Claim settlement information processing method and device, electronic equipment and storage medium

Abstract

The invention provides a claim settlement information processing method and device, electronic equipment and a storage medium, and relates to the technical field of image processing. The method comprises the following steps: acquiring image samples corresponding to a plurality of historical claim settlement cases, extracting an initial feature vector corresponding to each image sample by using a pre-trained network model, and constructing a historical image feature library based on the initial feature vector corresponding to each image sample; receiving a to-be-reviewed image of a current claim settlement case submitted by a user, and extracting a current feature vector corresponding to the to-be-reviewed image by using the network model; determining whether the to-be-reviewed image is an abnormal image according to the current feature vector corresponding to the to-be-reviewed image and a historical image feature library; if yes, risk early warning information is generated and output to a claim settlement auditing interface, the automation and intelligence level of after-sales receipt auditing is improved, and the receipt processing efficiency and the user experience are improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a method, apparatus, electronic device, and storage medium for processing claims information. Background Technology

[0002] In car after-sales claims, users typically need to submit visual evidence (such as photos of the faulty parts of the vehicle, detailed damage diagrams, repair process records, etc.) to support their claims. These visual materials are the core basis for determining liability and assessing losses.

[0003] Currently, the mainstream automotive after-sales claims review process primarily relies on manual verification of the authenticity of image evidence. Specifically, reviewers visually compare the content of images submitted for the current case and make subjective judgments based on historical experience to identify whether there are duplicate submissions or forgery / tampering.

[0004] However, the manual review model is inefficient and easily affected by individual subjective factors, resulting in inconsistent review standards and a high rate of misjudgment. Summary of the Invention

[0005] The purpose of this application is to address the shortcomings of the prior art by providing a claims information processing method, apparatus, electronic device, and storage medium to improve the accuracy, efficiency, and risk control level of after-sales document review.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for processing claims information, the method comprising: Image samples corresponding to multiple historical claims cases are obtained, and an initial feature vector corresponding to each image sample is extracted using a pre-trained network model. Based on the initial feature vector corresponding to each image sample, a historical image feature library is constructed. Receive the pending image of the current claim case submitted by the user, and use the network model to extract the current feature vector corresponding to the pending image; Based on the current feature vector corresponding to the image to be reviewed and the historical image feature library, it is determined whether the image to be reviewed is an abnormal image; If so, a risk warning message will be generated and output to the claims review interface.

[0007] Optionally, constructing a historical image feature library based on the initial feature vectors corresponding to each of the image samples includes: The initial feature vectors corresponding to the image samples are subjected to dimensionality reduction and quantization compression to generate an index sequence corresponding to the image samples. The identifier, index sequence, and metadata corresponding to the image sample are associated and stored in the historical image feature library.

[0008] Optionally, the step of performing dimensionality reduction and quantization compression on the initial feature vector corresponding to the image sample to generate the index sequence corresponding to the image sample includes: The initial feature vector corresponding to the image sample is reduced in dimensionality using deep principal component analysis to obtain the dimensionality-reduced features of the image sample. The dimensionality-reduced features of the image sample are then normalized to obtain the normalized features of the image sample. Sampling product quantization involves quantizing the normalized features corresponding to the image samples to obtain the index sequence corresponding to the image samples.

[0009] Optionally, determining whether the image to be reviewed is an abnormal image based on the current feature vector corresponding to the image to be reviewed and the historical image feature library includes: Based on the index sequence corresponding to the image to be reviewed and the index sequence corresponding to each image sample in the historical image feature library, at least one candidate image is obtained; The similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed is determined, and the image to be reviewed is identified as an abnormal image based on the metadata of the candidate image with the highest similarity.

[0010] Optionally, obtaining at least one candidate image based on the index sequence corresponding to the image to be examined and the index sequence corresponding to each image sample in the historical image feature library includes: The approximate nearest neighbor search method is used to search for at least one candidate cluster that is closest to the index sequence corresponding to the image to be examined from the index sequence corresponding to each of the image samples; At least one candidate image is obtained by taking the image samples corresponding to each candidate cluster as candidate images.

[0011] Optionally, determining the similarity between the initial feature vector corresponding to each of the candidate images and the current feature vector corresponding to the image to be examined includes: The cosine similarity is used to calculate the cosine similarity between the initial feature vector corresponding to the candidate image and the current feature vector corresponding to the image to be reviewed.

[0012] Optionally, identifying whether the image to be examined is an anomalous image based on the metadata of the candidate image with the highest similarity includes: If the similarity of the candidate image with the highest similarity is greater than a preset threshold, and the metadata of the candidate image indicates that the candidate image belongs to different claims cases, then the image to be reviewed is determined to be an abnormal image.

[0013] Secondly, embodiments of this application also provide a claims information processing device, the device comprising: The acquisition module is used to acquire image samples corresponding to multiple historical claims cases; The extraction module is used to extract the initial feature vectors corresponding to each image sample using a pre-trained network model. The construction module is used to construct a historical image feature library based on the initial feature vectors corresponding to each of the image samples; The receiving module is used to receive pending images of the current claim submitted by the user; The extraction module is further configured to use the network model to extract the current feature vector corresponding to the image to be reviewed; The determination module is used to determine whether the image to be reviewed is an abnormal image based on the current feature vector corresponding to the image to be reviewed and the historical image feature library; The output module is used to generate risk warning information and output the risk warning information to the claims review interface if the condition is met.

[0014] Optionally, the building module is specifically used for: The initial feature vectors corresponding to the image samples are subjected to dimensionality reduction and quantization compression to generate an index sequence corresponding to the image samples. The identifier, index sequence, and metadata corresponding to the image sample are associated and stored in the historical image feature library.

[0015] Optionally, the building module is specifically used for: The initial feature vector corresponding to the image sample is reduced in dimensionality using deep principal component analysis to obtain the dimensionality-reduced features of the image sample. The dimensionality-reduced features of the image sample are then normalized to obtain the normalized features of the image sample. Sampling product quantization involves quantizing the normalized features corresponding to the image samples to obtain the index sequence corresponding to the image samples.

[0016] Optionally, the determining module is specifically used for: Based on the index sequence corresponding to the image to be reviewed and the index sequence corresponding to each image sample in the historical image feature library, at least one candidate image is obtained; The similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed is determined, and the image to be reviewed is identified as an abnormal image based on the metadata of the candidate image with the highest similarity.

[0017] Optionally, the determining module is specifically used for: The approximate nearest neighbor search method is used to search for at least one candidate cluster that is closest to the index sequence corresponding to the image to be examined from the index sequence corresponding to each of the image samples; At least one candidate image is obtained by taking the image samples corresponding to each candidate cluster as candidate images.

[0018] Optionally, the determining module is specifically used for: The cosine similarity is used to calculate the cosine similarity between the initial feature vector corresponding to the candidate image and the current feature vector corresponding to the image to be reviewed.

[0019] Optionally, the determining module is specifically used for: If the similarity of the candidate image with the highest similarity is greater than a preset threshold, and the metadata of the candidate image indicates that the candidate image belongs to different claims cases, then the image to be reviewed is determined to be an abnormal image.

[0020] Thirdly, embodiments of this application provide an electronic device, which includes a memory for storing one or more programs and a processor. When the one or more programs are executed by the processor, the aforementioned claims information processing method is implemented.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned claims information processing method.

[0022] The beneficial effects of this application are: This application provides a claims information processing method, apparatus, electronic device, and storage medium. First, image samples attached to multiple closed historical claims cases are acquired. An initial feature vector corresponding to each image sample is extracted using a pre-trained network model. Based on the initial feature vectors of each image sample, a historical image feature library is constructed. Then, the same network model is used to extract the current feature vector corresponding to the image to be reviewed in the current claims case. Next, based on the current feature vector of the image to be reviewed and the historical image feature library, it is determined whether the image to be reviewed is an abnormal image. That is, using artificial intelligence, it can automatically detect whether the image submitted by the user has been used in historical claims records, even if the image has undergone common modifications. This achieves automated document recognition review, reduces the risk of erroneous or incorrect claims due to duplicate or forged evidence, and improves the automation and intelligence level of after-sales claims review. Finally, when an abnormal image is detected, risk warning information is generated and pushed to the front-end interface of the claims review system in real time, reducing the burden of manual review, improving claims processing efficiency and user experience, and providing traceable and verifiable technical support for the after-sales risk control system.

[0023] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, 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 this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a claims information processing method provided in this application embodiment; Figure 2 A flowchart illustrating another claims information processing method provided in this application embodiment; Figure 3 A flowchart illustrating another claims information processing method provided in this application embodiment; Figure 4 A flowchart illustrating another claims information processing method provided in this application embodiment; Figure 5 A flowchart illustrating another claims information processing method provided in this application embodiment; Figure 6 A schematic diagram of the structure of a claims information processing device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0028] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] Please see Figure 1 , Figure 1 This is a flowchart illustrating a claims information processing method provided in an embodiment of this application. The executing entity of this method can be an electronic device with data processing capabilities, such as a computer, server, or smartphone. Figure 1 As shown, the method includes: S101. Obtain image samples corresponding to multiple historical claims cases, and use the pre-trained network model to extract the initial feature vectors corresponding to each image sample, and construct a historical image feature library based on the initial feature vectors corresponding to each image sample.

[0030] Optionally, image samples attached to multiple closed historical claims can be obtained from the insurance company's historical claims system to form an image sample set. Each image sample stores: a case number and an image type (e.g., vehicle damage image, medical bill, medical record image, fire scene photo, etc.). For example, a car insurance case might include four photos of the vehicle's front left side, rear right side, engine compartment, and license plate area.

[0031] In this embodiment, before extracting the initial feature vectors corresponding to each image sample, a unified preprocessing can be performed on each image sample to eliminate noise and interference factors. The preprocessing includes the following: Size normalization: The size of each image sample is adjusted to 224×224 (to fit the input of the network model), and normalized using the following formula (1): (1) in, and The mean and standard deviation of the ImageNet dataset are used to ensure that the input distribution is consistent with that during training.

[0032] Then, a pre-trained network model (such as MobileNet) is used to extract high-level semantic features from each image sample. Specifically, each image sample is sequentially input into the network model, and the feature vector output by the "global average pooling layer" in the network model is taken as the initial feature vector. The initial feature vector is a high-dimensional vector of 1280 dimensions.

[0033] The initial feature vectors corresponding to the image samples can reflect the high-level semantic information of the images (such as damage patterns, ticket styles, text layouts, etc.) and have good discriminative capabilities. Therefore, a historical image feature library can be constructed based on the initial feature vectors corresponding to all image samples.

[0034] S102. Receive the pending image of the current claim case submitted by the user, and use the network model to extract the current feature vector corresponding to the pending image.

[0035] Optionally, the same preprocessing procedure as the image sample is performed on the image to be reviewed in the current claim case, namely, adjusting the size and standardizing (same as μ, σ), and the processed image to be reviewed is input into the same network model to obtain the current feature vector corresponding to the image to be reviewed.

[0036] S103. Based on the current feature vector corresponding to the image to be reviewed and the historical image feature library, determine whether the image to be reviewed is an abnormal image.

[0037] Abnormal images refer to images submitted repeatedly or that have been forged or altered.

[0038] In one feasible approach, the similarity between the current feature vector of the image to be reviewed and the initial feature vectors of each image sample in the historical image feature library can be used to determine whether the image to be reviewed is abnormal. If at least one image sample has a similarity greater than a similarity threshold with the image to be reviewed, it indicates that the content of the image sample is highly similar to that of the image to be reviewed, thus determining that the image to be reviewed is an abnormal image submitted repeatedly. This achieves automatic, efficient, and accurate determination of whether the image currently submitted by the user has been used in historical claims records. Even if the image has been modified in common ways such as cropping, rotating, color adjusting, compression, and adding watermarks, its association with historical images can still be effectively identified, reducing the risk of erroneous or incorrect claims due to duplicate or falsified evidence, and improving the automation and intelligence level of after-sales claims review.

[0039] S104. If so, generate risk warning information and output the risk warning information to the claims review interface.

[0040] The risk warning information includes risk level and risk type. Risk level includes: low / medium / high, and risk type includes: suspected image reuse, highly suspected forgery, and repeated submission across cases.

[0041] In one feasible approach, when an abnormal image is detected, a risk warning is generated and pushed to the front-end interface of the claims review system in real time. The warning is then displayed to the reviewer in the case details page as a pop-up or highlighted label, reducing the burden of manual review, improving claims processing efficiency and user experience, and providing traceable and verifiable technical support for the after-sales risk control system.

[0042] Therefore, the claims information processing method provided in this application can fundamentally improve the intelligence, accuracy and automation level of the light vehicle after-sales claims system in verifying the authenticity of image evidence, effectively curb fraudulent behavior, protect corporate assets, optimize user experience, and help the automotive industry move towards a new era of high-quality and high-efficiency intelligent services.

[0043] In summary, this application provides a claims information processing method. First, image samples attached to multiple closed historical claims cases are acquired. An initial feature vector corresponding to each image sample is extracted using a pre-trained network model. Based on the initial feature vectors of each image sample, a historical image feature library is constructed. Then, the same network model is used to extract the current feature vector corresponding to the image to be reviewed in the current claims case. Next, based on the current feature vector of the image to be reviewed and the historical image feature library, it is determined whether the image to be reviewed is an abnormal image. This utilizes artificial intelligence to automatically detect whether the image submitted by the user has been used in historical claims records, even if the image has undergone common modifications. This achieves automated document recognition, reducing the risk of erroneous or incorrect claims due to duplicate or falsified evidence, and improving the automation and intelligence level of after-sales claims review. Finally, when an abnormal image is detected, a risk warning is generated and pushed to the front-end interface of the claims review system in real time, reducing the burden of manual review, improving claims processing efficiency and user experience, and providing traceable and verifiable technical support for the after-sales risk control system.

[0044] Optionally, refer to Figure 2 As shown, in step S101 above, the historical image feature library is constructed based on the initial feature vector corresponding to the image sample, including: S201. Perform dimensionality reduction and quantization compression on the initial feature vectors corresponding to the image samples to generate the index sequence corresponding to the image samples.

[0045] S202. Associate the identifier, index sequence and metadata corresponding to the image sample and store them in the historical image feature library.

[0046] The metadata includes at least one of the following: image unique identifier, claim number, submission timestamp, vehicle identification number, or dealer code.

[0047] For example, the identifier corresponding to an image sample can be the name or ID of the image sample, used to accurately locate the image.

[0048] In deep learning-based image retrieval systems, dimensionality reduction and quantization compression of images are key technologies for improving storage efficiency, reducing transmission overhead, and accelerating similarity calculation. The goal is to significantly reduce the feature vector volume while preserving as much discriminative and semantic information as possible from the original features.

[0049] In one feasible approach, to transform each image sample from a high-dimensional, redundant deep learning feature representation into a compact, efficient, and rapidly comparable index form, and then bind it with key business information and store it in a vector database to form a "historical image feature library," this embodiment proposes further dimensionality reduction and quantization compression of the initial feature vectors corresponding to the image samples to generate index sequences corresponding to the image samples. This reduces the dimensionality to improve subsequent computational efficiency and storage performance. Furthermore, the image sample's identifier, index sequence, and metadata are associated; that is, an image sample can be represented as... The data is then written into a vector database to construct a historical image feature library.

[0050] Therefore, in this embodiment, the image samples are converted into feature vectors for dimensionality reduction and quantization encoding to obtain the index sequence corresponding to the image samples. A mapping relationship is established between the unique image identifier, the index sequence, and the metadata. The constructed mapping relationship is reliably stored in the vector database to support subsequent traceable risk audits, which is a prerequisite for achieving high-precision and low-latency retrieval.

[0051] Optionally, refer to Figure 3 As shown, step S201 above includes: S301. Using deep principal component analysis, the initial feature vector corresponding to the image sample is reduced in dimensionality to obtain the dimensionality-reduced features corresponding to the image sample. Then, the dimensionality-reduced features corresponding to the image sample are normalized to obtain the normalized features corresponding to the image sample.

[0052] Optionally, to remove background information while retaining the main foreground information, this embodiment proposes using Deep Principal Component Analysis (Deep-PCA) to project high-dimensional features to a low-dimensional subspace. Specifically, let the projection matrix be... Then the dimensionality-reduced features can be represented as ,in, The mean of the features in the training set.

[0053] Deep principal component analysis (DPMA) preserves key information by maximizing the projected variance, and its optimal W is determined by the covariance matrix. The former It consists of principal components.

[0054] To further improve retrieval performance, L2 normalization was performed on the dimensionality-reduced features to obtain the normalized features corresponding to the image samples. This allows the inner product to be used subsequently, which is equivalent to the cosine similarity, simplifying the calculation.

[0055] S302, Sample product quantization, quantizes the normalized features corresponding to the image samples to obtain the index sequence corresponding to the image samples.

[0056] Optionally, to achieve a higher compression ratio, the normalized features (i.e., floating-point vectors) corresponding to the image samples can be converted into a low-bit representation. Therefore, in this embodiment, it is proposed that the normalized features corresponding to the image samples can be quantized using Product Quantization (PQ) to obtain the index sequence corresponding to the image samples. Specifically, the normalized features are then... Divide into M subvectors ,...... Each subvector is then independently clustered using k-means, generating K cluster centers. These cluster centers constitute the codebook for that subvector. Let the codebook for the m-th subvector be... For each subvector, find its nearest cluster center, and use the index of that cluster center to represent the subvector. Concatenate the indices of all subvectors to obtain the index sequence of the original vector. Therefore, the normalized features corresponding to the image samples are approximated by the nearest codeword as shown in formula (2): (2) The entire vector is represented by an index sequence. This means that each index only needs The total storage is reduced to M×8 bits, reducing storage overhead and computational complexity.

[0057] The product quantization reconstruction error can be expressed as shown in the following formula (3): (3) Optionally, the floating-point low-dimensional features of a large number of image samples can be further compressed into a compact integer index sequence, which greatly saves storage space while maintaining sufficient recognition accuracy and supports fast approximate retrieval of massive images.

[0058] Optionally, refer to Figure 4 As shown, step S103 above includes: S401. Based on the index sequence corresponding to the image to be reviewed and the index sequence corresponding to each image sample in the historical image feature library, at least one candidate image is obtained.

[0059] Optionally, the image to be reviewed is subjected to the same dimensionality reduction and quantization compression processes as the image samples to generate an index sequence corresponding to the image to be reviewed.

[0060] S402. Determine the similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed, and identify whether the image to be reviewed is an abnormal image based on the metadata of the candidate image with the highest similarity.

[0061] In one feasible approach, the index sequence corresponding to the image to be examined can be compared with the index sequence corresponding to each image sample in the historical image feature library. The top-K most likely matching image samples with high index sequence overlap are used as candidate images. That is, this application does not compare the image to be examined with all image samples one by one, but quickly finds a batch of potentially related images through the "index sequence" - that is, "candidate images", realizing large-scale retrieval at the millisecond level. Then, after returning the top-K candidate images, a reordering is performed, and the original, unquantized high-dimensional feature vector is used to calculate the exact cosine similarity between the image to be examined and each candidate image. The specific similarity between the two is as shown in the following formula (4): (4) in, This is the current feature vector corresponding to the image to be reviewed. Let i be the initial feature vector corresponding to the i-th image sample. Let be the angle between two vectors, and its range be [ 1,1], A larger value indicates greater similarity; equivalently, Euclidean distance can also be used.

[0062] The images are then reordered according to their similarity scores, and the candidate images with the highest similarity are selected. Based on the metadata of the candidate images with the highest similarity, it is determined whether the images to be reviewed are abnormal images.

[0063] Therefore, in this embodiment, a batch of suspected duplicate images (candidate images) are quickly identified in a large number of image samples by using the "index sequence" corresponding to the image to be reviewed. Then, the accurate similarity is calculated using the original high-dimensional features, and combined with the metadata such as the claim number, vehicle identification number, and submission time of the candidate image with the highest similarity, it is determined whether the image to be reviewed is suspected of being reused or tampered with across cases. If such behavior exists, it is identified as an "abnormal image" and a risk warning is issued.

[0064] Optionally, compared to existing technical means of relying on manual review or simple image comparison in automobile after-sales claims, the solution provided in this application does not require human intervention in the entire comparison process. The system can automatically complete feature extraction, historical database retrieval, similarity calculation and risk level determination, and generate structured review suggestions, which greatly reduces the reliance on the experience of reviewers, improves the level of process standardization, has a high degree of automation, and reduces human intervention.

[0065] Optionally, refer to Figure 5As shown, step S401 above includes: S501. Using the approximate nearest neighbor search method, search for at least one candidate cluster that is closest to the index sequence corresponding to the image to be reviewed from the index sequence corresponding to each image sample.

[0066] S502. Take the image samples corresponding to each candidate cluster as candidate images to obtain at least one candidate image.

[0067] Optionally, considering that in large-scale scenarios (such as when the number of images in the database reaches N=10^7), using cosine similarity to measure the similarity between vectors would be too costly to perform N precise inner product or distance calculations for each query. Therefore, in this embodiment, to accelerate retrieval, an inverted index (IVF) combined with product quantization (PQ) is used. First, coarse clustering is used to narrow down the candidate range, and then PQ retrieval is performed within the candidate set, significantly reducing the computational load. Specifically, the vector database uses an IVF structure, first clustering the normalized features corresponding to all image samples to form k cluster centers. When using the approximate nearest neighbor search method, only the index sequence closest to the image to be examined is searched. There are 1 candidate cluster; at the same time, within the candidate cluster, the d-dimensional vector is segmented and encoded into a compact codebook index through product quantization, so that the distance can be quickly estimated by looking up the table, as shown in the following formula (5): (5) Where M is the number of subspaces. It is the codebook center of the m-th subvector.

[0068] After the approximate nearest neighbor search method returns the TOP-K (e.g., K=100) candidate images, it is often reordered and the exact cosine similarity is recalculated using the original high-dimensional features. Specifically, it is shown in the following formula (6): (6) and according to Arrange in descending order to improve the final recall accuracy.

[0069] In addition, some advanced methods introduce learnable similarity functions, such as optimizing the metric space through Siamese networks or contrastive learning, so that the similarity is defined as shown in the following formula (7): (7) in, This is the matching function parameterized for the neural network.

[0070] Optionally, this application, by leveraging an efficient near nearest neighbor retrieval algorithm and an optimized feature index structure, can complete the comparison of millions or even tens of millions of historical image databases within milliseconds, support high-concurrency claims requests, significantly shorten user waiting time, and improve overall service response speed.

[0071] Optionally, the similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be examined is determined, including: Cosine similarity is used to calculate the cosine similarity between the initial feature vector corresponding to the candidate image and the current feature vector corresponding to the image to be reviewed.

[0072] Optionally, in image retrieval, similarity calculation is a core step in measuring the semantic or visual relevance between the query image and candidate images in the database. The entire process relies on high-quality feature vector representations and is quantified through specific mathematical metrics. In this embodiment, firstly, let the current feature vector of the image to be reviewed, extracted by a pre-trained network model (such as MobileNet), be: The initial feature vector corresponding to the i-th image sample in the historical image database is: .

[0073] Where d is the feature dimension, and to ensure comparability, L2 normalization is usually performed on all features to obtain the normalized features corresponding to the image to be examined. ; and, the normalized features corresponding to the image samples are .

[0074] After normalization, all vectors lie on the unit hypersphere and have a constant magnitude of 1.

[0075] Based on this, cosine similarity is used to measure the similarity between vectors. It is defined as shown in the above formula (4). The cosine similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed is calculated. The higher the cosine similarity, the greater the similarity between the two images.

[0076] Optionally, based on the metadata of the candidate image with the highest similarity, it is determined whether the image to be reviewed is an anomalous image, including: If the similarity of the candidate image with the highest similarity is greater than a preset threshold, and the metadata of the candidate image indicates that the candidate image belongs to different claims cases, then the image to be reviewed is determined to be an abnormal image.

[0077] In one feasible approach, a preset threshold of 95% is used. If the similarity of the candidate image with the highest similarity is greater than 95%, it indicates that the content of the two images is highly similar. Further examination of the candidate image's metadata is required to determine whether the candidate image and the image under review belong to different claim cases. If the candidate image's metadata indicates that the candidate image belongs to different claim cases, it means that the image under review may be a duplicate submission (the same image used for multiple claims) or a slightly modified and reused image (such as cropping, color adjustment, adding watermarks, etc.). Therefore, the image under review can be determined to be an "abnormal image," triggering a risk warning. The risk warning information is then output to the claim review interface to prompt the reviewers to pay close attention to the current claim case.

[0078] Optionally, this application adopts a modular design and can be seamlessly embedded as an independent service or API into existing after-sales management systems, insurance claims platforms, or DMS (dealer management systems), allowing for flexible deployment. At the same time, its core technology can be quickly adapted to other image verification scenarios such as used car evaluation, repair record verification, and warranty audit, possessing strong versatility and business extensibility.

[0079] Optionally, this application can also combine the closed-loop review feedback to automatically label and retrain the model for misjudged or missed cases, so as to achieve dynamic optimization and iterative upgrading of recognition capabilities and ensure high accuracy and adaptability in long-term operation.

[0080] Optionally, this application proposes an artificial intelligence-based claims information processing method. Addressing the problem of repeated submissions or tampering of historical image evidence by users in existing after-sales claims processes, which makes manual review difficult, this application constructs a historical claims image feature library, utilizes a deep learning model to extract robust perceptual semantic feature vectors from the images, and performs feature encoding and similarity comparison on newly submitted images for review. Even if the images have undergone common modifications such as cropping, rotation, color adjustment, compression, and watermarking, the system can still accurately determine whether they are related to historical evidence. This method supports large-scale image retrieval at the millisecond level, can automatically mark high-risk repeated or suspected tampered images, and integrates the recognition results into the claims review process to assist human decision-making. This invention effectively improves the accuracy, efficiency, and risk control capabilities of claims review, reduces the risk of fraudulent claims, and simultaneously considers data privacy and system scalability. It is suitable for intelligent after-sales and damage assessment scenarios for car manufacturers, dealers, and insurance institutions.

[0081] Optionally, this application proposes an artificial intelligence-based claims information processing method, which produces beneficial effects including: 1. Significantly improves claims review efficiency and reduces labor costs. The system can intelligently search and risk-scorize the entire historical evidence database within seconds of a user submitting images, and automatically mark suspicious cases. This allows reviewers to focus on high-risk work orders, reducing the workload of routine image review by more than 80%, and improving overall claims processing time by 30%–50%. 2. Enhance corporate risk control capabilities and effectively curb fraudulent activities. By establishing a traceable record of image evidence usage and an intelligent early warning mechanism, this invention significantly reduces the risk of erroneous and incorrect claims due to duplicate claims and falsified evidence. Actual pilot data shows that after the system was implemented, the incidence of related types of fraud cases decreased by more than 60%, preventing economic losses of several million yuan annually.

[0082] 3. It possesses excellent versatility and business scalability. This technical framework is not only applicable to after-sales claims for light vehicles, but can also be seamlessly migrated to multiple related business scenarios such as insurance damage assessment, used car inspection, repair work order verification, and warranty claims, providing core capability support for building a unified intelligent automotive service risk control platform.

[0083] In summary, this application not only solves the technical pain point of difficulty in verifying the authenticity of image evidence in current automobile after-sales claims, but also achieves a systematic breakthrough in terms of accuracy, efficiency, security and scalability, and has outstanding practicality, advanced nature and industrial application prospects.

[0084] The following describes the claims information processing device and storage medium used to execute the claims information provided in this application. The specific implementation process and technical effects are described above and will not be repeated below.

[0085] Optionally, refer to Figure 6 The diagram shown is a structural schematic of a claims information processing device provided in this application. The device includes: The acquisition module 701 is used to acquire image samples corresponding to multiple historical claims cases; Extraction module 702 is used to extract the initial feature vector corresponding to each image sample using a pre-trained network model; Construction module 703 is used to construct a historical image feature library based on the initial feature vectors corresponding to each of the image samples; The receiving module 704 is used to receive the pending images of the current claim case submitted by the user; The extraction module 702 is further configured to use the network model to extract the current feature vector corresponding to the image to be reviewed; The determination module 705 is used to determine whether the image to be reviewed is an abnormal image based on the current feature vector corresponding to the image to be reviewed and the historical image feature library; The output module 706 is used to generate risk warning information and output the risk warning information to the claims review interface if the condition is met.

[0086] Optionally, the construction module 703 is specifically used for: The initial feature vectors corresponding to the image samples are subjected to dimensionality reduction and quantization compression to generate an index sequence corresponding to the image samples. The identifier, index sequence, and metadata corresponding to the image sample are associated and stored in the historical image feature library.

[0087] Optionally, the construction module 703 is specifically used for: The initial feature vector corresponding to the image sample is reduced in dimensionality using deep principal component analysis to obtain the dimensionality-reduced features of the image sample. The dimensionality-reduced features of the image sample are then normalized to obtain the normalized features of the image sample. Sampling product quantization involves quantizing the normalized features corresponding to the image samples to obtain the index sequence corresponding to the image samples.

[0088] Optionally, the determining module 705 is specifically used for: Based on the index sequence corresponding to the image to be reviewed and the index sequence corresponding to each image sample in the historical image feature library, at least one candidate image is obtained; The similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed is determined, and the image to be reviewed is identified as an abnormal image based on the metadata of the candidate image with the highest similarity.

[0089] Optionally, the determining module 705 is specifically used for: The approximate nearest neighbor search method is used to search for at least one candidate cluster that is closest to the index sequence corresponding to the image to be examined from the index sequence corresponding to each of the image samples; At least one candidate image is obtained by taking the image samples corresponding to each candidate cluster as candidate images.

[0090] Optionally, the determining module 705 is specifically used for: The cosine similarity is used to calculate the cosine similarity between the initial feature vector corresponding to the candidate image and the current feature vector corresponding to the image to be reviewed.

[0091] Optionally, the determining module 705 is specifically used for: If the similarity of the candidate image with the highest similarity is greater than a preset threshold, and the metadata of the candidate image indicates that the candidate image belongs to different claims cases, then the image to be reviewed is determined to be an abnormal image.

[0092] The above-described device is used to execute the method provided in the foregoing embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

[0093] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0094] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device can be integrated into a terminal device or a chip of a terminal device. The electronic device can be a computing device with data processing function, used to execute the claims information processing method provided in the above embodiment.

[0095] The electronic device includes: processor 801 and memory 802.

[0096] The memory 802 is used to store programs, and the processor 801 calls the programs stored in the memory 802 to execute the above method embodiments. The specific implementation and technical effects are similar, and will not be described again here.

[0097] Optionally, this application also provides a program product, such as a computer-readable storage medium, including a program that, when executed by a processor, performs the above-described method embodiments.

[0098] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0099] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0100] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0102] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A claim settlement information processing method characterized by comprising: The method includes: Image samples corresponding to multiple historical claims cases are obtained, and an initial feature vector corresponding to each image sample is extracted using a pre-trained network model. Based on the initial feature vector corresponding to each image sample, a historical image feature library is constructed. Receive the pending image of the current claim case submitted by the user, and use the network model to extract the current feature vector corresponding to the pending image; Based on the current feature vector corresponding to the image to be reviewed and the historical image feature library, it is determined whether the image to be reviewed is an abnormal image; If so, a risk warning message will be generated and output to the claims review interface.

2. The method according to claim 1, characterized in that, The construction of a historical image feature library based on the initial feature vectors corresponding to each of the image samples includes: The initial feature vectors corresponding to the image samples are subjected to dimensionality reduction and quantization compression to generate an index sequence corresponding to the image samples. The identifier, index sequence, and metadata corresponding to the image sample are associated and stored in the historical image feature library.

3. The method according to claim 2, characterized in that, The step of performing dimensionality reduction and quantization compression on the initial feature vector corresponding to the image sample to generate the index sequence corresponding to the image sample includes: The initial feature vector corresponding to the image sample is reduced in dimensionality using deep principal component analysis to obtain the dimensionality-reduced features of the image sample. The dimensionality-reduced features of the image sample are then normalized to obtain the normalized features of the image sample. Sampling product quantization involves quantizing the normalized features corresponding to the image samples to obtain the index sequence corresponding to the image samples.

4. The method according to claim 1, characterized in that, The step of determining whether the image to be reviewed is an abnormal image based on the current feature vector corresponding to the image to be reviewed and the historical image feature library includes: Based on the index sequence corresponding to the image to be reviewed and the index sequence corresponding to each image sample in the historical image feature library, at least one candidate image is obtained; The similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be reviewed is determined, and the image to be reviewed is identified as an abnormal image based on the metadata of the candidate image with the highest similarity.

5. The method according to claim 4, characterized in that, The step of obtaining at least one candidate image based on the index sequence corresponding to the image to be examined and the index sequence corresponding to each image sample in the historical image feature library includes: The approximate nearest neighbor search method is used to search for at least one candidate cluster that is closest to the index sequence corresponding to the image to be examined from the index sequence corresponding to each of the image samples; At least one candidate image is obtained by taking the image samples corresponding to each candidate cluster as candidate images.

6. The method according to claim 4, characterized in that, Determining the similarity between the initial feature vector corresponding to each candidate image and the current feature vector corresponding to the image to be examined includes: The cosine similarity is used to calculate the cosine similarity between the initial feature vector corresponding to the candidate image and the current feature vector corresponding to the image to be reviewed.

7. The method according to claim 4, characterized in that, The step of identifying whether the image to be examined is an abnormal image based on the metadata of the candidate image with the highest similarity includes: If the similarity of the candidate image with the highest similarity is greater than a preset threshold, and the metadata of the candidate image indicates that the candidate image belongs to different claims cases, then the image to be reviewed is determined to be an abnormal image.

8. A claims information processing device, characterized in that, The device includes: The acquisition module is used to acquire image samples corresponding to multiple historical claims cases; The extraction module is used to extract the initial feature vectors corresponding to each image sample using a pre-trained network model. The construction module is used to construct a historical image feature library based on the initial feature vectors corresponding to each of the image samples; The receiving module is used to receive pending images of the current claim submitted by the user; The extraction module is further configured to use the network model to extract the current feature vector corresponding to the image to be reviewed; The determination module is used to determine whether the image to be reviewed is an abnormal image based on the current feature vector corresponding to the image to be reviewed and the historical image feature library; The output module is used to generate risk warning information and output the risk warning information to the claims review interface if the condition is met.

9. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is executed by a processor to perform the method as described in any one of claims 1-7.

Images