Second, the optimal order of the relation cluster is not higher than the median value of the optimal order of all relation clusters. This ensures that the relation clusters entering the second layer link are neither just a very small isolated cluster, nor are they large in number but too far down in the overall consensus sequence. As a result, the second layer set of callable relation clusters retains relation clusters that have both a certain breadth of support and a certain ranking advantage. This set is essentially a batch of candidate relation clusters with the value of further expanding the callability from all relation clusters. In the second-level callable relation cluster set, the main relation cluster is selected in a fixed order of descending number of members, ascending optimal priority, ascending average consistency, and ascending total priority. The candidate objects in the main relation cluster form the third-level callable candidate set, and the main representative candidate is selected to form the progressive link expansion call result. The specific order for selecting the primary relation cluster can be as follows: first, compare the number of members, with a larger number of members taking precedence; if the number of members is the same, then compare the optimal order, with an earlier optimal order taking precedence; if they are still the same, then compare the average consistency, with a smaller average consistency taking precedence; if they are still the same, then compare the total order, with a smaller total order taking precedence. Through these four levels of comparison, the relation cluster that is ranked first is finally determined as the primary relation cluster. By continuously connecting the multi-layer screening, identity merging, and progressive shrinkage of relation clusters after the formal consensus sequence, the originally discrete candidates can be gradually compressed into structurally stable main relation clusters. This enables the link expansion process to simultaneously possess the ability to denoise layer by layer and focus on relations, thus avoiding the direct dominance of a single high-ranking candidate in the final result. For the primary representative candidates, we can first compare their order in the formal consensus sequence, with the earlier the order, the higher the priority; if the order is the same, we can compare the consistency quantity, with the smaller the consistency quantity, the higher the priority; if they are still the same, we can compare the initial coordination order, with the smaller the initial coordination order, the higher the priority; if they are still the same, we can prioritize the earlier one according to the original input order.
[0048] S5.2, Complete the progressive call expansion and construct the relational call conclusion as the output result, including, based on the progressive link expansion call result, determining the set to which the candidate objects in the third-level callable candidate set belong to the main relation cluster, determining the set to which the candidate objects in the second-level callable candidate set belong to the secondary relation cluster, and determining the set to which the candidate objects that did not enter the first-level callable candidate set belong to the exclusion cluster; By combining the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, the source relation category is determined for each candidate object. Then, by combining the consensus identity, formal consensus sequence order, link affiliation category, source relation category, consensus level, consistency quantity, and initial coordination order, a candidate-level relation conclusion table is formed. Output the cluster level, number of members, best order, total order and average consistency for each identity relation cluster, thus forming an identity relation cluster-level conclusion; Finally, the first, second, and third layer callable candidate sets, main relation clusters, main representative candidates, main relation cluster sets, secondary relation cluster sets, exclusion cluster sets, identity relation cluster level conclusions, and candidate level relation conclusion tables together constitute the relational call conclusion output results.
[0049] By combining constraints based on the number of members, optimal priority, total priority, and average consistency, the primary relation cluster can not only be locally prioritized but also achieve overall internal consistency and suppress external competition, thereby improving the stability of the final link. By synchronously outputting the primary relation cluster, secondary relation cluster, excluded cluster, and candidate-level and relation cluster-level conclusions, the identification result can be upgraded from a single-point judgment to a traceable relational call result, enhancing the consistency of subsequent link verification and continuous calls. This embodiment also provides a computer device applicable to the collaborative invocation method of person recognition information based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the collaborative invocation method of person recognition information based on artificial intelligence as proposed in the above embodiment. The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0050] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for collaborative invocation of human recognition information based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0051] In summary, this invention, through the continuous connection of the same-source relation domain and the cross-source relation domain, can transform the originally scattered identification results into a stable set of relations that can be progressively invoked, thereby improving the reliability and anti-interference capability of subsequent collaborative judgments. By solidifying the three-way judgment results into a unified judgment collaboration record, the results of the preceding relation screening and the subsequent identity judgment results can be closedly connected at the same object level, forming a stable judgment basis that can be progressively invoked. By synchronously outputting the main relation cluster, secondary relation cluster, exclusion cluster, and candidate-level and relation cluster-level conclusions, the identification results can be upgraded from single-point judgments to traceable relational invocation results, enhancing the consistency of subsequent link verification and continuous invocation.
[0052] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for collaboratively retrieving person recognition information based on artificial intelligence, characterized in that: include, For the upstream information sources of the character recognition information acquisition path, the similarity of the recognition information and candidate recognition information of the same upstream information source is calculated in the same source dual-path. At the same time, for the cross-source candidate recognition information of different upstream information sources, the cross-source similarity under different paths is calculated, and the same source initial local consistency relation domain and cross-source local consistency relation domain are constructed respectively. Bidirectional confirmation is performed on the initial local consensus relation domains of the same source and the local consensus relation domains of the cross-source. The joint sort value is calculated and the formal same-source confirmation relation domains and the formal cross-source confirmation relation domains are formed according to the corresponding joint sort value. Construct a shared identity original score vector, select the optimal identity category as the exclusive identity determination, and establish a determination collaboration record for the three determination paths; The consensus synthesis process is executed, the maximum joint identity probability is calculated, and a consensus level, consensus identity, and formal consensus order are generated for each candidate to form a formal consensus sequence. Based on the formal consensus sequence, candidate objects are screened in multiple layers and grouped into several identity relationship clusters according to consensus identity. The most stable main relationship cluster is selected and sent to the final link to complete the progressive call expansion and form a relational call conclusion as the output result.
2. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 1, characterized in that: The construction of the initial locally consistent relation domains of the same source and the locally consistent relation domains of different sources includes: A similarity ranking sequence is constructed for the similarity of two paths with the same origin. Candidate identification information is used as candidate objects, and a candidate ranking is constructed based on the similarity ranking of each candidate object. At the same time, the difference in the ranking values of the candidate rankings in the two paths is considered. A joint ranking value with the same origin is defined, and the mean of all joint ranking values with the same origin is calculated to define the initial local consistent relation domain. A ranking sequence is constructed for cross-source similarity. For each candidate ranking, the ranking values of different pathways and whether the ranking sequences of different pathways for cross-source candidate identification information are consistent are considered simultaneously. The cross-source joint ranking value is calculated, and the mean of all calculated cross-source joint ranking values is calculated to generate a cross-source locally consistent relation domain.
3. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 2, characterized in that: The bidirectional confirmation of the initial locally consistent relation domains of the same source and the locally consistent relation domains across sources includes, Based on the same-source initial local consistency relation domain and the cross-source local consistency relation domain, the candidate identification information of the same upstream information source is used as the query object. For each query object, reverse local relation domains and reverse cross-source relation domains are constructed. The similarity, ranking, and joint ranking values of the two paths are recalculated based on the query object. The mean of the corresponding joint ranking values is calculated to construct reverse initial local consistent relation domains and reverse cross-source initial local consistent relation domains respectively.
4. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 3, characterized in that: The formal same-source confirmation relation domain and the formal cross-source confirmation relation domain include, For each candidate identification information from the same upstream information source, the same-source mutual selection mark is determined. If the person identification information falls into the reverse initial local consistency relation domain of the candidate identification information, the same-source mutual selection mark is performed to form a formal same-source confirmation relation domain. Meanwhile, if the person identification information falls into the reverse cross-source initial local consistency relation domain of the candidate identification information, cross-source mutual selection marking is performed; Calculate the number of same-source bridging supports, and calculate the mean bridging support for all candidates in the cross-source initial relation domain; For candidate identification information from cross-source upstream information sources, multi-layer information filtering is performed to retain candidate identification information that has completed cross-source mutual selection marking, while also retaining candidate identification information whose same-source bridging support number is not lower than the average bridging support. Based on the same-source bridging support number and the cross-source joint ranking value, a formal cross-source confirmation ranking value is defined, and a formal cross-source confirmation relation domain is formed.
5. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 4, characterized in that: The selection of the optimal identity category as the exclusive identity determination includes, Based on the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain, a unified candidate set is formed, and the ranking values of the candidate objects in the unified candidate set in the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain are recorded. The sum of the ranking values of the candidate objects in the same-source confirmation relation domain and the ranking values in the cross-source confirmation relation domain is used as the initial cooperative ranking value of the candidate objects. For each candidate object in the unified candidate set, the shared identity determination path, the source-specific identity determination path, and the source attribute determination path are respectively called up. The shared identity determination path obtains the original score of the identity category to which the candidate object belongs through the upstream information source and outputs the shared identity original score vector. The source-specific identity determination pathway selects a determination model based on the actual source category of the object to be confirmed. It calculates the original score vector through normalized probability to obtain the source-specific identity determination probability vector, and selects the maximum determination probability vector as the optimal identity category, which is used as the exclusive identity determination. The maximum probability corresponding to the corresponding category is recorded.
6. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 5, characterized in that: The establishment of the three-path decision coordination record includes, The original score vector of shared identity is transformed into a shared identity determination probability vector, and the original score vector of source-specific identity determination is transformed into a source-specific identity determination probability vector. The original score vector of the source attribute is converted into the source determination probability, and the optimal category of shared identity, the optimal category of source-specific identity, and the optimal category of source attribute are determined respectively, thereby forming a determination collaboration record corresponding to each candidate object.
7. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 6, characterized in that: The calculation of the maximum joint identity probability includes: for each candidate object in the unified candidate set, calculating the joint raw score of the shared identity determination and the source-specific identity determination, converting the joint raw score into a joint identity probability, and selecting the maximum determination probability vector as the optimal category of joint identity; for each candidate object in the unified candidate set, calculating the consistency quantity of the joint identity determination with the shared identity determination through the cross-entropy algorithm.
8. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 7, characterized in that: The process of generating a consensus level, consensus identity, and formal consensus order for each candidate object to form a formal consensus sequence includes, Determine whether the optimal category of shared identity is consistent with the optimal category of source-specific identity, whether the optimal category of joint identity is consistent with the optimal category of shared identity, and whether the optimal category of source attribute is consistent with the actual source category of the current object to be confirmed. Based on this, generate three consistency flags and a total consistency count. Candidates are divided into four categories—strong consensus, medium consensus, weak consensus, and veto—based on the total number of consensuses. The optimal category of joint identity is used as the consensus identity of non-veto candidates. The candidates are then sorted in order of consensus level, total number of consensuses, consensus quantity, maximum joint identity probability, and initial coordination order to form a formal consensus sequence.
9. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 8, characterized in that: The selection of the most stable master relation cluster is sent to the final link, including: Based on the formal consensus sequence, a first-level screening is performed on all candidate objects, and the non-rejected candidate objects constitute the first-level callable candidate set. Based on consensus identity, candidate objects in the first-level callable candidate set are merged to construct identity relationship clusters, thereby obtaining the second-level callable relationship cluster set and the second-level callable candidate set. In the second-level callable relationship cluster set, the main relationship cluster is selected by comparing the number of members, and the candidate objects in the main relationship cluster form the third-level callable candidate set. The main representative candidate is selected by comparing the order in the formal consensus sequence, thus forming the progressive link expansion call result.
10. The method for collaborative retrieval of person recognition information based on artificial intelligence as described in claim 9, characterized in that: The process of completing the progressive call extension and forming a relational call conclusion as the output result includes: combining the formal same-source confirmation relation domain and the formal cross-source confirmation relation domain to determine the source relation category for each candidate object; and combining consensus identity, formal consensus sequence order, link affiliation category, source relation category, consensus level, consistency quantity, and initial coordination order to form a candidate-level relational conclusion table, which together constitutes the relational call conclusion output result.