Real estate registration verification method and system based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based real estate registration verification method and system, relating to the field of information verification technology. The method comprises: determining owner information, reference facial image data, and reference signature image data of the owner; obtaining comparative facial image data and comparative signature image data of the seller; obtaining first and second facial feature information, first and second signature feature information, and document feature information; obtaining first and second relationship information, and then determining property rights reference feature information and property rights verification feature information through a graph neural network model, thereby determining a verification result. According to the present invention, a graph neural network model can be used to verify whether the relationships between multiple owners have changed. If the authenticity of the sale is in doubt or the relationships between multiple owners have changed, a risk in the transaction can be indicated, thereby reducing the legal risks of the transaction and improving transaction security.

Description

Technical Field

[0001] The present invention relates to the field of information verification technology, and in particular to a real estate registration verification method and system based on artificial intelligence. Background Art

[0002] In related technologies, before a property's ownership transaction, the identities of the buyer and seller can be verified, and the property's own information can also be verified. However, for properties with multiple owners, the relationships between the owners may be more complicated, and the transaction of the property may also involve certain legal risks. However, related technologies lack a verification solution for the relationships between the owners. Summary of the Invention

[0003] The present invention provides an artificial intelligence-based real estate registration verification method and system, which can solve the technical problem that related technologies are difficult to verify the relationship between property owners, thereby making it difficult to reduce transaction risks.

[0004] According to a first aspect of the present invention, there is provided a real estate registration verification method based on artificial intelligence, comprising:

[0005] Obtain the property ownership certificate of the property to be transferred and determine the owner information of the property based on the property ownership certificate;

[0006] Obtaining the owner's reference facial image data and reference signature image data corresponding to the owner information through the backend database;

[0007] Obtaining comparison facial image data and comparison signature image data of the seller's personnel;

[0008] Obtaining, by means of a trained face recognition model, first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data;

[0009] Obtaining, through the trained signature recognition model, first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data;

[0010] Obtaining document feature information of property rights information documents through the trained document recognition model;

[0011] Obtaining first relationship information between a plurality of owners and second relationship information between a plurality of seller personnel;

[0012] Acquire property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model; and acquire property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model;

[0013] The verification result is obtained based on the property rights reference feature information and the property rights verification feature information.

[0014] According to the present invention, the property rights reference feature information is obtained based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model, and the property rights verification feature information is obtained based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model, including:

[0015] Processing the first relationship information using the trained relationship encoding model to obtain a first relationship feature, and inputting the first relationship feature into a first fully connected layer to obtain a first relationship weight;

[0016] Processing the second relationship information through the trained relationship encoding model to obtain a second relationship feature, and inputting the second relationship feature into the first fully connected layer to obtain a second relationship weight;

[0017] Obtaining a first graph structure based on the first relationship weight, the property rights ratio of the owner, the affiliation between the reference signature image data and the owner, the document feature information, the first facial feature information, and the first signature feature information;

[0018] Obtaining a second graph structure based on the second relationship weight, the property rights ratio of the seller's personnel, the affiliation between the compared signature image data and the seller's personnel, the document feature information, the second facial feature information, and the second signature feature information;

[0019] The first graph structure is processed using a trained graph neural network model to obtain property rights reference feature information;

[0020] The second graph structure is processed through the trained graph neural network model to obtain property rights verification feature information.

[0021] According to the present invention, a first graph structure is obtained based on the first relationship weight, the property rights ratio of the owner, the affiliation between the reference signature image data and the owner, the document feature information, the first facial feature information, and the first signature feature information, including:

[0022] Setting the document feature information, the first facial feature information, and the first signature feature information as first node information of a node in a first graph structure;

[0023] The owner's property rights ratio is set as the connection weight between the node corresponding to the document feature information and the node corresponding to the owner's first facial feature information in the first graph structure;

[0024] Setting the first relationship weights between different owners as the connection weights between nodes corresponding to the first facial feature information of different owners;

[0025] determining, based on the affiliation between the reference signature image data and the owner, a connection weight between a node corresponding to the first signature feature information and a node corresponding to the first facial feature information of the owner;

[0026] Obtaining a first adjacency matrix according to the connection weights;

[0027] Obtaining a first degree matrix according to the first adjacency matrix;

[0028] A first graph structure is obtained according to the first node information, the first adjacency matrix, and the first degree matrix.

[0029] According to the present invention, the method further comprises:

[0030] Obtaining a sample property information certificate, a first facial image sample of a sample owner corresponding to the sample property information certificate, a first signature image sample, and first sample relationship information between each sample owner;

[0031] Obtaining first sample facial feature information of a first facial image sample through a face recognition model;

[0032] Obtaining first sample signature feature information of the first signature image sample through the signature recognition model;

[0033] Obtaining sample document feature information of the sample property information document through the document recognition model;

[0034] Obtaining a first sample relationship weight corresponding to the first sample relationship information through a relationship encoding model;

[0035] Obtaining a first sample graph structure based on the sample ID feature information, the first sample facial feature information, the first sample signature feature information, the first sample relationship weight, and the property rights ratio of the sample owner;

[0036] Perform at least one of the following processes: using another facial image sample of the same sample owner to replace the first facial image sample and obtain facial feature information of the second sample, using another signature image sample of the same sample owner to replace the first signature image sample and obtain signature feature information of the second sample;

[0037] obtaining a second sample graph structure based on at least one of the second sample facial feature information and the second sample signature feature information;

[0038] Perform at least one of the following processes: replace the first facial image sample with a facial image sample of a different person and obtain third sample facial feature information, replace the first signature image sample with a signature image sample obtained by a different person signing the name of the sample owner and obtain third sample signature feature information, randomly modify the first sample relationship information between the sample owners and obtain the modified second sample relationship weight;

[0039] obtaining a third sample graph structure based on at least one of the third sample facial feature information, the third sample signature feature information, and the second sample relationship weight;

[0040] The first sample graph structure, the second sample graph structure, and the third sample graph structure are processed respectively by the graph neural network model to obtain first property rights sample feature information corresponding to the first sample graph structure, second property rights sample feature information corresponding to the second sample graph structure, and third property rights sample feature information corresponding to the third sample graph structure;

[0041] Obtaining a signature loss function according to the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information;

[0042] Determining a relationship weight attention coefficient based on the first sample relationship weight and the second sample relationship weight;

[0043] Obtaining a global loss function based on the first property rights sample feature information, the second property rights sample feature information, the third property rights sample feature information, and the relationship weight attention coefficient;

[0044] The face recognition model, signature recognition model, document recognition model, relation encoding model and graph neural network model are trained based on the global loss function and signature loss function.

[0045] According to the present invention, obtaining a signature loss function based on the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information includes:

[0046] Decoding the first sample signature feature information to obtain a first word vector of the text corresponding to the first signature feature information;

[0047] Decoding the second sample signature feature information to obtain a second word vector of the text corresponding to the second signature feature information;

[0048] Decoding the third sample signature feature information to obtain a third word vector of the text corresponding to the third signature feature information;

[0049] Splicing the first sample signature feature information and the second sample signature feature information to obtain a first splicing vector;

[0050] Splicing the first sample signature feature information and the third sample signature feature information to obtain a second splicing vector;

[0051] Input the first concatenated vector into the second fully connected layer to obtain first matching probability information;

[0052] Inputting the second concatenated vector into the second fully connected layer to obtain second matching probability information;

[0053] A signature loss function is obtained according to the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information.

[0054] According to the present invention, a signature loss function is obtained based on the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information, including:

[0055] According to the formula

[0056]

[0057] Get the signature loss function ,in, is the first word vector, is the second word vector, is the third word vector, sim is the similarity function, is the first matching probability information, is the second matching probability information, are preset parameters.

[0058] According to the present invention, determining a relationship weight attention coefficient according to the first sample relationship weight and the second sample relationship weight includes:

[0059] Determining, based on the first sample relationship weight and the second sample relationship weight, an average value of the first sample relationship weight corresponding to the unmodified first sample relationship information and a minimum value of the relationship weight change corresponding to the modified first sample relationship information;

[0060] The ratio of the average value of the first sample relationship weight corresponding to the unmodified first sample relationship information to the minimum value of the relationship weight change corresponding to the modified first sample relationship information is determined as the relationship weight attention coefficient.

[0061] According to the present invention, a global loss function is obtained based on the first property sample feature information, the second property sample feature information, the third property sample feature information and the relationship weight attention coefficient, including:

[0062] According to the formula

[0063]

[0064] Get the global loss function ,in, is the first property rights sample characteristic information, is the characteristic information of the second property rights sample, is the characteristic information of the third property rights sample, is the relationship weight attention coefficient, sim is the similarity function, are preset parameters.

[0065] According to a second aspect of the present invention, there is provided a real estate registration and verification system based on artificial intelligence, comprising:

[0066] The first acquisition module is used to obtain the property information certificate of the property to be transferred and determine the owner information of the property based on the property information certificate;

[0067] The second acquisition module is used to obtain the owner's reference facial image data and reference signature image data corresponding to the owner information through the background database;

[0068] A third acquisition module is used to obtain the comparison facial image data and comparison signature image data of the seller's personnel;

[0069] A face recognition module, configured to obtain first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data using a trained face recognition model;

[0070] A signature recognition module, configured to obtain first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data using a trained signature recognition model;

[0071] The document recognition model is used to obtain the document feature information of the property rights information document through the trained document recognition model;

[0072] a relationship module, configured to obtain first relationship information between a plurality of owners and second relationship information between a plurality of seller personnel;

[0073] a verification module, configured to obtain property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model, and to obtain property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model;

[0074] The verification result module is used to obtain the verification result based on the property rights reference feature information and the property rights verification feature information.

[0075] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0076] According to the present invention, not only can the seller's facial image data and signature image be verified to ensure the authenticity of the sale, but a graph neural network model can also be used to verify whether the relationships between multiple owners have changed. If the authenticity of the sale is in doubt or the relationships between multiple owners have changed, it can be indicated that the transaction is risky, thereby reducing the legal risks of the transaction and improving transaction security. Not only can the seller's facial image and signature be verified, but the relationships between the sellers can also be verified. During verification, the relationship weights between the sellers can be calculated, and the affiliation between the signature and the owner, as well as the owner's property rights ratio, can be determined, thereby objectively determining the adjacency matrix. The graph neural network model is then used to obtain property reference feature information that can comprehensively and accurately describe the comprehensive characteristics of the property and its owner, as well as property verification feature information that is the sum of the characteristics of the property and its sellers. The property reference feature information and property verification feature information are then compared. Therefore, during the comparison process, multiple information such as facial images, signatures, and relationship information can be comprehensively compared, improving the accuracy of the verification results and helping to reduce transaction risks. When determining the signature loss function, the construction of the signature loss function can be used to improve the model's ability to recognize the text of the signature. This allows the model to accurately recognize the signature of the same person even when there is a small error between signatures taken by the same person at different times, thereby improving the model's robustness. It can also distinguish the same name signed by different people, reducing the possibility of forged signatures. When determining the global loss function, the construction of the global loss function can be used to improve the model's ability to recognize facial images of the same owner taken at different times or signatures signed at different times, improving the model's robustness and the model's ability to recognize different facial images or forged signatures, as well as the risk of relationship changes, which helps reduce transaction risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A flowchart of a real estate registration and verification method based on artificial intelligence according to an embodiment of the present invention is exemplarily shown;

[0078] Figure 2 A schematic diagram exemplarily illustrates a method for obtaining a verification result according to an embodiment of the present invention;

[0079] Figure 3 A block diagram of an artificial intelligence-based real estate registration and verification system according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0080] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0081] Figure 1 A flowchart of a method for real estate registration and verification based on artificial intelligence according to an embodiment of the present invention is exemplarily shown. The method includes:

[0082] Step S1, obtaining the property ownership certificate of the property to be transferred, and determining the owner information of the property based on the property ownership certificate;

[0083] Step S2, obtaining the owner's reference facial image data and reference signature image data corresponding to the owner information through the backend database;

[0084] Step S3, obtaining the seller's personnel's comparison facial image data and comparison signature image data;

[0085] Step S4, obtaining first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data through the trained face recognition model;

[0086] Step S5, obtaining first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data through the trained signature recognition model;

[0087] Step S6: Obtaining the document feature information of the property rights information document through the trained document recognition model;

[0088] Step S7, obtaining first relationship information between multiple owners and second relationship information between multiple sellers;

[0089] Step S8: Acquire property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model; and acquire property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model;

[0090] Step S9: Obtain verification results based on the property rights reference feature information and the property rights verification feature information.

[0091] According to the artificial intelligence-based real estate registration verification method of an embodiment of the present invention, not only can the seller's facial image data and signature image be verified to ensure the authenticity of the sale, but also whether the relationship between multiple owners has changed through a graph neural network model. If the authenticity of the sale is in doubt or the relationship between multiple owners has changed, it can be prompted that the transaction is risky, thereby reducing the legal risk of the transaction and improving transaction security.

[0092] According to one embodiment of the present invention, in step S1, the property information certificate (e.g., a house ownership certificate, a real estate certificate, or a real estate certificate) may record the owner information of the property, such as the owner's name, the property ownership ratio, etc. This information may be read by OCR (optical character recognition) or other methods for subsequent processing, and the identification code of the property information certificate may also be obtained.

[0093] According to one embodiment of the present invention, in step S2, the information read from the property rights certificate can be used to query a backend database to obtain reference facial image data and reference signature image data for each owner corresponding to each piece of owner information. During property rights registration, reference facial image data can be captured for each owner, and each owner's signature can be captured and reference signature image data can be captured to facilitate subsequent verification during property rights changes.

[0094] According to one embodiment of the present invention, in step S3, when the owner intends to sell the property, the owner is referred to as the seller. Generally, the sale can only proceed if all owners listed in the property ownership certificate agree to the sale. Therefore, the seller must include all owners. To verify the seller, facial image data of the seller can be captured and their signature can be obtained to obtain comparative signature image data. This verification not only requires verification of the completeness of the seller (i.e., whether all owners are included), but also requires verification of whether the comparative facial image data and comparative signature image data match the reference facial image data and reference signature image data, respectively.

[0095] According to one embodiment of the present invention, in step S4, the face recognition model is a convolutional neural network model, such as a VGGNet or ResNet model. After training, it can be used to obtain first facial feature information from the reference facial image data and second facial feature information from the comparison facial image data. In this example, the first facial feature information and the second facial feature information are both in vector form. Similarly, in step S5, the signature recognition model is also a convolutional neural network model, such as a VGGNet or ResNet model. After training, it can be used to obtain first signature feature information from the reference signature image data and second signature feature information from the comparison signature image data. Similarly, in step S6, the document recognition model is also a convolutional neural network model, such as a VGGNet or ResNet model. After training, it can be used to obtain document feature information of the property information document. For example, an image of the property information document can be captured and feature information of the image can be obtained as the document feature information. Alternatively, the document recognition model can be a natural language processing model, such as an RNN model or an LSTM model, which can be used to determine semantic information of text in the property information document as the document feature information. The present invention does not impose any limitation on this. The first facial feature information, the second facial feature information, the first signature feature information, the second signature feature information, and the ID feature information are vectors of the same dimension.

[0096] According to one embodiment of the present invention, in step S7, first relationship information between multiple owners and second relationship information between multiple sellers may be obtained. For example, the first relationship information between two owners may include a parent-child relationship, a parent-child relationship, a husband-wife relationship, a partnership relationship, etc. Some of these relationships can be changed, such as a husband-wife relationship and a partnership relationship, while some cannot be changed, such as a parent-child relationship and a mother-child relationship. The second relationship information is similar to the first relationship information and will not be further described here.

[0097] According to one embodiment of the present invention, in step S8, the above information can be verified using a graph neural network model. The graph neural network model can be a graph convolutional neural network model, which is not limited by the present invention. For example, the document feature information, the first relationship information, the first facial feature information, and the first signature feature information can be processed using the graph neural network model to obtain property rights reference feature information. The document feature information, the second relationship information, the second facial feature information, and the second signature feature information can also be processed using the graph neural network model to obtain property rights verification feature information. The property reference feature information can reflect various information of the registered owner in the background database (including signature, facial image, first relationship information, etc.), and the property verification feature information can reflect various information used for verification obtained when the property is sold (including signature, facial image, second relationship information, etc.). If there is a large difference between the property verification feature information and the property reference feature information, it may be that there is a difference between the owner and the seller personnel performing the verification. For example, the seller personnel fails to include all owners, the facial image of the seller personnel does not match the facial image of the owner, the signature of the seller personnel does not match the signature of the owner, and the relationship between the seller personnel is different from the relationship between the owners (for example, the relationship has changed, such as divorce, partners no longer cooperate, etc.). In this case, the transaction may be risky (for example, the property is common property to be divided when the couple divorces, and there may be legal risks in trading the property). Risk warnings can be given in the verification results.

[0098] Figure 2 A schematic diagram of obtaining a verification result according to an embodiment of the present invention is exemplarily shown.

[0099] According to one embodiment of the present invention, in step S8, the property rights reference feature information is obtained based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information and the trained graph neural network model, and the property rights verification feature information is obtained based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information and the trained graph neural network model, including: processing the first relationship information through the trained relationship encoding model to obtain the first relationship feature, and inputting the first relationship feature into the first fully connected layer to obtain the first relationship weight; processing the second relationship information through the trained relationship encoding model to obtain the second relationship feature, and inputting the second relationship feature into the first fully connected layer to obtain the first relationship weight; The feature is input into the first fully connected layer to obtain the second relationship weight; the first graph structure is obtained according to the first relationship weight, the property rights ratio of the owner, the affiliation between the reference signature image data and the owner, the certificate feature information, the first facial feature information and the first signature feature information; the second graph structure is obtained according to the second relationship weight, the property rights ratio of the seller personnel, the affiliation between the comparison signature image data and the seller personnel, the certificate feature information, the second facial feature information and the second signature feature information; the first graph structure is processed by the trained graph neural network model to obtain property rights reference feature information; the second graph structure is processed by the trained graph neural network model to obtain property rights verification feature information.

[0100] According to one embodiment of the present invention, the first relationship information can be represented by a standardized text. For example, the first relationship information between owner A and owner B is "husband and wife relationship". A list of multiple relationships can also be provided for the owners to choose from. Based on the standardized text, the text information can be processed by a relationship encoding model, which can be a natural language processing model, such as an RNN model or an LSTM model, to obtain the semantic vector of the text information, that is, the first relationship feature. The first relationship feature can be input into the first fully connected layer, and after being processed by the activation function, the first relationship weight can be obtained. Based on the same method, the first relationship weights between the various owners can be obtained. Similarly, the second relationship weights between the various sellers can be determined.

[0101] According to one embodiment of the present invention, a first graph structure is obtained based on a first relationship weight, an owner's property rights ratio, a subordinate relationship between reference signature image data and the owner, certificate feature information, first facial feature information and first signature feature information, including: setting the certificate feature information, first facial feature information and first signature feature information as first node information of nodes in the first graph structure; setting the owner's property rights ratio as a connection weight between a node corresponding to the certificate feature information and a node corresponding to the owner's first facial feature information in the first graph structure; setting the first relationship weight between different owners as a connection weight between nodes corresponding to the first facial feature information of different owners; determining the connection weight between a node corresponding to the first signature feature information and a node corresponding to the owner's first facial feature information based on the subordinate relationship between reference signature image data and the owner; obtaining a first adjacency matrix based on the connection weight; obtaining a first degree matrix based on the first adjacency matrix; obtaining a first graph structure based on the first node information, the first adjacency matrix and the first degree matrix.

[0102] According to one embodiment of the present invention, in the first graph structure, the property information certificate, the reference facial image data and the reference signature image data of each owner can be used as nodes, the certificate feature information is the first node information of the node corresponding to the property information certificate, the first facial feature information is the first node information of the node corresponding to the owner's reference facial image data, and the first signature feature information is the first node information of the node corresponding to the owner's reference signature image data.

[0103] According to one embodiment of the present invention, edge-related information can also be determined in the first graph structure. If two nodes are connected, an edge exists between the two nodes. Furthermore, an edge not only indicates whether there is a relationship between the two nodes, but also determines whether the relationship between the two nodes is close through the edge's connection weight. Specifically, the owner's property rights ratio can be set as the connection weight between the node corresponding to the ID feature information and the node corresponding to the owner's first facial feature information in the first graph structure. For example, if owner A owns 50% of the property, the connection weight between the node corresponding to owner A's first facial feature information and the node corresponding to the ID feature information is 0.5. Furthermore, the first relationship weight between different owners can be set as the connection weight between the nodes corresponding to the different owners. For example, if the first relationship weight between owner A and owner B is 0.6, the connection weight between the nodes corresponding to owner A and owner B is also 0.6. Furthermore, based on the affiliation between the reference signature image data and the owner, the connection weight between the node corresponding to the first signature feature information and the node corresponding to the owner's first facial feature information can be determined. For example, if reference signature image data A is an image of owner A's signature, then reference signature image data A belongs to owner A, and the connection weight between the node corresponding to reference signature image data A and the node corresponding to owner A is 1. Nodes corresponding to other owners have no connection relationship with reference signature image data A, and the connection weight is 0. Nodes corresponding to document feature information also have no connection relationship with nodes corresponding to each reference signature image data, and the connection weight is 0. Through the above process, the connection weight of each edge can objectively express the relationship between the nodes.

[0104] According to one embodiment of the present invention, a first adjacency matrix can be determined based on the aforementioned connection weights. In the first adjacency matrix, the data in the i-th row and j-th column represents the connection weight between the i-th node and the j-th node. Both the i-th node and the j-th node can be any of the nodes corresponding to the aforementioned document feature information, the node corresponding to the owner's first facial feature information, and the node corresponding to the first signature feature information. After determining the first adjacency matrix, a first degree matrix can be obtained. For example, the i-th data on the diagonal of the first degree matrix is ​​equal to the sum of all data in the i-th row of the first adjacency matrix. Except for the data on the diagonal, all other data in the first degree matrix are 0, and i and j are both positive integers.

[0105] According to one embodiment of the present invention, based on the first node information, first adjacency matrix, and first degree matrix obtained above, a first graph structure can be obtained. That is, the first graph structure includes the first node information, the first adjacency matrix, and the first degree matrix. Similarly, a second graph structure can be obtained.

[0106] According to one embodiment of the present invention, a trained graph neural network model can process the first graph structure and obtain feature information for each node based on the first degree matrix, the first adjacency matrix, the first node information, and the parameters of the graph neural network model itself. That is, the node's original first node information and the first node information of its adjacent nodes are processed using the first degree matrix, the first adjacency matrix, and the parameters of the graph neural network model itself, thereby fusing the node's original first node information with the first node information of its adjacent nodes (nodes with a connected relationship) to obtain the node's feature information. Each node is processed as described above by the graph neural network model, thereby fusing it with the first node information of its adjacent nodes to obtain the node's feature information. In other words, the first node information is fused with the first node information of its surrounding neighboring nodes to update the feature information. Furthermore, the feature information of the node corresponding to the certificate feature information is the property rights reference feature information. Similarly, the second graph structure can be processed to obtain property rights verification feature information for the node corresponding to the certificate feature information.

[0107] According to one embodiment of the present invention, in step S9, a verification result may be obtained for the property rights reference feature information and the property rights verification feature information. For example, considering that even for the same person, there may be some differences in facial images and signatures taken at different times, a preset threshold value is set, for example, 0.8 and 0.5. If the similarity (for example, cosine similarity) between the property rights reference feature information and the property rights verification feature information is higher than 0.8, the verification result is that the legal risk is low. If the similarity between the property rights reference feature information and the property rights verification feature information is higher than 0.5 and less than or equal to 0.8, the verification result is that the legal risk is medium. If the similarity between the property rights reference feature information and the property rights verification feature information is less than or equal to 0.5, the verification result is that the legal risk is high.

[0108] In this way, not only the facial image and signature of the seller can be verified, but also the relationship between the sellers can be verified. During the verification, the relationship weights of the relationships between the sellers can be calculated, and the affiliation between the signature and the owner, as well as the owner's property rights ratio, can be determined, thereby objectively determining the adjacency matrix. The graph neural network model can be used to obtain property rights reference feature information that can comprehensively and accurately describe the comprehensive characteristics of the real estate and its property owners, as well as property rights verification feature information that is the sum of the characteristics of the real estate and its sellers. The property rights reference feature information and the property rights verification feature information are compared, so that during the comparison process, facial images, signatures, relationship information and other information can be comprehensively compared to improve the accuracy of the verification results and help reduce transaction risks.

[0109] According to one embodiment of the present invention, the above face recognition model, signature recognition model, document recognition model, relational encoding model, and graph neural network model may be trained before use. The method further includes:

[0110] Step S101: obtaining a sample property information certificate, a first facial image sample of a sample owner corresponding to the sample property information certificate, a first signature image sample, and first sample relationship information between each sample owner;

[0111] Step S102, obtaining first sample facial feature information of a first facial image sample through a face recognition model;

[0112] Step S103: obtaining first sample signature feature information of the first signature image sample through a signature recognition model;

[0113] Step S104, obtaining sample document feature information of the sample property rights information document through the document recognition model;

[0114] Step S105: obtaining a first sample relationship weight corresponding to the first sample relationship information through a relationship coding model;

[0115] Step S106, obtaining a first sample graph structure based on the sample ID feature information, the first sample facial feature information, the first sample signature feature information, the first sample relationship weight, and the property rights ratio of the sample owner;

[0116] Step S107, performing at least one of the following processes: using another facial image sample of the same sample owner to replace the first facial image sample and obtain second sample facial feature information; using another signature image sample of the same sample owner to replace the first signature image sample and obtain second sample signature feature information;

[0117] Step S108, obtaining a second sample graph structure based on at least one of the second sample facial feature information and the second sample signature feature information;

[0118] Step S109, performing at least one of the following processes: replacing the first facial image sample with a facial image sample of a different person and obtaining third sample facial feature information, replacing the first signature image sample with a signature image sample obtained by signing the name of the sample owner by a different person and obtaining third sample signature feature information, and randomly modifying the first sample relationship information between the sample owners and obtaining the modified second sample relationship weight;

[0119] Step S110, obtaining a third sample graph structure based on at least one of the third sample facial feature information, the third sample signature feature information, and the second sample relationship weight;

[0120] Step S111: Process the first sample graph structure, the second sample graph structure, and the third sample graph structure respectively through a graph neural network model to obtain first property rights sample feature information corresponding to the first sample graph structure, second property rights sample feature information corresponding to the second sample graph structure, and third property rights sample feature information corresponding to the third sample graph structure;

[0121] Step S112: obtaining a signature loss function based on the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information;

[0122] Step S113, determining a relationship weight attention coefficient according to the first sample relationship weight and the second sample relationship weight;

[0123] Step S114, obtaining a global loss function based on the first property rights sample feature information, the second property rights sample feature information, the third property rights sample feature information, and the relationship weight attention coefficient;

[0124] Step S115: train the face recognition model, signature recognition model, document recognition model, relational encoding model, and graph neural network model based on the global loss function and the signature loss function.

[0125] According to one embodiment of the present invention, in steps S101 to S106, the method for obtaining the first sample graph structure is similar to the method for obtaining the first graph structure described above, and will not be repeated here. In step S107, in order to improve the robustness and adaptability of the model, at least one of the following processes may be performed: using other facial image samples of the same sample owner to replace the first facial image sample and obtain the facial feature information of the second sample, using other signature image samples of the same sample owner to replace the first signature image sample and obtain the signature feature information of the second sample. In this way, the above model can be trained to accurately judge when facing errors in the facial images and signatures of the same person, that is, the facial images and signatures of the same person can still be accurately recognized. In addition, this method can also obtain more positive samples in a simple way, improve training intensity, and reduce training costs. Furthermore, in step S108, the second sample graph structure can be obtained in a similar manner to the method for obtaining the first sample graph structure described above.

[0126] According to one embodiment of the present invention, in step S109, in order to improve the model's ability to identify risks when the seller's personnel are inconsistent with the owner, the signatures are inconsistent, or the relationship between the owners changes, at least one of the following processes may be performed: using facial image samples of different persons to replace the first facial image sample and obtain third sample facial feature information, using signature image samples obtained by signing the names of the sample owners by different persons to replace the first signature image sample and obtain third sample signature feature information, randomly modifying the first sample relationship information between the sample owners and obtaining the modified second sample relationship weight, the second sample relationship weight is determined in a similar manner to the first relationship weight, and will not be repeated here. Using this method, more negative samples can be obtained in a simple way, the training intensity can be increased, the training cost can be reduced, and the model's ability to identify risks can be improved. Furthermore, in step S110, the third sample graph structure can be obtained in a similar manner to the above method of obtaining the first sample graph structure.

[0127] According to one embodiment of the present invention, in step S111, the first property rights sample feature information corresponding to the first sample graph structure, the second property rights sample feature information corresponding to the second sample graph structure, and the third property rights sample feature information corresponding to the third sample graph structure can be obtained through the graph neural network model. The acquisition method is similar to the above-mentioned property rights reference feature information and property rights verification feature information, and will not be repeated here.

[0128] According to one embodiment of the present invention, in step S112, a signature loss function is obtained based on the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information, including: decoding the first sample signature feature information to obtain a first word vector of the text corresponding to the first signature feature information; decoding the second sample signature feature information to obtain a second word vector of the text corresponding to the second signature feature information; decoding the third sample signature feature information to obtain a third word vector of the text corresponding to the third signature feature information; splicing the first sample signature feature information and the second sample signature feature information to obtain a first splicing vector; splicing the first sample signature feature information and the third sample signature feature information to obtain a second splicing vector; inputting the first splicing vector into the second fully connected layer to obtain first matching probability information; inputting the second splicing vector into the second fully connected layer to obtain second matching probability information; and obtaining a signature loss function based on the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information.

[0129] According to one embodiment of the present invention, the first, second, and third sample signature feature information can be decoded by a decoder (e.g., CTC decoding), and a recognition result of the text corresponding to the first signature feature information can be output. This recognition result is then processed by a natural language processing model to obtain a word vector for each character in the text corresponding to the first signature feature information. The word vectors for each character are then weighted and summed to obtain a first word vector. Similarly, the second and third word vectors can be obtained.

[0130] According to one embodiment of the present invention, the first word vector and the second word vector should be consistent in theory, and the first sample signature feature information used to represent the image features of the first signature image sample and the second sample signature feature information of the image features of other signatures of the same owner should be similar in theory. After splicing the first sample signature feature information and the second sample signature feature information, a first splicing vector can be obtained. The first word vector and the third word vector should be consistent in theory, but the first sample signature feature information used to represent the image features of the first signature image sample and the third sample signature feature information of the signature image sample obtained by signing the signature of the sample owner by different persons should be quite different in theory, so that the model can recognize the same signature signed by different people. Of course, the cosine similarity of the first sample signature feature information and the second sample signature feature information can also be directly used as the first matching probability information, and the cosine similarity of the first sample signature feature information and the third sample signature feature information can be used as the second matching probability information. The present invention does not impose any restrictions on this.

[0131] According to one embodiment of the present invention, a signature loss function is obtained based on the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information, including: obtaining the signature loss function according to formula (1) ,

[0132] (1)

[0133] in, is the first word vector, is the second word vector, is the third word vector, sim is the similarity function, is the first matching probability information, is the second matching probability information, are preset parameters.

[0134] According to one embodiment of the present invention, in formula (1), It can be equal to 1, or less than 1, for example, 0.99, and the present invention does not limit this. As mentioned above, the first word vector and the second word vector should be consistent in theory, and the first sample signature feature information and the second sample signature feature information should be similar in theory. Therefore, in theory, Should be close to 1, or close to , that is, the similarity between the first word vector and the second word vector (e.g., cosine similarity) Theoretically, it should be close to 1, and the first matching probability information indicating that the signature feature information of the first sample and the signature feature information of the second sample are similar should be close to 1 in theory, so that the model can accurately identify the signature of the same person even when there is a small error in the signatures of the same person at different times. Therefore, for and The error between them can be reduced during the training process to improve The value of can improve the model's recognition ability and robustness to signatures.

[0135] According to one embodiment of the present invention, as described above, the first word vector and the third word vector should be consistent in theory, but the first sample signature feature information and the third sample signature feature information should be quite different in theory. Therefore, in the training process, Reduced, thereby reducing the second matching probability information indicating that the first sample signature feature information and the third sample signature feature information are similar, indicating the similarity between the first word vector and the third word vector Improve (i.e., make Thus, the model can improve its ability to recognize signatures while distinguishing the same name signed by different people, thus reducing the possibility of signature forgery. and Adding them together, we can get the signature loss function.

[0136] In this way, the model's ability to recognize the text of the signature can be improved through the construction of the signature loss function, and the model can still accurately recognize the signature of the same person when there is a small error in the signatures of the same person at different times, thereby improving the robustness of the model and being able to distinguish the same name signed by different people, reducing the possibility of signature forgery.

[0137] According to one embodiment of the present invention, in step S113, the relationship weight attention coefficient is determined based on the first sample relationship weight and the second sample relationship weight, including: determining the first sample relationship weight average value corresponding to the unmodified first sample relationship information and the minimum relationship weight change value corresponding to the modified first sample relationship information based on the first sample relationship weight and the second sample relationship weight; and determining the ratio of the first sample relationship weight average value corresponding to the unmodified first sample relationship information to the minimum relationship weight change value corresponding to the modified first sample relationship information as the relationship weight attention coefficient.

[0138] According to one embodiment of the present invention, when randomly modifying the first sample relationship information between sample owners and obtaining the modified second sample relationship weights, the first relationship information of some owners is modified. For example, owner A and owner B were originally married, but during training, this relationship was modified to no relationship (considering owner A and owner B divorced). The relationship encoding model and the first fully connected layer may contain certain errors. Therefore, it is possible that the first sample relationship information of the owners is modified, but the change in the relationship weight between the owners is small, making it difficult for the model to recognize. Therefore, the first sample relationship weight corresponding to the first sample relationship information before the modification and the second sample relationship weight corresponding to the first sample relationship information after the modification can be determined. In other words, the difference between the first sample relationship weight and the second sample relationship weight can be determined. For example, the absolute value of the difference between the first sample relationship weight and the second sample relationship weight is taken as the relationship weight change. The minimum relationship weight change corresponding to the multiple modified first sample relationship information can be determined. During training, this minimum relationship weight change can be amplified, thereby widening the gap between the first sample relationship weight and the second sample relationship weight before and after the relationship modification, making it easier for the model to recognize the relationship change and improving the model's ability to distinguish between the relationship changes before and after. The ratio of the average value of the first sample relationship weight corresponding to the unmodified first sample relationship information to the minimum value of the relationship weight change corresponding to the modified first sample relationship information can be determined as the relationship weight attention coefficient, that is, the minimum value of the relationship weight change is used as the denominator, so that the relationship weight attention coefficient is reduced, which can widen the gap between the first sample relationship weight and the second sample relationship weight before and after the relationship modification.

[0139] According to one embodiment of the present invention, in step S114, a global loss function is obtained based on the first property sample feature information, the second property sample feature information, the third property sample feature information and the relationship weight attention coefficient, including: obtaining the global loss function according to formula (2) ,

[0140] (2)

[0141] in, is the first property rights sample characteristic information, is the characteristic information of the second property rights sample, is the characteristic information of the third property rights sample, is the relationship weight attention coefficient, sim is the similarity function, are preset parameters.

[0142] According to one embodiment of the present invention, It can be equal to 1 or less than 1, for example, 0.99, and this is not limited in the present invention. When obtaining the second property rights sample feature information, facial images or signatures of the same owner taken at different times or signed at different times are used. However, these facial images and signatures still belong to the same person. Therefore, the final obtained second property rights sample feature information should theoretically be similar to the first property rights sample feature information, thereby improving the robustness of the model. is the similarity (e.g., cosine similarity) between the first property sample feature information and the second property sample feature information. The gap can be narrowed to improve the similarity between the feature information of the first property sample and the feature information of the second property sample, thereby improving the robustness of the model.

[0143] According to one embodiment of the present invention, when obtaining the third property sample feature information, a facial image of a different person is used, or a forged signature is used. Therefore, the third property sample feature information finally obtained is theoretically not similar to the first property sample feature information, and the similarity between the two is The lower the value, the stronger the model's ability to identify different facial images or forged signatures. Furthermore, as mentioned above, reducing the relationship weight attention coefficient can widen the gap between the first sample relationship weight and the second sample relationship weight before and after the relationship modification, thereby improving the model's ability to identify risks caused by relationship changes. Therefore, the relationship weight attention coefficient can be multiplied by the similarity between the first property sample feature information and the third property sample feature information. During training, the product Decrease to reduce the relationship weight attention coefficient and the similarity between the third property sample feature information and the first property sample feature information, improve the model's ability to identify different facial images or forged signatures, and the ability to identify risks caused by relationship changes. and Adding them together, we can get the global loss function.

[0144] In this way, by constructing a global loss function, the model's ability to recognize facial images of the same owner taken at different times or signatures signed at different times can be improved, the model's robustness can be improved, and the model's ability to recognize situations where facial images are different or signatures are forged, as well as the ability to recognize risks caused by changes in relationships, can be improved, which is conducive to reducing transaction risks.

[0145] According to one embodiment of the present invention, in step S115, the face recognition model, signature recognition model, document recognition model, relational encoding model and graph neural network model, as well as the first fully connected layer can be trained through the global loss function and the signature loss function. For example, the parameters of the above multiple models can be updated through the global loss function and the signature loss function by back propagation, so as to train the multiple models. After multiple trainings and verifying in the verification set that the accuracy of the model meets the requirements (for example, the accuracy of the verification results reaches a predetermined standard), the training can be completed to obtain the trained face recognition model, the trained signature recognition model, the trained document recognition model, the trained relational encoding model and the trained graph neural network model, and use them in the processing of determining the verification results.

[0146] According to an embodiment of the present invention, the AI-based real estate registration verification method not only verifies the seller's facial image and signature data to ensure the authenticity of the sale, but also verifies whether the relationships between multiple owners have changed through a graph neural network model. If the authenticity of the sale is in doubt or the relationships between multiple owners have changed, it can indicate that the transaction is risky, thereby reducing legal risks and improving transaction security. This method verifies not only the seller's facial image and signature, but also the relationships between the sellers. During verification, the relationship weights between the sellers can be calculated. The affiliation between the signature and the owner, as well as the owner's property rights ratio, can also be determined, thereby objectively determining the adjacency matrix. The graph neural network model is then used to obtain property reference feature information that comprehensively and accurately describes the comprehensive characteristics of the property and its owner, as well as property verification feature information that represents the combined characteristics of the property and its sellers. The property reference feature information and property verification feature information are then compared. This allows for comprehensive comparison of multiple information, including facial images, signatures, and relationship information, to improve the accuracy of the verification results and help reduce transaction risks. When determining the signature loss function, the construction of the signature loss function can be used to improve the model's ability to recognize the text of the signature. This allows the model to accurately recognize the signature of the same person even when there is a small error between signatures taken by the same person at different times, thereby improving the model's robustness. It can also distinguish the same name signed by different people, reducing the possibility of forged signatures. When determining the global loss function, the construction of the global loss function can be used to improve the model's ability to recognize facial images of the same owner taken at different times or signatures signed at different times, improving the model's robustness and the model's ability to recognize different facial images or forged signatures, as well as the risk of relationship changes, which helps reduce transaction risks.

[0147] Figure 3A block diagram of an artificial intelligence-based real estate registration and verification system according to an embodiment of the present invention is exemplarily shown. The system includes:

[0148] The first acquisition module is used to obtain the property information certificate of the property to be transferred and determine the owner information of the property based on the property information certificate;

[0149] The second acquisition module is used to obtain the owner's reference facial image data and reference signature image data corresponding to the owner information through the background database;

[0150] A third acquisition module is used to obtain the comparison facial image data and comparison signature image data of the seller's personnel;

[0151] A face recognition module, configured to obtain first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data using a trained face recognition model;

[0152] A signature recognition module, configured to obtain first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data using a trained signature recognition model;

[0153] The document recognition model is used to obtain the document feature information of the property rights information document through the trained document recognition model;

[0154] a relationship module, configured to obtain first relationship information between a plurality of owners and second relationship information between a plurality of seller personnel;

[0155] a verification module, configured to obtain property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model, and to obtain property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model;

[0156] The verification result module is used to obtain the verification result based on the property rights reference feature information and the property rights verification feature information.

[0157] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0158] Those skilled in the art will appreciate that the embodiments of the present invention described above and shown in the accompanying drawings are intended to be illustrative only and are not intended to limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and illustrated in the embodiments. Any variations or modifications may be made to the embodiments of the present invention without departing from the principles described.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real estate registration verification method based on artificial intelligence, characterized in that: include: Obtain the property ownership certificate of the property to be transferred and determine the owner information of the property based on the property ownership certificate; Obtaining the owner's reference facial image data and reference signature image data corresponding to the owner information through the backend database; Obtaining comparison facial image data and comparison signature image data of the seller's personnel; Obtaining, by means of a trained face recognition model, first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data; Obtaining, through the trained signature recognition model, first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data; Obtaining document feature information of property rights information documents through the trained document recognition model; Obtaining first relationship information between a plurality of owners and second relationship information between a plurality of seller personnel; Acquire property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model; and acquire property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model; Obtaining verification results based on the property rights reference feature information and the property rights verification feature information; Acquire property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model; and acquire property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model, including: Processing the first relationship information using the trained relationship encoding model to obtain a first relationship feature, and inputting the first relationship feature into a first fully connected layer to obtain a first relationship weight; Processing the second relationship information through the trained relationship encoding model to obtain a second relationship feature, and inputting the second relationship feature into the first fully connected layer to obtain a second relationship weight; Obtaining a first graph structure based on the first relationship weight, the property rights ratio of the owner, the affiliation between the reference signature image data and the owner, the document feature information, the first facial feature information, and the first signature feature information; Obtaining a second graph structure based on the second relationship weight, the property rights ratio of the seller's personnel, the affiliation between the compared signature image data and the seller's personnel, the document feature information, the second facial feature information, and the second signature feature information; The first graph structure is processed using a trained graph neural network model to obtain property rights reference feature information; The second graph structure is processed through the trained graph neural network model to obtain property rights verification feature information.

2. The artificial intelligence-based real estate registration and verification method according to claim 1 is characterized in that: A first graph structure is obtained based on the first relationship weight, the property rights ratio of the owner, the affiliation between the reference signature image data and the owner, the document feature information, the first facial feature information, and the first signature feature information, including: Setting the document feature information, the first facial feature information, and the first signature feature information as first node information of a node in a first graph structure; The owner's property rights ratio is set as the connection weight between the node corresponding to the document feature information and the node corresponding to the owner's first facial feature information in the first graph structure; Setting the first relationship weights between different owners as the connection weights between nodes corresponding to the first facial feature information of different owners; determining, based on the affiliation between the reference signature image data and the owner, a connection weight between a node corresponding to the first signature feature information and a node corresponding to the first facial feature information of the owner; Obtaining a first adjacency matrix according to the connection weights; Obtaining a first degree matrix according to the first adjacency matrix; A first graph structure is obtained according to the first node information, the first adjacency matrix, and the first degree matrix.

3. The artificial intelligence-based real estate registration and verification method according to claim 1 is characterized in that: The method further comprises: Obtaining a sample property information certificate, a first facial image sample of a sample owner corresponding to the sample property information certificate, a first signature image sample, and first sample relationship information between each sample owner; Obtaining first sample facial feature information of a first facial image sample through a face recognition model; Obtaining first sample signature feature information of the first signature image sample through the signature recognition model; Obtaining sample document feature information of the sample property information document through the document recognition model; Obtaining a first sample relationship weight corresponding to the first sample relationship information through a relationship encoding model; Obtaining a first sample graph structure based on the sample ID feature information, the first sample facial feature information, the first sample signature feature information, the first sample relationship weight, and the property rights ratio of the sample owner; Perform at least one of the following processes: using another facial image sample of the same sample owner to replace the first facial image sample and obtain facial feature information of the second sample, using another signature image sample of the same sample owner to replace the first signature image sample and obtain signature feature information of the second sample; obtaining a second sample graph structure based on at least one of the second sample facial feature information and the second sample signature feature information; Perform at least one of the following processes: replace the first facial image sample with a facial image sample of a different person and obtain third sample facial feature information, replace the first signature image sample with a signature image sample obtained by a different person signing the name of the sample owner and obtain third sample signature feature information, randomly modify the first sample relationship information between the sample owners and obtain the modified second sample relationship weight; obtaining a third sample graph structure based on at least one of the third sample facial feature information, the third sample signature feature information, and the second sample relationship weight; The first sample graph structure, the second sample graph structure, and the third sample graph structure are processed respectively by the graph neural network model to obtain first property rights sample feature information corresponding to the first sample graph structure, second property rights sample feature information corresponding to the second sample graph structure, and third property rights sample feature information corresponding to the third sample graph structure; Obtaining a signature loss function according to the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information; Determining a relationship weight attention coefficient based on the first sample relationship weight and the second sample relationship weight; Obtaining a global loss function based on the first property rights sample feature information, the second property rights sample feature information, the third property rights sample feature information, and the relationship weight attention coefficient; The face recognition model, signature recognition model, document recognition model, relation encoding model and graph neural network model are trained based on the global loss function and signature loss function.

4. The artificial intelligence-based real estate registration and verification method according to claim 3 is characterized in that: Obtaining a signature loss function according to the first sample signature feature information, the second sample signature feature information, and the third sample signature feature information, including: Decoding the first sample signature feature information to obtain a first word vector of the text corresponding to the first signature feature information; Decoding the second sample signature feature information to obtain a second word vector of the text corresponding to the second signature feature information; Decoding the third sample signature feature information to obtain a third word vector of the text corresponding to the third signature feature information; Splicing the first sample signature feature information and the second sample signature feature information to obtain a first splicing vector; Splicing the first sample signature feature information and the third sample signature feature information to obtain a second splicing vector; Input the first concatenated vector into the second fully connected layer to obtain first matching probability information; Inputting the second concatenated vector into the second fully connected layer to obtain second matching probability information; A signature loss function is obtained according to the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information.

5. The artificial intelligence-based real estate registration and verification method according to claim 4 is characterized in that: According to the first word vector, the second word vector, the third word vector, the first matching probability information, and the second matching probability information, a signature loss function is obtained, including: According to the formula ； Get the signature loss function ,in, is the first word vector, is the second word vector, is the third word vector, sim is the similarity function, is the first matching probability information, is the second matching probability information, are preset parameters.

6. The artificial intelligence-based real estate registration and verification method according to claim 3 is characterized in that: Determining a relationship weight attention coefficient according to the first sample relationship weight and the second sample relationship weight includes: Determining, based on the first sample relationship weight and the second sample relationship weight, an average value of the first sample relationship weight corresponding to the unmodified first sample relationship information and a minimum value of the relationship weight change corresponding to the modified first sample relationship information; The ratio of the average value of the first sample relationship weight corresponding to the unmodified first sample relationship information to the minimum value of the relationship weight change corresponding to the modified first sample relationship information is determined as the relationship weight attention coefficient.

7. The artificial intelligence-based real estate registration and verification method according to claim 3 is characterized in that: According to the first property rights sample feature information, the second property rights sample feature information, the third property rights sample feature information and the relationship weight attention coefficient, a global loss function is obtained, including: According to the formula ； Get the global loss function ,in, is the first property rights sample characteristic information, is the characteristic information of the second property rights sample, is the characteristic information of the third property rights sample, is the relationship weight attention coefficient, sim is the similarity function, are preset parameters.

8. An artificial intelligence-based real estate registration and verification system, used to execute the method according to any one of claims 1 to 7, characterized in that: include: The first acquisition module is used to obtain the property information certificate of the property to be transferred and determine the owner information of the property based on the property information certificate; The second acquisition module is used to obtain the owner's reference facial image data and reference signature image data corresponding to the owner information through the background database; A third acquisition module is used to obtain the comparison facial image data and comparison signature image data of the seller's personnel; A face recognition module, configured to obtain first facial feature information of the reference facial image data and second facial feature information of the comparison facial image data using a trained face recognition model; A signature recognition module, configured to obtain first signature feature information of the reference signature image data and second signature feature information of the comparison signature image data using a trained signature recognition model; The document recognition model is used to obtain the document feature information of the property rights information document through the trained document recognition model; a relationship module, configured to obtain first relationship information between a plurality of owners and second relationship information between a plurality of seller personnel; a verification module, configured to obtain property rights reference feature information based on the document feature information, the first relationship information, the first facial feature information, the first signature feature information, and the trained graph neural network model, and to obtain property rights verification feature information based on the document feature information, the second relationship information, the second facial feature information, the second signature feature information, and the trained graph neural network model; The verification result module is used to obtain the verification result based on the property rights reference feature information and the property rights verification feature information.

Images