An insurance claim damage detection method, device, equipment and storage medium

Abstract

Embodiments of the present application provide an insurance claim damage detection method, device and equipment and storage medium, relating to the technical field of insurance claim. After obtaining the static damage image submitted by the user, the method can extract the damage appearance feature in the static damage image, and obtain the standard damage dynamic rule video and damage annotation data from the standard damage database according to the damage appearance feature. Then, taking the static damage image as the source image and the standard damage dynamic rule video as the driving video, the method generates a dynamic damage video through the first-order motion model guided by the damage annotation data, so as to generate a damage detection report according to the dynamic damage video. The method can intuitively present the key features of the damage through the dynamic damage video generated by the first-order motion model, accurately identify the real damage degree based on the video comparison of the dynamic rule, improve the accuracy and detection efficiency of the damage detection result, and can also be applied to the field of financial technology.

Description

Technical Field

[0001] This application relates to the field of insurance claims technology, and in particular to an insurance claims damage detection method, device, equipment and storage medium. Background Technology

[0002] In the field of insurance claims, damage detection is based on image data, and damage verification and assessment are conducted based on the damage detection results to process insurance claims. For example, in the claims process for personal insurance such as accident insurance and health insurance, damage verification and assessment are key steps that determine the efficiency of claims processing, customer satisfaction, and the effectiveness of fraud prevention.

[0003] During the injury detection process, users can submit photos of their injuries through the insurance claims system. Claims personnel then rely on their experience to compare the submitted static photos with the insurance company's internal injury standard image library to determine the severity and authenticity of the injury. Because judging the extent of injury based on experience is highly subjective, different claims personnel may have significantly different interpretations of the degree of injury corresponding to the depth of bruise color, leading to inconsistent damage assessment standards. Therefore, image recognition tools can be used to assist in extracting injury features from static photos. For example, color recognition algorithms can be used to analyze the RGB values ​​of the bruise and determine its depth, and size measurement tools can be used to calculate the area of ​​the swollen region.

[0004] However, image recognition tools can only extract static quantitative features, indicating the current color but not the duration of damage corresponding to that color. Furthermore, image recognition tools are susceptible to interference from shooting conditions, leading to significant differences in the RGB values ​​of the same damage under different lighting conditions. Size measurements can also deviate due to varying shooting distances, resulting in inaccurate feature extraction. Additionally, if the damaged area is obscured by clothing, the image recognition tool cannot determine if damage exists in the obscured area, requiring a new image capture and extending the damage assessment period. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method, apparatus, device, and storage medium for detecting damage in insurance claims, in order to solve the problem of low accuracy in damage detection.

[0006] According to a first aspect of this application, a method for detecting damage in insurance claims is provided, the method comprising: Obtain source images, including static damage images submitted by the user; Damage appearance features are extracted from the damaged region of the source image, and the damage appearance features include at least one of damage color features, damage morphology features, and region texture features. Based on the damage appearance characteristics, driving videos and damage annotation data are obtained from a standard damage database; the driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image; the damage annotation data includes damage-related anatomical points. A first-order motion model is used to generate a dynamic damage video based on the source image, the driving video, and the damage annotation data; the first-order motion model is a depth generation model; the first-order motion model is configured to extract dynamic pattern features from the driving video guided by the damage annotation data, and output a dynamic damage video based on the damage appearance features and the dynamic pattern features. A damage detection report is generated based on the dynamic damage video. The damage detection report includes key differences between the dynamic damage video and the driving video, as well as the associated claims information of the source image.

[0007] According to a second aspect of this application, an insurance claim damage detection device is provided, the device comprising: The image acquisition module is used to acquire source images, including static damage images submitted by the user. The appearance feature extraction module is used to extract damage appearance features from the damage area of ​​the source image, wherein the damage appearance features include at least one of damage color features, damage morphology features, and regional texture features. The driving guidance module is used to obtain driving video and damage annotation data from a standard damage database based on the damage appearance characteristics; the driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image; the damage annotation data includes damage-related anatomical points. The video generation module is used to generate a dynamic damage video based on the source image, the driving video, and the damage annotation data using a first-order motion model; the first-order motion model is a depth generation model; the first-order motion model is configured to extract dynamic pattern features from the driving video guided by the damage annotation data, and output a dynamic damage video based on the damage appearance features and the dynamic pattern features. The result output module is used to generate a damage detection report based on the dynamic damage video. The damage detection report includes key differences between the dynamic damage video and the driving video, as well as the associated claims information of the source image.

[0008] According to a third aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described insurance claim damage detection method.

[0009] According to a fourth aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described insurance claim damage detection method.

[0010] By employing the above technical solutions, embodiments of this application provide a method, apparatus, device, and storage medium for insurance claim damage detection. After acquiring a static damage image submitted by a user, the method can extract damage appearance features from the static damage image and retrieve standard damage dynamic pattern videos and damage annotation data from a standard damage database according to these features. Then, using the static damage image as the source image and the standard damage dynamic pattern video as the driving video, a first-order motion model is guided by the damage annotation data to generate a dynamic damage video, thereby generating a damage detection report based on the dynamic damage video. This method can intuitively present the key features of the damage through the dynamic damage video generated by the first-order motion model, and accurately identify the true degree of damage based on video comparison of dynamic patterns, improving the accuracy and efficiency of damage detection results. This method can also be applied to the fintech field.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the insurance claim damage detection method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the customer service platform structure provided in the embodiments of this application; Figure 3 This is a schematic diagram of the affine transformation optimization process provided in the embodiments of this application; Figure 4 A schematic diagram illustrating the medical adjustment process of the occlusion mask provided in this application embodiment; Figure 5 This is a schematic diagram illustrating the process of constructing a standard damage database as provided in an embodiment of this application. Figure 6 This is a schematic diagram of the structure of the insurance claim damage detection device provided in the embodiments of this application. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] In this embodiment, damage detection can be applied to the field of insurance claims, specifically by performing damage detection based on image data and verifying and assessing damage based on the detection results, thereby assisting insurance personnel in reviewing insurance claims. Damage verification and assessment are crucial steps in determining claims efficiency, customer satisfaction, and fraud prevention effectiveness. Furthermore, damage detection can be tailored to different types of damage depending on the type of insured object in the insurance business.

[0015] For example, in the claims process for personal insurance such as accident insurance and health insurance, the insured object is the user's person, and the type of injury detected is personal injury, such as bruises or sprains caused by external injuries. However, in the claims process for some property insurance such as car insurance, the insured object is property, and the type of injury detected is property damage, such as surface bumps or structural wear.

[0016] It should be noted that in some embodiments of this application, the injury detection process is illustrated using the personal insurance claims process as an example. Obviously, the described injury detection process can also be applied to the injury detection and claims processes for other types of insurance, and will not be shown one by one. In addition, the method can also be applied to the field of fintech.

[0017] Because injuries of different degrees will exhibit different appearance characteristics, and there can be a correlation between appearance characteristics and injury severity. For example, there is a positive correlation between the severity of injury and the area and depth of bruising; that is, the more severe the injury, the more pronounced the corresponding appearance characteristics. For instance, minor injuries result in smaller, lighter-colored bruises, while severe injuries result in larger, darker-colored bruises. Therefore, when verifying and assessing damage, appearance characteristics can be used to help determine the degree of injury.

[0018] In some embodiments, injury detection can be performed based on injury images submitted by the user. That is, during the injury detection process, the user can submit photos of the injury through the insurance claims system. Claims personnel then use their experience, such as over 5 years of experience in injury assessment, to compare the user-submitted static photos with the insurance company's internal injury standard image library, such as standard photos of mild bruising and diagrams of moderate joint swelling, to determine the severity and authenticity of the injury.

[0019] Judging the extent of damage based on experience is highly subjective. Different claims adjusters can have significantly different assessments of the degree of damage corresponding to the depth of bruise color, with differences reaching up to 40%, leading to inconsistent damage assessment standards. Furthermore, manual damage assessment methods cannot handle fraud related to dynamic features. For example, fresh bruises and bruises faked with paint may look similar at rest, but their dynamic changes are completely different, yet it is difficult to accurately distinguish them using static damage images. In addition, manual judgments are difficult to interpret, and users are unlikely to accept experience-based assessments, easily leading to skepticism.

[0020] In some embodiments, image recognition tools such as color analysis tools, size measurement tools, and simple AI image tools can also be used to assist in extracting damage features from static photographs. For example, color recognition algorithms can be used to analyze the RGB values ​​of bruises and determine the color depth, and size measurement tools can be used to calculate the area of ​​the swollen region.

[0021] However, damage detection methods that use image recognition tools to extract damage features from static photographs can only extract static quantitative features and cannot correlate with dynamic medical patterns. For example, the RGB value of a bruise (50, 20, 80) only indicates the current color and cannot determine the duration of the injury corresponding to that color. Fresh bruises may also appear similar in color within one day due to different skin types, leading to inaccurate recognition results from image recognition tools. Furthermore, image recognition tools are easily affected by shooting conditions. The RGB value of the same injury can vary by up to 30% under different lighting conditions, such as strong light or backlight. Size measurements can also deviate due to different shooting distances, further resulting in inaccurate feature extraction results. In addition, image recognition tools cannot handle occluded scenes. If the damaged area is obscured by clothing, such as a trouser leg covering a leg bruise, the image recognition tool cannot determine whether there is damage in the obscured area, requiring the customer to retake the photograph, thus extending the damage assessment period.

[0022] In some embodiments, medical imaging can also be used to assist in injury detection. Medical imaging may include X-rays, ultrasound images, etc., from certified medical institutions. For serious injuries such as fractures and joint dislocations, some insurance companies require customers to provide medical imaging reports to combine multiple injury images to determine the authenticity and severity of the injury.

[0023] However, the applicability of medical imaging-assisted detection is narrow, only applicable to severe injuries, and cannot cover most mild or moderate injuries such as bruising and mild swelling, resulting in low user acceptance of injury detection. Furthermore, because medical imaging examinations require users to visit certified medical institutions, the cost of testing increases, and the long turnaround time for results significantly prolongs the claims process, adding an average of 2-4 business days to each claim. In addition, medical imaging-assisted detection cannot reflect dynamic recovery; X-ray images only show the instantaneous state of bone injury and cannot determine the recovery trend, such as whether a fracture is healing normally, thus limiting its auxiliary role in claims conclusions.

[0024] In summary, the injury detection methods described in the above embodiments suffer from limitations in the claims process for personal insurance such as accident insurance and health insurance. These limitations stem from the single dimension of injury information and low accuracy in injury detection. Specifically, submitted injury certificates are often static photographs, such as snapshots of leg bruises or joint swelling. Static photographs only show the appearance of the injury at a specific moment and cannot reflect its dynamic characteristics. For example, claims adjusters cannot determine from static bruise photographs whether the bruise is fresh (injured within 1 day) or old (injured more than 5 days ago), nor can they assess whether joint swelling is accompanied by limited mobility (e.g., whether the knee flexion angle is below the normal range).

[0025] Furthermore, verifying the authenticity of injuries is difficult, leading to a high risk of fraud. Some users exploit the information blind spots in static photos to commit insurance fraud. They may present old injuries (such as bruises from 3 months ago) as fresh accidental injuries, exaggerate the extent of injury by taking photos at different angles (such as using close-up shots to make minor swelling appear more severe), or fabricate evidence by splicing together photos of other people's injuries. Because static photos cannot provide dynamic characteristics of changes over time, such as the color gradation of fresh injuries or the fading trend of swelling, claims adjusters rely solely on experience to make judgments, resulting in a fraud detection rate of less than 30% in such cases.

[0026] In addition, the cost of explaining claims conclusions is high. Damage assessment involves professional medical knowledge such as the correlation between the speed of bruising fading and the freshness of the injury, and the degree of injury corresponding to the angle of joint movement. Users may find it difficult to understand through text descriptions or static photo comparisons, which will prolong the explanation period of the claims process.

[0027] To address the issue of low accuracy in damage detection, some embodiments of this application provide a method for damage detection in insurance claims. This method can be applied to electronic devices with data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control computers. For ease of description, this application uses an electronic device as the executing entity for the method. It should be understood that the method can also be applied to other types of executing entities, which are not illustrated in all embodiments of this application. Figure 1As shown, the method includes: S101. Obtain the source image.

[0028] In the process of damage detection for insurance claims, source images can be obtained first. These source images include static damage images submitted by the user.

[0029] Source images can be obtained through multiple channels, including not only user-submitted static damage images, but also medical examination images, accident scene images, and surveillance images. These images can comprehensively reflect the damaged area from different dimensions, thereby enabling verification of the authenticity of multiple images and improving the accuracy of damage detection results.

[0030] Users can submit static damage images through the customer service platform (or system) of the insurance claims industry. The electronic device used to perform damage detection can then obtain these static damage images from the customer service platform. For example, ... Figure 2 As shown, the customer service platform can include a client-side interface, a claims processing interface, and a communication interface. By integrating system functions with business processes, the customer service platform is embedded into the entire life insurance claims process, achieving seamless connection between the client-side interface, the claims processing interface, and the communication interface.

[0031] The client application can be used for standardized photo capture, meaning that user interaction occurs on user-side devices such as computers and mobile terminals by running an application (APP) used in the insurance claims industry. During user interaction, guidance information can be provided to guide customers in taking injury photos according to standards. For example, static injury images should include normal skin within a 2cm radius around the injury; when photographing joints, both frontal and side views are required. The client application can also automatically detect photo quality, such as sharpness and injury integrity, based on the APP's built-in detection tools, ensuring that the source image meets the model input requirements.

[0032] The claims processing terminal can be used for dynamic loss assessment. That is, on the claims personnel's side device, by running an insurance claims industry APP to perform user interaction, the damage detection results and related data of the damage detection process are visualized to assist claims personnel in loss assessment.

[0033] The communication platform is used for visual explanation of claims conclusions. This involves sending animated comparison reports to clients and providing audio interpretations, allowing users to intuitively understand the claims process. For example, the animated comparison report sent via the communication platform might include a simulated injury animation on the left and a standard animation on the right, highlighting key differences and their medical basis. The audio interpretation can be a doctor's recorded voice message explaining why the injury falls into the corresponding injury category.

[0034] Based on the aforementioned customer service platform, multi-terminal interaction between the client, claims processing, and communication terminals can be achieved. Through this multi-terminal interaction, static damage images that better meet damage detection requirements can be obtained. Specifically, in some embodiments, when acquiring the source image, an initial damage image submitted by the user can be acquired first, and then a quality inspection can be performed on the initial damage image to detect image quality parameters. These image quality parameters include at least one of the following: the sharpness and integrity of the damaged area, and the shooting angle.

[0035] Then, the input quality requirements of the first-order motion model used in the subsequent detection process are obtained, and the detected image quality parameters are compared with the subsequent input quality requirements. If the image quality parameters meet the input quality requirements of the first-order motion model, it means that the image quality of the static damage image currently submitted by the user can be used for subsequent image processing. Therefore, the initial damage image can be used as the source image for subsequent damage detection.

[0036] If the image quality parameters do not meet the input quality requirements of the first-order motion model, it means that the image quality of the static damage image currently submitted by the user cannot be used for subsequent image processing. In this case, it is necessary to obtain the shooting standard information and generate user guidance data based on the shooting standard information.

[0037] For example, to meet the input requirements of a first-order motion model, user guidance data can be generated to guide the client in taking standardized photos. This involves automatically detecting the sharpness and damage integrity of the photos to determine image quality parameters. If the image quality parameters do not meet the input quality requirements of the first-order motion model, standard shooting information can be obtained through the claims app. For example, the standard shooting information requires that for the damaged area, 2cm of normal skin around the injury be included; for joint injuries, frontal and side views of the joint are required. User guidance data is then generated based on this standard shooting information and sent to the client to guide them in taking injury photos according to the standards, ensuring that the source image meets the model's input requirements.

[0038] S102. Extract damage appearance features from the damaged area of ​​the source image.

[0039] After acquiring the source image, feature extraction can be performed on the acquired source image to extract the appearance features of the damage area in the source image. These appearance features include at least one of damage color features, damage morphology features, and regional texture features. For example, an electronic device can use a static damage photograph submitted by a user as the source image (S) to extract appearance features of the damaged area, such as bruise color, swelling shape, and skin texture.

[0040] Damage appearance features can be extracted from source images using feature extraction models. These models are neural network models trained on damage sample data. They can analyze the color, shape, texture, and other properties of the source image to extract various features that represent the characteristics of the damage.

[0041] In some embodiments, in order to extract damage color features, the source image can be converted from the RGB color space to a color space more suitable for separating color information after calling the feature extraction model, and then the color distribution corresponding to the damage color can be analyzed by calculating the color histogram of the image.

[0042] For example, after acquiring a static image of damage submitted by a user, the image can be converted from the RGB color space to the HSV or LAB color space. Within the HSV or LAB color space, pixels in a specific color range can be extracted by setting a threshold, such as bruises appearing as purple or blue tones. Then, color histogram analysis is used to calculate the color histogram of the source image, and the characteristics of the bruise color are determined by analyzing the distribution of the histogram. Finally, by comparing the histograms of different color channels, the color distribution of the bruise area is identified.

[0043] In some embodiments, in order to extract damage morphology features, a feature extraction model can be invoked first, and the feature extraction model can be used to perform edge detection on the source image. Then, morphological operations can be performed to enhance the morphological features of the damaged area, and the damage morphology features can be obtained through region segmentation.

[0044] For example, to extract swelling morphology from static damage images, edge detection algorithms such as Canny edge detection can be used first to obtain the contour of the swollen region and analyze its morphology to identify the boundary of the swollen region. Then, morphological operations such as dilation and erosion are used to process the swollen region to enhance its morphological features. Specifically, dilation can fill small holes within the swollen region, making its contour more complete. Next, image segmentation algorithms such as thresholding and region growing are used to separate the swollen region from the background, obtaining the damage morphology features corresponding to the swelling morphology. The segmented swollen region can be further used for morphological analysis, such as calculating its area and perimeter.

[0045] In some embodiments, in order to extract regional texture features, the electronic device can calculate the gray-level co-occurrence matrix (GLCM) of the source image through a feature extraction model, perform local binary pattern (LBP) processing, and then use filters such as Gabor to extract texture features of the image in different directions and scales.

[0046] For example, to extract skin texture from a static damaged image, texture features can be extracted by calculating the gray-level co-occurrence matrix (GLCM) of the static damaged image after acquisition. GLCM provides feature parameters such as contrast, correlation, entropy, uniformity, and energy of the texture, reflecting the details and regularity of the skin texture. Then, based on the Local Binary Pattern (LBP) texture descriptor, texture information is extracted by analyzing the relationship between each pixel in the image and its neighboring pixels. Next, a Gabor filter is used to extract texture features of the static damaged image at different directions and scales. By convolving the static damaged image with the Gabor filter output, texture responses at different directions and scales can be obtained, facilitating the analysis of the details and directionality of the skin texture.

[0047] It should be noted that, in addition to the aforementioned damage color features, damage morphology features, and regional texture features, the damage appearance features extracted from the source image can also include other types of features, such as color contrast features, graphic target regions, and occlusion features. Different feature extraction algorithms can be used for different application scenarios to extract different types of damage appearance features from static damage images, which will not be shown one by one in the embodiments of this application.

[0048] S103. Obtain driving video and damage annotation data from the standard damage database based on the damage appearance characteristics.

[0049] After extracting damage appearance features from the source image, the damage type corresponding to the currently submitted static damage image can be determined based on these features. Then, driving video and damage annotation data can be obtained based on the damage type. The driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image.

[0050] Driving video and injury annotation data can be obtained from a pre-built standard injury database. This database can store various standard reference data representing different types of injuries. For example, it includes a dynamic medical standard injury database built in collaboration with insurance companies and certified medical institutions. This database can store various video files, such as videos of fresh bruises fading within 0-7 days and videos of movement trajectories 0-14 days after a knee sprain. These video files can be categorized and stored within the dynamic medical standard injury database according to injury type.

[0051] Similarly, to guide the model in learning key locations, damage annotation data can also be obtained from a standard damage database. This damage annotation data is obtained by professionals annotating sample damage images, and may include pre-annotated damage-related anatomical points. For example, a standard damage database may include an annotation database storing damage-related anatomical points annotated by orthopedic and dermatologists, such as the attachment points of the medial and lateral ligaments of the knee joint, bruise boundary feature points, etc., to guide the model to prioritize learning key damage locations and avoid misidentifying clothing textures and background patterns as key points.

[0052] In some embodiments, when retrieving driving video and damage annotation data from a standard damage database based on damage appearance features, the damage type can first be determined based on the damage appearance features, and then the standard damage database can be invoked. The standard damage database includes a standard damage dynamic library and an annotation database. Then, driving video of the same type is matched from the standard damage dynamic library according to the damage type, and damage annotation data of the same type is matched from the annotation database according to the damage type.

[0053] For example, when a driving video is needed, the injury type can be determined based on the injury's appearance features, and this type can be specified in the video data acquisition request. After identifying appearance features such as bruise color, swelling shape, and skin texture in a static injury image, the current injury type, i.e., a fresh bruise, can be determined based on a combination of quantitative indicators of these appearance features. After setting the fresh bruise in the video data acquisition request, a video data acquisition request can be sent to the medical standard injury dynamics library to obtain standard injury dynamic patterns video, i.e., fresh bruise fading video from 0 to 7 days, as the driving video (D), used to extract medical dynamic patterns of similar injuries.

[0054] S104. Using a first-order motion model, generate dynamic damage video based on the source image, driving video, and damage annotation data.

[0055] After acquiring the driving video and damage annotation data, the electronic device can call a first-order motion model (FOMM) to generate video. This means using the FOMM to generate a dynamic damage video based on the source image, driving video, and damage annotation data. The first-order motion model (FOMM) is a deep generative model. As a deep learning model, the FOMM can be used for image animation and video generation, animate static images, and achieve motion transfer, among other things.

[0056] A first-order motion model is configured to extract dynamic pattern features from a driving video, guided by damage annotation data, and output a dynamic damage video based on damage appearance features and dynamic pattern features. The first-order motion model can include a motion estimation module and an image generation module. The motion estimation module can separate the appearance and motion information of the target object through self-supervised learning and represent them as features. The motion estimation module can detect keypoints from the input driving video and perform local affine transformations, then use the displacements of these keypoints and the parameters of the local affine transformations as a representation of motion. The image generation module can model occlusions that occur during the target's motion, then extract appearance information from a given image, and combine it with the feature representations obtained by the motion estimation module to generate the final image or video.

[0057] like Figure 3 As shown, in some embodiments, when using a first-order motion model to generate a dynamic damage video based on the source image, driving video, and damage annotation data, the first-order motion model can first be guided by the damage annotation data to learn the key damage locations in the source image and driving video. Then, based on the learning results of the key damage locations, a composite Jacobian matrix is ​​generated. The composite Jacobian matrix includes an individual damage feature matrix and a standard dynamic feature matrix for similar damage points; the individual damage feature matrix is ​​the Jacobian matrix of key points in the source image; and the standard dynamic feature matrix for similar damage points is the Jacobian matrix of key points in the driving video.

[0058] Then, based on the composite Jacobi matrix, the morphological changes of the damage over time are captured to obtain dynamic regularity features. Then, based on the dynamic regularity features and the appearance features of the damage, multiple damage images are generated. Finally, the multiple damage images are fused according to the temporal relationship to generate a dynamic damage video.

[0059] For example, when performing keypoint detection and affine transformations, first-order motion models can be guided by medical annotation data to prioritize learning the critical locations of injuries, avoiding misidentification of clothing textures and background patterns as keypoints. Simultaneously, optimizing the Jacobian matrix calculation for local affine transformations allows for accurate capture of morphological changes in the injury over time, such as the daily color gradient of bruises and the skin contraction coefficient as swelling subsides, ensuring that the animation conforms to medical physiological principles, i.e., that the joint range of motion does not exceed physiological limits.

[0060] The composite Jacobian matrix adapted to the damage scenario can be expressed as:

[0061] in, The individual damage feature matrix is ​​the Jacobian matrix of key points in the static damage image submitted by the user, used to describe the individual damage features of the customer. This is the standard dynamic feature matrix for similar injuries, i.e., the Jacobian matrix of key points in a medical standard video, used to describe the standard dynamics of similar injuries. The composite Jacobian matrix can be obtained by multiplying the individual damage feature matrix and the standard dynamic feature matrix of similar damage, and then inverting the result. The composite Jacobian matrix can be used to quantify the dynamic difference between customer damage and standard damage, providing a precise mathematical basis for damage assessment.

[0062] In some embodiments, the occlusion mask can also be medically adjusted, such as... Figure 4 As shown, when generating a dynamic damage video using a first-order motion model based on the source image, driving video, and damage annotation data, an occlusion awareness generator is used to extract an occlusion mask from the source image and obtain a preset damage region priority rule. This damage region priority rule sets the visibility of damage regions in the source image. Based on the damage region priority rule, the occlusion mask is then supplemented and estimated to generate an optimized occlusion mask. Finally, based on the optimized occlusion mask, damage feature fusion is performed to obtain a fused mask damage feature map, and the dynamic damage video is generated based on this fused mask damage feature map.

[0063] For example, by optimizing the occlusion-aware generator, the occlusion mask ( The estimation process incorporates medical injury area priority rules, such as prioritizing the preservation of the visibility of skin surface injuries. Areas covered by clothing need to be marked and redrawn, but this needs to be verified in conjunction with medical common sense, in order to avoid the first-order motion model over-drawing of non-injury areas.

[0064] The occlusion mask and damage features can be fused according to the following formula:

[0065] In the formula, For the optimized occlusion mask, 1 can represent the visible damaged area, and 0 can represent the occluded area or the non-damaged area. For feature rollback operation, This is a damage feature map after fusion masking, used to ensure that the generated video animation only presents damage areas that conform to medical principles.

[0066] S105. Generate a damage detection report based on the dynamic damage video.

[0067] After generating a dynamic damage video using a first-order motion model, a damage detection report can be generated based on the dynamic damage video. This damage detection report includes key differences between the dynamic damage video and the driving video, as well as related claims information from the source image.

[0068] In some embodiments, when generating a damage detection report based on dynamic damage video, a first set of keyframes can be extracted from the dynamic damage video, and a second set of keyframes can be extracted from the driving video. Then, by comparing the first and second set of keyframes, key differences between the dynamic damage video and the driving video are generated. Finally, based on the first set of keyframes and the source image, the damage level is calculated, and related claims information can be queried according to the damage level.

[0069] For example, after generating a dynamic injury video using a first-order motion model, the dynamic injury video can be used as an injury recovery simulation animation, such as a 7-day fading animation corresponding to a user's bruise photo. Then, by using an animation comparison tool, the injury recovery simulation animation can be compared frame-by-frame with a standard injury animation, and differences can be marked. For example, if a customer's bruise only fades by 5% within 3 days, which is below the standard of 15% fading per day, it can be identified as a suspected old injury. Simultaneously, by combining the dynamic pattern characteristics corresponding to the keyframe set with the injury appearance characteristics in the static injury image, the injury level can be determined. For example, if the joint range of motion is <60°, the corresponding injury level is determined to be moderate injury. Then, based on the injury level, related claim information is determined. Related claim information can include related compensation standards. For example, if the joint range of motion is <60°, the current injury level is determined to meet the conditions for moderate injury compensation, thereby generating information related to moderate injury compensation and sending it to the claims department to assist claims personnel in quickly assessing the damage.

[0070] By applying the technical solutions of the above embodiments, the insurance claim injury detection method described in the above embodiments can achieve breakthroughs in three dimensions: efficiency, accuracy, and customer experience. The generated dynamic animation can intuitively present key characteristics of the injury, such as recovery trends and degree of activity restriction. Therefore, claims personnel do not need to repeatedly communicate with customers to supplement information, reducing the loss assessment time from an average of 2-3 working days to 0.5 working days, lowering labor costs by 30%, improving loss assessment efficiency, and reducing labor costs. The method automatically associates with compensation standards to generate loss assessment suggestion reports, reducing the manual judgment steps of claims personnel. For example, standard cases such as minor bruises can achieve semi-automated loss assessment, improving processing efficiency by 50%. The method can also improve the accuracy of loss assessment. Specifically, based on animation comparison using medical dynamics, it can accurately identify fraudulent behaviors such as old injuries being passed off as fresh injuries or exaggerating the extent of injury, increasing the fraud detection rate from 30% to 80%.

[0071] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide an insurance claim injury detection method. The difference between this method and the above embodiments is that it can realize the construction of a standard injury database, such as... Figure 5 As shown, the method includes: S201. Collect dynamic recovery videos of different types of damage; S202. Set category labels based on dynamically restored videos; S203. Classify the dynamic recovery videos according to the damage type to build a standard damage dynamic library; S204. Embed the classification labels into the video metadata of the dynamically restored video to build a label database.

[0072] When constructing a standardized injury database, dynamic recovery videos of different types of injuries can be collected. For example, to ensure the medical accuracy of the driving videos, a standardized dynamic injury database can be constructed in collaboration with certified medical institutions. In constructing this database, data collection can begin, such as collecting dynamic recovery videos of different types of injuries like bruising, swelling, and skin abrasions. The collected dynamic recovery videos can cover mild, moderate, and severe injuries, with more than 500 samples collected for each type of injury. These videos should include dynamic recovery videos of different ages, genders, and skin types, such as daily changes in mild leg bruising from 0 to 7 days, and activity trajectory videos from 0 to 14 days after a knee sprain.

[0073] Next, classification labels are set based on the dynamically recovered videos. These labels include damage type, damage stage, key dynamic features, and damage judgment criteria. The dynamically recovered videos are then categorized according to damage type to construct a standard dynamic damage library. Simultaneously, the classification labels are embedded into the video metadata of the dynamically recovered videos to construct a label database, forming a standard damage database.

[0074] For example, after acquiring dynamic recovery videos through data collection, attending physicians or medical experts can perform medical annotations on these videos. The medical annotations can include injury stages such as the acute, subacute, and recovery phases of the bruise; key dynamic features such as the bruise fading by 15% daily during the acute phase; and medical criteria such as no color change within 3 days indicating an old injury. By embedding these annotations into the video metadata, key frames in the dynamic recovery video can be annotated, forming medical annotation data.

[0075] In some embodiments, the constructed standard injury database can also be updated in real time, that is, the data is updated quarterly according to the latest "Personal Insurance Disability Assessment Standards" and clinical cases, adding rare injury types, such as contusions in special areas, to ensure the timeliness and accuracy of the dynamic database.

[0076] In some embodiments, as a specific implementation of the insurance claim damage detection method described in the above embodiments, some embodiments of this application also provide an insurance claim damage detection device, such as... Figure 6 As shown, the device includes: The image acquisition module is used to acquire source images, including static damage images submitted by the user. The appearance feature extraction module is used to extract damage appearance features from the damage area of ​​the source image, wherein the damage appearance features include at least one of damage color features, damage morphology features, and regional texture features. The driving guidance module is used to obtain driving video and damage annotation data from a standard damage database based on the damage appearance characteristics; the driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image; the damage annotation data includes damage-related anatomical points. The video generation module is used to generate a dynamic damage video based on the source image, the driving video, and the damage annotation data using a first-order motion model; the first-order motion model is a depth generation model; the first-order motion model is configured to extract dynamic pattern features from the driving video guided by the damage annotation data, and output a dynamic damage video based on the damage appearance features and the dynamic pattern features. The result output module is used to generate a damage detection report based on the dynamic damage video. The damage detection report includes key differences between the dynamic damage video and the driving video, as well as the associated claims information of the source image.

[0077] By applying the technical solutions of the above embodiments, the insurance claim damage detection device described in the above embodiments can acquire static damage images submitted by users through the image acquisition module, and extract damage appearance features from the static damage images through the appearance feature extraction module. The driving guidance module then acquires standard damage dynamic pattern videos and damage annotation data from the standard damage database according to the damage appearance features. The video generation module uses the static damage image as the source image and the standard damage dynamic pattern video as the driving video, and guides a first-order motion model to generate a dynamic damage video through the damage annotation data. The result output module then generates a damage detection report based on the dynamic damage video. The device can intuitively present the key features of the damage through the dynamic damage video generated by the first-order motion model, and accurately identify the actual degree of damage based on the video comparison of dynamic patterns, thereby improving the accuracy and efficiency of damage detection results.

[0078] It should be noted that other corresponding descriptions of the functional units involved in the insurance claim damage detection device provided in this application embodiment can be found in the corresponding descriptions in the insurance claim damage detection method provided in the above embodiments, and will not be repeated here.

[0079] This application also provides a computer device, specifically a personal computer, server, network device, etc. The computer device includes a bus, processor, memory, and communication interface, and may also include input / output interfaces and a display device. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores location information. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0080] Those skilled in the art will understand that the structure of the computer device described above is only a partial structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. A specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0081] In one embodiment, a computer-readable storage medium is also provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0082] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0083] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0084] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0085] Any references to memory, database, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc.

[0086] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0087] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0088] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0089] It should be noted that any software tools or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

[0090] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting damage in insurance claims, characterized in that, The method includes: Obtain source images, including static damage images submitted by the user; Damage appearance features are extracted from the damaged region of the source image, and the damage appearance features include at least one of damage color features, damage morphology features, and region texture features. Based on the damage appearance characteristics, driving videos and damage annotation data are obtained from a standard damage database; the driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image; the damage annotation data includes damage-related anatomical points. A first-order motion model is used to generate a dynamic damage video based on the source image, the driving video, and the damage annotation data; the first-order motion model is a depth generation model; the first-order motion model is configured to extract dynamic pattern features from the driving video guided by the damage annotation data, and output a dynamic damage video based on the damage appearance features and the dynamic pattern features. A damage detection report is generated based on the dynamic damage video. The damage detection report includes key differences between the dynamic damage video and the driving video, as well as the associated claims information of the source image.

2. The method according to claim 1, characterized in that, Obtain the source image, including: Obtain the initial damage image submitted by the user; The image quality parameters in the initial damaged image are detected, and the image quality parameters include at least one of the following: the sharpness and integrity of the damaged area and the shooting angle; If the image quality parameters do not meet the input quality requirements of the first-order motion model, obtain shooting standard information and generate user guidance data based on the shooting standard information.

3. The method according to claim 1, characterized in that, Based on the damage appearance characteristics, driving video and damage annotation data are obtained from a standard damage database, including: The damage type is determined based on the damage appearance characteristics; The standard damage database is invoked, which includes a standard damage dynamic library and a labeling database; Match driving videos of the same type from the standard damage dynamic library according to the damage type; According to the damage type, match the damage annotation data of the same type from the annotation database.

4. The method according to claim 3, characterized in that, The method further includes: Collect dynamic recovery videos of different types of injuries; The dynamic recovery video is classified and labeled, and the classification labels include damage type, damage stage, key dynamic features and damage judgment criteria. The dynamically restored videos are categorized according to the damage type to construct the standard damage dynamic library; The classification labels are embedded into the video metadata of the dynamically restored video to construct the label database.

5. The method according to claim 1, characterized in that, Using a first-order motion model, a dynamic damage video is generated based on the source image, the driving video, and the damage annotation data, including: The damage annotation data guides the first-order motion model to learn the key damage locations in the source image and the driving video; Based on the learning results of the key damage locations, a composite Jacobian matrix is ​​generated. The composite Jacobian matrix includes an individual damage feature matrix and a standard dynamic feature matrix of similar damage. The individual damage feature matrix is ​​the Jacobian matrix of key points in the source image. The standard dynamic feature matrix of similar damage is the Jacobian matrix of key points in the driving video. Based on the composite Jacobian matrix, the morphological changes of damage over time are captured to obtain the dynamic law characteristics; Multi-frame damage images are generated based on the dynamic pattern characteristics and the damage appearance characteristics; The damaged images from multiple frames are fused according to their temporal relationship to generate the dynamic damaged video.

6. The method according to claim 1, characterized in that, Using a first-order motion model, a dynamic damage video is generated based on the source image, the driving video, and the damage annotation data, including: An occlusion-aware generator is used to extract an occlusion mask from the source image. Obtain a preset damage region priority rule, wherein the damage region priority rule sets the visibility of the damage region in the source image; Based on the damage region priority rule, the occlusion mask is supplemented and estimated to generate an optimized occlusion mask; Based on the optimized occlusion mask, damage feature fusion is performed to obtain a fused mask damage feature map; The dynamic damage video is generated based on the fused mask damage feature map.

7. The method according to claim 1, characterized in that, A damage detection report is generated based on the dynamic damage video, including: Extract a first set of keyframes from the dynamic damage video, and extract a second set of keyframes from the driving video; By comparing the first set of keyframes and the second set of keyframes, key differences between the dynamic damage video and the driving video are generated. The damage level is calculated based on the first set of keyframes and the source image; Search for related claims information based on the stated damage level.

8. An insurance claim damage detection device, characterized in that, The device includes: The image acquisition module is used to acquire source images, including static damage images submitted by the user. The appearance feature extraction module is used to extract damage appearance features from the damage area of ​​the source image, wherein the damage appearance features include at least one of damage color features, damage morphology features, and regional texture features. The driving guidance module is used to obtain driving video and damage annotation data from a standard damage database based on the damage appearance characteristics; the driving video is a standard damage dynamic pattern video of the same type of damage in the static damage image; the damage annotation data includes damage-related anatomical points. The video generation module is used to generate a dynamic damage video based on the source image, the driving video, and the damage annotation data using a first-order motion model; the first-order motion model is a depth generation model; the first-order motion model is configured to extract dynamic pattern features from the driving video guided by the damage annotation data, and output a dynamic damage video based on the damage appearance features and the dynamic pattern features. The result output module is used to generate a damage detection report based on the dynamic damage video. The damage detection report includes key differences between the dynamic damage video and the driving video, as well as the associated claims information of the source image.

9. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images