Fraud image screening method and device based on key point filtering, equipment and medium

Abstract

The invention provides a fraud image screening method and device based on key point filtering, equipment and a medium, and relates to the technical field of image processing, and the method comprises the steps: separating a face in a first image through image segmentation, and obtaining a second image; performing vectorization processing on the second image, and retrieving from a database to obtain a similar image; key point extraction is carried out on the second image to obtain feature points; matching the feature points of the second image with the feature points of each similar image to obtain similar points; removing similar points on the edge contour of the human face; clustering the remaining similar points through a clustering algorithm, and extracting a similar main body region; and screening according to the number of similar points in the similar main body region to obtain a target image, and performing multi-dimensional information clustering on the target image to obtain a fraud evaluation image. According to the method, the similar pictures can be accurately mined, the number of the similar pictures is controlled, the manual auditing pressure is reduced, and the fraud image matching accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method, device, equipment and medium for screening fraudulent images based on key point filtering. Background Art

[0002] With the development of the Internet and AI technology in the financial field, the risk of fraud in transactions has also increased. Existing image anti-fraud systems compare the current image scene with existing fraud image scenes to identify potential fraudulent behavior. However, the existing technology still has the following problems in practical applications: Existing technologies often use overall image comparison to compare images, which is easily interfered by the face part, and the background image often contains too much noise or unimportant feature points, resulting in these noises interfering with effective feature extraction when extracting key points, and failing to accurately highlight the similar parts. These problems will result in a large number of inaccurate fraud images being screened out, and the accuracy of fraud risk judgment is low. Summary of the invention

[0003] The purpose of the present invention is to provide a method, device, equipment and medium for screening fraudulent images based on key point filtering to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows: In a first aspect, the present application provides a fraudulent image screening method based on key point filtering, comprising: Acquire a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image; Vectorizing the second image, and retrieving similar images from a database by calculating cosine similarity distance; Extracting key points from the second image and performing denoising processing to obtain feature points; Matching the feature points of the second image with the feature points of each similar image and performing denoising processing to obtain similar points; Extracting facial edge contour points from the second image, and removing similar points that coincide with the coordinates of the facial edge contour points; Clustering the remaining similar points by a clustering algorithm to obtain a first clustering result, extracting similar subject regions based on the first clustering result, and selecting similar points in the similar subject regions; Similar images are screened according to the number of similar points in similar subject areas to obtain a target image, multi-dimensional information clustering is performed on the target image to obtain a second clustering result, and a fraud assessment image is screened according to the second clustering result.

[0004] In a second aspect, the present application also provides a fraudulent image screening device based on key point filtering, comprising: A segmentation module is used to obtain a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image; A retrieval module, used for vectorizing the second image and retrieving similar images from a database by calculating a cosine similarity distance; An extraction module, used to extract key points from the second image and perform denoising to obtain feature points; A matching module, used for matching the feature points of the second image with the feature points of each similar image and performing denoising processing to obtain similar points; A removal module, used for extracting face edge contour points from the second image, and removing similar points whose coordinates coincide with the face edge contour points; A first clustering module is used to cluster the remaining similar points by a clustering algorithm to obtain a first clustering result, extract similar subject regions based on the first clustering result, and select similar points in the similar subject regions; The second clustering module is used to screen similar images according to the number of similar points in similar subject areas to obtain a target image, perform multi-dimensional information clustering on the target image to obtain a second clustering result, and screen the fraud assessment image according to the second clustering result. In a third aspect, the present application also provides a fraudulent image screening device using key point filtering, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the fraudulent image screening method using key point filtering as described above are implemented.

[0005] In a fourth aspect, the present application further provides a readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the fraudulent image screening method such as the above-mentioned key point filtering are implemented.

[0006] The beneficial effects of the present invention are: This application uses image segmentation technology to obtain image background information, extract key points in the image background information, filter the key points, perform similarity detection, and return multiple similar images; use pre-trained models in the process of image key point extraction and similarity matching, convert the confidence of feature points into probability, and through threshold screening, effectively avoid the time cost and resource consumption required for model fine-tuning, and obtain better results; by using the DBSCAN clustering algorithm for similar points in the image background, it can effectively remove noise points and abnormal points, thereby highlighting similar subjects in the background. This method can accurately mine similar pictures while controlling the number of similar pictures, reducing the pressure of manual review, and effectively judging whether there is a risk of fraud in the current image scene.

[0007] Other features and advantages of the present invention will be set forth in the following description, and in part will become apparent from the description, or may be understood by practicing embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 This is a flow chart of a fraudulent image screening method based on key point filtering in an embodiment of the present application; Figure 2 This is a structural diagram of a fraudulent image screening device based on key point filtering in an embodiment of the present application; Figure 3 This is a structural diagram of a fraudulent image screening device based on key point filtering in an embodiment of the present application.

[0010] Markings in the figure: 100-segmentation module; 200-retrieval module; 300-extraction module; 400-matching module; 500-removal module; 600-first clustering module; 700-second clustering module; 800-key point filtering fraud image screening device; 801-processor; 802-memory; 803-multimedia component; 804-I / O interface; 805-communication component. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0012] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0013] Existing image recognition anti-fraud technology has many drawbacks: When performing image verification, the recognition and verification is mainly carried out through the face image, and the background information of the image is not taken into consideration. When faced with new facial fraud methods, the key judgment basis is easy to be tampered with and the reliability is insufficient; When extracting key feature points of an image, since the face information is dominant, the extracted key points mainly represent the internal and edge features of the face, while the background features of the image are not fully utilized; There is a lack of effective denoising methods for the extracted key feature points, and redundant noise points remain, interfering with similarity judgment, thereby weakening the effectiveness of the anti-fraud system.

[0014] The existing technology ultimately requires manual review and judgment of the obtained similar images. If there are too many similar images, the manual review will be stressful.

[0015] In response to the above problems, the present invention extracts key feature points from the image background information, filters the key points, performs similarity detection and returns multiple similar images, and then further screens similar images through multidimensional data clustering. This can not only accurately match similar pictures, but also reduce the pressure of manual review and effectively determine whether there is a risk of fraud in the current image scene. Example 1

[0016] See also Figure 1 , the present application provides a fraudulent image screening method based on key point filtering, comprising: steps S100, S200, S300, S400, S500, S600 and S700; S100, obtaining a first image, separating a face in the first image by image segmentation, and masking the separated face area to obtain a second image; specifically comprising: S110, uses a hierarchical Transformer encoder to encode images to obtain multi-scale features; each layer of the encoder processes different feature scales to capture multi-scale features. First, the input image is divided into multiple non-overlapping small blocks (patches), and then each patch is embedded into a high-dimensional vector space through a linear layer. Each layer of Transformer Block processes the output of the previous layer and extracts more complex and high-level features layer by layer.

[0017] S120, using a multi-layer perceptron decoder to map the multi-scale features to the same dimension, and perform feature fusion, generating a segmentation map with the same size as the input image through an upsampling operation, and extracting a face region from the segmentation map; S130: Fill the pixels in the face area with black to obtain a second image.

[0018] Storing face mask images in the database in the form of arrays not only reduces memory consumption but also effectively protects customer privacy.

[0019] S200, vectorizing the second image, and retrieving a similar image from a database by calculating a cosine similarity distance; The database stores existing fraudulent images. This step is to preliminarily select images that are relatively similar to the current second image from the existing fraudulent images, and then perform feature point matching; The second image is represented by a vector using the pre-trained Swin Transformer model, and the resulting feature vector dimension is 768 dimensions, which is saved in the database. The cosine similarity distance between the feature vector of the current image and the feature vector of the fraudulent image in the database is calculated to obtain the top N most similar images and obtain similar images.

[0020] S300, extracting key points from the second image and performing denoising processing to obtain feature points; The key point information in the second image is extracted using the preset SuperPoint model, and the key point information includes the coordinate information of each key point. , the first confidence of the key point And the feature vector of the key point ; Since the training data of the SuperPoint model and the data of the fraudulent image have distribution differences, the key feature points extracted by SuperPoint may be biased. To avoid this problem, the present invention dynamically screens high-quality key points based on the first confidence level and the similarity of the feature vectors. Specifically: Normalizing the first confidence level to obtain a probability value for each key point; ; Among them, i is the serial number of the key point, is the probability value of key point i, is the maximum value of the first confidence of the key point, is the minimum value of the first confidence of the key point, is the first confidence of key point i; Based on the feature vectors of the key points, a similarity matrix between feature vectors is obtained, and the mean of the similarity between each key point and its k nearest key points is calculated and normalized to obtain a similarity score. ; The key points are screened based on the probability value and the similarity score to obtain feature points.

[0021] Combined probability value and similarity score , the final score of the key point is calculated as: ;

[0022] in, is the weight coefficient, is the final score.

[0023] Finally, the final score is screened using the set threshold to remove noise points in the second image, thereby obtaining high-quality key points as feature points for subsequent feature point matching.

[0024] S400, matching the feature points of the second image with the feature points of each similar image and performing denoising processing to obtain similar points, including: The process of feature point matching includes: extracting significant feature points from the image through feature point detection algorithms (such as Harris corner detection, SIFT, SURF, etc.), which are usually located in areas with rich textures or obvious edges in the image. Then, a descriptor is calculated for each feature point. The descriptor is a vector based on local image information. Gradient information descriptors such as SIFT and SURF, or binary descriptors such as BRIEF and ORB can be used. Next, the feature point descriptors in the two images are matched using a similarity metric to find the corresponding feature point pairs and complete the match. Finally, the feature point pairs are screened through the attention mechanism: S410, acquiring feature vectors and coordinate information of feature points, and processing the feature vectors and coordinate information of the same image through a self-attention mechanism to obtain a first processing result; Self-Attention is performed on each image separately, allowing each point to pay attention to the information of other points in its own image, so as to better understand the contextual relationship of local features in the entire image.

[0025] S420, using a cross attention mechanism to process feature vectors and coordinate information of the second image and similar images to obtain a second processing result; The features processed by Self-Attention will enter the Cross-Attention module to interact the features of the two images. This interaction helps to find possible corresponding points in the two images.

[0026] S430, calculating a second confidence level for each feature point based on the first processing result and the second processing result, and removing the feature points based on the second confidence level; Based on the importance scores of self-attention and cross-attention of each feature point, normalization or processing through the softmax function is performed to obtain the second confidence. Points with lower second confidence are considered unlikely to be correctly matched points, and these points are removed, thereby reducing the amount of data for subsequent processing and improving computational efficiency.

[0027] S440 , using a preset LightGlue model to calculate similarity for the remaining feature points, and screening the feature points based on the similarity to obtain similar points.

[0028] LightGlue is a graph neural network model designed for local feature matching. It can not only self-evaluate the accuracy of prediction results, but also improve processing efficiency by terminating calculations early and removing unmatched feature points while ensuring accuracy.

[0029] S500, extracting facial edge contour points from the second image, and removing similar points whose coordinates coincide with the facial edge contour points; Although the face information is masked in the above steps, there are still a large number of similar feature points at the face boundaries in the two images, which seriously affect the importance of the feature points in the image background. Therefore, the present invention uses image expansion technology to obtain the coordinate points of the outer contour of the face, and removes the similar points belonging to the area, and only retains the similar points in the background image.

[0030] The similar points obtained through the above steps can well identify the similar parts in the background of the two images. These similar points include not only similar main parts (such as slogans on the wall), but also some non-main parts (such as downlights). The number of similar feature points in the main part is large and densely distributed, while the number of similar feature points in the non-main part is small and discretely distributed. In order to further extract similar main parts, step S600 combines the spatial position information and local feature information of the similar points to cluster and screen them, and further extract similar subjects in the background image.

[0031] S600, clustering the remaining similar points by a clustering algorithm to obtain a first clustering result, extracting a similar subject region based on the first clustering result, and selecting similar points in the similar subject region; Acquire the feature vector and coordinate information of the similar points, and perform normalization processing on the feature vector and coordinate information of the similar points respectively; The processed feature vectors and coordinate information are spliced ​​to obtain the overall characteristics of the similar points; Based on the DBSCAN clustering algorithm, the overall characteristics of similar points are clustered, and the clusters are screened according to the number and distribution density of data points in the clusters to obtain similar main areas.

[0032] DBSCAN is a density-based clustering algorithm that is suitable for discovering clusters of arbitrary shapes and has the ability to handle noisy data. It defines clusters by the density of data points, relying on two main parameters: neighborhood radius (Eps) and minimum number of points (MinPts). The algorithm divides data points into three categories: core points, boundary points, and noise points. If the number of points in the Eps neighborhood of a point is not less than MinPts, the point is a core point; if the point is in the neighborhood of a core point but is not a core point itself, it is a boundary point; otherwise, it is a noise point.

[0033] The steps of DBSCAN clustering algorithm are as follows: Initialization: Select an arbitrary unvisited data point and mark it as "visited"; Core point check: Calculate the number of points in the Eps neighborhood of the point. If the number of points is greater than or equal to MinPts, the point is a core point, create a new cluster, and add all the points in the Eps neighborhood to the cluster; if the number of points is less than MinPts, the point is marked as a noise point (it may be reclassified as a boundary point in subsequent steps); Cluster expansion: Repeat the core point check for the points newly added to the cluster. If a new core point is found, merge the points in its neighborhood into the current cluster; Process the next point: Repeat the above steps for all unvisited points until all points have been visited; In the present invention, by setting two parameters, Eps and MinPts, similar points are clustered and only similar points in the clusters are retained, so that similar main parts can be obtained.

[0034] S700, screening similar images according to the number of similar points in similar subject areas to obtain a target image, performing multi-dimensional information clustering on the target image to obtain a second clustering result, and screening according to the second clustering result.

[0035] S710, sorting the N similar images according to the number of similar points from most to least, selecting the first M images with the most similar points, and obtaining the target image; S720, obtaining geographic coordinate information and shooting device information of the target image, normalizing the geographic coordinate information to obtain standard coordinate features, and performing one-hot encoding on the shooting device information to obtain standard device features; Normalization of geographic coordinate information: Geographic coordinate information is latitude and longitude, usually continuous values. The range of these coordinate values ​​may be relatively large, especially longitude from -180 to 180, latitude from -90 to 90. The normalization method is as follows: ;

[0036] Where: L is the original latitude value, is the maximum latitude, is the maximum latitude, is the normalized latitude value; Similarly: ; Where: G is the original longitude value, Maximum longitude, is the maximum longitude, is the normalized longitude value.

[0037] Shooting device information encoding: Use One-Hot encoding to convert the device category variable into binary form. Each category becomes a separate column with a value of 0 or 1.

[0038] In addition, if you need to use cosine similarity distance for clustering later, you should first normalize the cosine similarity distance and scale it to [0,1]. You can also use the Min-Max normalization method. S730, based on the number of similar points, cosine similarity distance, standard coordinate features and standard device features, the M images are clustered using a K-Means clustering algorithm to obtain the number of images in each cluster, and the clusters are screened according to the number of images to obtain fraud assessment images; specifically: Although different data features are used in clustering, the importance of each type of data feature is different, so it is necessary to set different weights for each feature to obtain the corresponding weight feature vector. The method is as follows: ;

[0039] in: is the weight feature vector, is the normalized data feature, is the corresponding weight, and j represents the jth data feature.

[0040] If the second image is clustered into a cluster alone, it is considered that the image has a low similarity with other fraudulent images and there is no fraud risk; If the second image belongs to the same cluster as the other images, it is considered to have a high similarity with the fraudulent images in the cluster, and the images in the cluster need to be provided to the business department for manual judgment. Example

[0041] See also Figure 2 In this embodiment, a fraud image screening device based on key point filtering includes: The segmentation module 100 is used to obtain a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image; A retrieval module 200 is used to vectorize the second image and retrieve similar images from a database by calculating a cosine similarity distance; An extraction module 300 is used to extract key points from the second image and perform denoising to obtain feature points; A matching module 400, configured to match the feature points of the second image with the feature points of each similar image and perform denoising processing to obtain similar points; A removal module 500 is used to extract facial edge contour points from the second image, and remove similar points that coincide with the coordinates of the facial edge contour points; A first clustering module 600 is used to cluster the remaining similar points by a clustering algorithm to obtain a first clustering result, extract similar subject regions based on the first clustering result, and select similar points in the similar subject regions; The second clustering module 700 is used to screen similar images according to the number of similar points in similar subject areas to obtain a target image, perform multi-dimensional information clustering on the target image to obtain a second clustering result, and screen the fraud assessment image according to the second clustering result.

[0042] As an optional implementation, the segmentation module 100 includes: An encoding unit 110, configured to encode the image using a hierarchical Transformer encoder to obtain multi-scale features; A decoding unit 120, configured to map the multi-scale features to the same dimension using a multi-layer perceptron decoder, perform feature fusion, generate a segmentation map of the same size as the input image through an upsampling operation, and extract a face region from the segmentation map; The filling unit 130 is used to fill the pixels in the face area with black to obtain a second image. Example

[0043] Corresponding to the above method embodiment, this embodiment also provides a fraudulent image screening device based on key point filtering. The fraudulent image screening device based on key point filtering described below and the fraudulent image screening method based on key point filtering described above can be referenced to each other.

[0044] Figure 3 8 is a block diagram of a fraud image screening device 800 based on key point filtering according to an exemplary embodiment. Figure 3As shown, the fraud image screening device 800 based on key point filtering includes a processor 801 and a memory 802. The fraud image screening device 800 based on key point filtering may also include one or more of a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805. Among them, the processor 801 is used to control the overall operation of the fraud image screening device 800 based on key point filtering to complete all or part of the steps in the above-mentioned fraud image screening method based on key point filtering. The memory 802 is used to store various types of data to support the operation of the fraud image screening device 800 based on key point filtering, and these data may include, for example, commands for any application or method operating on the fraud image screening device 800 based on key point filtering, as well as application-related data, such as contact data, messages sent and received, pictures, audio, video, etc. The memory 802 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, disk or optical disk.

[0045] The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone, and the microphone is used to receive external audio signals.

[0046] The received audio signal can be further stored in the memory 802 or sent through the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules can be keyboards, mice, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the fraud image screening device 800 based on key point filtering and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 can include: Wi-Fi module, Bluetooth module, NFC module.

[0047] In an exemplary embodiment, the device 800 for mutual signing and verification of digital files may be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned fraud image screening method based on key point filtering. Example

[0048] Corresponding to the above embodiment of the fraudulent image screening method based on key point filtering, this embodiment also provides a readable storage medium, and the readable storage medium described below and the fraudulent image screening method based on key point filtering described above can be referenced to each other.

[0049] A readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned fraudulent image screening method embodiment based on key point filtering.

[0050] The readable storage medium may specifically be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or other readable storage medium that can store program codes.

[0051] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0052] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A fraud image screening method based on key point filtering, characterized in that: include: Acquire a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image; Vectorizing the second image, and retrieving similar images from a database by calculating cosine similarity distance; Extracting key points from the second image and performing denoising processing to obtain feature points; Matching the feature points of the second image with the feature points of each similar image and performing denoising processing to obtain similar points; Extracting facial edge contour points from the second image, and removing similar points that coincide with the coordinates of the facial edge contour points; Clustering the remaining similar points by a clustering algorithm to obtain a first clustering result, extracting similar subject regions based on the first clustering result, and selecting similar points in the similar subject regions; Similar images are screened according to the number of similar points in similar subject areas to obtain a target image, multi-dimensional information clustering is performed on the target image to obtain a second clustering result, and a fraud assessment image is screened according to the second clustering result.

2. The fraudulent image screening method based on key point filtering according to claim 1 is characterized in that: Acquire a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image, including: A hierarchical Transformer encoder is used to encode the image to obtain multi-scale features; A multi-layer perceptron decoder is used to map the multi-scale features to the same dimension, and feature fusion is performed, a segmentation map having the same size as the input image is generated through an upsampling operation, and a face region is extracted from the segmentation map; The pixels in the face area are filled with black to obtain a second image.

3. The fraudulent image screening method based on key point filtering according to claim 1 is characterized in that: Extract key points from the second image and perform denoising to obtain feature points, including: Extract key point information in the second image using a preset SuperPoint model, wherein the key point information includes coordinate information of each key point, a first confidence level of the key point, and a feature vector of the key point; Normalizing the first confidence level to obtain a probability value for each key point; Based on the feature vectors of the key points, the mean of the similarities between each key point and its k nearest key points is calculated, and normalized to obtain a similarity score; The key points are screened based on the probability value and the similarity score to obtain feature points.

4. The fraudulent image screening method based on key point filtering according to claim 1 is characterized in that: The feature points of the second image are matched with the feature points of each similar image and denoised to obtain similar points, including: Acquire feature vectors and coordinate information of feature points, process the feature vectors and coordinate information of the same image through a self-attention mechanism, and obtain a first processing result; Using a cross attention mechanism to process feature vectors and coordinate information of the second image and similar images, to obtain a second processing result; Calculating a second confidence level for each feature point based on the first processing result and the second processing result, and removing the feature points based on the second confidence level; The preset LightGlue model is used to calculate the similarity of the remaining feature points, and the feature points are screened based on the similarity to obtain similar points.

5. The fraudulent image screening method based on key point filtering according to claim 1 is characterized in that: Clustering the remaining similar points by a clustering algorithm to obtain a first clustering result, and extracting similar subject areas based on the first clustering result, including: Acquire the feature vector and coordinate information of the similar points, and perform normalization processing on the feature vector and coordinate information of the similar points respectively; The processed feature vectors and coordinate information are spliced ​​to obtain the overall characteristics of the similar points; Based on the DBSCAN clustering algorithm, the overall characteristics of similar points are clustered, and the clusters are screened according to the number and distribution density of data points in the clusters to obtain similar main areas.

6. The fraudulent image screening method based on key point filtering according to claim 1 is characterized in that: The similar images are screened according to the number of similar points in the similar subject area to obtain a target image, multi-dimensional information clustering is performed on the target image to obtain a second clustering result, and the fraud assessment image is screened according to the second clustering result, including: Sort similar images by the number of similar points, select a preset number of images with the highest ranking, and obtain the target image; Acquire geographic coordinate information and shooting device information of the target image, perform normalization processing on the geographic coordinate information to obtain standard coordinate features, and perform one-hot encoding on the shooting device information to obtain standard device features; Based on the number of similar points, cosine similarity distance, standard coordinate features and standard device features, the K-Means clustering algorithm is used to cluster the target images to obtain the number of pictures in each cluster. The clusters are screened according to the number of pictures to obtain fraud assessment images.

7. A fraudulent image screening device based on key point filtering, characterized in that: include: A segmentation module is used to obtain a first image, separate a face in the first image by image segmentation, and mask the separated face area to obtain a second image; A retrieval module, used for vectorizing the second image and retrieving similar images from a database by calculating a cosine similarity distance; An extraction module, used to extract key points from the second image and perform denoising to obtain feature points; A matching module, used for matching the feature points of the second image with the feature points of each similar image and performing denoising processing to obtain similar points; A removal module, used for extracting face edge contour points from the second image, and removing similar points whose coordinates coincide with the face edge contour points; A first clustering module is used to cluster the remaining similar points by a clustering algorithm to obtain a first clustering result, extract similar subject regions based on the first clustering result, and select similar points in the similar subject regions; The second clustering module is used to screen similar images according to the number of similar points in similar subject areas to obtain a target image, perform multi-dimensional information clustering on the target image to obtain a second clustering result, and screen the fraud assessment image according to the second clustering result.

8. The fraudulent image screening device based on key point filtering according to claim 7 is characterized in that: The segmentation module comprises: The encoding unit is used to encode the image using a hierarchical Transformer encoder to obtain multi-scale features; A decoding unit, used to map the multi-scale features to the same dimension using a multi-layer perceptron decoder, perform feature fusion, generate a segmentation map of the same size as the input image through an upsampling operation, and extract a face region from the segmentation map; A filling unit is used to fill the pixels in the face area with black to obtain a second image.

9. A fraudulent image screening device based on key point filtering, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the fraudulent image screening method based on key point filtering as claimed in any one of claims 1 to 6 when executing the computer program.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the fraudulent image screening method based on key point filtering as described in any one of claims 1 to 6.

Images