A method and system for intelligent classification and automatic uploading of insurance claim investigation materials

Abstract

The application discloses an insurance claim survey material intelligent classification and automatic uploading method and system, relates to the technical field of insurance claim, and comprises the following steps: obtaining case data and pictures in a claim system by using DOM tree analysis and file stream camouflage technology; classifying multi-modal survey materials by a multi-scale feature fusion CNN combined with a spatial attention mechanism, and adaptively adjusting the classification results by adopting a dynamic threshold strategy; and automatically uploading the classified materials to a specified position of an insurance system by a concurrent control algorithm and an abnormal fuse mechanism. The automatic classification and uploading replace manual operation, parallel processing reduces waiting time, and processing speed is improved. The multi-scale feature fusion CNN combined with the spatial attention mechanism classifies the multi-modal survey materials, and the classification accuracy is significantly improved compared with traditional methods. The dynamic threshold mechanism and the integrated learning strategy enable the system to adapt to changes in material features caused by new fraud methods, and the system has good scalability.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of insurance claim settlement, in particular to a method and system for intelligent classification and automatic uploading of insurance claim settlement survey materials. BACKGROUND

[0002] In the field of insurance claim settlement, online survey is a core link that determines the efficiency and accuracy of claim settlement. In the traditional process, survey materials contain various data such as vehicle driving licenses, driving licenses, accident scene photos and other types of data materials. All data materials need to be manually classified, sorted and uploaded.

[0003] Chinese patent CN107993154A discloses a method for efficient multi-tasking of insurance claim settlement remote investigation. After receiving the notification of the accident, the insurer designates the evidence collector to conduct the survey on site. The evidence collector can be the reporter or the designated personnel of the insurer, understand the loss situation and the cause, and take photos of the accident scene. Chinese patent CN112308532A discloses a non-vehicle insurance insurance appraiser remote online video survey system and method, which includes a case structured information processing subsystem, a case online remote video acquisition subsystem and a system user. The system user uses a smart phone or a computer to interact with the case structured information processing subsystem and the case online remote video acquisition subsystem through a communication network, a telecommunications network or a mobile network to realize remote online video survey of insurance appraisers.

[0004] The above two patent documents both relate to insurance claim settlement automation technology, but their focus is on the evidence collection of on-site data, while the data after evidence collection still needs to be manually classified. However, manual operation is inefficient, the error rate of material classification is high, the system is closed, the recognition accuracy of identification cards in complex scenarios is insufficient, and other problems, which further reduces the efficiency and accuracy of insurance claim settlement. SUMMARY

[0005] To solve the above technical problems, the application provides the following technical solutions:

[0006] In a first aspect, the application provides a method for intelligent classification and automatic uploading of insurance claim settlement survey materials, comprising:

[0007] Case data and pictures in the claim settlement system are obtained by using DOM tree analysis and file stream camouflage technology;

[0008] Multi-modal survey materials are classified by using a multi-scale feature fusion CNN combined with a spatial attention mechanism, and a dynamic threshold strategy is used to adaptively adjust the classification results;

[0009] The classified materials are automatically uploaded to the designated location of the insurance system through a concurrent control algorithm and an abnormal fuse mechanism.

[0010] In one possible implementation, the method of obtaining case data and images from the claims system using DOM tree parsing and file stream masquerading techniques includes:

[0011] Real-time analysis of the online system's webpage structure to accurately locate key fields related to cases and ensure compatibility with differences between different system versions;

[0012] Capture the dynamically updated list of selected case images on the page, automatically trigger the processing flow, and display the recognition and classification status in real time;

[0013] The categorized materials are disguised as data in a format that the native upload component can accept; data is relayed through a browser plugin backend service to bypass CORS policy restrictions.

[0014] In one possible implementation, the classification of multimodal survey materials through multi-scale feature fusion CNN combined with a spatial attention mechanism includes:

[0015] To convert a color photograph to a grayscale image, the grayscale value is calculated using a weighted average method, as shown in the following formula:

[0016]

[0017] Where R, G, and B are the red, green, and blue channel values ​​of the color image, respectively, and Gray is the grayscale value of the corresponding pixel after conversion.

[0018] Gaussian filtering is used to remove image noise and improve image clarity. The Gaussian filtering algorithm is based on the Gaussian function formula:

[0019]

[0020] in: The center coordinates of the filter are The standard deviation is used to adjust the shape of the Gaussian function and the filtering strength.

[0021] The noise-reduced data is input into the constructed multi-scale feature fusion architecture to achieve the classification of multimodal exploration materials.

[0022] In one possible implementation, the multi-scale feature fusion architecture includes: a multi-scale feature extraction module, a spatial attention module, a feature fusion layer, a scene context feature extraction module, an object detection auxiliary classification module, and a texture feature enhancement module, wherein:

[0023] The multi-scale feature extraction module identifies targets of different sizes by using parallel convolutional kernels of different sizes. , , To capture multi-level features of the input image, a three-way parallel convolution branch is mainly used, as shown in the following formula:

[0024] 1×1 convolution extracts global features;

[0025] 3×3 convolution extracts mesoscale features;

[0026] 3×3 convolution extracts local features;

[0027] Channel dimension splicing;

[0028] Batch normalization processing;

[0029] in: This represents the input feature map, with dimension 1. , The kernel size is indicated by Convolution operation, Represents output feature maps at different scales. Indicates along the channel axis The shafts are spliced ​​together;

[0030] The spatial attention module employs a dual-path attention mechanism. It generates channel weights through a channel attention mechanism, generates a spatial weight map using a spatial attention path, and finally performs feature reweighting across the two paths. The formula is as follows:

[0031]

[0032]

[0033]

[0034] in: Indicates global average pooling. Represents a fully connected layer, with dimensions respectively. , r is the compression ratio. It is the sigmoid activation function. It is the ReLU activation function. Indicating attentional characteristics, , These represent the channel attention weight matrix and the spatial attention weight matrix, respectively. This represents element-wise multiplication. This represents element-wise addition.

[0035] The feature fusion layer performs cross-scale feature fusion, concatenating and reducing the dimensionality of multi-scale features with attention-weighted features. The feature maps weighted by the attention module are concatenated along the channel dimension to form a more comprehensive high-dimensional feature vector. A 1×1 convolutional layer is then used to reduce the dimensionality of this high-dimensional feature vector. Finally, a batch normalization layer accelerates model training convergence, as shown in the following formula:

[0036]

[0037]

[0038] in: It is an intermediate representation after feature fusion. This represents the input image downsampled using bilinear interpolation. These are features ultimately used for downstream tasks. Represents the corrected linear unit ;

[0039] The scene context feature extraction module extracts global scene features through a pre-trained Vision Transformer network, encoding the category, area proportion in the image, and spatial relative position of each global scene feature element into a scene context feature vector, as shown in the following formula:

[0040]

[0041]

[0042] Where: Input image Divided into indivual patches, through linear projection Add to the learnable embedding matrix Position encoding, using L-layer normalization to obtain the [CLS] tag. As a global feature;

[0043] The target detection auxiliary classification module uses a pre-trained target detection model YOLOv5 to locate and identify predefined specific objects strongly related to insurance claims in the photo. The output detection box position, object category, and its corresponding confidence score are used as evidence features to accurately assess the damage to the vehicle, as shown in the following formula:

[0044]

[0045]

[0046] in: This represents the i-th detection result. This refers to the detection head; MLP stands for Multilayer Perceptron. The bounding box coordinates are , For class probability distribution, For confidence level, The number of detection frames to retain;

[0047] The texture feature enhancement module uses Local Binary Pattern (LBP) to extract fine texture features from the image for accident details. The specific formula is as follows:

[0048]

[0049] in: For pixels in a grayscale image, binarize The value of the neighboring pixels. Generate a 256-bit histogram feature based on the center pixel value. .

[0050] In one possible implementation, the step of adaptively adjusting the classification result using a dynamic threshold strategy includes: adaptively adjusting the classification threshold based on the distribution characteristics of the training data using a dynamic threshold strategy, wherein the dynamic threshold calculation formula is:

[0051]

[0052] in, and Categories The mean and standard deviation of the predicted probability distribution. This is an adjustable parameter.

[0053] In one possible implementation, the final classification decision stage employs a stacking-based ensemble learning strategy, using a meta-classifier to learn the prediction results from combining multiple base classifiers.

[0054] In one possible implementation, the final classification decision stage employs a stacking-based ensemble learning strategy, using a meta-classifier to learn the prediction results combining multiple base classifiers, including:

[0055] The original training dataset is randomly divided into N disjoint subsets;

[0056] Perform N rounds of iterations. In the Nth round of iterations, the Nth subset is used as the validation set, and the remaining subsets are used as the training set. On this training set, the three base classifiers ResNet50, EfficientNetB3, and ViT-L / 16 are trained in their entirety.

[0057] The three trained base classifiers are used to predict the k-th validation set to obtain the predicted probability of each sample.

[0058] After N rounds of iteration, the predicted probability vectors generated by the base classifiers for each sample in the original training set are collected to form the training feature set of the meta classifier. The labels of this feature set are consistent with the labels of the original training set.

[0059] After completing cross-validation, using all the original training data, we retrained the three base classifiers ResNet50, EfficientNetB3, and ViT-L / 16 separately. These three finally trained models were used for subsequent predictions on new data.

[0060] Use the training feature set and corresponding labels generated in the previous step to train the meta-classifier;

[0061] When a brand-new, unseen survey image needs to be classified, base classifiers ResNet50, EfficientNetB3, and ViT-L / 16, along with a meta-classifier, are used for decision-making.

[0062] In one possible implementation, when a completely new, unseen survey image needs to be classified, a decision is made using base classifiers ResNet50, EfficientNetB3, and ViT-L / 16, and a meta-classifier, including:

[0063] First, the image is input into three pre-trained base classifiers: ResNet50, EfficientNetB3, and ViT-L / 16 to obtain three predicted probabilities.

[0064] Then these three probability values ​​are combined to form a new feature vector;

[0065] Finally, this new feature vector is input into the pre-trained meta-classifier, which outputs the final classification result, as shown in the formula:

[0066]

[0067] in: For the first A base classifier evaluates samples Category The predicted probability, These are the weights of the corresponding classifier.

[0068] In one possible implementation, the automatic uploading of the classified materials to a designated location in the insurance system via a concurrency control algorithm and an exception triggering mechanism includes:

[0069] Use FTP or HTTP protocols to transfer photos to the corresponding folder in the imaging system;

[0070] Build transmission paths and storage rules based on photo classification results;

[0071] During FTP transfer, an FTP connection is first established and user authentication is performed. After successful authentication, the corresponding folder is created in the image system according to the photo storage path.

[0072] Then, the photo data is transferred to the designated folder block by block using binary transmission mode. During the transmission process, the transmission progress is monitored in real time. If the transmission is interrupted, the interruption is automatically resumed.

[0073] During HTTP transmission, an HTTP request containing photo data and metadata information is constructed and sent to the specified interface of the imaging system using the POST or PUT method.

[0074] After receiving the request, the imaging system performs data verification and storage, and returns the storage result to the data transmission layer.

[0075] The data transmission layer determines whether the photo was successfully transmitted based on the returned result. If it fails, it resends the request.

[0076] Secondly, embodiments of this application provide a system for intelligent classification and automatic uploading of insurance claim investigation materials, including:

[0077] The acquisition module is used to obtain case data and images from the claims system using DOM tree parsing and file stream spoofing techniques;

[0078] The classification module is used to classify multimodal survey materials by combining multi-scale feature fusion CNN with spatial attention mechanism, and adopts dynamic threshold strategy to adaptively adjust the classification results;

[0079] The upload module is used to automatically upload the categorized materials to the designated location in the insurance system using a concurrency control algorithm and an exception triggering mechanism.

[0080] In this embodiment, automated classification and uploading replaces manual operation, parallel processing reduces waiting time, and processing speed is improved. Multi-scale feature fusion CNN combined with spatial attention mechanism classifies multimodal survey materials, significantly improving classification accuracy compared to traditional methods. Dynamic threshold mechanism and ensemble learning strategy enable the system to adapt to changes in material characteristics caused by new fraud methods, exhibiting good scalability. Attached Figure Description

[0081] Figure 1A flowchart illustrating an intelligent classification and automatic uploading method for insurance claim investigation materials provided in this application embodiment;

[0082] Figure 2 This is a schematic diagram of the structure of the multi-scale feature extraction module provided in the embodiments of this application;

[0083] Figure 3 A schematic diagram of an intelligent classification and automatic uploading system for insurance claim investigation materials provided in this application embodiment;

[0084] Figure 4 This is a schematic diagram of the framework of an electronic device provided in an embodiment of this application. Detailed Implementation

[0085] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0086] See Figure 1 The method for intelligent classification and automatic uploading of insurance claim investigation materials provided in this application includes:

[0087] S101 uses DOM tree parsing and file stream masquerading techniques to obtain case data and images from the claims system.

[0088] In this embodiment, DOM tree parsing technology is used to parse the online system's webpage structure in real time, accurately locating key fields such as case number and user ID, while remaining compatible with differences between different system versions. The system captures the dynamically updated list of selected case images on the page, automatically triggering the processing flow and displaying the identification and classification status in real time. The classified materials are disguised into a data format acceptable to the native upload component and transferred through a browser plugin's backend service, thus bypassing CORS policy restrictions.

[0089] S102 classifies multimodal survey materials by combining multi-scale feature fusion CNN with spatial attention mechanism, and adopts dynamic threshold strategy to adaptively adjust the classification results.

[0090] Converting color photographs to grayscale images reduces the amount of image data and lowers the computational complexity of subsequent processing. The grayscale value is calculated using a weighted average method, as shown in the formula below:

[0091]

[0092] Where R, G, and B are the red, green, and blue channel values ​​of the color image, respectively, and Gray is the grayscale value of the corresponding pixel after conversion. This formula is used to convert a color image to a grayscale image.

[0093] Gaussian filtering is used to remove image noise and improve image clarity. Gaussian filtering is based on the Gaussian function, and its formula is:

[0094]

[0095] in, The center coordinates of the filter are The standard deviation is used to adjust the shape and filtering strength of the Gaussian function. Image smoothing and noise suppression are achieved by convolving the Gaussian filter with the grayscale image.

[0096] The denoised data is input into the constructed multi-scale feature fusion architecture to classify multimodal exploration materials. In this embodiment, the multi-scale feature fusion architecture used in the feature extraction and classification layers, combined with a spatial attention mechanism, significantly improves the classification accuracy of multimodal exploration materials.

[0097] The multi-scale feature fusion architecture in this embodiment includes: a multi-scale feature extraction module, a spatial attention module, a feature fusion layer, a scene context feature extraction module, an object detection auxiliary classification module, and a texture feature enhancement module.

[0098] The multi-scale feature extraction module is designed to effectively identify targets of different sizes (such as large-area dents on car bodies and small-area text on documents), such as... Figure 2 As shown, the multi-scale feature extraction module employs parallel convolutional kernels of different sizes. , , Multi-level feature capture of the input image mainly employs a three-way parallel convolution branch, expressed by the following formula:

[0099] (1×1 convolution extracts global features)

[0100] (3×3 convolution to extract mesoscale features)

[0101] (3×3 convolution to extract local features)

[0102] (Channel dimension splicing)

[0103] (Batch normalization processing)

[0104] in, This represents the input feature map, with dimension 1. , The kernel size is indicated by Convolution operation, Represents output feature maps at different scales. Indicates along the channel axis The shafts are spliced ​​together.

[0105] The spatial attention module optimizes the traditional SE module, primarily employing a dual-path attention mechanism to enhance useful features and suppress noisy or irrelevant features. Channel weights are generated through a channel attention mechanism, a spatial weight map is generated using a spatial attention path, and finally, the features are reweighted using both paths. The formula is as follows:

[0106] (Channel attention path)

[0107] (Spatial attention path)

[0108] (Feature reweighting)

[0109] in, Indicates global average pooling. Represents a fully connected layer, with dimensions respectively. , (r is the compression ratio) It is the sigmoid activation function. It is the ReLU activation function. Indicating attentional characteristics, , These represent the channel attention weight matrix and the spatial attention weight matrix, respectively. This represents element-wise multiplication. This indicates element-wise addition.

[0110] The feature fusion layer primarily performs cross-scale feature fusion, concatenating and reducing the dimensionality of multi-scale features with attention-weighted features. The feature maps weighted by the attention module are concatenated along the channel dimension to form a more comprehensive high-dimensional feature vector. To control model complexity and further refine features, a 1×1 convolutional layer is then used to reduce the dimensionality of this high-dimensional feature vector. Finally, a batch normalization layer accelerates model training convergence and improves the model's generalization ability. The final output of this layer is a highly refined feature vector used for subsequent classification decisions.

[0111] The mathematical expression for this architecture is:

[0112]

[0113]

[0114] in, It is an intermediate representation after feature fusion. This represents the input image downsampled using bilinear interpolation. These are features ultimately used for downstream tasks. Represents the corrected linear unit .

[0115] The scene context feature extraction module extracts global scene features, such as "roads," "vehicles," "buildings," and "green belts," using a pre-trained Vision Transformer network. It encodes the category of these elements, their area proportion in the image, and their spatial relative positions (e.g., "vehicles on the road") into a scene context feature vector. This vector provides a macroscopic basis for determining whether a photograph accurately reflects the accident environment. The specific formula is as follows:

[0116]

[0117]

[0118] Mainly by input image Divided into indivual patches, through linear projection Add to the learnable embedding matrix Position encoding, using L-layer normalization to obtain the [CLS] tag. As a global feature.

[0119] The object detection-assisted classification module uses a pre-trained object detection model, YOLOv5, to locate and identify predefined, specific objects strongly relevant to insurance claims in photos. These object categories can include "vehicle dents," "paint scratches," "glass cracks," and "license plate numbers," among others. The model's output—the location of the detection boxes, the object category, and its corresponding confidence score—serves as a strong set of evidentiary features for accurately assessing the vehicle's damage. The formula is as follows:

[0120]

[0121]

[0122] in, This represents the i-th detection result. This refers to the detection head; MLP stands for Multilayer Perceptron. The bounding box coordinates are , For class probability distribution, For confidence level, The number of detection boxes to retain (Top-K sorted by confidence level).

[0123] The texture feature enhancement module targets accident details (such as fine cracks, material surface wear, and deformation), features that may be lost in high-level features of conventional convolutional networks. This method employs algorithms such as Local Binary Patterns (LBP) or Histogram of Oriented Gradients (HOG) to extract fine texture features from images. For example, LBP describes local texture by comparing the grayscale relationships between pixels and their neighbors, exhibiting good robustness to illumination changes. These texture feature descriptors are a powerful complement to deep learning features. The specific formulas are as follows:

[0124]

[0125] The main operation involves calculating the pixel values ​​of the grayscale image. 8-neighborhood difference, binarization The value of the neighboring pixels. Generate a 256-bit histogram feature based on the center pixel value. .

[0126] This embodiment uses a dynamic threshold strategy instead of a fixed threshold (such as 0.5), adaptively adjusting the classification threshold based on the distribution characteristics of the training data. The dynamic threshold calculation formula is:

[0127]

[0128] in, and Categories The mean and standard deviation of the predicted probability distribution. This is an adjustable parameter (usually set to 1.0-1.5).

[0129] To achieve optimal classification accuracy and model robustness, this embodiment employs a Stacking-based ensemble learning strategy in the final classification decision stage. This strategy uses a "meta-classifier" to learn how to intelligently combine the predictions of multiple "base-classifiers," rather than using simple weighted voting. This method more effectively utilizes the differences between different models, thus achieving superior performance compared to any single model. The specific steps are as follows:

[0130] The goal of this step is to generate training data for the meta-classifier while strictly avoiding data leakage. This embodiment uses K-Fold Cross-Validation, taking K=5 as an example:

[0131] (1) Data partitioning: The original training dataset is randomly divided into 5 non-overlapping subsets (Fold 1 to Fold 5).

[0132] (2) Cross-training and prediction: Five rounds of iteration are performed. In the k-th iteration, the k-th subset is used as the validation set. The remaining four subsets are used as the training set, and the three base classifiers ResNet50, EfficientNetB3, and ViT-L / 16 are trained on this training set respectively. The three trained base classifiers are used to predict the k-th validation set to obtain the predicted probability of each sample. For example, for each image in the validation set, a vector consisting of the probabilities predicted by the three models is obtained.

[0133] (3) Constructing the meta-classifier training set: After 5 rounds of iteration, each sample in the original training set has a predicted probability vector generated by the base classifier. These vectors are then combined to form the training feature set of the meta-classifier (LightGBM). The labels of this feature set are consistent with the labels of the original training set.

[0134] (4) After completing cross-validation, use all the original training data to retrain the three base classifiers ResNet50, EfficientNetB3, and ViT-L / 16. These three finally trained models will be used for subsequent predictions on new data (test set).

[0135] The LightGBM meta-classifier is trained using the meta-feature set and corresponding labels generated in the previous step. LightGBM will learn how to make a final, more accurate judgment based on the prediction outputs of the three base classifiers. For example, it might learn that when ResNet50 and ViT both give high probabilities, the results are very reliable; while when EfficientNetB3 disagrees with the other two, the conclusions of the former two should be trusted more.

[0136] When a brand-new, unseen survey image needs to be classified, the decision-making process is as follows:

[0137] (1) First, input the image into the three final base classifiers (ResNet50, EfficientNetB3, ViT-L / 16) trained in step 4 to obtain three predicted probabilities.

[0138] (2) Then, combine these three probability values ​​to form a new feature vector.

[0139] (3) Finally, input this new feature vector into the already trained LightGBM meta-classifier, and LightGBM outputs the final classification result (e.g., "ID photo" or "on-site photo").

[0140] The formula is as follows:

[0141]

[0142] in, For the first A base classifier evaluates samples Category The predicted probability, These are the weights of the corresponding classifier.

[0143] S103 automatically uploads the classified materials to the designated location in the insurance system using a concurrent control algorithm and an abnormal circuit breaker mechanism.

[0144] Based on the classification results, the data transmission layer uses either FTP or HTTP protocols to transfer photos to the corresponding folder in the imaging system. The selection of the transmission protocol is automatically adapted according to the imaging system's support and the network environment. If the imaging system supports the FTP protocol and the network environment requires high file transfer speeds, the FTP protocol is selected first. If the imaging system only supports the HTTP protocol, or if data needs to be transferred in an environment with high network security requirements, the HTTP protocol is used, and SSL / TLS encryption is enabled.

[0145] The transmission path and storage rules are constructed based on the photo classification results. For example, if a photo is classified as an ID photo, it will be transmitted to the ID photo folder in the image system, with the storage path format being "image system root directory / ID photo / case number / photo file name"; if it is a crime scene photo, it will be transmitted to the crime scene photo folder, with the storage path format being "image system root directory / crime scene photo / case number / photo shooting time / photo file name". During transmission, the authentication, encryption, and other security mechanisms stipulated in the protocol are followed to ensure the accurate and secure transmission of photo data.

[0146] During FTP transfer, an FTP connection is first established and user authentication is performed. After successful authentication, the corresponding folder is created in the image system based on the photo storage path (if the folder does not exist). Then, using binary transfer mode, the photo data is transferred block by block to the designated folder. The transfer progress is monitored in real time, and if an interruption occurs, the transfer is automatically resumed.

[0147] During HTTP transmission, an HTTP request containing photo data and metadata (such as case number, photo category, etc.) is constructed and sent to the designated interface of the image system using the POST or PUT method. Upon receiving the request, the image system verifies and stores the data, returning the storage result to the data transmission layer. The data transmission layer determines whether the photo transmission was successful based on the returned result; if it fails, the request is resent, with a maximum of three attempts.

[0148] It should be noted that this implementation continuously optimizes the model weights through an online learning module, enabling the system to dynamically adapt to changes in new fraud methods and material characteristics.

[0149] In this embodiment, the concurrent upload count is dynamically adjusted based on system load to achieve automatic queuing optimization. When a system response timeout or error is detected, the system automatically pauses and records the breakpoint to prevent data loss. A sidebar operation panel is intelligently overlaid on the original interface without modifying the original system. The entire process uses a browser plugin as a carrier, intelligently overlaying operation functions onto the original system page to achieve end-to-end automation from data acquisition to classification and uploading, significantly improving claims processing efficiency and accuracy.

[0150] Corresponding to the intelligent classification and automatic uploading method for insurance claim investigation materials provided in the above embodiments, this application also provides an embodiment of an intelligent classification and automatic uploading system for insurance claim investigation materials.

[0151] See Figure 3 The intelligent classification and automatic uploading system 20 for insurance claim investigation materials in this embodiment includes:

[0152] Module 201 is used to obtain case data and images from the claims system using DOM tree parsing and file stream masquerading techniques.

[0153] The classification module 202 is used to classify multimodal survey materials by combining multi-scale feature fusion CNN with spatial attention mechanism, and adopts dynamic threshold strategy to adaptively adjust the classification results.

[0154] Upload module 203 is used to automatically upload the classified materials to the designated location in the insurance system through a concurrency control algorithm and an abnormal circuit breaker mechanism.

[0155] Corresponding to the above embodiments, this application also provides an electronic device.

[0156] See Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 300 may include a processor 301, a memory 302, and a communication unit 303. These components communicate via one or more buses. Those skilled in the art will understand that the electronic device structure shown in the figures does not constitute a limitation on the embodiments of this application. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0157] The communication unit 303 is used to establish a communication channel, thereby enabling the electronic device to communicate with other devices.

[0158] The processor 301 serves as the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 302, and calls data stored in the memory to perform various functions and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 301 may consist only of a central processing unit (CPU). In this embodiment, the CPU may have a single processing core or include multiple processing cores.

[0159] Memory 302 is used to store the execution instructions of processor 301. Memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0160] When the execution instructions in memory 302 are executed by processor 301, the electronic device 300 is able to perform some or all of the steps in the above method embodiments.

[0161] Corresponding to the above embodiments, this application also provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, wherein when the program runs, it can control the device where the computer-readable storage medium is located to execute some or all of the steps in the above method embodiments. In specific implementation, the computer-readable storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0162] Corresponding to the above embodiments, this application also provides a computer program product containing executable instructions that, when executed on a computer, cause the computer to perform some or all of the steps in the above method embodiments.

[0163] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0164] The above description is merely a specific embodiment of this application. Any variations 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 protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An insurance claim survey material intelligent classification and automatic uploading method, characterized in that, The application relates to a method for automatically classifying multi-modal survey materials in an insurance claim system. Case data and pictures in the claim system are obtained by using DOM tree analysis and file stream camouflage technology, including: Real-time analysis of online system webpage structure, accurate positioning of key fields related to cases and compatibility of different system version differences; Capture the list of selected case pictures updated dynamically, automatically trigger the processing flow, and display the identification and classification in real time; Camouflage the classified materials into a data format that can be received by the original upload component; Break through the CORS policy restriction by using the browser plug-in background service to transfer data; Classify multi-modal survey materials by using a multi-scale feature fusion CNN combined with a spatial attention mechanism, and adopt a dynamic threshold strategy to adaptively adjust the classification results; The multi-scale feature fusion CNN combined with the spatial attention mechanism for classifying the multi-modal survey materials comprises the following steps: Convert the color photo into a grayscale image, and calculate the grayscale value by using a weighted average method, the formula being as follows: ； Wherein R, G and B are the red, green and blue channel values of the color image respectively, and Gray is the grayscale value of the corresponding pixel point after conversion; Remove image noise by using a Gaussian filter algorithm to improve image definition, and the Gaussian filter is based on a Gaussian function formula: ； wherein: is the filter center coordinate, is the standard deviation, used to regulate the shape of the Gaussian function and the filter strength; Input the data after noise reduction into a constructed multi-scale feature fusion architecture to realize classification of the multi-modal survey materials, and the multi-scale feature fusion architecture comprises a multi-scale feature extraction module, a spatial attention module, a feature fusion layer, a scene context feature extraction module, a target detection auxiliary classification module and a texture feature enhancement module; The multi-scale feature extraction module identifies targets of different sizes, adopts different size convolution kernels in parallel , , The multi-level feature capture of the input image mainly adopts three parallel convolution branches, and the formula is as follows: 1x1 convolution extracts global features; 3x3 convolutions extract mid-scale features; 3x3 convolution extracts local features; , channel dimension concatenation; batch normalization processing; wherein: represents an input feature map, with dimension , represents a convolution operation with a kernel size of , represents an output feature map with different scales, represents concatenation along the channel axis axis; The spatial attention module adopts a double-path attention mechanism, generates channel weights by using a channel attention mechanism, generates a spatial weight map by using a spatial attention path, and finally performs double-path feature reweighting; the formula being as follows: ； ； ； wherein: represents a global average pooling, represents a fully connected layer with dimensions , , r is the compression ratio, is a sigmoid activation function, is a ReLU activation function, represents an attention feature, , respectively represent a channel attention weight matrix and a spatial attention weight matrix, represents an element-wise multiplication, represents an element-wise addition; The feature fusion layer performs cross-scale feature fusion, splices and reduces dimensions of the multi-scale features and the features weighted by the attention mechanism, splices the feature maps weighted by the attention mechanism in the channel dimension to form a more comprehensive high-dimensional feature vector, and then uses a 1x1 convolution layer to reduce the dimensions of the high-dimensional feature vector; the formula being as follows: ； ； wherein: is the intermediate representation after feature fusion, represents the input image down-sampled by bilinear interpolation, is the final feature for downstream tasks, represents a rectified linear unit ; The scene context feature extraction module extracts global scene features by using a pre-trained Vision Transformer network, encodes the categories of global scene feature elements, the area proportion in the picture and the spatial relative positions therebetween into a scene context feature vector; The target detection auxiliary classification module uses a pre-trained target detection model YOLOv5 to locate and identify specific objects related to insurance claim in the photo, and outputs the detection frame position, object category and corresponding confidence score as evidence features for accurately evaluating the damage condition of the vehicle; the formula being as follows: ； ； wherein: represents the i-th detection result, represents a detection head, and MLP represents a multi-layer perception, the bounding box coordinates are , is a class probability distribution, is a confidence, is the number of retained detection boxes; The texture feature enhancement module extracts fine texture features of the image by using a local binary pattern LBP for accident details; The classified materials are automatically uploaded to the designated position of the insurance system by using a concurrent control algorithm and an abnormal fuse mechanism.

2. The method for intelligent classification and automatic uploading of insurance claim survey materials according to claim 1, characterized in that, The dynamic threshold strategy is adopted to adaptively adjust the classification result, including: adopting a dynamic threshold strategy to adaptively adjust the classification threshold according to the distribution characteristics of the training data, and the dynamic threshold calculation formula is: ； where and are the mean and standard deviation of the predictive probability distribution of class respectively, is an adjustable parameter.

3. The method for intelligent classification and automatic uploading of insurance claim survey materials according to claim 1, characterized in that, The classified materials are automatically uploaded to the designated location of the insurance system through the concurrent control algorithm and the abnormal melting mechanism, including: Photos are transmitted to the corresponding folder of the image system using FTP or HTTP protocol; According to the photo classification result, the transmission path and storage rules are constructed; In the FTP transmission process, first, establish an FTP connection, perform user authentication, and after authentication, create the corresponding folder in the image system according to the photo storage path; Then, using binary transmission mode, the photo data is transmitted to the specified folder block by block, and in the transmission process, the transmission progress is monitored in real time, and if the transmission is interrupted, the breakpoint transmission is automatically performed; In the HTTP transmission process, an HTTP request containing photo data and metadata information is constructed, and the POST or PUT method is used to send the request to the specified interface of the image system; After the image system receives the request, it performs data verification and storage, and returns the storage result to the data transmission layer; The data transmission layer determines whether the photo is successfully transmitted according to the return result, and if it fails, it sends the request again.

4. An insurance claim survey material intelligent classification and automatic uploading system, characterized in that, Including: An acquisition module is used to obtain case data and pictures in the claim system by using DOM tree parsing and file stream camouflage technology, including: Real-time parsing of online system webpage structure, accurate positioning of key fields related to cases and compatibility of different system version differences; Capture the case picture selection list updated dynamically, automatically trigger the processing flow, and display the recognition and classification in real time; The classified materials are disguised as data formats that can be received by the original upload component; Through the browser plug-in background service, the data is transferred, breaking through the CORS strategy restriction; A classification module is used to classify multi-modal survey materials by multi-scale feature fusion CNN combined with spatial attention mechanism, and to adaptively adjust the classification result by dynamic threshold strategy; The multi-scale feature fusion CNN combined with spatial attention mechanism is used to classify multi-modal survey materials, including: The color photo is converted into a gray image, and the weighted average method is used to calculate the gray value, and the formula is as follows: ； Where: R, G, B are the red, green, and blue channel values of the color image, and Gray is the gray value of the corresponding pixel point after conversion; Gaussian filter algorithm is used to remove image noise and improve image clarity. The Gaussian filter is based on the Gaussian function formula: ； wherein: is the filter center coordinate, is the standard deviation, used to regulate the shape of the Gaussian function and the filter strength; The denoising processed data is input into the constructed multi-scale feature fusion architecture to realize the classification of multi-modal survey materials. The multi-scale feature fusion architecture includes: multi-scale feature extraction module, spatial attention module, feature fusion layer, scene context feature extraction module, target detection auxiliary classification module and texture feature enhancement module; The multi-scale feature extraction module identifies targets of different sizes, adopts different size convolution kernels in parallel , , Multi-level feature capture of the input image mainly adopts three parallel convolution branches, and the formula is as follows: 1x1 convolution extracts global features; 3x3 convolutions extract mid-scale features; 3x3 convolution extracts local features; , channel dimension concatenation; batch normalization processing; wherein: represents an input feature map, with dimensions , represents a convolution operation with a kernel size of , represents an output feature map of different scales, represents concatenation along the channel axis axis; The spatial attention module adopts a double-path attention mechanism, generates channel weights through channel attention mechanism, generates spatial weight maps using spatial attention path, and finally performs double-path feature reweighting; The formula is as follows: ； ； ； wherein: represents a global average pooling, represents a fully connected layer with dimensions , , r is the compression ratio, is a sigmoid activation function, is a ReLU activation function, represents an attention feature, , respectively represent a channel attention weight matrix and a spatial attention weight matrix, represents an element-wise multiplication, represents an element-wise addition; The feature fusion layer performs cross-scale feature fusion, splices and reduces the dimension of multi-scale features and attention-weighted features, the feature map weighted by the attention module is spliced in the channel dimension to form a more comprehensive high-dimensional feature vector, and then a 1x1 convolution layer is used to reduce the dimension of the high-dimensional feature vector, as follows: ； ； wherein: is the intermediate representation after feature fusion, represents the input image down-sampled by bilinear interpolation, is the final feature for downstream tasks, represents a rectified linear unit ; The scene context feature extraction module extracts global scene features through a pre-trained Vision Transformer network, encodes the category, area ratio in the picture and spatial relative position between global scene feature elements into a scene context feature vector; The target detection auxiliary classification module uses a pre-trained target detection model YOLOv5 to locate and identify specific objects that are strongly related to insurance claims in the picture, and outputs the detection frame position, object category and corresponding confidence score as evidence features for accurately evaluating the damage of the vehicle, as follows: ； ； wherein: denotes the i-th detection result, denotes a detection head, and MLP denotes a multi-layer perceptron, the bounding box coordinates are , is a class probability distribution, is a confidence score, is the number of retained bounding boxes; The texture feature enhancement module extracts fine texture features of the image by using a local binary pattern (LBP) for accident details. The uploading module is used for automatically uploading the classified materials to the designated position of the insurance system through a concurrent control algorithm and an abnormal fuse mechanism.

Images