Automated loss assessment system in collaboration with computer vision and anti-fraud

Abstract

The application provides an automatic loss assessment system for computer vision and anti-fraud cooperation, relates to the technical field of insurance supervision and management, and comprises a data acquisition module, a reliability score module, a value score calculation module and a threshold adjustment and decision module. The application automatically adjusts the strictness of the review according to the size of the loss amount, and the greater the amount, the more stringent the review. This dynamic adjustment can not only ensure the quick processing of small cases, but also strengthen the risk control of high-value cases, achieving the intelligent review effect of different people and different amounts in the conventional scheme, thereby improving the efficiency and effectively identifying fraud.

Description

Technical Field

[0001] This invention relates to the field of insurance supervision and management technology, specifically to an automated loss assessment system that combines computer vision with anti-fraud measures. Background Technology

[0002] In automated vehicle accident damage assessment systems, determining the presence of fraud is crucial. Its core significance lies in overcoming the fundamental limitation of traditional technologies that focus solely on the extent of vehicle damage while neglecting the authenticity of materials. Without fraud detection, even if the system can accurately identify damaged areas and repair costs, it cannot distinguish whether these images themselves have been tampered with, staged, or misused. This could lead to fraudulent claims being approved directly, resulting in the loss of insurance funds. By introducing fraud detection, only genuine and credible cases are processed quickly, thereby improving processing efficiency while effectively intercepting fraud risks and ensuring the fairness and reliability of damage assessment results.

[0003] In existing technology, CN114462553A discloses an image annotation and feature extraction method and system for auto insurance anti-fraud. This technology includes: an auto insurance feature table construction module, an image acquisition module, an annotation function module, and a feature extraction module. The annotation function module includes a multi-label category annotation module, a vehicle damage location annotation module, and a face annotation module. The feature extraction module is used to extract features from each annotated dataset. The aforementioned technical solution mainly focuses on establishing image feature annotation and extraction for auto insurance anti-fraud, making the extracted image features more objective, generating structured auto insurance data that can be used for cross-validation, and improving data quality.

[0004] However, the aforementioned existing technologies rely solely on image recognition to determine the extent of vehicle damage, which easily overlooks the authenticity of the photos themselves. For example, whether the photos are clear, whether they have been tampered with, whether the time and location of the shooting are reasonable, and whether the user has any abnormal uploading behavior. This makes it difficult for conventional solutions to identify cases of insurance fraud through forged images or deliberately staged photos. At the same time, conventional solutions mostly adopt fixed review standards, treating all cases the same regardless of the amount of damage assessed. This can lead to overly strict review of small cases, affecting efficiency, while large cases may be under-reviewed, making it difficult to achieve a flexible balance between efficiency and risk control.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an automated loss assessment system that integrates computer vision and anti-fraud measures to address the problems mentioned in the background section. This invention automatically adjusts the rigor of the review process based on the amount of the loss; the larger the amount, the stricter the review. This dynamic adjustment ensures rapid processing of small cases while strengthening risk control for high-value cases. It achieves intelligent review effects that are difficult to achieve with conventional methods, tailored to different individuals and different amounts, thereby improving efficiency and more effectively identifying fraud.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An automated damage assessment system that combines computer vision and anti-fraud measures includes:

[0009] Data acquisition module: Acquires vehicle accident video materials and simultaneously acquires reliability-related feature data, including image quality features, content consistency features, metadata features, and user behavior features;

[0010] Reliability scoring module: Processes the above-collected data sets to establish a reliability scoring conversion model. The model is designed based on a supervised learning framework for machine learning and is trained using a neural network algorithm. Reliability-related feature data is input into the reliability scoring conversion model, and reliability scoring parameters are directly output. The reliability scoring parameters include image quality score, content consistency score, metadata score, and user behavior score.

[0011] Value scoring calculation module: Constructs a weighted aggregation formula, based on the transformed reliability scoring parameters input into the weighted aggregation formula, which uses a linear combination to calculate and obtain the value score;

[0012] Threshold Adjustment and Decision Module: Sets a reliability judgment threshold and a loss assessment correction coefficient. The reliability judgment threshold is an initial benchmark value used to reflect the system's tolerance for fraud. The reliability judgment threshold is adjusted using the loss assessment correction coefficient to obtain a reliability correction threshold. Then, the value score obtained by the value scoring calculation module is compared with the reliability correction threshold, and a loss assessment result judgment report is output to generate a decision scheme on whether to conduct manual review.

[0013] Furthermore, the image quality features are used to assess the image's sharpness, contrast, and integrity, ensuring that accident details are clearly identifiable; the content consistency features are used to detect signs of tampering or forgery; the metadata features are additional information about the image, including shooting time, geographical location, and file attributes, and are used to verify the authenticity and temporal consistency of the accident; user behavior features analyze upload behavior patterns, including upload frequency, operating habits, and interaction history, and are used to identify abnormal activity or fraudulent intent.

[0014] Furthermore, in the reliability scoring module, the reliability scoring conversion model includes an input layer, a hidden layer, and an output layer. During the training phase, the reliability scoring conversion model is trained by learning the complex nonlinear relationship between reliability-related feature data from a large number of historical cases and the labels of cases that are ultimately proven to be real or fraudulent.

[0015] After training, for a new accident case, the structured feature data of the case is input into the model. The neural network performs forward propagation calculations through its multi-layer network structure, and finally generates four key scoring parameters directly in the output layer: image quality score, content consistency score, metadata score, and user behavior score.

[0016] Furthermore, the functional mapping relationship of the reliability scoring transformation model is as follows:

[0017]

[0018] Where X is the input feature vector: ,in, arrive This represents the various reliability-related characteristic data after initial alignment; The nonlinear function learned by the reliability scoring transformation model. These are the output image quality score, content consistency score, metadata score, and user behavior score, respectively.

[0019] Then, feature normalization processing is performed on the output image quality score, content consistency score, metadata score, and user behavior score. First, all four scores calculated in the current case are traversed, and the maximum and minimum values ​​of each score in the historical sample database or the current batch of data are calculated. Then, the min-max normalization method is used to perform a linear transformation on each original score: after subtracting the minimum value of the corresponding reliability score parameter from each score, the original score is precisely compressed to the closed interval [0, 1] by dividing by the range.

[0020] Furthermore, outlier filtering is performed on the reliability score parameters obtained from the conversion:

[0021] If any reliability score parameter obtained from the conversion meets the following conditions, it is considered an outlier:

[0022]

[0023] in:

[0024] S represents a single reliability score parameter value;

[0025] μ is the average value of the reliability score dataset samples in the massive historical cases, and the reliability score dataset includes the image quality score dataset, content consistency score dataset, metadata score dataset and user behavior score dataset.

[0026] k is a constant threshold coefficient used to define the boundary range of outliers;

[0027] σ is the sample standard deviation of the dataset corresponding to the reliability score parameter value, used to quantify the dispersion of the data;

[0028] And the formula for obtaining the sample mean μ is: Where m is the total number of samples in the reliability scoring dataset, This represents the value of a single data point with index i in the reliability score dataset.

[0029] Furthermore, the weighted aggregation formula in the value scoring calculation module is as follows:

[0030]

[0031] in:

[0032] V represents the value score, used to comprehensively evaluate the reliability of vehicle accident video footage. A higher value score indicates a lower probability of fraud in the footage, and a higher image quality score. Quality dimensions reflecting image sharpness and noise level; content consistency score. Metadata scoring reflects the logical consistency between the location and shape of vehicle damage in the images and the accident description. Reflects the credibility of metadata based on shooting time, location, and equipment information; user behavior rating. It reflects the degree of abnormality in user upload behavior characteristics;

[0033] These are the weighting coefficients for image quality score, content consistency score, metadata score, and user behavior score, respectively. The relative importance of each weighting coefficient is determined based on the significance of the corresponding reliability scoring parameters and is further established according to expert experience, while also satisfying the following conditions: .

[0034] Furthermore, in the threshold adjustment and decision module, the reliability correction threshold is calculated using the following formula:

[0035]

[0036] in:

[0037] This is the reliability correction threshold, which is the threshold used for final comparison after adjustment based on the loss assessment amount.

[0038] The reliability judgment threshold is a preset initial baseline value of the system, which is used to reflect the basic tolerance for fraudulent behavior. The reliability judgment threshold is set as a constant based on historical data or business experience.

[0039] The loss assessment correction coefficient is a dynamic adjustment factor that is positively correlated with the loss assessment amount. Its value increases as the loss assessment amount increases, reflecting the risk control principle that the larger the amount, the stricter the review.

[0040] Furthermore, the selection methods for the damage assessment correction coefficient include the linear proportional method and the piecewise interval method, wherein the process of the linear proportional method is as follows:

[0041] First, the system obtains the raw data of the estimated repair cost from the vehicle accident damage assessment process. Then, the estimated repair cost is substituted into a preset linear calculation formula. The core of this formula is a fixed proportional coefficient obtained through training with historical data. The system performs a simple multiplication operation between this coefficient and the damage assessment cost, thereby directly generating a value λ that is proportional to the amount. The entire process is based entirely on numerical calculation, which allows λ to increase linearly as the input amount increases.

[0042] Furthermore, the segmented interval method is used to select the loss correction coefficient. The process is as follows:

[0043] The frequency of fraud cases and the distribution of involved amounts within different monetary ranges were statistically analyzed. Combined with the pass rate of legitimate cases in each range, segmentation points for monetary amounts were defined. Cases within each range exhibited internal consistency in risk characteristics, but also showed variations across the ranges. After segmenting the ranges, a fixed risk level was assigned to each monetary range based on its historical fraud risk level and the operational tolerance requirements. The value or a preset benchmark adjustment coefficient forms a set of adjustments based on the amount. The larger the value, the better the step-like mapping relationship, so that the adjustment of the reliability judgment threshold can fit the actual risk situation of different monetary levels.

[0044] Furthermore, in the threshold adjustment and decision-making module, the process of comparing the value score obtained by the value score calculation module with the reliability correction threshold includes:

[0045] like This indicates that the overall reliability score of the materials in the vehicle accident case has reached the standard set by the system, and the probability of fraud is low. The system allows the damage assessment result to be automatically approved without human intervention.

[0046] like If the overall reliability score of the materials in the current case does not meet the system's set standard, there is a potential risk of fraud. The system will prevent the loss assessment result from being automatically approved and generate a manual review instruction.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This invention comprehensively judges the authenticity of photos from multiple perspectives, including photo clarity, presence of retouching, correspondence between shooting time and location, and user upload habits, making the judgment more comprehensive. Furthermore, this invention automatically adjusts the rigor of the review based on the amount of damages, with stricter scrutiny for larger amounts. This dynamic adjustment ensures rapid processing of small cases while strengthening risk control for high-value cases, achieving a personalized and amount-based intelligent review effect that is difficult to achieve in conventional solutions. This improves efficiency while more effectively identifying fraud. Attached Figure Description

[0049] Figure 1 This is a block diagram of the automated damage assessment system that combines computer vision and anti-fraud technology according to the present invention.

[0050] Figure 2 This is a flowchart illustrating the operation of the automated damage assessment system that combines computer vision and anti-fraud measures, as described in this invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0052] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0053] Example:

[0054] Please see Figures 1-2 The present invention provides the following technical solutions:

[0055] An automated damage assessment system that combines computer vision and anti-fraud measures includes:

[0056] Data acquisition module: Acquires vehicle accident video materials and simultaneously acquires reliability-related feature data, including image quality features, content consistency features, metadata features, and user behavior features;

[0057] The image quality features are used to evaluate the sharpness, contrast and integrity of the image, including the image brightness, noise level, resolution, color reproduction and whether there is blur or occlusion, to ensure that accident details such as the location and extent of vehicle damage and the surrounding environment can be clearly identified.

[0058] The content consistency feature is used to detect possible signs of tampering or forgery. By analyzing the logical consistency between the vehicle damage parts and shapes in the image and the accident description, as well as identifying operation traces such as splicing, copying and moving, blurring, adding or deleting elements, it is determined whether the image content truly reflects the original appearance of the accident.

[0059] Metadata features are additional information for images, including shooting time, geographical location (such as GPS coordinates), and file attributes (such as file size, format, modification time, device model, camera parameters, original creation time, etc.), which are used to verify the authenticity of the accident and the consistency of the time sequence. For example, it checks whether the shooting time matches the reporting time, whether the geographical location matches the accident scene, and whether there are any anomalies in the file modification records.

[0060] User behavior feature analysis analyzes upload patterns, including upload frequency, operational habits (such as upload time, device type, and operation path), interaction history (such as historical police reports, multiple claims, and user credit rating), and potentially associated IP addresses, device fingerprints, network environment, etc., to identify abnormal activity or fraudulent intent, such as high-frequency uploads within the same time period, use of virtual devices, or proxy IPs. In internet consumer finance or factoring service scenarios, the system can further identify cross-platform fraud risks by analyzing the correlation between a user's historical loan records and current police reports.

[0061] The reliability scoring module performs preprocessing operations such as cleaning, normalization, missing value imputation, and feature alignment on the collected multi-dimensional feature data, including image quality, content consistency, metadata, and user behavior, to ensure that the data format is consistent and meets the model input requirements. Based on a machine learning supervised learning framework, the reliability scoring conversion model employs a deep neural network architecture with multiple hidden layers. Through forward and backward propagation mechanisms, it utilizes samples labeled with real or fraudulent tags from a massive amount of historical cases for end-to-end training to minimize the error between the predicted score and the real label, thereby learning the complex nonlinear relationship between features and case authenticity. After training, the preprocessed feature data is input into the model, and after multi-layer network computation, it directly outputs reliability scoring parameters in four dimensions: image quality score, content consistency score, metadata score, and user behavior score. Each scoring parameter is a continuous value between 0 and 1, quantifying the credibility of the image material from different perspectives.

[0062] The reliability scoring conversion model employs a typical multi-layer neural network structure, consisting of an input layer, several hidden layers, and an output layer. During the training phase, the model is based on a massive amount of historical cases, using reliability-related feature data extracted from these cases as input. Simultaneously, it uses the authenticity labels of known cases that have ultimately been proven genuine or fraudulent as supervisory signals. Through multiple neurons and non-linear activation functions in the hidden layers, the model automatically learns the deep interaction relationships and complex patterns between features. Furthermore, it continuously optimizes the network weights using a backpropagation algorithm, gradually approximating the complex non-linear mapping relationship between feature data and authenticity labels. This enables the model to accurately assess the multi-dimensional reliability of new cases.

[0063] After training, for a new accident case, the same feature extraction and preprocessing process as in the training phase is followed to organize the reliability-related feature data of the case into a structured feature vector, which is then input into the trained neural network model. The model performs forward propagation computation through its network structure containing multiple hidden layers. Each hidden layer performs a linear transformation on the input data and applies a non-linear activation function, extracting and combining higher-level abstract features layer by layer. Finally, in the output layer, the model uses an appropriate activation function to map the computation results into continuous values ​​between 0 and 1, directly generating reliability scoring parameters for four key dimensions: image quality score, content consistency score, metadata score, and user behavior score. These scores quantify the reliability level of the case materials from multiple perspectives, including image clarity, content authenticity, metadata credibility, and user behavior compliance.

[0064] The functional mapping relationship of the reliability scoring conversion model is as follows:

[0065]

[0066] Where X is the input feature vector: ,in, arrive This represents the various reliability-related characteristic data after initial alignment; The nonlinear function learned by the reliability scoring transformation model. These are the output image quality score, content consistency score, metadata score, and user behavior score, respectively.

[0067] Feature normalization is performed on the output image quality score, content consistency score, metadata score, and user behavior score. This step aims to eliminate the differences in dimensions and orders of magnitude that may exist between the four scoring dimensions due to different original calculation methods, and to ensure that each score is comparable and additive in the subsequent comprehensive analysis.

[0068] Specifically, the process first iterates through all four calculated scores in the current case, and then calculates the maximum and minimum values ​​for each score in the historical sample database or the current batch of data, thus obtaining the value range for each dimension. Next, a min-max standardization method is used to linearly transform each original score: subtracting the minimum value of the corresponding reliability score parameter from each score, and then dividing by the range, precisely compresses the original scores to a closed interval of 0 to 1, while preserving the relative order and proportional relationships between the original data. If all scores for a certain dimension are equal (i.e., the maximum value equals the minimum value), then all scores for the corresponding reliability score parameter are uniformly mapped to 0.5 or kept unchanged to ensure the stability and feasibility of the normalization process.

[0069] Outlier filtering is performed on the reliability score parameters obtained from the conversion:

[0070] If any reliability score parameter obtained from the conversion meets the following conditions, it is considered an outlier:

[0071]

[0072] in:

[0073] x represents a single reliability score parameter value, signifying a single reliability score parameter value for a specific dimension (such as image quality, content consistency, etc.) in the current case. It is the object being examined, and its value is compared to the historical average level. absolute difference This reflects the degree to which the current score deviates from the normal typical value;

[0074] μ is the average value of the reliability score dataset samples in the massive historical cases, and the reliability score dataset includes the image quality score dataset, content consistency score dataset, metadata score dataset and user behavior score dataset.

[0075] k is a constant threshold coefficient used to define the boundary range of outliers, determined by the system based on business experience, risk control requirements, or 3 The principle is predetermined. It defines the boundary multiple for outlier detection, that is, the maximum allowable deviation is k times the standard deviation;

[0076] σ is the sample standard deviation of the dataset corresponding to the reliability score parameter value, used to quantify the dispersion of the data;

[0077] And the formula for obtaining the sample mean μ is: Where m is the total number of samples in the reliability scoring dataset, This represents the value of a single data point with index i in the reliability score dataset.

[0078] Value scoring calculation module: Constructs a weighted aggregation formula, based on the transformed reliability scoring parameters input into the weighted aggregation formula, which uses a linear combination to calculate and obtain the value score;

[0079] The weighted aggregation formula in the value scoring calculation module is:

[0080]

[0081] in:

[0082] V represents the value score, used to comprehensively evaluate the reliability of vehicle accident video footage. A higher value indicates a lower probability of fraud and improved image quality. The quality dimension reflects the image's sharpness and noise level. The higher the score, the better the image quality and the easier it is to identify accident details, thus providing a reliable visual basis for subsequent analysis.

[0083] Content consistency score This score reflects the logical consistency between the location and shape of vehicle damage in the images and the accident description. A higher score indicates greater consistency between the image content and the accident statement, less obvious signs of tampering, and a higher level of authenticity. Therefore, Value rating There is a positive correlation, meaning that the better the image quality, the higher the overall reliability.

[0084] Metadata scoring This score reflects the credibility of metadata based on shooting time, location, and equipment information. A higher score indicates a greater consistency between the metadata and the actual circumstances of the incident (e.g., whether the time matches, the location matches, and whether the equipment was commonly used), and there are no signs of abnormal modification. Value rating There is a positive correlation, meaning that the more reliable the metadata, the higher the overall reliability.

[0085] User behavior rating This score reflects the degree of abnormality in a user's upload behavior. A higher score indicates that the user's behavior is more consistent with normal patterns, with no abnormal activities such as frequent uploads, use of proxy IPs, or switching between multiple devices, and the less obvious the fraudulent intent. Therefore, Value rating There is a positive correlation, meaning that the more standardized the user behavior, the higher the overall reliability.

[0086] These are the weighting coefficients for image quality score, content consistency score, metadata score, and user behavior score, respectively. The relative importance of each weighting coefficient is determined based on the significance of the corresponding reliability scoring parameters and is further established according to expert experience.

[0087] User behavior characteristics directly reflect abnormal traces of operators during the material uploading process. Behavioral characteristics are the most sensitive fraud indicators, and assigning them the highest weight can quickly capture high-risk cases.

[0088] Metadata (such as shooting time, GPS location, device model, and file modification history) provides a digital fingerprint of the image. Metadata has a high threshold for tampering and is highly objective, capable of independently verifying the authenticity of the image and its consistency with the scene; therefore, its weight should be second only to user behavior characteristics.

[0089] Content consistency features focus on the internal logical consistency of an image, relying on the accuracy of the image recognition model for judgment, and are susceptible to misjudgments due to factors such as shooting angle and lighting. Image quality scoring, on the other hand, primarily measures basic quality dimensions such as image sharpness, blurriness, and exposure. While low-quality images may affect manual review, quality itself does not directly equate to fraud; therefore, content consistency features and image quality features have the lowest weight. Thus, it is necessary to meet certain requirements. In actual deployment, the weighting coefficients can also be set with reference to the risk preferences of different financial businesses. For example, for investor identity verification scenarios involving publicly offered securities investment funds, the system will strengthen the weight of user behavior scores according to relevant business rules to meet regulatory requirements.

[0090] Threshold Adjustment and Decision Module: This module sets a basic reliability threshold, serving as the system's initial baseline to reflect its basic tolerance for fraudulent activities. It also introduces a loss assessment correction coefficient, which is linked to the specific loss assessment amount. This coefficient dynamically adjusts the threshold to reflect the risk control principle that higher amounts require stricter review.

[0091] Specifically, the system first obtains the estimated repair cost from the damage assessment stage. Then, it determines the corresponding damage assessment correction coefficient based on this cost and uses this coefficient to adjust the reliability judgment threshold, thereby calculating the reliability correction threshold. Subsequently, the system compares the value score output by the value scoring module with this reliability correction threshold. If the value score reaches or exceeds the correction threshold, it indicates that the case materials are highly reliable, and the system allows the damage assessment result to be automatically approved and directly enter the claims process. If the value score is lower than the correction threshold, it indicates a potential fraud risk, and the system will prevent the damage assessment result from being automatically approved and generate a manual review instruction. Finally, the system outputs a damage assessment result report containing information such as the judgment result, score details, threshold settings, and adjustments.

[0092] The formula for calculating the reliability correction threshold is:

[0093]

[0094] in:

[0095] The reliability correction threshold, which is the threshold used for final comparison after adjusting the loss assessment amount, directly determines whether a case can automatically pass the review. A higher score means stricter requirements for material reliability; only cases with higher overall scores are exempt from manual review. Conversely, a lower score means lower requirements for material reliability. The lower the threshold, the more lenient the review standards. Compared with the base threshold Proportional to the adjustment factor Proportional;

[0096] The reliability threshold is a preset initial baseline value used by the system to reflect the basic tolerance for fraudulent behavior. It is set as a constant based on historical data or business experience. The level of [something] determines the rigor of the basic review; if A higher setting means that regular cases will require a higher reliability score to pass, which helps reduce fraud underreporting, but may increase the amount of manual review required for normal cases; if... A lower threshold indicates a higher automation pass rate, but may introduce more fraud risks.

[0097] The loss assessment correction coefficient is a dynamic adjustment factor that is positively correlated with the loss assessment amount. Its value increases as the loss assessment amount increases, reflecting the risk control principle that the larger the amount, the stricter the review.

[0098] The selection methods for the damage assessment correction coefficient include the linear proportional method and the piecewise interval method. The process of the linear proportional method is as follows:

[0099] The system automatically acquires the estimated repair cost as raw data from the vehicle accident damage assessment process, ensuring the accuracy and real-time nature of the input. Subsequently, the system substitutes this estimated repair cost into a pre-defined linear calculation formula. This formula, trained on massive amounts of historical data, uses a fixed proportional coefficient that reflects the statistical correlation between the assessed damage cost and the risk of fraud. The system then performs a simple multiplication of this coefficient and the assessed damage cost to directly generate a value λ that is proportional to the amount. The entire process is based entirely on numerical calculations, requiring no manual intervention. This allows λ to increase linearly with the input amount, thus smoothly enhancing the rigor of review for high-value cases. This method is simple and intuitive, allowing for rapid system response, and is suitable for scenarios where the assessed damage cost and the risk of fraud exhibit an approximately linear relationship.

[0100] Selecting the loss assessment correction coefficient using the segmented interval method The process is as follows:

[0101] Firstly, based on in-depth mining and analysis of massive amounts of historical case data, the system statistically analyzes the frequency of fraud cases and the specific distribution of the amounts involved within different monetary ranges. Simultaneously, it combines this with the pass rate performance of legitimate cases within each monetary range to comprehensively assess the risk characteristics of each segment. Through data analysis, it identifies critical points where fraud risk changes significantly, and accordingly scientifically delineates monetary segmentation nodes. This ensures that cases within each segment exhibit a high degree of internal consistency in risk characteristics, while different segments demonstrate clear risk differences.

[0102] Based on the segmentation of transaction ranges, the system assigns a fixed λ value or a preset benchmark adjustment coefficient to each range, taking into account the historical fraud risk level corresponding to each range and the risk tolerance requirements for different amounts of cases. This creates a tiered mapping relationship where the higher the amount, the larger the λ value. This segmented range method allows the adjustment of reliability judgment thresholds to accurately match the actual risk situation of different amount levels. Stricter review standards are implemented in high-amount ranges, while standards are appropriately relaxed in low-amount ranges, achieving a balance between risk control and processing efficiency. Furthermore, in high-risk, high-amount ranges, the system can introduce a smart contract and smart synchronization verification mechanism. By comparing the on-chain stored smart contract code with the business logic, the system ensures the matching of the damage assessment amount and the repair plan, ultimately achieving a closed loop of trusted transactions.

[0103] The process of comparing the value score obtained by the value scoring calculation module with the reliability correction threshold includes:

[0104] like If the overall reliability score of the materials in the vehicle accident case reaches the standard set by the system, the probability of fraud is low. The system will automatically determine the case as a credible case, allowing the damage assessment result to be automatically passed and directly transferred to the subsequent claims payment process without human intervention, thereby greatly improving processing efficiency.

[0105] like If the overall material reliability score of the current case does not meet the system's set standard, there is a potential risk of fraud or material quality problems. The system will automatically block the loss assessment result from passing and immediately generate a manual review instruction containing specific risk warnings and score details. The case will be pushed to the manual review queue, where professional reviewers will further verify the image materials, loss assessment amount, and various scores to ensure the accuracy and reliability of the loss assessment result.

[0106] This decision-making scheme is not only used for claims review, but can also serve as a certificate for subsequent credible transactions, supporting the automated execution of related financial services (such as factoring financing or fund unit redemption).

[0107] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0108] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0109] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An automated damage assessment system that integrates computer vision and anti-fraud technologies, characterized in that: include: Data acquisition module: Acquires vehicle accident video materials and simultaneously acquires reliability-related feature data, including image quality features, content consistency features, metadata features, and user behavior features; Reliability scoring module: Processes the above-collected data sets to establish a reliability scoring conversion model. The model is designed based on a supervised learning framework for machine learning and is trained using a neural network algorithm. Reliability-related feature data is input into the reliability scoring conversion model, and reliability scoring parameters are directly output. The reliability scoring parameters include image quality score, content consistency score, metadata score, and user behavior score. Value scoring calculation module: Constructs a weighted aggregation formula, based on the transformed reliability scoring parameters input into the weighted aggregation formula, which uses a linear combination to calculate and obtain the value score; Threshold Adjustment and Decision Module: Sets a reliability judgment threshold and a loss assessment correction coefficient. The reliability judgment threshold is an initial benchmark value used to reflect the system's tolerance for fraud. The reliability judgment threshold is adjusted using the loss assessment correction coefficient to obtain a reliability correction threshold. Then, the value score obtained by the value scoring calculation module is compared with the reliability correction threshold, and a loss assessment result judgment report is output to generate a decision scheme on whether to conduct manual review.

2. The automated damage assessment system combining computer vision and anti-fraud as described in claim 1, characterized in that: The image quality features are used to assess the sharpness, contrast, and integrity of the image to ensure that accident details are clearly identifiable; the content consistency features are used to detect signs of tampering or forgery. Metadata features are image-attached information, including shooting time, geographical location, and file attributes. These metadata features are used to verify the authenticity and temporal consistency of the incident. User behavior features analyze upload behavior patterns, including upload frequency, operating habits, and interaction history. These user behavior features are used to identify abnormal activities or fraudulent intentions.

3. The automated damage assessment system combining computer vision and anti-fraud as described in claim 1, characterized in that: In the reliability scoring module, the reliability scoring conversion model includes an input layer, a hidden layer, and an output layer. During the training phase, the reliability scoring conversion model is trained by learning the complex nonlinear relationship between reliability-related feature data from a large number of historical cases and the labels of cases that are ultimately proven to be real or fraudulent. After training, for a new accident case, the structured feature data is input into the reliability scoring conversion model. The neural network performs forward propagation calculations through its multi-layer network structure, and finally directly generates four key scoring parameters at the output layer: image quality score, content consistency score, metadata score, and user behavior score.

4. The automated damage assessment system combining computer vision and anti-fraud as described in claim 3, characterized in that: The functional mapping relationship of the reliability scoring conversion model is as follows: ； Where X is the input feature vector: ,in, arrive This represents the various reliability-related characteristic data after initial alignment; The nonlinear function learned by the reliability scoring transformation model. These are the output image quality score, content consistency score, metadata score, and user behavior score, respectively. Then, feature normalization processing is performed on the output image quality score, content consistency score, metadata score, and user behavior score. First, all four scores calculated in the current case are traversed, and the maximum and minimum values ​​of each score in the historical sample database or the current batch of data are calculated. Then, the min-max normalization method is used to perform a linear transformation on each original score: after subtracting the minimum value of the corresponding reliability score parameter from each score, the original score is precisely compressed to the closed interval [0, 1] by dividing by the range.

5. The automated damage assessment system combining computer vision and anti-fraud as described in claim 4, characterized in that: Outlier filtering is performed on the reliability score parameters obtained from the conversion: If any reliability score parameter obtained from the conversion meets the following conditions, it is considered an outlier: ； in: S represents a single reliability score parameter value; μ is the average value of the reliability score dataset samples in the massive historical cases, and the reliability score dataset includes the image quality score dataset, content consistency score dataset, metadata score dataset and user behavior score dataset. k is a constant threshold coefficient used to define the boundary range of outliers; σ is the sample standard deviation of the dataset corresponding to the reliability score parameter value, used to quantify the dispersion of the data; And the formula for obtaining the sample mean μ is: Where m is the total number of samples in the reliability score dataset. This represents the value of a single data point with index i in the reliability score dataset.

6. The automated damage assessment system combining computer vision and anti-fraud as described in claim 1, characterized in that: The weighted aggregation formula in the value scoring calculation module is: ； in: V represents the value score, used to comprehensively evaluate the reliability of vehicle accident video footage. A higher value score indicates a lower probability of fraud in the footage, and a higher image quality score. Quality dimensions reflecting image sharpness and noise level; content consistency score. Metadata scoring reflects the logical consistency between the location and shape of vehicle damage in the images and the accident description. Reflects the credibility of metadata based on shooting time, location, and equipment information; user behavior rating. It reflects the degree of abnormality in user upload behavior characteristics; These are the weighting coefficients for image quality score, content consistency score, metadata score, and user behavior score, respectively. The relative importance of each weighting coefficient is determined based on the significance of the corresponding reliability scoring parameters and is further established according to expert experience, while also satisfying the following conditions: .

7. The automated damage assessment system combining computer vision and anti-fraud as described in claim 1, characterized in that: In the threshold adjustment and decision module, the formula for calculating the reliability correction threshold is: ； in: This is the reliability correction threshold, which is the threshold used for final comparison after adjustment based on the loss assessment amount. The reliability judgment threshold is a preset initial baseline value of the system, which is used to reflect the basic tolerance for fraudulent behavior. The reliability judgment threshold is set as a constant based on historical data or business experience. The loss assessment correction coefficient is a dynamic adjustment factor that is positively correlated with the loss assessment amount. Its value increases as the loss assessment amount increases, reflecting the risk control principle that the larger the amount, the stricter the review.

8. The automated damage assessment system combining computer vision and anti-fraud as described in claim 7, characterized in that: The selection methods for the damage assessment correction coefficient include the linear proportional method and the piecewise interval method. The process of the linear proportional method is as follows: First, the raw data of the estimated repair cost is obtained from the vehicle accident damage assessment process. Then, the estimated repair cost is substituted into a preset linear calculation formula. The core of the preset linear calculation formula is a fixed ratio coefficient obtained by training through historical data. The system performs a simple multiplication operation between a fixed ratio coefficient and the assessed loss amount to directly generate a value λ that is proportional to the amount. The entire process is based on numerical calculation, which allows λ to grow linearly as the input amount increases.

9. The automated damage assessment system combining computer vision and anti-fraud as described in claim 7, characterized in that: The segmented interval method is used to select the damage assessment correction coefficient. The process is as follows: The frequency of fraud cases and the distribution of involved amounts within different monetary ranges were statistically analyzed. Combined with the pass rate of legitimate cases in each range, segmentation points for monetary amounts were defined. Cases within each range exhibited internal consistency in risk characteristics, but also showed variations across the ranges. After segmenting the ranges, a fixed risk level was assigned to each monetary range based on its historical fraud risk level and the operational tolerance requirements. The value or a preset benchmark adjustment coefficient forms a set of adjustments based on the amount. The larger the value, the better the step-like mapping relationship, so that the adjustment of the reliability judgment threshold can fit the actual risk situation of different monetary levels.

10. The automated damage assessment system combining computer vision and anti-fraud as described in claim 9, characterized in that: The threshold adjustment and decision-making module includes the following process: comparing the value score obtained by the value score calculation module with the reliability correction threshold. like This indicates that the overall reliability score of the materials in the vehicle accident case has reached the standard set by the system, and the probability of fraud is low. The system allows the current damage assessment result to be automatically approved without manual intervention. like This indicates that the overall reliability score of the materials in the current case has not met the system's set standards, and there is a potential risk of fraud. The system will prevent the current loss assessment result from being automatically approved and will generate a manual review instruction.

Images