Insurance image classification method and device, computer equipment and storage medium

Abstract

The invention relates to the field of artificial intelligence and finance, and discloses an insurance image classification method and device, computer equipment and a storage medium, and the method comprises the steps: carrying out the block coding of an input insurance image, and generating an image block embedding vector sequence; performing feature extraction on the image block embedded vector sequence through a dynamic multi-level attention mechanism, and generating input self-adaptive multi-scale attention features; inputting the multi-scale attention features into a self-adaptive expert mixing layer, and dynamically selecting and fusing calculation results of an expert module according to feature contents; and based on the output features of the expert mixing layer, generating category probability distribution of the target image through a classification head, and outputting an insurance image classification result corresponding to the category probability distribution. According to the technical scheme, the identification precision of the insurance image in the key category can be improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and finance, and in particular to an insurance image classification method, device, computer equipment and storage medium. Background Art

[0002] In image processing scenarios in the financial sector, businesses such as insurance claims and property appraisals place stringent demands on the accuracy and efficiency of image classification. For example, auto insurance claims require rapid identification of vehicle damage types (such as scratches and dents), while property insurance requires accurate classification of building structural damage (such as cracks and leaks). However, existing image classification methods based on Visual Transformers (ViT) face significant challenges in financial scenarios:

[0003] Insufficient feature generalization: Insurance images (such as vehicle damage and building cracks) have characteristics such as variable perspectives, complex backgrounds, and a small proportion of local key areas. Traditional static attention mechanisms are unable to dynamically focus on key areas of different scales.

[0004] The conflict between computing resources and accuracy: Financial services require extremely high real-time processing, but existing Mixture of Experts (MoE) methods, due to fixed routing strategies, lead to redundant computations and are unable to meet the response speed requirements of high-concurrency claims scenarios.

[0005] Negative impact of data distribution imbalance: High-value claims cases account for a very small proportion in the dataset. The traditional cross-entropy loss function tends to bias the model towards the majority class, resulting in an increased missed detection rate for the key minority class.

[0006] While existing techniques, such as bilateral adaptive attention, have made progress in general image classification, they are not optimized for the multimodal nature of financial images and lack a synergistic design between class imbalance and computational efficiency. Therefore, there is an urgent need for a financial insurance image classification method that can improve the recognition accuracy of key categories while ensuring real-time performance. Summary of the Invention

[0007] The present invention provides an insurance image classification method, apparatus, computer equipment and storage medium, aiming to ensure real-time performance while improving the recognition accuracy of insurance images in key categories.

[0008] In a first aspect, a method for classifying insurance images is provided, comprising the following steps:

[0009] S10, performing block encoding on the input insurance image to generate an image block embedding vector sequence;

[0010] S20, extracting features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features;

[0011] S30, inputting the multi-scale attention features into an adaptive expert mixture layer, dynamically selecting and fusing the calculation results of the expert modules according to the feature content;

[0012] S40. Based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution of the target image, and an insurance image classification result corresponding to the category probability distribution is output.

[0013] In a second aspect, an insurance image classification device is provided, comprising:

[0014] An image block encoding module is used to perform block encoding on the input insurance image and generate an image block embedding vector sequence;

[0015] A feature extraction module is used to extract features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features;

[0016] An expert mixture routing module, configured to input the multi-scale attention features into an adaptive expert mixture layer, and dynamically select and fuse the calculation results of the expert modules according to the feature content;

[0017] The classification result output module is used to generate a category probability distribution of the target image through the classification head based on the output features of the expert mixture layer, and output the insurance image classification result corresponding to the category probability distribution.

[0018] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned insurance image classification method when executing the computer program.

[0019] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned insurance image classification method are implemented.

[0020] In the scheme implemented by the above-mentioned insurance image classification method, apparatus, computer device, and storage medium, the input insurance image is block-encoded to generate a sequence of image block embedding vectors; features are extracted from the image block embedding vector sequence using a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features; the multi-scale attention features are input into an adaptive expert mixture layer, which dynamically selects and fuses the calculation results of expert modules based on feature content; based on the output features of the expert mixture layer, a category probability distribution of the target image is generated through a classification head, and the insurance image classification result corresponding to the category probability distribution is output. In the scheme provided by the present invention, block encoding and a dynamic multi-level attention mechanism are used to achieve adaptive feature extraction of local damage and global structure in insurance images, thereby improving fine-grained classification accuracy; combined with the sparse routing strategy of the adaptive expert mixture layer, only expert modules related to the current input are activated during the inference phase, so that the model maintains high accuracy while significantly improving computational efficiency compared to traditional methods; the complete process optimization from feature extraction to classification output is particularly suitable for the robustness requirements of complex images in financial scenarios such as insurance claims, and the classification accuracy is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0022] Figure 1 This is a schematic diagram of an application environment of the insurance image classification method according to one embodiment of the present invention;

[0023] Figure 2 This is a flow chart of a method for classifying insurance images according to an embodiment of the present invention;

[0024] Figure 3 is a schematic diagram of a feature extraction process in one embodiment of the present invention;

[0025] Figure 4 This is a flow chart of expert hybrid routing in one embodiment of the present invention;

[0026] Figure 5 This is a schematic diagram of a process of model training in one embodiment of the present invention;

[0027] Figure 6 1 is a schematic diagram of a process for optimizing a loss function according to an embodiment of the present invention;

[0028] Figure 7 is a schematic diagram of an insurance image classification device according to an embodiment of the present invention;

[0029] Figure 8 is a structural diagram of a computer device in one embodiment of the present invention;

[0030] Figure 9 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] The insurance image classification method provided by the embodiment of the present invention is applied in the following fields: Figure 1 In an application environment, the client communicates with the server through a network. The client can take an insurance image and send it to the server to view the classification results returned by the server. The server encodes the input insurance image in blocks to generate a sequence of image block embedding vectors; a dynamic multi-level attention mechanism is used to extract features from the image block embedding vector sequence to generate input-adaptive multi-scale attention features; the multi-scale attention features are input into an adaptive expert mixture layer, which dynamically selects and fuses the calculation results of expert modules based on the feature content; based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution for the target image, and the insurance image classification result corresponding to the category probability distribution is output.

[0033] In the solution provided by the present invention, adaptive feature extraction of local damage and global structure in insurance images is achieved through block coding and dynamic multi-level attention mechanism, thereby improving the accuracy of fine-grained classification; combined with the sparse routing strategy of the adaptive expert mixture layer, only the expert modules related to the current input are activated in the inference stage, so that the model maintains high accuracy while significantly improving the computational efficiency compared with traditional methods; the complete process optimization from feature extraction to classification output is particularly suitable for the robustness requirements of complex images in financial scenarios such as insurance claims, and the classification accuracy is significantly improved. Among them, the client can be but is not limited to various personal computers, laptops, smart phones, tablets and portable wearable devices. The server can be implemented with an independent server or a server cluster composed of multiple servers. The present invention is described in detail below through specific embodiments.

[0034] See also Figure 2 As shown, Figure 2 A flowchart of the insurance image classification method provided by an embodiment of the present invention includes the following steps:

[0035] S10. Block-encode the input insurance image to generate an image block embedding vector sequence.

[0036] In this embodiment, the input insurance image is segmented into 16×16 patching sequences, and linear projection is used to generate an embedding vector of dimension D. For an insurance image of 512×512 pixels, the output sequence length N=1024.

[0037] S20. Perform feature extraction on the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features.

[0038] Among them, Figure 3 As shown, step S20 specifically includes:

[0039] S21, performing multi-level attention calculation on the input image block embedding vector sequence, where each level corresponds to feature extraction of different granularity;

[0040] S22, analyzing the semantic content of the input insurance image through a lightweight controller and dynamically assigning weights to each layer;

[0041] S23. Aggregate multi-scale features according to hierarchical weights to generate multi-scale attention features that fuse fine-grained and global features.

[0042] The dynamic multi-level attention mechanism in step S20 is optimized by computation, specifically including:

[0043] The attention calculation of low-weight layers is sparsely approximated, and redundant feature calculations are dynamically skipped based on the layer weights to reduce computing resource consumption.

[0044] In this embodiment, in the dynamic multi-level attention mechanism, for each level l∈{1,…,L}, the query (Q), key (K), and value (V) projection matrices are calculated:

[0045]

[0046] in, is a learnable parameter, D h = D / L is the head dimension, and X is the input insurance image. Through hierarchical projection, the model can capture features of different granularities in parallel.

[0047] Receive global average pooling features through NAS controller Output layer weights:

[0048]

[0049] The temperature coefficient τ is initially set to 1.0 and linearly decays to 0.1 during training to promote discretization. This design allows the model to automatically enhance low-level weights (e.g., c1 = 0.4) for vehicle damage images (requiring fine-grained analysis) and enhance high-level weights (e.g., c4 = 0.5) for panoramic building images (requiring global understanding).

[0050] The final attention output is the weighted sum of the results at each level:

[0051]

[0052] Compared with the traditional multi-head attention mechanism (MHA), the computational complexity of the dynamic multi-level attention mechanism (DMLA) is kept as O(N 2 D), but the actual FLOPs are reduced by about 30% (because the low-weight layers can be approximated).

[0053] S30: Input the multi-scale attention features into the adaptive expert mixture layer, and dynamically select and fuse the calculation results of the expert modules according to the feature content.

[0054] Among them, Figure 4 As shown, step S30 specifically includes:

[0055] S31, inputting the multi-scale attention features into an adaptive expert mixture layer;

[0056] S32. Generate a gating signal based on the multi-scale attention features and dynamically select the top-K expert modules related to the current input; each expert module has the processing capability of different feature patterns;

[0057] S33. Perform weighted fusion on the outputs of the selected expert modules to generate expert mixed features.

[0058] The expert module in step S30 is sparsely activated, specifically including:

[0059] During the inference phase, only the expert modules selected by the gated signal are retained to participate in the calculation, and the parameters of the inactivated expert modules are frozen to reduce memory usage.

[0060] In this embodiment, the DMLA output feature is received Calculate the expert choice probability:

[0061]

[0062] Here, ε~N(0,0.01) is the exploration noise, and k=2 ensures computational sparsity. The key improvement is to use the layer weights c of the DMLA as additional inputs to the gating network to achieve joint attention-routing optimization.

[0063] Each expert e∈{1,…,E} is defined as:

[0064]

[0065] The final output is the weighted sum of the selected experts' outputs:

[0066]

[0067] This design enables the expert modules to present a natural specialization division of labor (for example, Expert3 is good at processing metal deformation features, and Expert5 focuses on glass crack identification).

[0068] S40. Based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution of the target image, and an insurance image classification result corresponding to the category probability distribution is output.

[0069] In a specific embodiment, if Figure 5 As shown, the present invention is based on a visual Transformer model and also includes the following visual Transformer model training steps:

[0070] S101. Construct a focus loss function to optimize model parameters by dynamically adjusting class weights and hard sample weights. Hard samples are samples that are difficult to learn correctly during model training and result in high loss. These samples typically have complex features or are on the edge of a class, but are not incorrectly labeled data.

[0071] S102, applying sparsity constraints to the gating network to balance the load distribution of the expert module;

[0072] S103. Jointly train the dynamic multi-level attention mechanism and the adaptive expert mixture layer to achieve end-to-end optimization.

[0073] Among them, Figure 6 As shown, it also includes optimization of the focal loss function, including:

[0074] S201. Reversely adjust the loss weight according to the frequency of insurance image categories to improve the recognition sensitivity of high-value minority classes;

[0075] S202, applying exponential weight decay to difficult samples based on prediction confidence;

[0076] S203. Constraining the gating network parameters through regularization terms to prevent the expert module from degrading.

[0077] In this embodiment, an improved focal loss is used to target the long-tail distribution of insurance image data:

[0078]

[0079] in, f i is the frequency of category i to achieve balanced weighting; γ = 2 increases the weight of difficult samples; L1 regularization term λ = 0.01 constrains the sparsity of the gating network.

[0080] In this embodiment, block coding and a dynamic multi-level attention mechanism are used to achieve adaptive feature extraction of local damage and global structure in insurance images, thereby improving the accuracy of fine-grained classification. Combined with the sparse routing strategy of the adaptive expert mixture layer, only the expert modules related to the current input are activated during the inference phase, so that the model maintains high precision while significantly improving computational efficiency compared to traditional methods. The complete process optimization from feature extraction to classification output is particularly suitable for the robustness requirements of complex images in financial scenarios such as insurance claims, and the classification accuracy is significantly improved.

[0081] In a specific embodiment, an application example of a car insurance claim settlement scenario is provided: intelligent damage assessment for vehicle front-end collision damage.

[0082] Input: A 960×1280 pixel image of the accident scene uploaded via a mobile app, including:

[0083] V-shaped wrinkle deformation of the front anti-collision beam (damage to core safety components determines the vehicle's scrap level);

[0084] The right front lamp assembly was broken (appearance damage affected the repair cost calculation);

[0085] The water tank bracket is slightly displaced (hidden mechanical damage, need to refer to the maintenance manual for judgment);

[0086] DMLA dynamic multi-level attention processing:

[0087] Damage area value classification:

[0088] The hierarchical weight distribution c = [0.4, 0.3, 0.2, 0.1] prioritizes parsing the powertrain association structure;

[0089] The cross-layer attention score in the crumple area of ​​the anti-collision beam reached 0.96 (triggering the vehicle safety structure failure threshold);

[0090] The displacement of the water tank bracket is fused with multi-scale features, and an abnormal offset of 0.5 mm is identified at the third level;

[0091] IA-MoE intelligent routing decision:

[0092] Expert modules work together to activate the anti-collision beam damage token:

[0093] Expert4 (body structure safety expert): compares square crumple zone design parameters and calculates repair cost / residual value rate;

[0094] Expert 6 (Used Car Valuation Expert): Predicts the impact of accident records on the resale price of a vehicle over the next three years;

[0095] The right front light assembly token is routed to: Expert2 (OEM / aftermarket cost expert): This identifies the source of the accessory supply chain based on the light assembly code.

[0096] Water tank bracket token trigger: Expert5 (mechanical linkage analysis expert): Verify the risk of heat dissipation system installation benchmark deviation based on CAD drawings;

[0097] Output and insurance and financial decisions:

[0098] Intelligent damage assessment classification: Determined as "structural total loss" (91% confidence level), triggering the following financial actions:

[0099] Automatically match the "presumed total loss" clause in insurance terms and initiate the vehicle residual value auction process;

[0100] Generate differentiated claim settlement options: 1. Pay according to the insured amount (vehicle recovery required); 2. Continue to use the vehicle after repair (renew the policy at a reduced insured amount);

[0101] Explainable Financial Parameters:

[0102] The deformation of the anti-collision beam is associated with a 37% repair / residual value ratio (contribution 65%);

[0103] Identification of original lighting components reduces the risk of fraudulent claims by 15% (by comparing historical fraud case database);

[0104] Business Process Reengineering:

[0105] The 3D damage reconstruction model output by DMLA is connected to the 4S shop maintenance system to generate real-time:

[0106] Dynamic maintenance quotation (including labor costs and parts price fluctuation data);

[0107] Reinsurance risk rating report (predicting the probability of subsequent losses based on damage types).

[0108] In this auto insurance claims scenario, hierarchical attention is used to locate key financially sensitive damages, expert modules are linked to quantify insurance losses, and a dynamic financial decision tree is output to achieve full-link value transformation from visual damage to insurance terms, residual value management, and risk pricing.

[0109] like Figure 7 As shown, an embodiment of the present invention further provides an insurance image classification device, comprising:

[0110] The image block encoding module 10 is used to perform block encoding on the input insurance image to generate an image block embedding vector sequence.

[0111] The feature extraction module 20 is used to extract features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features.

[0112] The feature extraction module 20 is specifically used for:

[0113] Performing multi-level attention calculation on the input image block embedding vector sequence, where each level corresponds to feature extraction of different granularity;

[0114] A lightweight controller analyzes the semantic content of the input insurance image and dynamically assigns weights to each layer.

[0115] Multi-scale features are aggregated according to hierarchical weights to generate multi-scale attention features that are a fusion of fine-grained and global features.

[0116] The expert mixture routing module 30 is used to input the multi-scale attention features into the adaptive expert mixture layer, and dynamically select and fuse the calculation results of the expert modules according to the feature content.

[0117] The expert hybrid routing module 30 is specifically used for:

[0118] Inputting the multi-scale attention features into an adaptive expert mixture layer;

[0119] Generate gating signals based on multi-scale attention features and dynamically select the top-K expert modules related to the current input; each expert module has the processing capability of different feature patterns;

[0120] The outputs of the selected expert modules are weightedly fused to generate expert hybrid features.

[0121] The classification result output module 40 is used to generate a category probability distribution of the target image through the classification head based on the output features of the expert mixture layer, and output the insurance image classification result corresponding to the category probability distribution.

[0122] In a specific embodiment, a model training module is also included, specifically including:

[0123] The focus loss function construction unit is used to construct the focus loss function and optimize the model parameters by dynamically adjusting the category weights and difficult sample weights;

[0124] The sparsity constraint unit is used to impose sparsity constraints on the gating network and balance the load distribution of the expert module;

[0125] An end-to-end optimization unit is used to jointly train the dynamic multi-level attention mechanism and the adaptive expert mixture layer to achieve end-to-end optimization.

[0126] The model training module also includes a focus loss function optimization unit, which is specifically used to:

[0127] Inversely adjust the loss weight based on the frequency of insurance image categories to improve the recognition sensitivity of high-value minority classes;

[0128] Apply exponential weight decay to difficult samples based on prediction confidence;

[0129] The regularization term is used to constrain the parameters of the gating network to prevent the expert module from degenerating.

[0130] The specific limitations of the insurance image classification device can be found in the limitations of the insurance image classification method described above and will not be further elaborated here. Each module within the aforementioned insurance image classification device may be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules may be embedded in or independent of a processor within a computer device in hardware form, or may be stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0131] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, memory, network interface and database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the service side of an insurance image classification method.

[0132] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, memory, a network interface, a display screen, and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the functions or steps on the client side of a method for classifying insurance images.

[0133] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0134] S10, performing block encoding on the input insurance image to generate an image block embedding vector sequence;

[0135] S20, extracting features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features;

[0136] S30, inputting the multi-scale attention features into an adaptive expert mixture layer, dynamically selecting and fusing the calculation results of the expert modules according to the feature content;

[0137] S40. Based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution of the target image, and an insurance image classification result corresponding to the category probability distribution is output.

[0138] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0139] S10, performing block encoding on the input insurance image to generate an image block embedding vector sequence;

[0140] S20, extracting features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features;

[0141] S30, inputting the multi-scale attention features into an adaptive expert mixture layer, dynamically selecting and fusing the calculation results of the expert modules according to the feature content;

[0142] S40. Based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution of the target image, and an insurance image classification result corresponding to the category probability distribution is output.

[0143] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0144] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0145] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0146] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for classifying insurance images, characterized in that: The following steps are involved: S10, performing block encoding on the input insurance image to generate an image block embedding vector sequence; S20, extracting features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features; S30, inputting the multi-scale attention features into an adaptive expert mixture layer, dynamically selecting and fusing the calculation results of the expert modules according to the feature content; S40. Based on the output features of the expert mixture layer, a classification head is used to generate a category probability distribution of the target image, and an insurance image classification result corresponding to the category probability distribution is output.

2. The insurance image classification method according to claim 1, characterized in that: Step S20 specifically includes: Performing multi-level attention calculation on the input image block embedding vector sequence, where each level corresponds to feature extraction of different granularity; A lightweight controller analyzes the semantic content of the input insurance image and dynamically assigns weights to each layer. Multi-scale features are aggregated according to hierarchical weights to generate multi-scale attention features that are a fusion of fine-grained and global features.

3. The insurance image classification method according to claim 1, characterized in that: Step S30 specifically includes: Inputting the multi-scale attention features into an adaptive expert mixture layer; Generate gating signals based on multi-scale attention features and dynamically select the top-K expert modules related to the current input; each expert module has the processing capability of different feature patterns; The outputs of the selected expert modules are weightedly fused to generate expert hybrid features.

4. The insurance image classification method according to claim 1, characterized in that: It also includes the model training steps, including: Construct a focal loss function to optimize model parameters by dynamically adjusting category weights and difficult sample weights; Imposing sparsity constraints on the gating network to balance the load distribution of the expert module; Jointly train dynamic multi-level attention mechanisms and adaptive mixture-of-experts layers to achieve end-to-end optimization.

5. The insurance image classification method according to claim 4, characterized in that: It also includes optimization of the focal loss function, including: Inversely adjust the loss weight based on the frequency of insurance image categories to improve the recognition sensitivity of high-value minority classes; Apply exponential weight decay to difficult samples based on prediction confidence; The regularization term is used to constrain the parameters of the gating network to prevent the expert module from degenerating.

6. The insurance image classification method according to claim 1, characterized in that: The dynamic multi-level attention mechanism in step S20 is optimized through computation, specifically including: The attention calculation of low-weight layers is sparsely approximated, and redundant feature calculations are dynamically skipped based on the layer weights to reduce computing resource consumption.

7. The insurance image classification method according to claim 1, characterized in that: The expert module in step S30 is sparsely activated, specifically including: During the inference phase, only the expert modules selected by the gated signal are retained to participate in the calculation, and the parameters of the inactivated expert modules are frozen to reduce memory usage.

8. An insurance image classification device, characterized in that: include: An image block encoding module is used to perform block encoding on the input insurance image and generate an image block embedding vector sequence; A feature extraction module is used to extract features from the image block embedding vector sequence through a dynamic multi-level attention mechanism to generate input-adaptive multi-scale attention features; An expert mixture routing module, configured to input the multi-scale attention features into an adaptive expert mixture layer, and dynamically select and fuse the calculation results of the expert modules according to the feature content; The classification result output module is used to generate a category probability distribution of the target image through the classification head based on the output features of the expert mixture layer, and output the insurance image classification result corresponding to the category probability distribution.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the insurance image classification method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the insurance image classification method according to any one of claims 1 to 7 are implemented.

Images