A Multimodal AI-Based Image Classification Method and System for Insurance Claims

Abstract

This invention relates to a method and system for classifying insurance claim images based on multimodal AI, belonging to the field of data processing. The method includes: a target edge node acquiring a target task, which requests the classification of a set of insurance claim images related to a target user's insurance claim business; the target edge node analyzing the insurance claim image set through an agent deployed on the target edge node to obtain a task descriptor of the metadata of the insurance claim image set output by the agent; the target edge node analyzing the task descriptor of the metadata through a benefit function to obtain a comprehensive benefit value for the target task, which represents the multi-dimensional evaluation result of the target task; and the target edge node triggering, based on the comprehensive benefit value of the target task, operations at the target edge node itself, through multi-edge node collaboration, or at a central node, to classify the insurance claim image set using a multimodal AI model.

Description

Technical Field

[0001] This invention belongs to the field of data processing and relates to a method and system for classifying insurance claims images based on multimodal AI. Background Technology

[0002] As the insurance industry undergoes profound digital transformation, the automation and intelligentization of the claims process have become crucial. Currently, most insurance companies use centralized cloud servers to process various images generated during the claims process, such as accident scene photos and medical documents. This architecture transmits massive amounts of image data from branches or user terminals across various locations to a remote data center for unified AI analysis.

[0003] However, the network transmission of image data (especially high-definition photos and videos) is significantly time-consuming, directly impacting the end-to-end timeliness of claims processing. The processing complexity of claims images varies greatly, ranging from simple document recognition to complex damage assessment, resulting in extremely uneven demands on computing resources. This leads to drastic fluctuations in the load on the central server, easily creating processing bottlenecks during peak business periods. Summary of the Invention

[0004] In view of this, in order to solve the above problems, the present invention provides an insurance claims image classification method and system based on multimodal AI.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Firstly, a method for classifying insurance claims images based on multimodal AI is provided. The insurance claims image classification system includes multiple edge nodes and a central node. The method is applied to any target edge node among the multiple edge nodes, including: the target edge node acquiring a target task, which requests to classify a set of insurance claims images related to the insurance claims business of a target user; the target edge node analyzing the set of insurance claims images through an agent deployed on the target edge node to obtain a task descriptor of the metadata of the set of insurance claims images output by the agent; the target edge node analyzing the task descriptor of the metadata through a benefit function to obtain a comprehensive benefit value of the target task, which represents the multi-dimensional evaluation result of the target task; and the target edge node triggering an operation to classify the set of insurance claims images using a multimodal AI model, either locally on the target edge node, through multi-edge node collaboration, or at the central node, based on the comprehensive benefit value of the target task.

[0007] Therefore, the technical solution of this invention analyzes the insurance claim image set by setting up an intelligent agent at each edge node, generating a comprehensive "metadata task descriptor," and calculating a "comprehensive benefit value" based on this descriptor using a benefit function. This design transforms the edge node from merely a data collection or simple filtering endpoint into a rudimentary form of intelligence capable of self-evaluation and decision-making regarding tasks. Its core effect lies in the fact that the system, for the first time at the architectural level, implements a mechanism for differentiated routing decisions based on the characteristics of the task itself (quantified in the descriptor). Specifically, it creatively transforms the selection of subsequent processing nodes (local, collaborative, or centralized) from fixed rules or simple strategies (such as "upload all" or "process all locally") into a dynamic, adaptive, and intelligent triggering process. From a system methodology perspective, this provides a novel, task-self-evaluation-based, elastic scheduling approach to address the latency, privacy, and computational load imbalance issues caused by centralized processing in insurance claims, improving the overall rationality and intelligence of the system's resource utilization.

[0008] Optionally, the task descriptor of the metadata describes the following aspects of the target task through structured data or natural language: target task complexity, target data sensitivity, and target processing urgency. The target task complexity is used to characterize the complexity of the insurance claims image set containing insurance claims business data, and the target data sensitivity is used to characterize the sensitivity of the insurance claims image set containing insurance claims business data.

[0009] This demonstrates that the abstract, experience-dependent judgment criteria in insurance claims processing (such as case complexity, privacy risks, and processing priorities) have been transformed into structured, quantifiable technical parameters. This design ensures that subsequent intelligent routing decisions are no longer based on simple rules or single indicators, but rather on a deep, quantitative understanding of the business essence. It provides a crucial data interface for the deep integration of business rules and technical systems, significantly improving the rationality of system decisions.

[0010] Optionally, the intelligent agent includes an image structure analysis channel, a text information extraction channel, a metadata reading channel, and a feature fusion and quantization unit. The target edge node analyzes the insurance claim image set through the intelligent agent deployed on the target edge node to obtain the task descriptor of the metadata of the insurance claim image set output by the intelligent agent. This includes: the target edge node calling the image structure analysis channel to extract visual features from the insurance claim image set and evaluate the complexity based on visual features to obtain the initial task complexity; the target edge node calling the text information extraction channel to extract and analyze structured and unstructured text from the insurance claim image set to obtain the initial data sensitivity; the target edge node calling the metadata reading channel to parse the file attribute information of the insurance claim image set to obtain the initial processing urgency; and the target edge node calling the feature fusion and quantization unit to perform weighted fusion and normalization calculations on the initial task complexity, initial data sensitivity, and initial processing urgency according to the insurance business rule mapping table to obtain the task descriptor of the metadata.

[0011] Therefore, by referencing the comprehensive cognitive methods humans use to process claims materials (parallel analysis of visual, textual, and metadata data), multi-level, fine-grained perception of image information is achieved. In particular, the feature fusion and quantization unit performs weighted and normalized operations based on an insurance business rule mapping table. This is not a simple feature concatenation, but rather encodes domain knowledge (such as the sensitive weights of specific injury terms and the urgency coefficients of case types) as part of the fusion algorithm. This design solves the adaptability problem of general image analysis models in the insurance field, generating highly adaptable and accurately quantified descriptors for subsequent routing decisions.

[0012] Optionally, the target edge node analyzes the task descriptor of the metadata through a benefit function to obtain the comprehensive benefit value of the target task, including: the target edge node analyzes the task descriptor of the metadata through a benefit function, the estimated latency of the central node, the load of the target edge node, and the collaboration potential information of multiple edge nodes to obtain the comprehensive benefit value of the target task.

[0013] Optionally, the benefit function, the task descriptor of the metadata, the estimated latency of the central node, the load of the target edge node, the collaboration potential information of multiple edge nodes, and the comprehensive benefit value of the target task satisfy the following relationship:

[0014] Benefit_Score = U_weight / [ω1*C_score*Delay_center+ω2*S_level_cost+ω3*Edge_Load+ω4*Collaboration_Potential]; where Benefit_Score represents the overall benefit value of the target task, U_weight represents the urgency of the target processing, C_score represents the complexity of the target task, S_level_cost represents the sensitivity of the target data, Delay_center represents the estimated delay of the central node, Edge_Load represents the load of the target edge nodes, Collaboration_Potential represents the collaboration potential information of multiple edge nodes, and ω1, ω2, ω3, and ω4 are preset weights.

[0015] Therefore, by designing the above formula, the urgency of processing is placed in the numerator as a benefit, while complexity, sensitivity costs, load, and collaboration potential are placed in the denominator as costs. This unifies the three mutually constraining goals of "privacy, real-time performance, and computational optimization" into a computable optimization framework. Each term in the formula has a clear physical and business meaning, and the strategy is adjusted through weighting coefficients (ω1-ω4). This design enables the system to quantitatively compare the expected benefits of different routing strategies within milliseconds, improving the accuracy and rationality of decisions regarding "triggering local, collaborative, or centralized processing."

[0016] Optionally, the collaboration potential information of multiple edge nodes satisfies the following relationship:

[0017] Collaboration_Potential=(ω41*Q_net)*(ω42*M_comp)*(ω43*F_task)*(ω44*T_trust);

[0018] Wherein, Q_net is the network quality factor, which characterizes the average latency and packet loss rate between the target edge node and a potential edge node for collaboration; M_comp is the computing power matching factor, which characterizes the available computing power of the potential edge node; F_task is the task matching factor, which characterizes the degree of matching between the tasks that the potential edge node can handle and the target task; T_trust is the trust and cost factor, which characterizes the trust level of the potential edge node; among multiple edge nodes, the edge node that collaborates with the target edge node is the N edge node with the highest value of collaboration potential information, where N is an integer greater than or equal to 1; ω41, ω42, ω43, ​​and ω44 are preset weights.

[0019] Therefore, it can be seen that by systematically integrating four dimensions—network quality, aggregated computing power, task-node semantic matching degree, and trust cost—into a comprehensive potential value in edge computing task routing, and especially by introducing the similarity calculation (F_task) between "node capability profile" and "task description vector," semantic-based intelligent service discovery and matching are achieved, rather than just load balancing. This formula transforms the originally vague judgment of "whether collaboration is possible" into a refined and comparable numerical value, enabling the advanced strategy of "multi-edge node collaboration" to move from a concept to a stable and efficient technical implementation, significantly improving the success rate and efficiency of complex task processing.

[0020] Optionally, based on the comprehensive benefit value of the target task, the target edge node triggers an operation to classify the insurance claim image set using a multimodal AI model, either locally at the target edge node, through multi-edge node collaboration, or at the central node. This includes: if the comprehensive benefit value of the target task is greater than or equal to a first threshold, the target edge node classifies the insurance claim image set using its local lightweight multimodal AI model to obtain a classified insurance claim image set; if the comprehensive benefit value of the target task is less than the first threshold but greater than or equal to a second threshold, the target edge node classifies the insurance claim image set using a multimodal AI model through multi-edge node collaboration to obtain a classified insurance claim image set; if the comprehensive benefit value of the target task is less than the second threshold but greater than or equal to a third threshold, the target edge node extracts a set of strong semantic and non-privacy feature vectors from the insurance claim image set and sends them to the central node. The target edge node receives a set of feature vectors from the central node, which is analyzed by a multimodal AI model to determine the category of the insurance claim image set. If the overall benefit value of the target task is less than a second threshold but greater than or equal to a third threshold, the target edge node extracts a set of strong semantic and non-privacy feature vectors from the insurance claim image set and sends the set of feature vectors to the central node. It then receives the set of feature vectors from the central node, which is analyzed by a multimodal AI model to determine the category of the insurance claim image set. If the overall benefit value of the target task is less than the third threshold, the target edge node encrypts the insurance claim image set to obtain an encrypted insurance claim image set and sends it to the central node. The central node then receives the encrypted insurance claim image set, which is analyzed by a multimodal AI model to determine the category of the insurance claim image set.

[0021] This demonstrates that the system constructs a continuous and smooth task processing hierarchy, transitioning from "completely edge processing" to "completely centralized processing." Each level of strategy corresponds to an optimal cost-benefit balance: local final review pursues maximum efficiency; federated collaboration utilizes distributed computing power; feature uploading balances privacy and accuracy; and encrypted uploading ensures the processing of complex cases. This layered and progressive routing mechanism enables the system to precisely guide tasks to the most suitable computing resource layer based on their real-time "benefit pressure," achieving global dynamic optimization of processing accuracy, response speed, privacy security, and resource consumption.

[0022] Optionally, if the load on the central node is high, the target edge node lowers the first and second thresholds; if the current analysis mode is anti-fraud mode, the target edge node lowers the third threshold.

[0023] This demonstrates that the system's routing strategy possesses business environment awareness and adaptive capabilities. By dynamically lowering thresholds based on central load (causing more tasks to remain at the edge) or based on anti-fraud mode requirements (causing more tasks to be uploaded to the central for in-depth analysis), the system can automatically adapt to different scenarios such as business peaks and special risk control. This achieves online and flexible linkage between technical scheduling logic and upper-layer business strategies, upgrading the entire system from a static automation tool to an intelligent and elastic network capable of responding to changes in business conditions, significantly improving the system's practicality and robustness.

[0024] Optionally, the method further includes sending the target edge node to the classified insurance claim image set to the business system.

[0025] Secondly, an insurance claims image classification system is provided. This system includes multiple edge nodes and a central node. Any one of the edge nodes is configured as follows: the target edge node acquires a target task, which requests the classification of a set of insurance claims images related to a target user's insurance claims business; the target edge node analyzes the insurance claims image set through an agent deployed on the target edge node, obtaining a task descriptor of the metadata of the insurance claims image set output by the agent; the target edge node analyzes the task descriptor of the metadata through a benefit function, obtaining a comprehensive benefit value for the target task, which represents the multi-dimensional evaluation result of the target task; based on the comprehensive benefit value of the target task, the target edge node triggers operations at the target edge node itself, through multi-edge node collaboration, or at the central node, to classify the insurance claims image set using a multimodal AI model.

[0026] The objectives and other advantages of this invention can be realized and obtained through the following description. Attached Figure Description

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0028] Figure 1 This is a schematic diagram of the architecture of an insurance claims image classification system provided by the present invention;

[0029] Figure 2 A flowchart illustrating an insurance claims image classification method based on multimodal AI provided by this invention;

[0030] Figure 3 A scene diagram illustrating an insurance claims image classification method based on multimodal AI provided by this invention;

[0031] Figure 4 This is a schematic diagram of the structure of a processing device provided by the present invention. Detailed Implementation

[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. The accompanying drawings are for illustrative purposes only, representing only schematic diagrams and not actual physical objects, and should not be construed as limiting the present invention. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent actual products.

[0033] This application provides an insurance claim image classification system, which includes multiple edge nodes and a center node.

[0034] Any of the multiple edge nodes and central nodes can be in the form of a terminal, or a chip or chip system that can be set in that terminal. This terminal can also be referred to as a user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent, or user device. In the embodiments of this application, the terminal device can be a mobile phone, tablet computer, computer with wireless transceiver capabilities, virtual reality (VR) terminal device, augmented reality (AR) terminal device, wireless terminal in industrial control, wireless terminal in self-driving, wireless terminal in remote medical care, wireless terminal in smart grid, wireless terminal in transportation safety, wireless terminal in smart city, wireless terminal in smart home, vehicle-mounted terminal, RSU with terminal functionality, etc. The terminal device of this application may also be an on-board module, on-board component, on-board chip, or on-board unit that is built into a vehicle as one or more components or units. The vehicle can implement the method provided in this application through the built-in on-board module, on-board component, on-board chip, or on-board unit.

[0035] Multiple edge nodes can be geographically distributed edge devices, while the central node is located in the data center.

[0036] In this communication system, taking any one of the multiple edge nodes as an example, the target edge node obtains the target task, which requests the classification of a set of insurance claim images related to the target user's insurance claim business. The target edge node analyzes the set of insurance claim images through an agent deployed on the target edge node, obtaining a task descriptor of the metadata of the set of insurance claim images output by the agent. The target edge node analyzes the task descriptor of the metadata through a benefit function, obtaining a comprehensive benefit value of the target task, which represents the multi-dimensional evaluation result of the target task. Based on the comprehensive benefit value of the target task, the target edge node triggers the operation of classifying the set of insurance claim images using a multimodal AI model, either locally on the target edge node, through multi-edge node collaboration, or at the central node. The method is described in detail below.

[0037] like Figure 2 As shown, an insurance claim image classification method based on multimodal AI is provided. This method is applied to any one of the multiple edge nodes mentioned above. The specific process of this method is as follows:

[0038] S201, the target edge node obtains the target task, and the target task requests to classify the set of insurance claim images of the target user's insurance claim business.

[0039] For example, when an insured user reports an accident through the insurance company's mobile application and uploads photos of the accident scene and details of the damage, these images are dispatched via the cloud to a corresponding edge node (such as the branch office server that processes claims in that area); or, when an insurance surveyor takes photos at the accident scene using a dedicated terminal, that terminal acts as an edge node to directly receive and generate processing tasks; or, when third-party institutions such as hospitals and repair shops that cooperate with the insurance company submit medical images or repair quotations related to claims in batches through data interfaces, the gateway server connecting to these institutions, acting as an edge node, will also receive the corresponding processing tasks.

[0040] The objective task explicitly includes a request to classify and process a "set of insurance claim images" related to a specific claim case for a "target user." This image set may contain multiple images, such as photos of vehicle damage, property damage, medical expense lists, medical diagnostic certificates, and photos of identification documents.

[0041] S202, the target edge node analyzes the insurance claim image set through the agent deployed on the target edge node, and obtains the task descriptor of the metadata of the insurance claim image set output by the agent.

[0042] Combination Figure 3 As shown, the task descriptor of the metadata describes the target task using structured data or natural language, including: target task complexity, target data sensitivity, and target processing urgency. Target task complexity characterizes the complexity of the insurance claims image set containing insurance claims business data, and target data sensitivity characterizes the sensitivity of the insurance claims image set to the insurance claims business data. This transforms abstract, experience-dependent judgment criteria in insurance claims business (such as case complexity, privacy risks, and processing priorities) into structured, quantifiable technical parameters. This design ensures that subsequent intelligent routing decisions are no longer based on simple rules or single indicators, but rather on a deep quantitative understanding of the business essence. It provides a key data interface for the deep integration of business rules and technical systems, significantly improving the rationality of system decisions.

[0043] For example, the intelligent agent includes an image structure analysis channel, a text information extraction channel, a metadata reading channel, and a feature fusion and quantization unit. Based on this, the target edge node calls the image structure analysis channel to perform visual feature extraction and complexity evaluation based on the visual features of the insurance claim image set, obtaining an initial task complexity score. In one possible implementation, the image structure analysis channel first extracts visual features from the insurance claim image set, using a lightweight convolutional neural network model to analyze the texture, layout, and object composition of the images, such as determining whether the image is a regular table or a photograph containing multiple pages, distortion, or a complex scene, and outputs a quantified initial task complexity score accordingly. The target edge node then calls the text information extraction channel to extract and analyze structured and unstructured text from the insurance claim image set, obtaining an initial data sensitivity score. In one possible implementation, the text information extraction channel processes the image using an optical character recognition engine to extract all text information and then uses natural language processing technology to parse the text: on the one hand, it identifies structured fields such as "amount," "date," and "diagnosis result," and on the other hand, it analyzes unstructured descriptive text. This channel has a built-in sensitive keyword rule base and context analysis model. By matching keywords (such as specific disease names or high-value figures) and evaluating their contextual information, it calculates the initial data sensitivity level. The target edge node calls the metadata reading channel to parse the file attribute information of the insurance claim image set to obtain the initial processing urgency. In another possible implementation, the metadata reading channel works in parallel. It parses the system attributes of the image files (such as file size, creation time, and format) and metadata embedded in the file header or obtained through associated business data (such as the image source device, uploaded geographical location, associated insurance type code, and report timestamp). Based on preset timeliness rules (e.g., personal injury cases reported at night may have higher urgency), it generates an initial processing urgency index. Finally, the target edge node invokes the feature fusion and quantization unit to perform weighted fusion and normalization calculations on the initial task complexity, initial data sensitivity, and initial processing urgency according to the insurance business rule mapping table, obtaining a task descriptor for the metadata. In one possible implementation, the initial outputs of all these channels are fed into the feature fusion and quantization unit. The core of this unit is an insurance business rule mapping table, which defines the weight coefficients and normalization calculation rules for each dimension (complexity, sensitivity, urgency) under different business scenarios. For example, for "auto insurance personal injury" cases, this mapping table may assign a higher weight to "data sensitivity." Based on the specific claim case type, the unit invokes the corresponding rule set to perform weighted fusion and normalization calculations on the three initial indicators, ultimately outputting a compact, structured metadata task descriptor containing multi-dimensional quantified values, providing accurate and unified input for subsequent routing decisions.

[0044] Therefore, by referencing the comprehensive cognitive methods humans use to process claims materials (parallel analysis of visual, textual, and metadata data), multi-level, fine-grained perception of image information is achieved. In particular, the feature fusion and quantization unit performs weighted and normalized operations based on an insurance business rule mapping table. This is not a simple feature concatenation, but rather encodes domain knowledge (such as the sensitive weights of specific injury terms and the urgency coefficients of case types) as part of the fusion algorithm. This design solves the adaptability problem of general image analysis models in the insurance field, generating highly adaptable and accurately quantified descriptors for subsequent routing decisions.

[0045] S203, the target edge node analyzes the task descriptor of the metadata through the benefit function to obtain the comprehensive benefit value of the target task. The comprehensive benefit value represents the multi-dimensional evaluation result of the target task.

[0046] The target edge node obtains the comprehensive benefit value of the target task by analyzing the task descriptor of the metadata, the estimated latency of the central node, the load of the target edge node, and the collaboration potential information of multiple edge nodes through the benefit function analysis.

[0047] For example, the benefit function, the task descriptor of metadata, the estimated latency of the central node, the load of the target edge node, the collaboration potential information of multiple edge nodes, and the comprehensive benefit value of the target task satisfy the following relationship:

[0048] Benefit_Score = U_weight / [ω1*C_score*Delay_center + ω2*S_level_cost + ω3*Edge_Load + ω4*Collaboration_Potential]; where Benefit_Score represents the overall benefit value of the target task, U_weight represents the urgency of the target processing, C_score represents the complexity of the target task, S_level_cost represents the sensitivity of the target data, Delay_center represents the estimated latency of the central node, which is the average processing latency of a single task estimated in real time by the central node based on the current queue length and computing power, Edge_Load represents the load of the target edge nodes, such as the overall load rate of the target edge nodes' current CPU, memory, and other resources. Collaboration_Potential represents the collaboration potential information of multiple edge nodes, as detailed below. ω1, ω2, ω3, and ω4 are preset weights.

[0049] In other words, the four costs in the denominator of the above formula correspond to constraints in different dimensions.

[0050] The first cost, ω1*C_score*Delay_center, is the complexity-driven central processing cost. The product of C_score and Delay_center means that the more complex the task, the longer the expected processing time at the central node, and the higher its "cost," thus incentivizing the system to avoid directly uploading it to the center. The second cost, ω2*S_level_cost, is the data privacy compliance cost. Target data sensitivity is a function strongly correlated with data sensitivity; the higher the sensitivity, the higher the value, increasing non-linearly. This cost directly reflects the mandatory constraints on privacy protection. When sensitivity is extremely high, this cost dominates the denominator, forcing decisions towards local or anonymized processing to ensure that original sensitive data does not easily leave the country. The third cost, ω3*Edge_Load, is the local resource consumption cost. Edge_Load is the combined load rate of the target edge node's current CPU, memory, and other resources. This is used to achieve load balancing among edge nodes, preventing a single node from overloading due to handling too many tasks. The fourth cost, ω4*Collaboration_Potential, represents the potential costs and risks of distributed collaboration. Although named "Potential," a higher value in the denominator indicates better available collaborative resources and lower collaborative risks, thus lowering the overall cost of collaboration. It quantifies the advantage of using surrounding nodes to form a temporary federation for computation, and its calculation comprehensively considers network conditions, computing power matching, task suitability, and trust level.

[0051] The weighting coefficients ω1 to ω4 in the formula are not fixed, but dynamically adjusted by the central node based on the global strategy. For example, during periods when the company emphasizes data security, ω2 can be increased to amplify the impact of privacy costs; during peak business periods, to ensure overall throughput, ω3 can be increased to more actively distribute edge load. This design allows the routing strategy to be closely linked with upper-layer business objectives.

[0052] Therefore, by designing the above formula, the urgency of processing is placed in the numerator as a benefit, while complexity, sensitivity costs, load, and collaboration potential are placed in the denominator as costs. This unifies the three mutually constraining goals of "privacy, real-time performance, and computational optimization" into a computable optimization framework. Each term in the formula has a clear physical and business meaning, and the strategy is adjusted through weighting coefficients (ω1-ω4). This design enables the system to quantitatively compare the expected benefits of different routing strategies within milliseconds, improving the accuracy and rationality of decisions regarding "triggering local, collaborative, or centralized processing."

[0053] Optionally, the collaboration potential information of multiple edge nodes satisfies the following relationship:

[0054] Collaboration_Potential=(ω41*Q_net)*(ω42*M_comp)*(ω43*F_task)*(ω44*T_trust);

[0055] In this context, ω41, ω42, ω43, ​​and ω44 are preset weights, and Q_net is the network quality factor. The network quality factor is used to characterize the average latency and packet loss rate between the target edge node and a potential edge node used for collaboration. For example, the network quality factor is usually quantified as Q_net = exp(-λ*Avg_Latency)*(1-Packet_Loss_Rate), where Avg_Latency is the average round-trip latency between the target node and the potential edge node, Packet_Loss_Rate is the packet loss rate, and λ is an adjustment coefficient. This function shows that the impact of latency on network quality is non-linear: when the latency is low, its decay effect on Q_net is small; once the latency exceeds a certain threshold, Q_net will decrease exponentially. This accurately reflects the sensitivity of real-time collaborative tasks to network latency, so as to effectively filter out edge nodes that have strong computing power but unstable network connections. M_comp is the computing power matching factor, which is used to characterize the available computing power of potential edge nodes. The computing power matching factor satisfies M_comp=log(1+Σ(Resource_Weight_i*Available_Resource_i)), where i represents the i-th type of available computing power (hereinafter referred to as resource). Resource Available_Resource_i includes not only general CPU and memory, but may also cover GPU or NPU computing power that is crucial to AI tasks. Resource_Weight_i is a weighting coefficient set for different resource types (for example, the weight of GPU is much higher than that of CPU). By taking the logarithm, this design can significantly distinguish the strength of computing power, and also prevent the disproportionate excessive influence of a certain resource value on the overall evaluation, thus ensuring the balance of the evaluation. F_task is the task fit factor, which characterizes the degree of fit between the tasks that potential edge nodes can handle and the target task. For example, the task fit factor satisfies F_task=Cosine_Similarity(V_task,V_node), where V_task is a task semantic vector generated by an embedding model from the target task descriptor (C_score, S_level, U_weight, etc.); V_node is a "capability profile" vector formed by potential collaborative nodes in historical tasks. This profile is learned through the metadata of the node's historical tasks (such as the types of images it is good at handling, the types of risks it often handles, and the accuracy indicators it has achieved). By calculating the cosine similarity between the two, the system can intelligently discover those nodes that have rich experience in handling similar or similar complexity and sensitivity tasks, thereby greatly improving the success rate and processing quality of collaboration.In addition, T_trust is a trust and cost factor, which can be used to characterize the trust level of potential edge nodes. For example, the trust and cost factors satisfy the following relationship: T_trust=(Trust_Level / Max_Trust_Level)*(1-Cost_Ratio). Trust_Level is dynamically evaluated based on the node's security authentication and the reliability of its historical collaborative behavior (such as task completion rate and result authenticity). Cost_Ratio may involve the internal shared costs or external settlement fees that need to be paid to use the node's computing power.

[0056] Based on this, the edge nodes that collaborate with the target edge node in a multi-edge node network are the N edge nodes with the highest values ​​of collaboration potential information, where N is an integer greater than or equal to 1.

[0057] Therefore, it can be seen that by systematically integrating four dimensions—network quality, aggregated computing power, task-node semantic matching degree, and trust cost—into a comprehensive potential value in edge computing task routing, and especially by introducing the similarity calculation (F_task) between "node capability profile" and "task description vector," semantic-based intelligent service discovery and matching are achieved, rather than just load balancing. This formula transforms the originally vague judgment of "whether collaboration is possible" into a refined and comparable numerical value, enabling the advanced strategy of "multi-edge node collaboration" to move from a concept to a stable and efficient technical implementation, significantly improving the success rate and efficiency of complex task processing.

[0058] S204, based on the comprehensive benefit value of the target task, the target edge node triggers an operation to classify the insurance claim image set through a multimodal AI model, either locally at the target edge node, through multi-edge node collaboration, or at the central node.

[0059] Combination Figure 3 As shown, if the comprehensive benefit value of the target task is greater than or equal to the first threshold (denoted as θ1), optionally, it also includes the operation of classifying the insurance claim image set through the local lightweight multimodal AI model of the target edge node if the task is simple (low C_score), non-sensitive (low S_level), and the target edge node has a light load, so that the target edge node completes the final classification and review by the local lightweight model.

[0060] Combination Figure 3As shown, if the comprehensive benefit value of the target task is less than the first threshold but greater than or equal to the second threshold (denoted as θ2), and optionally also includes tasks that are complex (high in C_score) but decomposable (e.g., a material containing multiple sub-images) and have high real-time requirements, then the target edge node, through multi-edge node collaboration, performs classification processing on the insurance claim image set using a multimodal AI model to obtain the classified insurance claim image set. For example, the target edge node splits the insurance claim image set into multiple subsets, and each edge node in the multi-edge node performs classification processing on one subset using its own local lightweight multimodal AI model, and then returns the result to the target edge node. In addition, the multi-edge node includes the target edge node, and the target edge node processes its own subset. In this way, the task is decomposed by the dominant node (i.e., the target edge node), and then summarized by multiple edge nodes after federated inference.

[0061] Combination Figure 3 As shown, if the comprehensive benefit value of the target task is less than the second threshold and greater than or equal to the third threshold (denoted as θ3), optionally, it also includes task sensitivity (high S_level) or requires the central deep model capability, but is not extremely complex. Then, the target edge node extracts a set of strong semantic and non-privacy feature vectors from the insurance claim image set and sends the set of feature vectors to the central node to receive the set of classified feature vectors obtained by the central node through multimodal AI model analysis. For example, the target edge node uses a pre-built strong semantic feature extraction model (usually a pruned deep neural network encoder) to perform forward computation on the insurance claim image set. This ensures that the feature vectors output from the last layer highly represent the business semantic information in the images that is crucial for claim classification (e.g., features corresponding to "damage level" and "part type" in vehicle damage images; features corresponding to "expense items" and "medical institution level" in medical documents). Simultaneously, through model design (such as information bottleneck constraints) and post-processing techniques (such as perturbing or zeroing out specific dimensions of the feature vectors), it actively filters or blurs information directly related to personal identity and sensitive privacy (such as feature activation corresponding to textual or visual elements like names, ID numbers, and detailed addresses), thereby generating a set of "strong semantic and non-privacy feature vectors." Subsequently, the target edge node sends this set of feature vectors, rather than the original image data, to the central node via a secure channel. Upon receiving these feature vectors, the central node invokes its deployed, larger, and more precise multimodal AI analysis model. Because the central model possesses stronger understanding capabilities and a richer knowledge base, it can perform in-depth analysis and classification of these anonymized but business-semantic feature vectors, and return the results to the target edge nodes. Thus, the target edge nodes determine the set of insurance claim images to be classified based on the set of classified feature vectors.

[0062] Combination Figure 3 As shown, if the comprehensive benefit value of the target task is less than the third threshold, it may optionally include: the task is extremely complex and novel (C_score is extremely high), then the target edge node encrypts the insurance claim image set, obtains the encrypted insurance claim image set, sends the encrypted insurance claim image set to the central node, and receives the classified insurance claim image set obtained by the central node through multimodal AI model analysis of the encrypted insurance claim image set.

[0063] This demonstrates that the system constructs a continuous and smooth task processing hierarchy, transitioning from "completely edge processing" to "completely centralized processing." Each level of strategy corresponds to an optimal cost-benefit balance: local final review pursues maximum efficiency; federated collaboration utilizes distributed computing power; feature uploading balances privacy and accuracy; and encrypted uploading ensures the processing of complex cases. This layered and progressive routing mechanism enables the system to precisely guide tasks to the most suitable computing resource layer based on their real-time "benefit pressure," achieving global dynamic optimization of processing accuracy, response speed, privacy security, and resource consumption.

[0064] Optionally, if the central node experiences high load, the target edge node lowers the first and second thresholds; if the current analysis mode is anti-fraud mode, the target edge node lowers the third threshold. This demonstrates that the system's routing strategy possesses business environment awareness and adaptive capabilities. By dynamically lowering thresholds based on central load (causing more tasks to remain at the edge) or based on anti-fraud mode requirements (causing more tasks to be uploaded to the central for in-depth analysis), the system can automatically adapt to different scenarios such as peak business periods and specialized risk control. This achieves online and flexible linkage between technical scheduling logic and upper-layer business strategies, upgrading the entire system from a static automation tool to an intelligent, elastic network capable of responding to changes in business conditions, significantly improving the system's practicality and robustness.

[0065] Optionally, the method further includes: the target edge node sending the classified insurance claim image set to the business system, so that the business system asynchronously updates the global knowledge base.

[0066] In summary, the technical solution of this invention analyzes the insurance claim image set by setting up an intelligent agent at each edge node, generating a comprehensive "metadata task descriptor," and calculating a "comprehensive benefit value" based on this descriptor using a benefit function. This design transforms the edge node from merely a data collection or simple filtering endpoint into a rudimentary form of intelligence capable of self-evaluation and decision-making regarding tasks. Its core effect lies in the fact that the system, for the first time at the architectural level, implements a mechanism for differentiated routing decisions based on the characteristics of the task itself (quantified in the descriptor). Specifically, it creatively transforms the selection of subsequent processing nodes (local, collaborative, or centralized) from fixed rules or simple strategies (such as "upload all" or "process all locally") into a dynamic, adaptive, and intelligent triggering process. From a system methodology perspective, this provides a novel, task-self-evaluation-based, elastic scheduling approach to address the latency, privacy, and computational load imbalance issues caused by centralized processing in insurance claims, improving the overall rationality and intelligence of the system's resource utilization.

[0067] Figure 4 This is a schematic diagram of the structure of a processing device provided in an embodiment of this application. Exemplarily, the processing device may be a terminal, or a chip (system) or other component or assembly that can be disposed on the terminal. Figure 4 As shown, the processing device 200 may include a processor 201. Optionally, the processing device 200 may also include a memory 202 and / or a transceiver 203. The processor 201 is coupled to the memory 202 and the transceiver 203, for example, via a communication bus.

[0068] The following is combined with Figure 4 A detailed description of each component of the processing equipment 200 is provided below:

[0069] The processor 201 is the control center of the processing device 200. It can be a single processor or a collective term for multiple processing elements. For example, the processor 201 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement the embodiments of this application, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0070] Optionally, the processor 201 can perform various functions of the processing device 200 by running or executing software programs stored in the memory 202 and calling data stored in the memory 202, such as performing the above-mentioned functions. Figure 4 The method shown.

[0071] In a specific implementation, as one example, the processor 201 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0072] In a specific implementation, as one example, the processing device 200 may also include multiple processors, for example... Figure 4 The processor 201 shown is an example. Each of the processors 201 can be a single-core processor or a multi-core processor. Here, "processor" can refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0073] The memory 202 is used to store the software program that executes the solution of this application, and is controlled by the processor 201 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0074] Optionally, the memory 202 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 202 may be integrated with the processor 201 or exist independently, and may be connected via the interface circuitry of the processing device 200. Figure 4 (Not shown in the image) is coupled to processor 201, but this embodiment does not specifically limit this.

[0075] Transceiver 203 is used for communication with other processing devices. For example, if processing device 200 is a terminal, transceiver 203 can be used to communicate with a network device or with another terminal device. As another example, if processing device 200 is a network device, transceiver 203 can be used to communicate with a terminal or with another network device.

[0076] Optionally, transceiver 203 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0077] Optionally, the transceiver 203 can be integrated with the processor 201, or it can exist independently and be connected via the interface circuit of the processing device 200. Figure 4 (Not shown in the image) is coupled to processor 201, but this embodiment does not specifically limit this.

[0078] Understandable, Figure 4 The structure of the processing device 200 shown does not constitute a limitation on the processing device. Actual processing devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0079] Furthermore, the technical effects of the processing device 200 can be referred to the technical effects of the method described in the above method embodiments, and will not be repeated here.

[0080] It should be understood that the processor in the embodiments of this application can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0081] It should also be understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0083] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0084] In this application, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0085] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0087] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0089] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0091] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. 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 scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for classifying insurance claim images based on multimodal AI, characterized in that, The insurance claims image classification system includes multiple edge nodes and a center node. The method is applied to any one of the multiple edge nodes as a target edge node, including: The target edge node acquires the target task, and the target task requests to classify the set of insurance claim images for the target user's insurance claim business. The target edge node analyzes the insurance claim image set through an agent deployed on the target edge node, and obtains a task descriptor of the metadata of the insurance claim image set output by the agent; The target edge node analyzes the task descriptor of the metadata through a benefit function to obtain the comprehensive benefit value of the target task. The comprehensive benefit value represents the multi-dimensional evaluation result of the target task. The target edge node, based on the comprehensive benefit value of the target task, triggers an operation to classify the insurance claim image set using a multimodal AI model, either locally at the target edge node, through multi-edge node collaboration, or at the central node. The task descriptor of the metadata describes the target task using structured data or natural language, including the following: target task complexity, target data sensitivity, and target processing urgency. The target task complexity characterizes the complexity of the insurance claims image set containing insurance claims business data, and the target data sensitivity characterizes the sensitivity of the insurance claims image set to containing insurance claims business data. The intelligent agent includes an image structure analysis channel, a text information extraction channel, a metadata reading channel, and a feature fusion and quantization unit. The target edge node analyzes the insurance claim image set through the intelligent agent deployed on the target edge node, and obtains a task descriptor of the metadata of the insurance claim image set output by the intelligent agent, including: The target edge node calls the image structure analysis channel to extract visual features from the insurance claim image set and evaluate the complexity based on the visual features to obtain the initial task complexity. The target edge node calls the text information extraction channel to extract and analyze the structured and unstructured text of the insurance claim image set, and obtains the initial data sensitivity. The target edge node calls the metadata reading channel to parse the file attribute information of the insurance claim image set and obtain the initial processing urgency. The target edge node calls the feature fusion and quantization unit to perform weighted fusion and normalization calculations on the initial task complexity, the initial data sensitivity, and the initial processing urgency according to the insurance business rule mapping table, and obtains the task descriptor of the metadata.

2. The method according to claim 1, characterized in that, The target edge node analyzes the task descriptor of the metadata using a benefit function to obtain the comprehensive benefit value of the target task, including: The target edge node analyzes the task descriptor of the metadata, the estimated latency of the central node, the load of the target edge node, and the collaboration potential information of multiple edge nodes through a benefit function to obtain the comprehensive benefit value of the target task.

3. The method according to claim 2, characterized in that, The benefit function, the task descriptor of the metadata, the estimated latency of the central node, the load of the target edge node, the collaboration potential information of the multiple edge nodes, and the comprehensive benefit value of the target task satisfy the following relationship: Benefit_Score = U_weight / [ω1*C_score*Delay_center + ω2*S_level_cost + ω3*Edge_Load + ω4*Collaboration_Potential]; where Benefit_Score represents the overall benefit value of the target task, U_weight represents the urgency of the target processing, C_score represents the complexity of the target task, S_level_cost represents the data sensitivity of the target, Delay_center represents the estimated delay of the center node, Edge_Load represents the load of the target edge node, Collaboration_Potential represents the collaboration potential information of the multiple edge nodes, and ω1, ω2, ω3 and ω4 are preset weights.

4. The method according to claim 3, characterized in that, The collaboration potential information of the multi-edge nodes satisfies the following relationship: Collaboration_Potential=(ω41*Q_net)*(ω42*M_comp)*(ω43*F_task)*(ω44*T_trust); Wherein, Q_net is the network quality factor, which characterizes the average latency and packet loss rate between the target edge node and a potential edge node for collaboration; M_comp is the computing power matching factor, which characterizes the available computing power of the potential edge node; F_task is the task matching factor, which characterizes the degree of matching between the task that the potential edge node can handle and the target task; T_trust is the trust and cost factor, which characterizes the trust level of the potential edge node; among the multiple edge nodes, the edge node that collaborates with the target edge node is the N edge node with the highest value of the collaboration potential information, where N is an integer greater than or equal to 1; ω41, ω42, ω43, ​​and ω44 are preset weights.

5. The method according to claim 3, characterized in that, The target edge node, based on the comprehensive benefit value of the target task, triggers operations at the target edge node itself, through multi-edge node collaboration, or at the central node, to classify the insurance claim image set using a multimodal AI model, including: If the comprehensive benefit value of the target task is greater than or equal to the first threshold, the target edge node performs classification processing on the insurance claim image set through the local lightweight multimodal AI model of the target edge node to obtain the classified insurance claim image set. If the comprehensive benefit value of the target task is less than the first threshold and greater than or equal to the second threshold, then the target edge node will classify the insurance claim image set by means of multi-edge node collaboration and by means of a multimodal AI model. If the comprehensive benefit value of the target task is less than the second threshold and greater than or equal to the third threshold, the target edge node extracts a set of strong semantic and non-privacy feature vectors from the insurance claim image set and sends the set of feature vectors to the central node to receive the set of classified feature vectors obtained by the central node through multimodal AI model analysis, thereby determining the classified insurance claim image set based on the set of classified feature vectors. If the comprehensive benefit value of the target task is less than the second threshold and greater than or equal to the third threshold, then the target edge node extracts a set of strong semantic and non-privacy feature vectors from the insurance claim image set, sends the set of feature vectors to the central node, and receives the set of classified feature vectors obtained by the central node through analysis of the set of feature vectors using a multimodal AI model, so as to determine the classified insurance claim image set based on the set of classified feature vectors. If the comprehensive benefit value of the target task is less than the third threshold, the target edge node encrypts the insurance claim image set to obtain the encrypted insurance claim image set, sends the encrypted insurance claim image set to the central node, and receives the classified insurance claim image set obtained by the central node through multimodal AI model analysis of the encrypted insurance claim image set.

6. The method according to claim 5, characterized in that, If the load on the central node is high, the target edge node lowers the first threshold and the second threshold. If the current analysis mode is anti-fraud mode, the target edge node lowers the third threshold.

7. The method according to claim 6, characterized in that, The method further includes: The target edge node sends the classified set of insurance claim images to the business system.

8. An image classification system for insurance claims, characterized in that, The insurance claims image classification system includes multiple edge nodes and a center node, wherein any one of the multiple edge nodes is configured as follows: The target edge node acquires the target task, and the target task requests to classify the set of insurance claim images for the target user's insurance claim business. The target edge node analyzes the insurance claim image set through an agent deployed on the target edge node, and obtains a task descriptor of the metadata of the insurance claim image set output by the agent; The target edge node analyzes the task descriptor of the metadata through a benefit function to obtain the comprehensive benefit value of the target task. The comprehensive benefit value represents the multi-dimensional evaluation result of the target task. The target edge node, based on the comprehensive benefit value of the target task, triggers an operation to classify the insurance claim image set using a multimodal AI model, either locally at the target edge node, through multi-edge node collaboration, or at the central node. The task descriptor of the metadata describes the target task using structured data or natural language, including the following: target task complexity, target data sensitivity, and target processing urgency. The target task complexity characterizes the complexity of the insurance claims image set containing insurance claims business data, and the target data sensitivity characterizes the sensitivity of the insurance claims image set to containing insurance claims business data. The intelligent agent includes an image structure analysis channel, a text information extraction channel, a metadata reading channel, and a feature fusion and quantization unit. The target edge node analyzes the insurance claim image set through the intelligent agent deployed on the target edge node, and obtains a task descriptor of the metadata of the insurance claim image set output by the intelligent agent, including: The target edge node calls the image structure analysis channel to extract visual features from the insurance claim image set and evaluate the complexity based on the visual features to obtain the initial task complexity. The target edge node calls the text information extraction channel to extract and analyze the structured and unstructured text of the insurance claim image set, and obtains the initial data sensitivity. The target edge node calls the metadata reading channel to parse the file attribute information of the insurance claim image set and obtain the initial processing urgency. The target edge node calls the feature fusion and quantization unit to perform weighted fusion and normalization calculations on the initial task complexity, the initial data sensitivity, and the initial processing urgency according to the insurance business rule mapping table, and obtains the task descriptor of the metadata.

Images