A fraud detection and identification method and device based on image background similarity

Abstract

The present invention discloses a fraud detection and identification method and device based on image background similarity. The method includes obtaining image data uploaded by each user for loan application, filtering normal background images in the image data based on a scene classification model to obtain suspected fraud scene images; extracting background feature information in the suspected fraud scene images based on a feature extraction model; clustering the suspected fraud scene images according to the background feature information to obtain ID card scene clustering results and headshot scene clustering results; integrating and outputting the ID card scene clustering results and headshot scene clustering results. The present invention utilizes scene classification to improve clustering effects, utilizes a feature extraction model to effectively extract image background features, utilizes HDBSCAN to cluster the extracted image features of suspected fraud scenes, ensures the effect after clustering, and facilitates fraud detection and fraud gang identification.

Description

Technical Field

[0001] The present invention relates to the technical field of fraud detection and fraud gang identification, and in particular to a fraud detection and identification method and device based on image background similarity. Background Art

[0002] In recent years, with the emergence of the Internet, Internet finance has developed rapidly. Although the booming development of Internet finance has injected new vitality into the financial industry, its development process has also exposed some problems and risks, and brought great challenges to financial management. At present, Internet fraud caused by Internet finance has become very professional and industrialized. In actual business, applying for a loan requires uploading a photo of your ID card and a personal headshot, and the backgrounds of the pictures uploaded through intermediaries are very similar, such as in an office, outside a building, or in a hotel. For the identification of such fraud gangs, it is extremely difficult to achieve if all photos uploaded by customers are manually reviewed. Due to the diversity and irregularity of the image samples themselves, the difficulty of image review increases sharply with the increase in the number of images to be reviewed. In order to assist manual review, an algorithm for clustering by image background similarity is proposed. As an application of image classification, scene classification has developed relatively maturely. Scene classification includes not only coarse-grained classification of indoor and outdoor scenes, but also fine-grained classification such as classrooms, hospitals, and basketball courts. In actual scenarios, there are very few user images applied for by fraud gangs, and the number of normal images is much higher than the number of fraud gang scene images. This will bring a lot of noise to the clustering objects, resulting in low clustering accuracy. Summary of the invention

[0003] The purpose of the present invention is to provide a fraud detection and identification method and device based on image background similarity in view of the deficiencies in the prior art.

[0004] To achieve the above objectives, in a first aspect, the present invention provides a fraud detection and identification method based on image background similarity, comprising:

[0005] Obtaining the picture data uploaded by each user in the loan application, the picture data including the ID card picture and the headshot picture;

[0006] filtering the normal background images in the image data based on the scene classification model to obtain suspected fraud scene images, wherein the suspected fraud scene images include suspected fraud identity card images and suspected fraud headshot images;

[0007] Extracting background feature information from the suspected fraud scene image based on a feature extraction model;

[0008] Clustering the suspected fraud scene pictures according to the background feature information to obtain ID card scene clustering results and headshot scene clustering results, wherein the ID card scene clustering results and headshot scene clustering results include suspected fraud ID card pictures and suspected fraud headshot pictures of several clusters in turn;

[0009] According to the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, the account number overlap IOU is calculated by pairwise comparison:

[0010]

[0011] in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where both e and f are natural numbers greater than zero;

[0012] Combine the account number sets whose overlap IOU is greater than the set overlap threshold Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f;

[0013] The account number is and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results.

[0014] Furthermore, the ID card scene clustering results and the headshot scene clustering results are used as normal clusters, and the suspected fraud scene images that cannot be clustered into the ID card scene clustering results and the headshot scene clustering results are used as noise clusters. The clustering results are re-filtered and integrated using the maximum intra-class distances of the noise clusters and the normal clusters, specifically including: calculating the maximum intra-class distances of the noise clusters and the normal clusters respectively. The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient.

[0015] Furthermore, the value of B is 0.8.

[0016] Furthermore, the overlap threshold is 0.5.

[0017] In a second aspect, the present invention provides a fraud detection and identification device based on image background similarity, comprising:

[0018] An acquisition module is used to acquire the image data uploaded by each user in the loan application, wherein the image data includes an ID card image and a headshot image;

[0019] A filtering module, for filtering the normal background images in the image data based on the scene classification model to obtain suspected fraud scene images, wherein the suspected fraud scene images include suspected fraud identity card images and suspected fraud headshot images;

[0020] A feature extraction module, used to extract background feature information in the suspected fraud scene image based on a feature extraction model;

[0021] A clustering module, for clustering suspected fraud scene pictures according to the background feature information to obtain ID card scene clustering results and headshot scene clustering results, wherein the ID card scene clustering results and headshot scene clustering results include suspected fraud ID card pictures and suspected fraud headshot pictures of several clusters in turn;

[0022] The clustering result integration output module is used to calculate the IOU of the account number based on the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, by pairwise comparison:

[0023]

[0024] in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where e and f are both natural numbers greater than zero; the account number sets whose overlap IOU is greater than the set overlap threshold are combined Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f; and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results.

[0025] Furthermore, the clustering result integration output module uses the ID card scene clustering results and the headshot scene clustering results as normal clusters, and uses the suspected fraud scene images that cannot be clustered into the ID card scene clustering results and the headshot scene clustering results as noise clusters, and uses the maximum intra-class distances of the noise clusters and the normal clusters to re-filter and integrate the clustering results, specifically including: calculating the maximum intra-class distance of the noise cluster for the noise cluster and the normal cluster respectively. The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient.

[0026] Furthermore, the value of B is 0.8.

[0027] Furthermore, the overlap threshold is 0.5.

[0028] Beneficial effects: 1. The present invention utilizes scene classification to filter normal background images, obtain suspected fraudulent background images, and avoid a large number of noise images that lead to very poor clustering results.

[0029] 2. The present invention uses a deep neural network (MobileNetV2) to train a scene classification model to divide the scenes into suspected fraud scenes and normal scenes. The suspected fraud scenes can be defined by observing pictures with similar backgrounds uploaded by overdue accounts.

[0030] 3. The present invention uses a deep neural network (ResNet50) and fraud scene image training to obtain a feature extraction model. The training set is a set of images with similar backgrounds. Each group of image clusters with similar backgrounds is labeled with the same label. Then, the parameters of the model of the entire network are obtained by training the entire network, which can effectively extract the background features of the image.

[0031] 4. The present invention uses HDBSCAN to cluster the extracted image features of suspected fraud scenes. HDBSCAN is insensitive to noise points, ensuring the effect after clustering.

[0032] 5. The present invention uses the maximum intra-class distance of the noise class and the normal class to re-filter and integrate the clustering results, avoiding interference from noise-class images and improving the accuracy of the output. It serves as an aid to manual processing and facilitates fraud detection and fraud gang identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of a fraud detection and recognition method based on image background similarity;

[0034] Figure 2 It is the filtered normal background picture and the effect picture of the suspected fraud scene picture;

[0035] Figure 3 This is a schematic diagram of the effect after clustering;

[0036] Figure 4 It is a block diagram of the principle of a fraud detection and recognition device based on image background similarity. DETAILED DESCRIPTION

[0037] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. The embodiments are implemented based on the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0038] like Figure 1 As shown, an embodiment of the present invention provides a fraud detection and identification method based on image background similarity, comprising:

[0039] Obtain the image data uploaded by each user in the loan application, which includes the ID card image and the headshot image.

[0040] Based on the scene classification model, normal background images in the image data are filtered to obtain suspected fraud scene images, including suspected fraud ID card images and suspected fraud headshot images. Since fraud scene images account for a small part of all image data, if all image data are clustered, there will be a lot of noise, resulting in very poor clustering results. Therefore, before using image similarity for clustering, the suspected fraud scenes are first identified, thereby filtering out most normal images.

[0041] As a preferred embodiment of the present invention, the present invention uses the deep learning network MobilenetV2 to perform binary classification on suspected fraud scenes and normal scenes. Through the pictures uploaded by historical overdue customers, it can be found that if overdue customers apply for loans through intermediaries, the scenes are usually office scenes. For example, the background of ID card photos is mostly the desktop, and there will be tables, papers, keyboards, etc. on the desktop; while normal scenes are usually outdoors, shopping malls, homes, etc. Based on the above experience, the scenes are classified into two categories, and the suspected fraud scenes and normal scenes are manually labeled before training. Figure 2 The ID card pictures uploaded by customers showing normal scenes and suspected fraud scenes are used. The present invention uses the pre-trained classification MobilenetV2 model of ImageNet itself and performs migration training based on this model. At the same time, data light enhancement, random cutting data enhancement, and random flipping data enhancement are added. This network calculates the difference between the predicted value and the true value through binary cross entropy, calculates the loss, and reduces the loss through back propagation, and finally obtains the network parameters. Through suspected scene detection, most normal pictures can be filtered out.

[0042] The background feature information in the above-mentioned suspected fraud scene pictures is extracted based on the feature extraction model. As a preferred embodiment, the present invention also uses a deep learning model to classify the above-mentioned suspected fraud scene pictures, and the network output obtained in the previous layer of the classifier is the picture feature. The present invention uses the picture classification task to train the feature extraction model. Considering that the multi-classification task is more complex than the binary classification task, the present invention preferably uses ResNet50 as the basic architecture, which is more complex, has more parameters, and has stronger fitting ability. For the suspected fraud scene pictures, the pictures with similar backgrounds are first marked as the same category as the training set through manual review. In order to increase the amount of data, the pictures in the training set are randomly cut, randomly flipped, and randomly transformed for perspective as data enhancement, and the parameters of the ImageNet pre-trained classification ResNet50 model are used as the starting parameters for migration training. The loss function uses the cross entropy loss, and the gradient descent method is used during training. When the loss is minimized, the entire parameters of the model are obtained. Then, using this model, the output of the previous layer of the classifier can be taken to obtain the 2048-dimensional background features of the picture.

[0043] The suspected fraud scene pictures are clustered according to the background feature information to obtain the ID card scene clustering results and the headshot scene clustering results, wherein the ID card scene clustering results include several clusters of suspected fraud ID card pictures, and the headshot scene clustering results include several clusters of headshot scene clustering results.

[0044] As a preferred embodiment, for the extracted image background features, the present invention uses HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) for clustering. After the binary classification model, most non-fraudulent users have been filtered, but the clustered data samples still contain some noise data. Ordinary clustering algorithms are very sensitive to noise points. If a noise data is exactly between two clusters, the two clusters may be connected together, resulting in incorrect clustering. Density-based clustering (such as DBSCAN, HDBSCAN) can classify sparse samples with low density as noise, and cluster other non-noise samples at the same time. HDBSCAN extends DBSCAN by converting DBSCAN into a hierarchical clustering algorithm, and then uses the extraction plane clustering technology based on clustering stability. The biggest difference from traditional DBSCAN is that HDBSCAN can handle clustering problems with different densities. For the filtered fraud sample data, the sample density of similar background features is different, and in addition to the fraud gang data, there are also some sparse noise data in the clustered samples. The most important contribution of HDBSCAN is that it redefines the distance between samples. In most clustering algorithms, the distance between samples uses the Euclidean distance, but the Euclidean distance will lead to the curse of dimensionality, and the Euclidean distance cannot reflect the density of samples well, which leads to the clustering results being significantly disturbed by noise. In HDBSCAN, the core distance is first defined. The core distance of a sample is the distance between the sample and the kth nearest neighbor sample. The core distance can be expressed as:

[0045]

[0046] Where x represents the sample, represents the kth nearest neighbor sample of sample x, Represents x and Based on this, HDBSCAN defines the mutual reach distance between samples: :

[0047]

[0048] Among them, a and b represent samples respectively, and d(a,b) is the distance between a and b. HDBSCAN redefines the sample distance. The mutual reach distance makes the distance within the high-density cluster unaffected. At the same time, the distance between the noise sample point and the sample points of other clusters is increased, so that the single-link clustering can be more robust to noise samples. After defining the mutual reach distance, the samples can be connected by edges, and the mutual reach distance is the weight of the sample edge. In this way, by selecting the threshold for pruning, the minimum spanning tree can be established. Thus, the cluster hierarchy is constructed by threshold selection, the clustering tree is compressed, and the clusters are extracted to obtain the final Figure 3 The clusters shown.

[0049] According to the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, we compare the two combinations to calculate the IOU of the account numbers:

[0050]

[0051] in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where e and f are both natural numbers greater than zero. The account number sets whose overlap IOU is greater than the set overlap threshold are combined Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f. and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results. The output results will output pictures with potential similarity, which can be used to assist manual review and find fraud gangs. The overlap threshold can be set to 0.5.

[0052] In the embodiment of the present invention, the ID card scene clustering results and the headshot scene clustering results are used as normal clusters, and the suspected fraud scene pictures that cannot be clustered into the ID card scene clustering results and the headshot scene clustering results are used as noise clusters. The clustering results are re-filtered and integrated using the maximum intra-class distances of the noise clusters and the normal clusters, specifically including: calculating the maximum intra-class distances of the noise clusters and the normal clusters respectively. The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient. The value of B is preferably 0.8. Clusters with a value higher than 0.8 usually contain many noisy images.

[0053] like Figure 4 As shown, based on the above embodiments, those skilled in the art can understand that the present invention also provides a fraud detection and identification device based on image background similarity, including an acquisition module 1, a filtering module 2, a feature extraction module 3, a clustering module 4 and a clustering result integration output module 5.

[0054] The acquisition module 1 is used to acquire the image data uploaded by each user in the loan application, and the image data includes the ID card image and the headshot image.

[0055] The filtering module 2 is used to filter the normal background images in the image data based on the scene classification model to obtain suspected fraud scene images, which include suspected fraud ID card images and suspected fraud headshot images. Since fraud scene images account for a small part of all image data, if clustering is performed on all image data, there will be a lot of noise, resulting in very poor clustering results. Therefore, before clustering using image similarity, the suspected fraud scene is first identified, thereby filtering out most of the normal images.

[0056] As a preferred embodiment of the present invention, the present invention uses the deep learning network MobilenetV2 to perform binary classification on suspected fraud scenes and normal scenes. Through the pictures uploaded by historical overdue customers, it can be found that if overdue customers apply for loans through intermediaries, the scenes are usually office scenes. For example, the background of ID card photos is mostly the desktop, and there will be tables, papers, keyboards, etc. on the desktop; while normal scenes are usually outdoors, shopping malls, homes, etc. Based on the above experience, the scenes are classified into two categories, and the suspected fraud scenes and normal scenes are manually labeled before training. Figure 2 The ID card pictures uploaded by customers showing normal scenes and suspected fraud scenes are used. The present invention uses the pre-trained classification MobilenetV2 model of ImageNet itself and performs migration training based on this model. At the same time, data light enhancement, random cutting data enhancement, and random flipping data enhancement are added. This network calculates the difference between the predicted value and the true value through binary cross entropy, calculates the loss, and reduces the loss through back propagation, and finally obtains the network parameters. Through suspected scene detection, most normal pictures can be filtered out.

[0057] The feature extraction module 3 is used to extract the background feature information in the above-mentioned suspected fraud scene pictures based on the feature extraction model. As a preferred embodiment, the present invention also uses a deep learning model to classify the above-mentioned suspected fraud scene pictures, and the network output obtained in the previous layer of the classifier is the picture background feature. The present invention uses the picture classification task to train the feature extraction model. Considering that the multi-classification task is more complex than the binary classification task, the present invention preferably uses ResNet50 as the basic architecture, which is more complex, has more parameters, and has stronger fitting ability. For suspected fraud scene pictures, first mark the pictures of similar backgrounds as the same category as the training set through manual review. In order to increase the amount of data, the pictures in the training set are randomly cut, randomly flipped, and randomly transformed for perspective as data enhancement, and the parameters of the ImageNet pre-trained classification ResNet50 model are used as the starting parameters for migration training. The loss function uses the cross entropy loss, and the gradient descent method is used during training. When the loss is minimized, the entire parameters of the model are obtained. Then, using this model, the output of the previous layer of the classifier can be taken to obtain the 2048-dimensional background features of the picture.

[0058] Clustering module 4 is used to cluster suspected fraud scene pictures according to background feature information to obtain ID card scene clustering results and headshot scene clustering results, wherein the ID card scene clustering results include several clusters of suspected fraud ID card pictures, and the headshot scene clustering results include several clusters of headshot scene clustering results.

[0059] As a preferred embodiment, the present invention uses HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) for clustering. HDBSCAN extends DBSCAN by converting DBSCAN into a hierarchical clustering algorithm, and then uses the technology of extracting plane clustering based on clustering stability. The biggest difference from traditional DBSCAN is that HDBSCAN can handle clustering problems with different densities. The core of the clustering algorithm is single-link clustering, but single-link clustering is very sensitive to noise. If a noise data point is exactly between two clusters, the two clusters may be connected together, resulting in incorrect clustering. The most important contribution of HDBSCAN is that it redefines the distance between samples. In most clustering algorithms, the distance between samples uses Euclidean distance, but Euclidean distance will lead to dimensionality disaster, and Euclidean distance cannot reflect the density of points well, which leads to the clustering result being significantly disturbed by noise. In HDBSCAN, the core distance is first defined. The core distance of a sample is the distance between the sample and the kth nearest neighbor sample. The core distance can be expressed as:

[0060]

[0061] Where x represents the sample, represents the kth nearest neighbor sample of sample x, Represents x and Based on this, HDBSCAN defines the mutual reach distance between samples: :

[0062]

[0063] Among them, a and b represent samples respectively, and d(a,b) is the distance between a and b. HDBSCAN redefines the sample distance. The mutual reach distance makes the distance within the high-density cluster unaffected. At the same time, the distance between the noise sample point and the sample points of other clusters is increased, so that the single-link clustering can be more robust to noise samples. After defining the mutual reach distance, the samples can be connected by edges, and the mutual reach distance is the weight of the sample edge. In this way, by selecting the threshold for pruning, the minimum spanning tree can be established. Thus, the cluster hierarchy is constructed by threshold selection, the clustering tree is compressed, and the clusters are extracted to obtain the final Figure 3 The clusters shown.

[0064] The clustering result integration output module 5 is used to calculate the IOU of the account number based on the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, by pairwise comparison:

[0065]

[0066] in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where e and f are both natural numbers greater than zero. The account number sets whose overlap IOU is greater than the set overlap threshold are combined Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f. and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results. The output results will output pictures with potential similarity, which can be used to assist manual review and find fraud gangs. The overlap threshold can be set to 0.5.

[0067] The clustering result integration output module 5 of the embodiment of the present invention also takes the ID card scene clustering result and the headshot scene clustering result as normal clusters, and takes the suspected fraud scene pictures that cannot be clustered into the ID card scene clustering result and the headshot scene clustering result as noise clusters, and uses the maximum intra-class distance of the noise cluster and the normal cluster to re-filter and integrate the clustering results, specifically including: calculating the maximum intra-class distance of the noise cluster and the normal cluster for the noise cluster and the normal cluster respectively. The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient. The value of B is preferably 0.8. Clusters with a value higher than 0.8 usually contain many noisy images.

[0068] In the embodiment provided by the present invention, the ID card picture and the headshot picture are obtained with the consent of the user, and all the user data obtained complies with the relevant policies and legal provisions.

[0069] The above is only a preferred embodiment of the present invention. It should be noted that for ordinary technicians in this technical field, other parts not specifically described belong to the prior art or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

Claims

1. A fraud detection and identification method based on image background similarity, characterized in that: include: Obtaining the picture data uploaded by each user in the loan application, the picture data including the ID card picture and the headshot picture; filtering the normal background images in the image data based on the scene classification model to obtain suspected fraud scene images, wherein the suspected fraud scene images include suspected fraud identity card images and suspected fraud headshot images; Extracting background feature information from the suspected fraud scene image based on a feature extraction model; Clustering the suspected fraud scene pictures according to the background feature information to obtain ID card scene clustering results and headshot scene clustering results, wherein the ID card scene clustering results and headshot scene clustering results include suspected fraud ID card pictures and suspected fraud headshot pictures of several clusters in turn; According to the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, the account number overlap IOU is calculated by pairwise comparison: ； in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where both e and f are natural numbers greater than zero; Combine the account number sets whose overlap IOU is greater than the set overlap threshold Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f; The account number is and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results; The ID card scene clustering results and the headshot scene clustering results are used as normal clusters, and the suspected fraud scene pictures that cannot be clustered into the ID card scene clustering results and the headshot scene clustering results are used as noise clusters. The clustering results are re-filtered and integrated using the maximum intra-class distances of the noise clusters and the normal clusters, specifically including: calculating the maximum intra-class distances of the noise clusters and the normal clusters respectively. The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient.

2. The fraud detection and identification method based on image background similarity according to claim 1, characterized in that: The value of B is 0.

8.

3. The fraud detection and identification method based on image background similarity according to claim 1, characterized in that: The overlap threshold is 0.

5.

4. A fraud detection and identification device based on image background similarity, characterized in that: include: An acquisition module is used to acquire the image data uploaded by each user in the loan application, wherein the image data includes an ID card image and a headshot image; A filtering module, for filtering the normal background images in the image data based on the scene classification model to obtain suspected fraud scene images, wherein the suspected fraud scene images include suspected fraud identity card images and suspected fraud headshot images; A feature extraction module, used to extract background feature information in the suspected fraud scene image based on a feature extraction model; A clustering module, for clustering suspected fraud scene pictures according to the background feature information to obtain ID card scene clustering results and headshot scene clustering results, wherein the ID card scene clustering results and headshot scene clustering results include suspected fraud ID card pictures and suspected fraud headshot pictures of several clusters in turn; The clustering result integration output module is used to calculate the IOU of the account number based on the account numbers corresponding to the ID card scene clustering results and the headshot scene clustering results, by pairwise comparison: ； in, is the account number set of the e-th cluster in the ID card scene clustering result, is the account number set of the fth cluster in the headshot scene clustering result, where e and f are both natural numbers greater than zero; the account number sets whose overlap IOU is greater than the set overlap threshold are combined Determine the same cluster and take the set in the account number set combination With Collection The union of gets the account number set , and grouped by account number As the result of the account number clustering, m is a natural number ranging from 1 to e, and n is a natural number ranging from 1 to f; and Add the account number set in the class cluster The corresponding and missing suspected fraudulent ID card pictures and suspected fraudulent headshot pictures are output as clustering results; The clustering result integration output module uses the ID card scene clustering results and the headshot scene clustering results as normal clusters, and uses the suspected fraud scene images that cannot be clustered into the ID card scene clustering results and the headshot scene clustering results as noise clusters, and uses the maximum intra-class distances of the noise clusters and the normal clusters to re-filter and integrate the clustering results, specifically including: calculating the maximum intra-class distance of the noise cluster for the noise cluster and the normal cluster respectively The maximum distance within the class and the normal class , the intra-class distance in the normal class cluster Greater than the set distance threshold Cluster removal, where the distance threshold is set , B is the threshold coefficient.

5. The fraud detection and identification device based on image background similarity according to claim 4, characterized in that: The value of B is 0.

8.

6. The fraud detection and identification device based on image background similarity according to claim 4, characterized in that: The overlap threshold is 0.5.

Images