An image-based risk mining method, device, storage medium and equipment

Abstract

The application provides an image-based risk mining method and device, a storage medium and equipment. A plurality of existing business orders are obtained, and the business orders include user images; background image features of the user images in each existing business order are extracted to obtain an image feature library; the background image features in the image feature library are clustered to obtain a plurality of image feature classes, and each image feature class corresponds to a background image feature set; error classes and / or non-risk classes in the plurality of image feature classes are filtered out to obtain risk classes, thereby constituting a risk image feature library. The categories in the obtained risk image feature library have riskiness, and are consistent with the business demand of anti-fraud risk identification. The categories in the risk image feature library are filtered, have riskiness, and the categories are not excessively fragmented.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an image-based risk mining method, apparatus, storage medium, and device. Background Technology

[0002] In existing facial recognition verification processes and similar business scenarios, the system guides and collects 1 to N facial images of the user to determine whether the user is a real person (not a photograph) and whether the user matches their declared identity, not someone else. In this process, the value of background information (non-facial or non-human body parts) in the images is not effectively utilized. Besides facial information in the images, background information also has significant value.

[0003] During our operations, we discovered that many orders involving identity spoofing or intermediary services exhibit clustering, originating from relatively fixed geographical locations and rooms. For orders processed by the same intermediary, the backgrounds of the user images in these high-risk orders show similarities. Therefore, by identifying orders with similar backgrounds in their images, we can help uncover unknown intermediary offices and / or the locations of fraud rings; we can also determine online whether a new business order originates from a known intermediary office and / or other group, thus proactively identifying risks.

[0004] In related technologies, image information is classified using KNN classification, rather than true clustering. Essentially, it assigns the current image sample to an existing category in a database using KNN classification. If a category cannot be found due to similarity thresholds or other reasons, a new category is created for the current image sample. However, this category construction method has low relevance to business needs, and creating a new category when no category can be found leads to overly fragmented category splitting. Furthermore, it fails to fully uncover business risks because it cannot explore the relationships between categories in the database.

[0005] In summary, the urgent problem to be solved in this field is how to effectively utilize background information in facial images to uncover business risks. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides an image-based risk mining method, apparatus, storage medium, and device.

[0007] In a first aspect, embodiments of the present invention provide an image-based risk mining method, comprising:

[0008] Obtain multiple existing business orders, wherein the business orders include user images;

[0009] Extract background image features from user images in each existing business order to obtain an image feature library;

[0010] The background image features in the image feature library are clustered to obtain multiple image feature classes, and each image feature class corresponds to a set of background image features.

[0011] Error classes and / or risk-free classes are filtered out from the multiple image feature classes to obtain risk classes, thus forming a risk image feature library.

[0012] In some implementations, the step of extracting background image features from user images in each existing business order to obtain an image feature library includes:

[0013] For each existing business order, perform portrait segmentation on the user image and extract the corresponding background image;

[0014] Determine whether each extracted background image meets the preset feature information conditions;

[0015] For background images that meet the preset feature information conditions, background image features are extracted using a pre-trained image feature extraction model to obtain an image feature library.

[0016] In some implementations, the preset feature information conditions include the background area ratio being greater than a first preset threshold and / or the texture features of the background image meeting preset requirements.

[0017] In some implementations, the method further includes:

[0018] Using the cluster centers of risk classes obtained from multiple existing business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated set of background image features corresponding to the risk classes, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

[0019] In some implementations, the method further includes:

[0020] Securing multiple high-risk business orders;

[0021] Extract background image features from user images in the risky business orders to obtain a risky image feature library;

[0022] Cluster the background image features in the risk image feature library to obtain at least one risk class;

[0023] Using the cluster centers of risk classes obtained from multiple risky business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated set of background image features corresponding to the risk classes, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

[0024] In some implementations, clustering background image features in the image feature library using a clustering algorithm with a specified number of categories includes:

[0025] For each background image feature in the image feature library, determine the most similar cluster center;

[0026] Background image features whose similarity to the most similar cluster center meets a third preset threshold are extracted from the image feature library and added to the candidate list of the risk class corresponding to the most similar cluster center;

[0027] For each risk class, a temporary cluster center is calculated based on the background image feature set of that risk class and the background image features in the candidate list.

[0028] Calculate the similarity between the background image features in the candidate list of the risk class and the temporary cluster centers;

[0029] Background image features in the candidate list whose similarity to the temporary cluster center meets the fourth preset threshold are updated to the risk class and the cluster center of the risk class is updated. Background image features in the candidate list whose similarity to the temporary cluster center does not meet the fourth preset threshold are put back into the image feature library.

[0030] Repeat the above clustering process until all risk classes no longer change or the preset number of iterations is reached.

[0031] In some implementations, the method further includes:

[0032] Obtain new business orders;

[0033] Extract background image features from user images in new business orders;

[0034] The extracted background image features are matched with the background image features in the risk image feature library to obtain N background image features that are most similar to the extracted background image features;

[0035] The risk class to which the N background image features belong is determined, and the risk class with the most frequent occurrences is identified as the risk class corresponding to the background image features of the user image in the new business order.

[0036] In some implementations, the step of matching the extracted background image features with background image features in the risk image feature library to obtain N background image features most similar to the extracted background image features includes:

[0037] Calculate the similarity between the extracted background image features and each background image feature in the risk image feature library to obtain the top M most similar background image features;

[0038] The background image features with similarity below a preset threshold are filtered out from the M background image features to obtain the final N most similar background image features.

[0039] In some implementations, when there are multiple user images in the new business order, the extracted background image features are matched with background image features in the risk image feature library to obtain N background image features most similar to the extracted background image features, including:

[0040] The background image features extracted from each user image in the multiple user images are matched with the background image features in the risk image feature library to obtain the N most similar background image features corresponding to the background image features extracted from each user image;

[0041] The N most similar background image features extracted from each user's image are aggregated, sorted, and the top N background image features are taken as the N most similar background image features.

[0042] Secondly, embodiments of the present invention provide an image-based risk mining device, comprising:

[0043] The acquisition module is used to acquire multiple existing business orders, wherein the business orders include user images;

[0044] The extraction module is used to extract background image features from user images in each existing business order to obtain an image feature library;

[0045] The clustering module is used to cluster the background image features in the image feature library to obtain multiple image feature classes, and each image feature class corresponds to a set of background image features;

[0046] The database construction module is used to filter out error classes and / or risk-free classes from the multiple image feature classes to obtain risk classes, thus forming a risk image feature library.

[0047] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by one or more processors, implements the method described in the first aspect.

[0048] Fourthly, embodiments of the present invention provide a computer device including a memory and one or more processors, wherein the memory stores a computer program, and the computer program, when executed by the one or more processors, implements the method described in the first aspect.

[0049] One or more embodiments of the present invention can bring at least the following beneficial effects:

[0050] The image-based risk mining method provided by this invention extracts background image features from user images in various existing business orders to obtain an image feature library. The background image features within the image feature library are then clustered to obtain multiple image feature classes, each corresponding to a set of background image features. Error classes and / or risk-free classes are filtered out from these multiple image feature classes to obtain a risk image feature library composed of risky classes. The categories in the obtained risk image feature library are risky, consistent with the business requirements for anti-fraud risk identification. The categories in the risk image feature library are filtered to ensure they are risky. The categories are not overly fragmented. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the image-based risk mining method provided in an embodiment of the present invention;

[0053] Figure 2 This is an example of constructing a risk image feature library provided in an embodiment of the present invention;

[0054] Figure 3 This is an optimized example of the risk image feature library provided in the embodiments of the present invention;

[0055] Figure 4 This is an online search example provided in the embodiments of the present invention;

[0056] Figure 5 This is a schematic diagram of an image-based risk mining device provided in an embodiment of the present invention;

[0057] Figure 6 This is a business order example provided in an embodiment of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0059] In existing facial recognition verification processes and similar business scenarios, the system guides and collects 1 to N facial images of the user to determine if the user is a real person (not a photograph) and consistent with their declared identity, not someone else. In this process, the value of background information (non-facial or body parts) in the images is not effectively utilized. Besides facial information, background information also has significant value. In practice, many orders involving identity fraud or intermediaries exhibit clustering, originating from relatively fixed geographical locations and rooms. For orders handled by the same intermediary, these high-risk orders show similar locations (e.g., GPS information) and similar backgrounds in the images. Therefore, by identifying orders with clustered locations and similar backgrounds, it is possible to effectively uncover unknown intermediary offices and / or fraud rings; it is also possible to determine online whether a new order originates from a known intermediary office and / or other group, thus identifying risks in advance.

[0060] However, in some actual business operations, a large number of business orders do not contain GPS information, or the GPS information is deliberately tampered with, or the GPS information is inaccurate due to indoor conditions. Therefore, embodiments of the present invention provide a mining strategy that relies solely on images as a supplement to address the risk mining problem of business orders without location information.

[0061] It should be understood that a business order can be a record formed by all the necessary information for a user's loan application. The information in a business order may include a unique order ID (multiple applications from the same user will result in different orders), and 1 to N images of the user or information sufficient to obtain the images (such as download links). In some cases, the information in a business order may also include the applicant ID (user_id), GPS information (latitude and longitude) at the time of application, and other information.

[0062] Real-time online photo clustering schemes in related technologies are based on image similarity, but their essence is KNN classification, not true clustering. That is, they assign the current sample to an existing category in a database using KNN classification; if a category cannot be found due to similarity thresholds or other reasons, a new category is constructed using the current image. Therefore, they have the following drawbacks:

[0063] 1) Low relevance between category construction and business objectives. Category construction and attribution in the library are based solely on image feature similarity, unrelated to the business objective of whether the category involves intermediary or agency services. Therefore, although the library contains numerous categories and similar images within each category, the scenarios and orders corresponding to these images pose no business risk and are thus considered categories without business significance.

[0064] 2) Adding new categories based on search results is not ideal. For example, multiple images taken from different angles at the same physical location may be split into two categories due to the limitations of the feature model, as their similarity is not high. This fragmentation of categories results in a small number of samples for each category. Consequently, the KNN classification process may fail to satisfy the majority of categories, leading to errors (KNN is essentially a voting process, assigning the category with the most votes. Due to the imperfections of feature vectors, some incorrect matches are inevitable. Therefore, if there are only one or two correct answers among the first K categories, it is easy to assign the result to the category with the most incorrect votes).

[0065] 3) Unable to discover related categories in the database, thus failing to fully uncover risky orders. For example, even if a new image is also taken at physical location A and the search hits category C1, it is still impossible to discover related category C2 images that were also taken at physical location A, resulting in missed risky orders.

[0066] This application addresses the problems of the aforementioned category construction methods having low relevance to business operations, creating new categories when no belonging category can be found in the search, resulting in overly fragmented category splitting, and failing to fully uncover business risks due to the inability to explore the correlation between categories in the library. It provides an image-based risk mining method, apparatus, storage medium, and device that effectively utilizes background information in user images to uncover business risks.

[0067] Example 1

[0068] like Figure 1 As shown, this embodiment provides an image-based risk mining method, including the following steps:

[0069] Step S101: Obtain multiple existing business orders, which include user images.

[0070] It should be understood that the information in a business order is structured information. In addition to user images, multiple existing business orders may also include a unique order ID to identify the business order, for example... Figure 6 As shown, each row represents an independent business order. The fields in the diagram are merely examples, but typically include at least these fields or equivalent information. The order number is the order ID, and the image download URL is the download address of the user's image. Among multiple existing business orders, there may be one or more risky business orders to be identified. The number of existing business orders obtained should be substantial enough to extract background image features and cluster into multiple risk classes. This allows for the identification of risks in existing and new business orders, the review of clustered business orders suspected of being related to intermediary or agency services, and the implementation of anti-fraud based on (background) images. The number of existing business orders can be set as needed; this embodiment does not impose a specific limit.

[0071] Step S102: Extract background image features from user images in each existing business order to obtain an image feature library. In some implementations, step S102 may further include the following sub-steps:

[0072] Step S102a: Perform portrait segmentation on the user image in each existing business order and extract the corresponding background image.

[0073] This embodiment does not limit the specific portrait segmentation algorithm and its training method. For example, this embodiment can use the ExtremeC3Net model in the prior art to achieve portrait segmentation.

[0074] Step S102b: Determine whether each extracted background image meets the preset feature information conditions.

[0075] This embodiment uses feature information conditions to determine the image quality of each extracted background image in order to determine whether the image contains sufficient information.

[0076] In some implementations, the preset feature information conditions may include the background area ratio being greater than a first preset threshold and / or the texture features of the background image meeting preset requirements.

[0077] It should be understood that in practical applications, image quality can be determined by either the condition that the background area ratio is greater than a first preset threshold, or by the condition that the texture features of the background image meet preset requirements. Alternatively, both conditions can be used simultaneously; that is, only when both conditions are met is the background image considered to satisfy the preset feature information conditions. Using two conditions for image quality determination can yield higher-quality clustered image samples, resulting in more accurate clustering results.

[0078] The phrase "background area ratio greater than the first preset threshold" can refer to the fact that the background area occupies a larger proportion of the entire image than the first preset threshold. This first preset threshold is set according to requirements, for example, 50%. In this case, if the background area occupies a larger proportion of the entire image than 50%, it indicates that the background in the user image can provide enough information for subsequent clustering. Conversely, if it is less than or equal to 50%, it indicates that the background in the user image is too small, and the user image reflects the user more than the environment, which is insufficient to provide enough information for subsequent clustering.

[0079] Determining whether a background image has rich textures can be based on whether its texture features meet preset requirements. For example, a white wall cannot provide information about the scene. In some cases, meeting the preset requirements for background image texture features may include:

[0080] Calculate the variance of pixel values, the number of feature points, and / or the number of feature lines in the background image;

[0081] Weighted summation of pixel value variance, number of feature points, and / or number of feature lines;

[0082] If the summation result is greater than the second preset threshold, then the texture features of the background image meet the preset requirements.

[0083] In this model, a pixel value variance close to 0 indicates a relatively simple texture in the background image, while a larger pixel value variance indicates a richer texture. Feature points can be corner points, and feature lines can be lines in the background. By weighted summing one or more of the pixel value variance, the number of feature points, and the number of feature lines, the texture features of the background image are further judged based on the summation result to determine whether they meet the preset requirements. The corresponding weights of pixel value variance, the number of feature points, and the number of feature lines can be set according to needs and are not limited here.

[0084] Step S102c: For a background image that meets the preset feature information conditions, use a pre-trained image feature extraction model to extract background image features and obtain an image feature library.

[0085] Specifically, background image features can refer to background image feature vectors. This embodiment does not limit the specific image feature extraction model and its training method, nor does it limit the dimension of the background image feature vector or whether it is a floating-point number or a binary number. For example, the DELG model in the prior art can be used to extract a 2048-dimensional floating-point feature vector.

[0086] Step S103: Cluster the background image features in the image feature library to obtain multiple image feature classes, each of which corresponds to a set of background image features.

[0087] Specifically, a clustering algorithm is used to cluster all image features in the image feature library to obtain multiple image feature classes. This embodiment does not limit the specific clustering algorithm, but it needs to use a clustering algorithm that does not pre-specify the number of classes. For example, the DBSCAN clustering algorithm can be used. In some cases, multiple clustering algorithms can also be combined to improve the effect. For example, the DBSCAN clustering algorithm can be used first, and then the Birch algorithm can be used to split the excessively large classes.

[0088] Optionally, this embodiment allows for some business orders whose background image features, after clustering, do not belong to any image feature class. Such categories do not belong to the risk class and can be directly ignored.

[0089] Step S104: Filter out error classes and / or risk-free classes from multiple image feature classes to obtain risk classes, thus forming a risk image feature library.

[0090] After obtaining multiple image feature classes, other methods can be used to verify the business orders and corresponding images corresponding to each image feature class for further review, thereby filtering out erroneous samples and / or risk-free classes in the clustering results. For example, if a certain class K corresponds to 5 images, but these 5 images are due to the same user submitting orders 5 times at home, there is no fraud risk. The real risk lies in the situation where multiple users submit orders from the same location. This verification and filtering step can be automatic, such as filtering based on the order's user ID; it can also be manually intervened, such as manually removing obviously erroneous images and their orders from the clustering results. This step can also be combined with other risk control methods, such as user profiling and historical records, for filtering. This embodiment does not limit the implementation method of the filtering.

[0091] The final result is a risk image feature library, including background image feature sets for each risk class, with each set corresponding to samples from a specific risk class within a cluster. This enables risk mining of existing business orders, constructing a risk category library that can be used for risk classification of existing business orders or for online risk retrieval of new business orders.

[0092] To further improve the accuracy of risk mining, further analysis of risk orders can be conducted. Firstly, this can supplement the environment for constructing the aforementioned risk image feature library. Since the aforementioned risk image feature library is based on image feature clustering without prior knowledge, it may contain omissions. Supplementing with the following risk order step can yield a more complete risk order database—an updated risk image feature library. Secondly, existing clues (multiple risk orders) obtained through other means can also be used to mine risks.

[0093] In some implementations, the background image features in the image feature library obtained in step S102 can be clustered again to avoid missing risky business orders in the previous clustering and filtering process. To improve the accuracy of the re-clustering, the cluster centers of each risk class obtained by filtering after clustering based on multiple existing business orders are used as seeds. Therefore, the number of clusters is specified, that is, the same as the number of risk classes obtained in the previous clustering. Furthermore, the above method may also include:

[0094] Step S201a: Using the cluster centers of risk classes obtained from multiple existing business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library obtained in step S102 to obtain the updated risk class corresponding to the background image feature set, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

[0095] In some implementations, a set of user images corresponding to known risk orders can be provided. Each set of user images may include one or more user images. By extracting background image features from these user images, a corresponding risk image feature library can be obtained. Clustering the background image features within this risk image feature library also yields at least one risk class. Using the cluster centers of each risk class obtained in this implementation as seeds, the background image features in the image feature library obtained in step S102 are clustered again. This can also prevent the previous clustering and filtering process from missing risky business orders. Similarly, the number of categories in this re-clustering is specified, that is, the same as the number of risk classes obtained in the previous clustering. Furthermore, the above method may also include:

[0096] Step S201b1: Obtain multiple risky business orders. These multiple risky business orders are business orders with known risks. The specific number is determined according to actual needs, aiming to achieve a good clustering effect.

[0097] Step S201b2: Extract background image features from user images in risky business orders to obtain a risk image feature library. The extraction of background image features from user images in risky business orders can be performed using a similar method to steps S102a to S102c, and will not be elaborated further here.

[0098] Step S201b3: Cluster the background image features in the risk image feature library to obtain at least one risk class.

[0099] Step S201b4: Using the cluster centers of risk classes obtained from multiple risky business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated set of background image features corresponding to the risk classes, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

[0100] This embodiment does not limit the clustering algorithm to a single specified number of categories; for example, it can be the k-means++ algorithm. In one implementation of this embodiment, an improved k-means++ algorithm is used as the clustering algorithm for the specified number of categories to cluster the background image features in the image feature library, including:

[0101] Step 1: Initialize the clusters using the cluster centers of the risk classes (classes of known risk orders) obtained from multiple risky business orders to obtain K initial clusters.

[0102] Step 2: Determine the most similar cluster center for each background image feature in the image feature library.

[0103] Step 3: Extract background image features whose similarity to the most similar cluster center meets the third preset threshold from the image feature library and add them to the candidate list of the risk class corresponding to the most similar cluster center, but do not update the cluster center for the time being; background image features that do not meet the third preset threshold are not processed.

[0104] Step 4: For each risk class (update K classes sequentially), calculate the temporary cluster center of the risk class based on the background image feature set of the risk class and the background image features in the candidate list.

[0105] Step 5: Calculate the similarity between the background image features in the candidate list of the risk class and the temporary cluster center. Update the background image features in the candidate list whose similarity with the temporary cluster center meets the fourth preset threshold to the risk class and update the cluster center of the risk class. Put the background image features in the candidate list whose similarity with the temporary cluster center does not meet the fourth preset threshold back into the image feature library (remove them from the candidate list).

[0106] Repeat steps 2 through 5 of the above clustering process until all risk classes (K classes) no longer change or the preset number of iterations is reached.

[0107] It should be understood that the aforementioned clustering process allows background image features to not enter any class, indicating that the background image features are not risky, thereby improving the accuracy of risk identification and avoiding the waste of computing resources.

[0108] Based on the aforementioned construction or updating of the risk image feature library, new business orders can also be retrieved online to assess their risk in real time. Specifically, this involves extracting image features from the user image corresponding to the new order image, matching them using image retrieval techniques and background image features from the risk image feature library, and returning the N most similar business orders. This embodiment does not limit the retrieval method or pre- and post-processing methods. For example, the existing Faiss retrieval engine can be used. Therefore, the method of this embodiment may further include:

[0109] Step S301: Obtain a new business order.

[0110] Step S302: Extract the background image features of the user image in the new business order.

[0111] Step S303: Match the extracted background image features with the background image features in the risk image feature library to obtain the N background image features that are most similar to the extracted background image features.

[0112] In some implementations, step S303 may include:

[0113] Step S303a: Calculate the similarity between the extracted background image features and each background image feature in the risk image feature library, and obtain the top M most similar background image features;

[0114] Step S303b: Filter out background image features with similarity less than the preset metric threshold from the M background image features to obtain the final N most similar background image features, where M≥N.

[0115] In some cases, a new business order may contain multiple user images (i.e., the user took multiple images). In such cases, the extracted background image features are matched with background image features in a risk image feature library to obtain the N background image features most similar to the extracted background image features, including:

[0116] The background image features extracted from each user image are matched with the background image features in the risk image feature library to obtain the N most similar background image features corresponding to the background image features extracted from each user image (the extraction method of the N most similar background image features corresponding to the background image features extracted from each user image can be referred to steps S303a to S303b); and

[0117] The N most similar background image features extracted from each user's image are aggregated, sorted, and the top N background image features are taken as the N most similar background image features.

[0118] Step S304: Count the risk classes to which the N background image features belong, and determine the risk class corresponding to the background image features of the user image in the new business order as the risk class to which the most frequent risk class belongs.

[0119] Based on the KNN principle, new business orders are assigned to the category with the most matches. The corresponding business order and its user image are then returned.

[0120] If these N background image features belong to multiple risk classes, then for each of these risk classes, calculate the average similarity between the background image features extracted based on the new business order and each background image feature in that risk class. Determine the risk class with the highest average similarity as the risk class corresponding to the background image features of the user image in the new business order. Then return the corresponding business order and its user image.

[0121] Example 2

[0122] This embodiment provides two specific scenario examples.

[0123] Scenario 1: Real-time assessment of the risk level of new business orders

[0124] In practical use, the risk image feature library is periodically rebuilt based on historical business orders, and the construction process refers to... Figure 2 The process involves extracting newly added risky orders over a given period and adding them to the database. As new business orders arrive, the risk image feature database will be continuously updated. This example does not limit the timing period. For instance, it can be fully rebuilt daily or every 7 days. Optionally... Figure 2 The resulting risk image feature library can be used as follows: Figure 3 The process was optimized to obtain an optimized risk image feature library.

[0125] After the build is complete, refer to Figure 4 The purpose of conducting online searches is to assess the risk of new orders in real time.

[0126] Scenario 2: Based on existing clues (obtained through other means), uncover risky orders.

[0127] In practice, various anti-fraud risk control technologies are often used in combination, and multiple institutions in the same industry frequently communicate and collaborate in the field of anti-fraud. Therefore, a scenario exists where anti-fraud personnel obtain a batch of risky orders through other risk control methods or other industry peers. Based on these known risky orders, unknown risky orders can be identified from historical orders. See [link to relevant documentation] for the identification method. Figure 3 The identified risky orders can be further processed in conjunction with other anti-fraud and risk control measures; this embodiment does not impose any limitations on this.

[0128] Example 3

[0129] like Figure 5 As shown, this embodiment provides an image-based risk mining device, including:

[0130] The acquisition module 501 is used to acquire multiple existing business orders, which include user images.

[0131] Extraction module 502 is used to extract background image features from user images in each existing business order to obtain an image feature library;

[0132] Clustering module 503 is used to cluster background image features in the image feature library to obtain multiple image feature classes, each image feature class corresponding to a set of background image features;

[0133] The database construction module 504 is used to filter out error classes and / or risk-free classes from multiple image feature classes to obtain risk classes, thus forming a risk image feature library.

[0134] In some implementations, the extraction module 502 is used to extract background image features from user images in various existing business orders. When obtaining the image feature library, it may include:

[0135] For each existing business order, perform portrait segmentation on the user image and extract the corresponding background image; determine whether each extracted background image meets the preset feature information conditions; for the background image that meets the preset feature information conditions, use a pre-trained image feature extraction model to extract background image features and obtain an image feature library.

[0136] The preset feature information conditions may include the area ratio of the background being greater than a first preset threshold and / or the texture features of the background image meeting preset requirements.

[0137] In some cases, the texture features of the background image meet preset requirements, which may include:

[0138] Calculate the variance of pixel values, the number of feature points, and / or the number of feature lines in the background image;

[0139] Weighted summation of pixel value variance, number of feature points, and / or number of feature lines;

[0140] If the summation result is greater than the second preset threshold, then the texture features of the background image meet the preset requirements.

[0141] In some implementations, the background image features in the aforementioned image feature library can be clustered again to avoid missing risky business orders in the previous clustering and filtering process. To improve the accuracy of the re-clustering, the cluster centers of each risk class obtained by filtering after clustering multiple existing business orders are used as seeds. Therefore, the number of clusters is specified, that is, the same as the number of risk classes obtained by the previous clustering. Furthermore, the above-mentioned device may also include: a mining module for re-clustering the background image features in the aforementioned image feature library.

[0142] Specifically, the mining module is used for:

[0143] Using the cluster centers of risk classes obtained from multiple existing business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the aforementioned image feature library to obtain the updated risk class corresponding to the background image feature set, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

[0144] Or, for:

[0145] Multiple risky business orders are obtained; background image features of user images in the risky business orders are extracted to obtain a risky image feature library; the background image features in the risky image feature library are clustered to obtain at least one risk class; using the cluster centers of the risk classes obtained from multiple risky business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated risk class corresponding to the background image feature set, thereby constituting an updated risky image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risky image feature library.

[0146] In one implementation of this embodiment, an improved k-means++ algorithm is used as the clustering algorithm for a specified number of categories to cluster the background image features in the image feature library, including:

[0147] Initialize using the cluster centers of risk classes (classes of known risk orders) obtained from multiple risky business orders as seeds to obtain K initial clusters;

[0148] For each background image feature in the image feature library, the most similar cluster center is determined. Background image features whose similarity to the most similar cluster center meets the third preset threshold are taken out of the image feature library and added to the candidate list of the risk class corresponding to the most similar cluster center, but the cluster center is not updated for the time being; background image features that do not meet the third preset threshold are not processed.

[0149] For each risk class (K classes are updated sequentially), a temporary cluster center is calculated based on the background image feature set and the background image features in the candidate list for that risk class. The similarity between the background image features in the candidate list and the temporary cluster center is calculated. Background image features in the candidate list whose similarity to the temporary cluster center meets the fourth preset threshold are updated to that risk class, and the cluster center of that risk class is updated. Background image features in the candidate list whose similarity to the temporary cluster center does not meet the fourth preset threshold are put back into the image feature library (removed from the candidate list).

[0150] Repeat the above clustering process until all risk classes (K classes) no longer change or the preset number of iterations is reached.

[0151] It should be understood that the aforementioned clustering process allows background image features to not enter any class, indicating that the background image features are not risky, thereby improving the accuracy of risk identification and avoiding the waste of computing resources.

[0152] The apparatus in this embodiment may further include: an online retrieval module, used for:

[0153] Obtain a new business order and extract the background image features of the user image in the new business order; match the extracted background image features with the background image features in the risk image feature library to obtain the N background image features that are most similar to the extracted background image features; count the risk classes to which the N background image features belong, and determine the risk class corresponding to the background image features of the user image in the new business order as the risk class to which the N background image features belong the most.

[0154] In some implementations, obtaining the N background image features most similar to the extracted background image features may include: calculating the similarity between the extracted background image features and each background image feature in the risk image feature library to obtain the top M most similar background image features; filtering out background image features with similarity less than a preset metric threshold from the M background image features to obtain the final N most similar background image features, where M≥N.

[0155] In some cases, a new business order may contain multiple user images (i.e., the user took multiple images). In such cases, the extracted background image features are matched with background image features in a risk image feature library to obtain the N background image features most similar to the extracted background image features, including:

[0156] The background image features extracted from each user image in multiple user images will be matched with the background image features in the risk image feature library to obtain the N most similar background image features corresponding to the background image features extracted from each user image (the extraction method of the N most similar background image features corresponding to the background image features extracted from each user image can be referred to above); and the N most similar background image features corresponding to the background image features extracted from each user image will be summarized and sorted, and the top N background image features will be taken as the N most similar background image features.

[0157] If these N background image features belong to multiple risk classes, then for each of these risk classes, calculate the average similarity between the background image features extracted based on the new business order and each background image feature in that risk class. Determine the risk class with the highest average similarity as the risk class corresponding to the background image features of the user image in the new business order. Then return the corresponding business order and its user image.

[0158] It should be understood that the apparatus of this embodiment has all the beneficial effects of the method embodiment.

[0159] Those skilled in the art will understand that the above-described modules or steps can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. This invention is not limited to any specific hardware and software combination.

[0160] Example 4

[0161] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by one or more processors, implements the method or apparatus of the foregoing embodiment.

[0162] In this embodiment, the computer-readable 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 Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The method can be referred to in the foregoing embodiments of the present invention, and will not be repeated here.

[0163] Example 5

[0164] This embodiment provides a computer device, including a memory and one or more processors. The memory stores a computer program, and when the computer program is executed by one or more processors, it implements the method or apparatus of the foregoing embodiment.

[0165] In this embodiment, the processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the methods in the above embodiments. The methods implemented when a computer program running on the processor is executed can be referred to the foregoing embodiments of the present invention, and will not be repeated here.

[0166] In embodiments of the present invention, the categories in the obtained risk image feature library are risky, consistent with the business requirements of anti-fraud risk identification. The categories in the risk image feature library are screened to ensure they are risky. The categories are not overly fragmented. Because clustering algorithms are used to obtain the categories, there is a high probability that two categories will not belong to the same scenario with identical perspectives. Furthermore, risky orders are more thoroughly mined; once a match is found, the vast majority of risky orders in that scenario can be identified. Secondary mining based on known risky orders can further recall more risky orders, thereby improving the retrieval effect and avoiding risk omissions.

[0167] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative.

[0168] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0169] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. An image-based risk mining method, characterized in that, include: Obtain multiple existing business orders, wherein the business orders include user images; Extract background image features from user images in each existing business order to obtain an image feature library; The background image features in the image feature library are clustered to obtain multiple image feature classes, and each image feature class corresponds to a set of background image features. Error classes and / or risk-free classes are filtered out from the multiple image feature classes to obtain risk classes, thus forming a risk image feature library; The method further includes: Using the cluster centers of risk classes obtained from multiple existing business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library, resulting in an updated set of background image features corresponding to the updated risk classes, thereby constituting an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library; or, Securing multiple high-risk business orders; Extract background image features from user images in the risky business orders to obtain a risky image feature library; Cluster the background image features in the risk image feature library to obtain at least one risk class; Using the cluster centers of risk classes obtained from multiple risky business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated set of background image features corresponding to the risk classes, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

2. The image-based risk mining method according to claim 1, characterized in that, The step of extracting background image features from user images in each existing business order to obtain an image feature library includes: For each existing business order, perform portrait segmentation on the user image and extract the corresponding background image; Determine whether each extracted background image meets the preset feature information conditions; For background images that meet the preset feature information conditions, background image features are extracted using a pre-trained image feature extraction model to obtain an image feature library.

3. The image-based risk mining method according to claim 2, characterized in that, The preset feature information conditions include the background area ratio being greater than a first preset threshold and / or the texture features of the background image meeting preset requirements.

4. The image-based risk mining method according to claim 1, characterized in that, The clustering algorithm using a specified number of clusters to cluster background image features in the image feature library includes: For each background image feature in the image feature library, determine the most similar cluster center; Background image features whose similarity to the most similar cluster center meets a third preset threshold are extracted from the image feature library and added to the candidate list of the risk class corresponding to the most similar cluster center; For each risk class, a temporary cluster center is calculated based on the background image feature set of that risk class and the background image features in the candidate list. Calculate the similarity between the background image features in the candidate list of the risk class and the temporary cluster centers; Background image features in the candidate list whose similarity to the temporary cluster center meets the fourth preset threshold are updated to the risk class and the cluster center of the risk class is updated. Background image features in the candidate list whose similarity to the temporary cluster center does not meet the fourth preset threshold are put back into the image feature library. Repeat the above clustering process until all risk classes no longer change or the preset number of iterations is reached.

5. The image-based risk mining method according to claim 1, characterized in that, Also includes: Obtain new business orders; Extract background image features from user images in new business orders; The extracted background image features are matched with the background image features in the risk image feature library to obtain N background image features that are most similar to the extracted background image features; The risk class to which the N background image features belong is determined, and the risk class with the most frequent occurrences is identified as the risk class corresponding to the background image features of the user image in the new business order.

6. The image-based risk mining method according to claim 5, characterized in that, The step of matching the extracted background image features with the background image features in the risk image feature library to obtain the N background image features most similar to the extracted background image features includes: Calculate the similarity between the extracted background image features and each background image feature in the risk image feature library to obtain the top M most similar background image features; The background image features with similarity below a preset threshold are filtered out from the M background image features to obtain the final N most similar background image features.

7. The image-based risk mining method according to claim 5, characterized in that, When there are multiple user images in the new business order, the extracted background image features are matched with the background image features in the risk image feature library to obtain N background image features that are most similar to the extracted background image features, including: The background image features extracted from each user image in the multiple user images are matched with the background image features in the risk image feature library to obtain the N most similar background image features corresponding to the background image features extracted from each user image; The N most similar background image features extracted from each user's image are aggregated, sorted, and the top N background image features are taken as the N most similar background image features.

8. An image-based risk assessment device, characterized in that, include: The acquisition module is used to acquire multiple existing business orders, wherein the business orders include user images; The extraction module is used to extract background image features from user images in each existing business order to obtain an image feature library; The clustering module is used to cluster the background image features in the image feature library to obtain multiple image feature classes, and each image feature class corresponds to a set of background image features; The database construction module is used to filter out error classes and / or risk-free classes from the multiple image feature classes to obtain risk classes and form a risk image feature library. The mining module is used to cluster the background image features in the image feature library using cluster centers of risk classes obtained from multiple existing business orders as seeds, and employing a clustering algorithm with a specified number of categories to obtain an updated set of background image features corresponding to the risk classes, thereby constituting an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library; or... The mining module is used to obtain multiple risky business orders; Extract background image features from user images in the risky business orders to obtain a risky image feature library; Cluster the background image features in the risk image feature library to obtain at least one risk class; Using the cluster centers of risk classes obtained from multiple risky business orders as seeds, a clustering algorithm with a specified number of categories is used to cluster the background image features in the image feature library to obtain the updated set of background image features corresponding to the risk classes, thereby forming an updated risk image feature library; wherein, the specified number of categories is the same as the number of risk classes in the risk image feature library.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, It includes a memory and one or more processors, wherein a computer program is stored on the memory, and the computer program, when executed by the one or more processors, implements the method as described in any one of claims 1 to 7.

Images