Auxiliary damage marking method and device, computer equipment and storage medium
By performing damage labeling and model training on the vehicle pictures to be identified in vehicle insurance claims, a target identification model is formed, which solves the problem of insufficient accuracy of vehicle damage labeling and improves the accuracy and consistency of labeling.
Patent Information
- Application Number
- CN202411512361.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2025-02-18
AI Technical Summary
In the prior art, there is a problem of insufficient accuracy in vehicle damage marking during vehicle insurance claims, resulting in uncertainty in claims processing, maintenance plan formulation and risk assessment.
By obtaining the pictures of the vehicles to be identified, performing damage labels and repair labels, obtaining labels, and using these labels to train and update the preconfigured damage identification model to form a target identification model. This model is used to damage labeling images to be labeled, and to calculate the checksum accuracy through historical labeling labels, and adjust the model parameters to improve the labeling accuracy.
It improves the accuracy and consistency of vehicle damage labeling, reduces errors and fatigue of manual labeling, and enhances the reliability and efficiency of damage identification.
Smart Images
Figure CN119478504A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of data processing and financial technology, and in particular to an auxiliary damage marking method, device, equipment and medium. Background Art
[0002] In financial scenarios, especially when it comes to vehicle insurance business (for example, insurance business for passenger cars, trucks, and motorcycles), auxiliary identification of vehicle damage is a crucial link. In the process of vehicle insurance claims, accurate identification of vehicle damage is of great significance to both insurance companies and insured persons. The accuracy of vehicle damage marking directly affects the subsequent claims process, maintenance plan formulation, risk assessment and other key aspects.
[0003] At present, common auxiliary damage annotation methods mainly rely on manual operation combined with some basic image editing tools. In the process of vehicle insurance claims, when receiving a damage case, the claims adjuster needs to obtain photos of the accident vehicle, and then use experience and relevant standards to annotate the damage location and degree. However, the vehicle insurance business involves a large number of vehicle damage cases, each of which contains multiple types of structured (such as vehicle model, damage location code, etc.) and unstructured data (such as images of damage locations, text descriptions, etc.). With the continuous increase in business volume, manual annotation faces many challenges. Due to the huge amount of data, claims adjusters are prone to fatigue and negligence during the processing process. On the other hand, for some complex damage situations, such as involving multiple parts and various types of damage, it is difficult to annotate them comprehensively and accurately. The experience and judgment standards of different claims adjusters are different, which also makes the annotation results lack consistency, resulting in inaccurate annotation results. Summary of the invention
[0004] The present invention provides an auxiliary damage marking method, device, computer equipment and medium to improve the accuracy of auxiliary damage marking.
[0005] In a first aspect, an auxiliary damage marking method is provided, comprising:
[0006] Obtain a picture of the vehicle to be identified, and mark the damage and repair of the corresponding vehicle parts in the picture of the vehicle to be identified to obtain a marking label;
[0007] The preconfigured damage identification model is trained using the annotation tags to obtain a training model, the cross entropy of the training model is calculated, model parameters of the training model are updated based on the cross entropy to obtain a parameter update model, the recall rate of the parameter update model is calculated, and when the recall rate meets a preset value, the parameter update model is used as a target identification model;
[0008] Query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling tag corresponding to the current image to be labeled with damage, use the historical labeling tag to perform labeling verification on the labeling result, and obtain a verification labeling result;
[0009] Querying the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculating the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results;
[0010] When the identification accuracy does not meet the preset accuracy, the target identification model is adjusted in parameters to obtain an adjusted model, and the adjusted model is verified for accuracy. When the accuracy verification result meets the preset result, the adjusted model is used to mark damages on the vehicle image to be identified.
[0011] In a second aspect, an auxiliary damage marking device is provided, comprising:
[0012] A maintenance marking module is used to obtain a picture of a vehicle to be identified, mark the damage of the corresponding vehicle parts in the picture of the vehicle to be identified, and perform maintenance marking to obtain a marking label;
[0013] A model training module, used to perform model training on a preconfigured damage identification model using the annotation tags to obtain a training model, calculate the cross entropy of the training model, update the model parameters of the training model based on the cross entropy to obtain a parameter update model, calculate the recall rate of the parameter update model, and when the recall rate meets a preset value, use the parameter update model as a target identification model;
[0014] A label verification module is used to query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling label corresponding to the current image to be labeled with damage, and use the historical labeling label to perform labeling verification on the labeling result to obtain a verified labeling result;
[0015] A model accuracy calculation module is used to query the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculate the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results;
[0016] The damage marking module is used to adjust the parameters of the target identification model to obtain an adjusted model when the identification accuracy does not meet the preset accuracy, verify the accuracy of the adjusted model, and use the adjusted model to mark the damage of the vehicle image to be identified when the accuracy verification result meets the preset result.
[0017] In a third aspect, a computer device is provided, including 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 objection speech mining method when executing the computer program.
[0018] 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 auxiliary damage marking method are implemented.
[0019] Compared with the problems described in the background technology, the present invention can provide raw data for subsequent annotation and model training by obtaining a picture of a vehicle to be identified, and perform damage annotation and repair annotation on the corresponding vehicle parts in the picture of the vehicle to be identified. The obtained annotation label can represent the annotation information in a specific format (label) to facilitate subsequent model training and use, and also provide a basis for subsequent verification and accuracy calculation; the preconfigured damage identification model can be trained with the annotation label, and the obtained training model can be trained with a large amount of annotation data, so that the model can gradually master the law of vehicle damage identification, thereby improving its accuracy and reliability, and the cross entropy of the training model can be calculated to determine the prediction effect of the model, and then the model can be adjusted in real time or the model can be used to perform damage annotation; querying the current picture to be marked with damage can use the historical annotation labels of the previously annotated pictures in the same case to verify the annotation results of the model and correct possible errors, wherein the current picture to be marked with damage refers to the picture in the vehicle damage annotation process, The vehicle picture that needs to be marked with damage at present, and the historical marking label is used to mark and verify the marking result, and the verification marking result can use the marking information of other pictures in the same case contained in the historical marking label to improve the accuracy of the marking result and reduce the situation of wrong marking; query the number of historical labels of the historical marking label and the current number of labels of the target identification model. The historical marking labels of the previously marked pictures in the same case can be used to verify the marking result of the model, correct possible errors, and calculate the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification marking result. The performance of the model can be fully understood, including the accuracy of the model, reuse, etc., so as to provide a basis for further optimization of the model; when the identification accuracy does not meet the preset accuracy, the parameters of the target identification model are adjusted to obtain the adjusted model. By adjusting the parameters, the performance of the model can be improved and the analysis accuracy of the model can be improved, so that it can better meet the actual needs. The present invention can improve the accuracy of auxiliary damage marking. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.
[0021] Figure 1 is a schematic diagram of an application environment of an auxiliary damage marking method in one embodiment of the present invention;
[0022] Figure 2is a schematic diagram of a process of an auxiliary damage marking method in one embodiment of the present invention;
[0023] Figure 3 is a schematic diagram of a structure of an auxiliary damage marking device in one embodiment of the present invention;
[0024] Figure 4 is a schematic diagram of a structure of a computer device in one embodiment of the present invention;
[0025] Figure 5 It is another structural schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION
[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0027] The auxiliary damage marking method provided by the embodiment of the present invention can be applied in the following aspects: Figure 1In an application environment, the client communicates with the server through a network. The server can obtain a picture of a vehicle to be identified, perform damage annotation and repair annotation on the corresponding vehicle parts in the picture of the vehicle to be identified, and obtain annotation labels; use the annotation labels to train a preconfigured damage identification model to obtain a training model, calculate the cross entropy of the training model, and based on the cross entropy, update the model parameters of the training model to obtain a parameter update model, calculate the recall rate of the parameter update model, and when the recall rate meets a preset value, use the parameter update model as the target identification model; query the current picture to be labeled for damage, use the target identification model to perform damage annotation on the current picture to be labeled for damage, obtain the annotation result, and query the current The historical annotation labels corresponding to the previous damaged annotated pictures are used to perform annotation verification on the annotation results to obtain verification annotation results; the number of historical labels of the historical annotation labels and the current number of labels of the target identification model are queried, and the identification accuracy of the target identification model is calculated according to the number of historical labels, the current number of labels and the verification annotation results; when the identification accuracy does not meet the preset accuracy, the parameters of the target identification model are adjusted to obtain an adjustment model, and the accuracy of the adjustment model is verified. When the accuracy verification result meets the preset result, the adjustment model is used to perform damage annotation on the vehicle picture to be identified. Among them, the client can be but is not limited to various personal computers, laptops, smart phones, tablet computers 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.
[0028] See also Figure 2 As shown, Figure 2 A flowchart of an auxiliary damage marking method provided by an embodiment of the present invention includes the following steps:
[0029] S1. Obtain a picture of a vehicle to be identified, and perform damage marking and repair marking on corresponding vehicle parts in the picture of the vehicle to be identified to obtain a marking label.
[0030] The embodiment of the present invention can provide raw data for subsequent annotation and model training by obtaining the vehicle image to be identified. The vehicle image to be identified refers to the vehicle-related image that needs to be analyzed and annotated for damage during the vehicle damage identification process.
[0031] It should be noted that the vehicle images to be identified may be images of vehicles such as passenger cars, trucks, and motorcycles.
[0032] Optionally, the image of the vehicle to be identified can be obtained through vehicle damage cases in the vehicle insurance business.
[0033] Furthermore, the embodiment of the present invention can present the annotation information in a specific format (label) by performing damage annotation and repair annotation on the corresponding vehicle parts in the vehicle image to be identified to obtain the annotation labels, so as to facilitate subsequent model training and use, and also provide a basis for subsequent verification and accuracy calculation.
[0034] As an embodiment of the present invention, the method of performing damage marking and repair marking on the corresponding vehicle parts in the vehicle image to be identified to obtain the marking label includes: extracting the damaged parts of the damaged vehicle corresponding to the vehicle image to be identified, identifying damage characteristics of the damaged parts, constructing a damage label for the damaged vehicle based on the damage characteristics, and constructing a repair label matching the damage label, marking the damaged vehicle for damage using the damage label, and marking the damaged vehicle for repair using the repair label to obtain the marking label.
[0035] Optionally, the damaged part can be extracted through an image frame constructed by Java, and the damage label can be obtained by matching the damage features with damage features in a database to obtain corresponding specific damage conditions, such as bumper damage, headlight damage and other features. Based on the damage conditions, a corresponding damage label is constructed, such as a headlight damage label, and a corresponding maintenance label is constructed based on the damage label, which is a headlight maintenance label.
[0036] Furthermore, in another optional embodiment of the present invention, the identifying the damage characteristics of the damaged part includes: performing image decomposition on the damage image corresponding to the damaged part to obtain a decomposed image, extracting wavelet coefficients of the decomposed image, performing image fusion on the decomposed image based on the wavelet coefficients to obtain a fused image, detecting image corner points of the fused image, and using the image corner points to identify the damage characteristics of the damaged part.
[0037] Optionally, the decomposed image may be obtained by performing wavelet decomposition on the damage image corresponding to the damage site, the wavelet coefficients may be extracted by wavelet basis functions, and the detection of image corner points of the fused image may be achieved by a corner point detection technology.
[0038] S2. Perform model training on a preconfigured damage identification model using the annotation tags to obtain a training model, calculate the cross entropy of the training model, update model parameters of the training model based on the cross entropy to obtain a parameter update model, calculate the recall rate of the parameter update model, and when the recall rate meets a preset value, use the parameter update model as a target identification model.
[0039] The embodiment of the present invention uses the annotation labels to train the preconfigured damage identification model, and the training model can be trained with a large amount of annotated data, so that the model can gradually grasp the rules of vehicle damage identification, thereby improving its accuracy and reliability. Among them, the preconfigured damage identification model refers to a model for identifying vehicle damage conditions, whose structure and parameters have been preliminarily set before performing a damage identification task based on a vehicle image. The preconfigured damage identification model can be configured through a convolutional neural network. The preconfigured damage identification model includes structures such as a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer is used to extract features of the image, the pooling layer is used to reduce data dimensions, and the fully connected layer is used to classify or regress the extracted features.
[0040] As an embodiment of the present invention, the method of using the annotated labels to perform model training on a preconfigured damage identification model to obtain a training model includes: dividing the annotated labels into training data to obtain a training set and a test set, using the training set to perform model training on the preconfigured damage identification model to obtain a preliminary training model, and using the test set to test and verify the preliminary training model to obtain a training model.
[0041] Optionally, the marking labels are divided into training data to obtain a training set and a test set by using 10 percent of the images corresponding to the marking labels as a test set and 90 percent as a training set.
[0042] In the embodiment of the present invention, the cross entropy of the training model is calculated to determine the prediction effect of the model, and then the model is adjusted in real time or the model is used to mark damage. The cross entropy refers to an indicator that measures the difference between the model prediction result and the real label.
[0043] As an embodiment of the present invention, the calculating the cross entropy of the training model includes: calculating the cross entropy of the training model using the following formula:
[0044] β=-∑ i px i logqx i
[0045] Among them, β represents the cross entropy, x i Represents the i-th data in the training model, px i Represents x i The true value of qx i Represents x i The predicted value of .
[0046] Furthermore, the embodiment of the present invention updates the model parameters of the training model based on the cross entropy to obtain a parameter update model. By continuously adjusting the model parameters, we can reduce the value of the cross entropy and improve the accuracy of the model.
[0047] As an embodiment of the present invention, based on the cross entropy, the model parameters of the training model are updated to obtain a parameter update model, including: calculating the gradient of the influence of the cross entropy on the parameters of the training model, based on the parameter influence gradient, reversely adjusting the parameters of the training model to obtain adjustment parameters, iteratively optimizing the adjustment parameters to obtain optimized parameters, and calculating the second cross entropy of the training model corresponding to the optimized parameters according to the optimized parameters, and when the second cross entropy meets a preset value, using the training model corresponding to the optimized parameters as the parameter update model.
[0048] The parameter impact gradient refers to an indicator that measures the difference between the probability distribution predicted by the model and the probability distribution of the true label.
[0049] Optionally, the calculation of the gradient of the influence of the cross entropy on the parameters of the training model can be calculated through a loss function, and based on the parameter influence gradient, the parameters of the training model are adjusted in reverse to obtain the adjusted parameters, which can be achieved through a back propagation algorithm operation, and the iterative optimization of the adjusted parameters to obtain the optimized parameters can be achieved through repeated iterative operations, and the calculation of the second cross entropy of the training model corresponding to the optimized parameters based on the optimized parameters can be calculated through a stochastic gradient descent algorithm. It should be noted that when the second cross entropy satisfies the preset value, it means that the second cross entropy satisfies the preset value, and the preset value can be set to 0.1, or can be set according to the actual application scenario.
[0050] Furthermore, the embodiment of the present invention can evaluate the ratio of the number of positive examples correctly identified by the parameter update model to the actual number of positive examples by calculating the recall rate of the parameter update model, so as to determine the damage analysis capability of the parameter update model. The recall rate refers to an indicator that measures the model's ability to identify positive examples.
[0051] Exemplarily, the recall rate can be calculated by the following formula:
[0052]
[0053] Among them, F represents the recall rate, A represents the number of positive examples predicted by the parameter update model, and B represents the actual number of positive examples.
[0054] It should be noted that the recall rate is 0.7, which can also be set according to actual application scenarios.
[0055] S3. Query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling tag corresponding to the current image to be labeled with damage, use the historical labeling tag to perform labeling verification on the labeling result, and obtain a verified labeling result.
[0056] In the embodiment of the present invention, by querying the current image to be annotated with damage, the historical annotation labels of the previously annotated images in the same case can be used to verify the annotation results of the model and correct possible errors. The current image to be annotated with damage refers to the vehicle image that needs to be annotated with damage in the vehicle damage annotation process.
[0057] Furthermore, the embodiment of the present invention utilizes the target identification model to perform damage annotation on the current image to be labeled with damage, and the annotation results can be obtained by annotating new images to further test the generalization ability of the model, that is, whether the model can accurately perform damage annotation when facing unseen images.
[0058] As an embodiment of the present invention, the target identification model is used to perform damage annotation on the current image to be labeled for damage to obtain an annotation result, including: resizing the current image to be labeled for damage to obtain an adjusted image, normalizing the adjusted image to obtain a normalized image, using the target identification model to perform damage prediction on the normalized image to obtain damage prediction parameters, performing parameter label classification on the damage prediction parameters to obtain classification labels, and determining the annotation result according to the classification labels.
[0059] Optionally, the image to be marked with damage is resized to obtain a resized image, which means resizing the image so that it meets the size specifications required by the model input layer. For example, if the model input requires the image size to be 224*224 pixels, and the original image size is different, a corresponding scaling operation is required, which can be adjusted by an image resizing script constructed by java, the normalized image can be obtained by adjusting the pixel value of the resized image to the interval of 0 to 1, and the damage prediction parameter is classified by parameter labels to obtain classification labels, which can be obtained by clustering through clustering functions.
[0060] Furthermore, the embodiment of the present invention utilizes the historical annotation tags to perform annotation verification on the annotation results, and the verified annotation results can utilize the annotation information of other images in the same case contained in the historical annotation tags to improve the accuracy of the annotation results and reduce the occurrence of erroneous annotations.
[0061] S4. Query the number of historical tags of the historically annotated tags and the current number of tags of the target identification model, and calculate the identification accuracy of the target identification model according to the number of historical tags, the current number of tags and the verification and annotation results.
[0062] The embodiment of the present invention can verify the labeling results of the model using the historical labeling labels of previously labeled images in the same case by querying the historical labeling number of the historical labeling labels and the current label number of the labels of the target identification model to correct possible errors.
[0063] Furthermore, the embodiment of the present invention can comprehensively understand the performance of the model, including the accuracy and reuse of the model, by calculating the identification accuracy of the target identification model according to the historical label number, the current label number and the verification and annotation results, thereby providing a basis for further optimization of the model.
[0064] Optionally, the process of calculating the identification accuracy of the target identification model based on the number of historical labels, the current number of labels and the verification and annotation results is as follows: determine the numerator and denominator. Numerator: determine the number of correct annotations from the verification and annotation results. For example, if the verification and annotation results are recorded in the form of a mark indicating whether the annotation is correct or not (such as 1 for correct and 0 for incorrect), then the number of statistics marked as 1 is the value of the numerator; denominator: usually the sum of the number of historical labels and the number of current labels. However, if there is a situation of repeated calculation (for example, there are overlapping parts of historical labels and current labels that require special processing), adjustments must be made based on specific annotation rules and actual conditions. Based on the numerator and the above, the identification accuracy of the target identification model is calculated using the following formula:
[0065] Identification accuracy = numerator (number of correct annotations) / denominator (total number of annotations) × 100%
[0066] S5. When the identification accuracy does not meet the preset accuracy, the target identification model is adjusted in parameters to obtain an adjusted model, and the accuracy of the adjusted model is verified. When the accuracy verification result meets the preset result, the adjustment model is used to mark the damage of the vehicle image to be identified.
[0067] The embodiment of the present invention adjusts the parameters of the target identification model when the identification accuracy rate does not meet the preset accuracy rate, and obtains an adjusted model. By adjusting the parameters, the performance of the model can be improved and the analysis accuracy of the model can be improved, so that it can better meet actual needs.
[0068] It should be noted that when the identification accuracy rate does not meet the preset accuracy rate, it means that the calculation result of the identification accuracy rate is less than the preset accuracy rate (80%).
[0069] As an embodiment of the present invention, the method of adjusting parameters of the target identification model to obtain an adjusted model includes: analyzing the model performance capability of the target identification model, constructing a model adjustment plan for the target identification model based on the model performance capability, and adjusting parameters of the target identification model based on the model adjustment plan to obtain an adjusted model.
[0070] Optionally, the analysis of the model performance capability of the target identification model can be performed by comparing the labeling results output by the model with the actual labeling labels (verification labeling results) to analyze the model's incorrect labeling situations, such as checking which vehicle parts are often incorrectly labeled, whether the degree of damage is incorrectly judged, or the damage part is incorrectly identified, etc. The process of constructing a model adjustment plan for the target identification model based on the model performance capability is as follows: if it is found that the model does not learn certain features sufficiently, consider increasing the complexity of the model, such as increasing the number of layers or neurons in the neural network; if the model has an overfitting problem (performs well on training data but poorly on new data), regularization methods, such as L1 or L2 regularization, can be used to limit the size of model parameters to prevent overfitting; if the model converges too slowly, hyperparameters such as the learning rate can be adjusted. For example, the learning rate can be appropriately lowered. If the current learning rate is 0.1, you can try to adjust it to 0.01, etc.; the process of adjusting the parameters of the target identification model based on the model adjustment scheme to obtain the adjusted model is as follows: if the number of layers or neurons of the neural network is increased, the newly added parameters need to be reinitialized and the model needs to be retrained; if the regularization method is used, the corresponding regularization term needs to be added to the loss function, and the gradient and update parameters of the model need to be recalculated; if the hyperparameters such as the learning rate are adjusted, it is only necessary to use the new hyperparameter values for model training in the subsequent training process.
[0071] Furthermore, the embodiment of the present invention can verify whether the adjusted model meets the requirements by performing accuracy verification on the adjusted model. If the requirements are met, the adjusted model is used for damage annotation.
[0072] Optionally, the accuracy verification of the adjustment model may be verified by calculating a regression value of the adjustment model.
[0073] In the embodiment of the present invention, when the accuracy verification result meets the preset result, the damage of the vehicle picture to be identified is marked by using the adjustment model to obtain an auxiliary damage identification result with higher accuracy.
[0074] It should be understood that the size of the serial numbers of the steps in the above embodiment does not mean 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 embodiment of the present invention. At the same time, it can improve the efficiency of auxiliary damage identification to meet user needs.
[0075] In one embodiment, an auxiliary damage labeling device is provided, and the auxiliary damage labeling device corresponds one-to-one with the auxiliary damage labeling method in the above embodiment. Figure 3 As shown, the objection speech mining device includes a maintenance marking module 201, a model training module 202, a label verification module 203, a model accuracy calculation module 204 and a damage marking module 205. The detailed description of each functional module is as follows:
[0076] The maintenance marking module 201 is used to obtain a picture of a vehicle to be identified, mark the damage and maintenance of the corresponding vehicle parts in the picture of the vehicle to be identified, and obtain a marking label;
[0077] The model training module 202 is used to perform model training on the preconfigured damage identification model using the annotation tags to obtain a training model, calculate the cross entropy of the training model, update the model parameters of the training model based on the cross entropy to obtain a parameter update model, calculate the recall rate of the parameter update model, and when the recall rate meets a preset value, use the parameter update model as a target identification model;
[0078] The label verification module 203 is used to query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling label corresponding to the current image to be labeled with damage, and use the historical labeling label to perform labeling verification on the labeling result to obtain a verified labeling result;
[0079] The model accuracy calculation module 204 is used to query the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculate the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results;
[0080] The damage marking module 205 is used to adjust the parameters of the target identification model to obtain an adjusted model when the identification accuracy does not meet the preset accuracy, and to verify the accuracy of the adjusted model. When the accuracy verification result meets the preset result, the adjustment model is used to mark the damage of the vehicle image to be identified.
[0081] In one embodiment, the repair labeling module 201 performs damage labeling and repair labeling on corresponding vehicle parts in the vehicle image to be identified to obtain a labeling label, which is used to: extract the damaged part of the damaged vehicle corresponding to the vehicle image to be identified, identify the damage characteristics of the damaged part, construct a damage label for the damaged vehicle based on the damage characteristics, and construct a repair label matching the damage label, use the damage label to label the damaged vehicle for damage, and use the repair label to label the damaged vehicle for repair, thereby obtaining a labeling label.
[0082] In one embodiment, the repair annotation module 201, when executing the identification of the damage features of the damaged part, is used to: perform image decomposition on the damage image corresponding to the damaged part to obtain a decomposed image, extract wavelet coefficients of the decomposed image, perform image fusion on the decomposed image based on the wavelet coefficients to obtain a fused image, detect image corner points of the fused image, and use the image corner points to identify the damage features of the damaged part.
[0083] In one embodiment, the model training module 202 performs model training on a preconfigured damage identification model using the annotation labels to obtain a training model, which is used to: divide the annotation labels into training data to obtain a training set and a test set, perform model training on the preconfigured damage identification model using the training set to obtain a preliminary training model, and obtain a training model after testing and verifying the preliminary training model using the test set.
[0084] In one embodiment, the model training module 202 performs model parameter update on the training model to obtain a parameter update model, which is used to: calculate the gradient of the cross entropy's influence on the parameters of the training model, based on the parameter influence gradient, reversely adjust the parameters of the training model to obtain adjustment parameters, iteratively optimize the adjustment parameters to obtain optimization parameters, calculate the second cross entropy of the training model corresponding to the optimization parameters based on the optimization parameters, and when the second cross entropy meets a preset value, use the training model corresponding to the optimization parameters as the parameter update model.
[0085] In one embodiment, the label verification module 203 performs damage labeling on the current image to be labeled with damage using the target identification model to obtain a labeling result, which is used to: resize the current image to be labeled with damage to obtain a resized image, normalize the resized image to obtain a normalized image, use the target identification model to perform damage prediction on the normalized image to obtain damage prediction parameters, perform parameter label classification on the damage prediction parameters to obtain classification labels, and determine the labeling result according to the classification labels.
[0086] In one embodiment, the damage labeling module 205 performs parameter adjustment on the target identification model to obtain an adjusted model, which is used to: analyze the model performance capability of the target identification model, construct a model adjustment plan for the target identification model based on the model performance capability, and adjust the parameters of the target identification model based on the model adjustment plan to obtain an adjusted model.
[0087] For the specific definition of the auxiliary damage annotation device, please refer to the definition of the auxiliary damage annotation method above, which will not be repeated here. Each module in the above-mentioned auxiliary damage annotation device can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0088] 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 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an 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 the 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 a function or step on the server side of an auxiliary damage marking method.
[0089] 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 5 As shown. The computer device includes a processor, a 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 the 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 the computer program is executed by the processor, it implements the functions or steps of a client side of an auxiliary damage annotation method
[0090] 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, wherein when the processor executes the computer program, the following steps are implemented:
[0091] Obtain a picture of the vehicle to be identified, and mark the damage and repair of the corresponding vehicle parts in the picture of the vehicle to be identified to obtain a marking label;
[0092] The preconfigured damage identification model is trained using the annotation tags to obtain a training model, the cross entropy of the training model is calculated, model parameters of the training model are updated based on the cross entropy to obtain a parameter update model, the recall rate of the parameter update model is calculated, and when the recall rate meets a preset value, the parameter update model is used as a target identification model;
[0093] Query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling tag corresponding to the current image to be labeled with damage, use the historical labeling tag to perform labeling verification on the labeling result, and obtain a verification labeling result;
[0094] Querying the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculating the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results;
[0095] When the identification accuracy does not meet the preset accuracy, the target identification model is adjusted in parameters to obtain an adjusted model, and the adjusted model is verified for accuracy. When the accuracy verification result meets the preset result, the adjusted model is used to mark damages on the vehicle image to be identified.
[0096] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:
[0097] Obtain a picture of the vehicle to be identified, and mark the damage and repair of the corresponding vehicle parts in the picture of the vehicle to be identified to obtain a marking label;
[0098] The preconfigured damage identification model is trained using the annotation tags to obtain a training model, the cross entropy of the training model is calculated, model parameters of the training model are updated based on the cross entropy to obtain a parameter update model, the recall rate of the parameter update model is calculated, and when the recall rate meets a preset value, the parameter update model is used as a target identification model;
[0099] Query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling tag corresponding to the current image to be labeled with damage, use the historical labeling tag to perform labeling verification on the labeling result, and obtain a verification labeling result;
[0100] Querying the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculating the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results;
[0101] When the identification accuracy does not meet the preset accuracy, the target identification model is adjusted in parameters to obtain an adjusted model, and the adjusted model is verified for accuracy. When the accuracy verification result meets the preset result, the adjusted model is used to mark damages on the vehicle image to be identified.
[0102] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant descriptions on the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.
[0103] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and 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. As an illustration and not limitation, RAM is available in many 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).
[0104] Those skilled in the art can clearly understand that for the convenience and simplicity 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.
[0105] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; 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 replace some of the technical features therein by equivalents; and 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 be included in the protection scope of the present invention. It should be noted that if software tools or components other than those of the company appear in the embodiments of this application, they are only used for illustration and do not represent actual use.
Claims
1. An auxiliary damage marking method, characterized in that: include: Obtain a picture of the vehicle to be identified, and mark the damage and repair of the corresponding vehicle parts in the picture of the vehicle to be identified to obtain a marking label; The preconfigured damage identification model is trained using the annotation tags to obtain a training model, the cross entropy of the training model is calculated, model parameters of the training model are updated based on the cross entropy to obtain a parameter update model, the recall rate of the parameter update model is calculated, and when the recall rate meets a preset value, the parameter update model is used as a target identification model; Query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling tag corresponding to the current image to be labeled with damage, use the historical labeling tag to perform labeling verification on the labeling result, and obtain a verification labeling result; Querying the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculating the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results; When the identification accuracy does not meet the preset accuracy, the target identification model is adjusted in parameters to obtain an adjusted model, and the adjusted model is verified for accuracy. When the accuracy verification result meets the preset result, the adjusted model is used to mark damages on the vehicle image to be identified.
2. The auxiliary damage marking method according to claim 1, characterized in that: The step of marking the corresponding vehicle parts in the image of the vehicle to be identified with damage and repair marks to obtain a marking label includes: Extracting the damaged part of the damaged vehicle corresponding to the image of the vehicle to be identified; identifying injury characteristics of the injury site; constructing a damage label of the damaged vehicle based on the damage feature, and constructing a repair label matching the damage label; The damaged vehicle is marked with the damage label, and the damaged vehicle is marked with the repair label to obtain a marked label.
3. The auxiliary damage marking method according to claim 2, characterized in that: The identifying the injury characteristics of the injury site includes: Decomposing the damaged image corresponding to the damaged part to obtain a decomposed image; extracting wavelet coefficients of the decomposed image; Based on the wavelet coefficients, performing image fusion on the decomposed images to obtain a fused image; Detecting image corner points of the fused image; The image corner points are used to identify damage features of the damaged area.
4. The auxiliary damage marking method according to claim 1, characterized in that: The method of using the annotation labels to perform model training on the preconfigured damage identification model to obtain a training model includes: Divide the labeled labels into training data to obtain a training set and a test set; Using the training set to perform model training on the preconfigured damage identification model to obtain a preliminary training model; After the preliminary training model is tested and verified using the test set, a training model is obtained.
5. The auxiliary damage marking method according to claim 1, characterized in that: The updating of model parameters of the training model based on the cross entropy to obtain a parameter update model includes: Calculating the gradient of the cross entropy's effect on the parameters of the training model; Based on the parameter influence gradient, reversely adjust the parameters of the training model to obtain adjustment parameters; Iteratively optimizing the adjustment parameters to obtain optimized parameters; According to the optimization parameters, calculating the second cross entropy of the training model corresponding to the optimization parameters; When the second cross entropy meets a preset value, the training model corresponding to the optimization parameter is used as a parameter updating model.
6. The auxiliary damage marking method according to claim 1, characterized in that: The method of using the target identification model to perform damage annotation on the current image to be annotated with damage to obtain an annotation result includes: Resizing the current image to be marked with damage to obtain a resized image; Normalizing the resized image to obtain a normalized image; Using the target identification model to perform damage prediction on the normalized image to obtain damage prediction parameters; Performing parameter label classification on the damage prediction parameters to obtain classification labels; A labeling result is determined according to the classification label.
7. The auxiliary damage marking method according to claim 1, characterized in that: The step of adjusting parameters of the target identification model to obtain an adjusted model includes: analyzing a model performance capability of the target identification model; Based on the performance capability of the model, construct a model adjustment scheme for the target identification model; The target identification model is parameter adjusted based on the model adjustment scheme to obtain an adjusted model.
8. An auxiliary damage marking device, characterized in that: include: A maintenance marking module is used to obtain a picture of a vehicle to be identified, mark the damage of the corresponding vehicle parts in the picture of the vehicle to be identified, and perform maintenance marking to obtain a marking label; A model training module, used to perform model training on a preconfigured damage identification model using the annotation tags to obtain a training model, calculate the cross entropy of the training model, update the model parameters of the training model based on the cross entropy to obtain a parameter update model, calculate the recall rate of the parameter update model, and when the recall rate meets a preset value, use the parameter update model as a target identification model; A label verification module is used to query the current image to be labeled with damage, use the target identification model to perform damage labeling on the current image to be labeled with damage, obtain a labeling result, query the historical labeling label corresponding to the current image to be labeled with damage, and use the historical labeling label to perform labeling verification on the labeling result to obtain a verified labeling result; A model accuracy calculation module is used to query the number of historical labels of the historically annotated labels and the current number of labels of the target identification model, and calculate the identification accuracy of the target identification model according to the number of historical labels, the current number of labels and the verification and annotation results; The damage marking module is used to adjust the parameters of the target identification model to obtain an adjusted model when the identification accuracy does not meet the preset accuracy, verify the accuracy of the adjusted model, and use the adjusted model to mark the damage of the vehicle image to be identified when the accuracy verification result meets the preset result.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the auxiliary damage marking 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 auxiliary damage marking method according to any one of claims 1 to 7 are implemented.