WO2022134354A1 - 车损检测模型训练、车损检测方法、装置、设备及介质 - Google Patents

车损检测模型训练、车损检测方法、装置、设备及介质 Download PDF

Info

Publication number
WO2022134354A1
WO2022134354A1 PCT/CN2021/083782 CN2021083782W WO2022134354A1 WO 2022134354 A1 WO2022134354 A1 WO 2022134354A1 CN 2021083782 W CN2021083782 W CN 2021083782W WO 2022134354 A1 WO2022134354 A1 WO 2022134354A1
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle damage
sample
image
result
loss value
Prior art date
Application number
PCT/CN2021/083782
Other languages
English (en)
French (fr)
Inventor
陈攀
刘莉红
刘玉宇
肖京
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2022134354A1 publication Critical patent/WO2022134354A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the field of artificial intelligence classification models, and in particular, to a vehicle damage detection model training, vehicle damage detection method, device, computer equipment and storage medium.
  • the present application provides a vehicle damage detection model training, vehicle damage detection method, device, computer equipment and storage medium, which can extract a target area and determine a first loss value, and then extract and rearrange the target area to identify the target area. Combine the vehicle damage shape results of the large-grained damage results and the authenticity results of the authenticity identification, and determine the second loss value and the third loss value, so as to obtain the total loss value, and then train according to the total loss value until convergence, in order to prevent Overfitting in recognition, focusing on learning the fine-grained degree of vehicle damage morphological features, and learning to reduce the influence of features that interfere with noise, improve the accuracy of recognition, reduce costs, and improve training efficiency.
  • a vehicle damage detection model training method comprising:
  • the vehicle damage sample set includes vehicle damage sample images, and one vehicle damage sample image is associated with a vehicle damage label group;
  • the vehicle damage label group includes a vehicle damage rectangle area and a vehicle damage shape label type ;
  • the vehicle damage sample images are extracted and rearranged to obtain a plurality of vehicle damage sample partial images corresponding to the vehicle damage sample images; A authenticity label association;
  • the target area perform partial vehicle damage shape recognition on all the partial images of the vehicle damage samples corresponding to the vehicle damage sample images, and identify the sample results corresponding to the partial images of the vehicle damage samples; the sample results Including vehicle damage shape results and true and false results;
  • a second loss value is determined, and at the same time determining a third loss value according to the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample;
  • the initial parameters of the initial detection model are iteratively updated, until the total loss value reaches the preset convergence condition, the initial detection after convergence
  • the model is recorded as a trained vehicle damage detection model.
  • a vehicle damage detection method comprising:
  • a vehicle damage detection model training device comprising:
  • the obtaining module is used to obtain a vehicle damage sample set; the vehicle damage sample set includes vehicle damage sample images, and one vehicle damage sample image is associated with a vehicle damage label group; the vehicle damage label group includes a vehicle damage rectangular area and a vehicle damage label group. Type of vehicle damage form label;
  • an input module for inputting the vehicle damage sample image into an initial detection model containing initial parameters
  • an extraction module configured to extract the global vehicle damage morphological feature in the vehicle damage sample image, identify the target area result, and determine a first loss value according to the target area result and the vehicle damage rectangular area;
  • the rearrangement module is configured to extract and rearrange the vehicle damage sample image according to the target area in the target area result to obtain a plurality of vehicle damage sample partial images corresponding to the vehicle damage sample image;
  • the partial image of the vehicle damage sample is associated with a authenticity label;
  • the first recognition module is configured to perform local vehicle damage shape recognition on all the vehicle damage sample partial images corresponding to the vehicle damage sample images according to the target area results, and identify the corresponding vehicle damage sample partial images.
  • Sample results; the sample results include vehicle damage morphological results and authenticity results;
  • the determining module is configured to determine the first vehicle damage shape according to the vehicle damage shape label type corresponding to the vehicle damage sample image and the vehicle damage shape result corresponding to the vehicle damage sample partial image corresponding to the vehicle damage sample image. Two loss values, and at the same time determine a third loss value according to the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample;
  • a loss module configured to determine a total loss value according to the first loss value, the second loss value and the third loss value
  • a training module configured to iteratively update the initial parameters of the initial detection model when the total loss value does not reach a preset convergence condition, until the total loss value reaches the preset convergence condition, after convergence
  • the initial detection model is recorded as the vehicle damage detection model that has been trained.
  • a vehicle damage detection device comprising:
  • the receiving module is used to receive the vehicle damage detection instruction and obtain the vehicle damage image
  • the second recognition module is configured to input the vehicle damage image into the vehicle damage detection model trained by the above-mentioned vehicle damage detection model training method, and perform vehicle damage shape recognition on the vehicle damage image through the vehicle damage detection model, and obtain Final result; the final result characterizes the type of vehicle damage morphology in the vehicle damage image.
  • a computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, the processor implementing the following steps when executing the computer-readable instructions:
  • the vehicle damage sample set includes vehicle damage sample images, and one vehicle damage sample image is associated with a vehicle damage label group;
  • the vehicle damage label group includes a vehicle damage rectangle area and a vehicle damage shape label type ;
  • the vehicle damage sample images are extracted and rearranged to obtain a plurality of vehicle damage sample partial images corresponding to the vehicle damage sample images; A authenticity label association;
  • the target area perform partial vehicle damage shape recognition on all the partial images of the vehicle damage samples corresponding to the vehicle damage sample images, and identify the sample results corresponding to the partial images of the vehicle damage samples; the sample results Including vehicle damage shape results and true and false results;
  • a second loss value is determined, and at the same time determining a third loss value according to the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample;
  • the initial parameters of the initial detection model are iteratively updated, until the total loss value reaches the preset convergence condition, the initial detection after convergence
  • the model is recorded as a trained vehicle damage detection model.
  • a computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor further implements the following steps when executing the computer-readable instructions:
  • the final result represents The types of vehicle damage patterns in the vehicle damage images are shown.
  • the vehicle damage detection model training method, device, computer equipment and storage medium provided by this application are obtained by obtaining a vehicle damage sample set; the vehicle damage sample set includes a vehicle damage sample image, a vehicle damage sample image and a vehicle damage label group association; the vehicle damage label group includes a vehicle damage rectangular area and a vehicle damage form label type; input the vehicle damage sample image into an initial detection model containing initial parameters; extract the global vehicle damage form in the vehicle damage sample image feature, identify the result of the target area, and determine the first loss value according to the result of the target area and the rectangular area of vehicle damage; according to the target area in the result of the target area, extract and analyze the sample image of the vehicle damage.
  • the present application provides a vehicle damage detection model training method, which realizes the recognition of the target area in the vehicle damage sample image through the global vehicle damage morphological feature recognition. , and then destructively rearrange the target area in the target area result to obtain the partial image of the vehicle damage sample containing the associated authenticity label, and perform the partial vehicle damage shape recognition on the partial image of the vehicle damage sample, and obtain the vehicle damage combined with the result of the target area.
  • the morphological results and the true and false results of authenticity identification are used to determine the total loss value for training, which realizes the identification of the target area results of the large-grained damage dimension through the learning of the global vehicle damage morphological characteristics, and combines the results of the target area to learn the large-grained damage dimension.
  • the local vehicle damage morphology recognition of the local area under the fine-grained damage dimension, and through the rearrangement of the destructive vehicle damage sample local image to prevent over-fitting in the recognition, and learning the noise authenticity feature to identify the authenticity of the noise which can reduce the interference of noise in the identification process, improve the accuracy and reliability of vehicle damage shape detection, and the present application can improve the accuracy and quality of vehicle damage shape recognition.
  • the vehicle damage detection method, device, computer equipment and storage medium provided by the present application are obtained by acquiring vehicle damage images;
  • the vehicle damage detection model performs vehicle damage shape recognition on the vehicle damage image to obtain the final result;
  • the final result represents the vehicle damage shape type in the vehicle damage image.
  • the application improves the recognition speed, thereby improving the accuracy of the vehicle damage detection model.
  • the accuracy and reliability of vehicle damage type identification improves the efficiency of damage determination, reduces costs, and improves customer satisfaction.
  • FIG. 1 is a schematic diagram of an application environment of a vehicle damage detection model training method or a vehicle damage detection method in an embodiment of the present application;
  • FIG. 2 is a flowchart of a method for training a vehicle damage detection model in an embodiment of the present application
  • FIG. 3 is a flowchart of a vehicle damage detection method in an embodiment of the present application.
  • FIG. 4 is a schematic block diagram of a vehicle damage detection model training device in an embodiment of the present application.
  • FIG. 5 is a schematic block diagram of a vehicle damage detection device in an embodiment of the present application.
  • FIG. 6 is a schematic diagram of a computer device in an embodiment of the present application.
  • the vehicle damage detection model training method provided by the present application can be applied in the application environment as shown in FIG. 1 , wherein the client (computer device) communicates with the server through the network.
  • the client computer equipment
  • the server includes but is not limited to various personal computers, notebook computers, smart phones, tablet computers, cameras and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a method for training a vehicle damage detection model is provided, and its technical solution mainly includes the following steps S10-S80:
  • the vehicle damage sample set includes a vehicle damage sample image, and one vehicle damage sample image is associated with a vehicle damage label group; the vehicle damage label group includes a vehicle damage rectangular area and a vehicle damage shape Label type.
  • the vehicle damage sample set is a set of vehicle damage sample images
  • the vehicle damage sample images are historically collected vehicle damage images
  • the vehicle damage sample images are used as samples for training a vehicle damage detection model.
  • a vehicle damage sample image is associated with a vehicle damage label group
  • the vehicle damage label group includes the vehicle damage rectangular area and the vehicle damage shape label type
  • the vehicle damage label group is the same as the vehicle damage label group.
  • the vehicle damage label group defines the coordinate range of the vehicle damage area in the vehicle damage sample image corresponding to it and which type of vehicle damage shape label it belongs to.
  • the vehicle damage rectangle area is the area of the coordinate range of the vehicle damage area
  • the vehicle damage form label type is the fine-grained damage type of the fine-grained damage dimension under the large-grained damage type of each granular damage dimension defined for the vehicle damage form
  • the large-grained damage type is the root node of the first layer in the tree structure of the defined damage type classification, that is, the damage type of the large category divided in all damage type classifications
  • the fine-grained damage type is defined as
  • the damage types subdivided under the large-grained damage type for example, the large-grained damage type includes three large-grained damage types such as scratches, dents, and cracks, and the fine-grained damage types are promising.
  • the large-grained damage type is depression. Concave, wrinkle, dead fold and other types, the vehicle damage shape label type includes the subdivision damage type under the large-grained damage type, for example: the vehicle damage shape label type includes a "sag-fold" label.
  • the initial detection model includes the initial parameters, and the initial parameters may be preset initial parameter values.
  • the global vehicle damage morphological feature is extracted from the input image, and the target area containing damage and the damage type of the large-grained damage dimension in the image are identified, and the global vehicle damage morphological feature is the same as the above.
  • the characteristics of the damaged area related to the large-grained damage type such as the features of concave, fold, dead fold, etc., extract the global vehicle damage morphological features in the vehicle damage sample image, that is, extract the large-scale damage in the vehicle damage sample image.
  • the granular damage type has the characteristics of global and common damage.
  • the local area with the global vehicle damage morphological characteristics is framed, so as to identify the The target area result, the target area result includes the target area and the large-grain damage result, and the target area result is the rectangular area range and the large-grain damage dimension that identify the vehicle damage pattern in the vehicle damage sample image.
  • Damage type the target area is an identified rectangular area with vehicle damage shape
  • the large-grain damage result is the result of the identified damage type with the large-grain damage dimension.
  • the result of the target area and the rectangular area of vehicle damage are input into the first loss function, and the first loss value is calculated by the first loss function, and the first loss value indicates the identified The difference between the target area result and the vehicle damage rectangular area.
  • the initial detection model includes a target area detection model based on the SSD and FPN model architecture
  • the target area detection model is used to extract the global vehicle damage morphological feature in the vehicle damage sample image, and identify the target area result
  • the network structure of the target area detection model is the network structure constructed based on the SSD and the FPN model architecture.
  • the FPN model architecture is an architecture in which each layer in the SSD architecture is fused by applying the FPN model architecture in the neural network structure of the SSD.
  • the FPN is the feature pyramid, and the FPN model architecture is to compress (downsample) first, then increase A large (up-sampling), and then the same-level fusion architecture, the SSD model architecture is a target detection model for target area generation and multi-feature information fusion.
  • extracting and re-arranging the partial image of the vehicle damage sample from the vehicle damage sample image and the process of extracting and rearranging is to extract the target area in the identified target area result to obtain:
  • the original partial sample image, and the original partial sample image is associated with the true label in the true and false labels, and the original partial sample image is segmented and rearranged to obtain a destructive sample image, and the destructive sample image is obtained.
  • a sample image is associated with a pseudo-label of the authenticity labels.
  • the partial image of the vehicle damage sample includes the original partial sample image and the destructive sample image
  • the original partial sample image is the target area directly obtained from the vehicle damage sample image according to the result of the target area
  • the extracted image, the destructive sample image is an image that has been rearranged on the original local sample image
  • the authenticity label includes a true label and a false label
  • the authenticity label is a label for distinguishing authenticity
  • the true labels are labels assigned to the unprocessed original images
  • the pseudo labels are labels assigned to the rearranged images.
  • the vehicle damage sample image is extracted and rearranged, and a plurality of vehicle damage sample images corresponding to the vehicle damage sample image are obtained.
  • Local images of vehicle damage samples including:
  • the extraction process is a process of copying from the vehicle damage sample image according to the range of the target area, thereby obtaining the original local sample image corresponding to the vehicle damage sample image, and establishing the The association relationship between the original local sample image and the true label in the true and false labels.
  • S402 Segment and rearrange the original local sample image to obtain a destructive sample image, and associate the destructive sample image with the false label in the true and false labels.
  • the process of dividing and rearranging is a process of dividing the original local sample image according to a preset area, and then rearranging it to obtain an image of the same size, and the preset area can be set according to requirements
  • the size of the preset area may be determined according to the size of the target area in the identified target area result, or the size may be determined according to the large-grain damage result in the identified target area result
  • the The preset area is an M ⁇ N area, wherein the M ⁇ N area is smaller than the size of the target area.
  • the default size of the preset area is an area of 5 ⁇ 4, and the divided A region of the same size as the preset region is determined as the region to be arranged, all the regions to be arranged are randomly arranged, and rearranged into an image of the same size as the original local sample image, so that all the regions after segmentation and rearrangement processing
  • the original partial sample image, a plurality of the destructive sample images corresponding to the original partial sample image can be obtained, and the association relationship between the destructive sample image and the false label in the true and false labels can be established.
  • the destructive sample image is an image with visual noise, and the destructive sample image can prevent overfitting of local vehicle damage pattern recognition and provide data for noise true and false feature learning.
  • the present application realizes that according to the target area in the target area result, the vehicle damage sample image is extracted to obtain the original local sample image corresponding to the vehicle damage sample image, and the original local sample image and the vehicle damage sample image are obtained.
  • the true label association in the true and false labels, and the original partial sample image is segmented and rearranged to obtain a destructive sample image, and the destructive sample image is associated with the false label in the true and false labels, so , the original local sample images with true labels and the destructive sample images with false labels can be obtained, which improves the accuracy of subsequent identification of vehicle damage patterns, and prevents overfitting in the identification of local vehicle damage patterns. Subsequent identification of local vehicle damage patterns provides data sources.
  • S50 Perform local vehicle damage shape recognition on all the vehicle damage sample partial images corresponding to the vehicle damage sample images according to the target area results, and identify the sample results corresponding to the vehicle damage sample partial images;
  • the sample results include vehicle damage morphological results and true and false results.
  • the target area result further includes the large-grained damage result corresponding to the local pattern of the vehicle damage sample, where the large-grained damage result is the result of the identified damage type with the large-grained damage dimension,
  • the identification of the local vehicle damage shape is to select a fine-grained convolution layer corresponding to the large-grained damage result, and extract the local vehicle damage shape feature on the local image of the vehicle damage sample through the selected fine-grained convolution layer. and the noise authenticity feature extraction, the process of extracting the local vehicle damage morphological feature and the noise authenticity feature respectively.
  • the local vehicle damage morphological feature is the relevant feature of the fine-grained damage type under the large-grained damage type
  • the noise authenticity feature is the feature of noise that causes the damage shape to be irregular or unsmooth
  • the local vehicle damage morphological identification includes fine-grained damage.
  • Vehicle damage identification and authenticity identification in the identification process, combine the extracted local vehicle damage morphological feature extraction and the correlation relationship between the noise authenticity features, and mutually influence each other in fine-grained vehicle damage identification and authenticity identification. Adjustment The parameters in the fine-grained convolution layer continuously improve the accuracy of fine-grained vehicle damage identification and authenticity identification.
  • Recognition is to identify the authenticity features of visual noise, the noise authenticity features, the vehicle damage shape result is the result of the fine-grained damage dimension damage type determined after the fine-grained vehicle damage identification, and the authenticity The result is the result of whether there is noise determined after the authenticity identification, and the authenticity result indicates the authenticity of whether there is noise in the partial image of the vehicle damage sample.
  • the initial detection model further includes a vehicle damage shape detection model based on the ResNet50 model architecture, the network structure of the vehicle damage shape detection model is the network structure of the ResNet50 model architecture, and the vehicle damage shape detection model is used to realize
  • the target area result performs partial vehicle damage shape recognition on all the vehicle damage sample partial images corresponding to the vehicle damage sample images, and identifies a model of the sample results corresponding to the vehicle damage sample partial images.
  • the partial vehicle damage pattern recognition is performed on all the partial images of the vehicle damage samples corresponding to the vehicle damage sample images, and identification of the shape of the vehicle damage is performed.
  • the large-grained damage result is the result of the identified damage type with the large-grained damage dimension
  • the local vehicle damage morphological feature is the relevant feature of the fine-grained damage type under the large-grained damage type
  • the The initial detection model includes a plurality of the fine-grained convolutional layers corresponding to the major granularity damage types one-to-one, and the hierarchical structures of the fine-grained convolutional layers may be the same or different.
  • the hierarchical structure of the fine-grained convolutional layer may be the hierarchical structure of the convolutional layer of the ResNet50 model, and the local feature vector map is convolved through the fine-grained convolutional layer, thereby extracting the local feature vector map.
  • the morphological characteristics of the vehicle damage, and the authenticity characteristics of the noise are extracted.
  • S503 Perform fine-grained vehicle damage identification according to the extracted local vehicle damage morphological features to obtain the vehicle damage morphological result, and perform authenticity identification according to the extracted noise authenticity features to obtain the authenticity result.
  • the fine-grained vehicle damage identification refers to the identification of local features in the fine-grained damage dimension
  • the authenticity identification refers to the identification of authentic and fake features of visual noise
  • the vehicle damage shape result is to pass through the The result of the damage type of the fine-grained damage dimension determined after the fine-grained vehicle damage identification.
  • the process of the vehicle damage morphological result of the fine-grained damage type under the granular damage type, the authenticity result is the result of whether there is noise determined after the authenticity identification, and the authenticity result indicates that the vehicle Whether there is noise in the local image of the damaged sample is true or false.
  • S504 Determine the vehicle damage shape result and the authenticity result as the sample result. Understandably, the vehicle damage shape result and the true and false results are marked as the sample results.
  • the present application realizes the selection of a fine-grained convolution layer corresponding to the large-grained damage result; the target area result includes the large-grained damage result; the selected fine-grained convolutional layer is used to perform the partial image of the vehicle damage sample. Extracting local vehicle damage morphological features, and simultaneously extracting noise authenticity features; performing fine-grained vehicle damage identification according to the extracted local vehicle damage morphological features, obtaining the vehicle damage morphological results, and extracting the noise authenticity features according to the extraction. Perform authenticity identification to obtain the true and false results; determine the vehicle damage shape result and the true and false results as the sample results, so that the fine-grained convolution layer corresponding to the large-grained damage result is selected.
  • S60 Determine a second loss value according to the type of the vehicle damage form label corresponding to the vehicle damage sample image and the vehicle damage form result corresponding to the vehicle damage sample partial image corresponding to the vehicle damage sample image , and at the same time, a third loss value is determined according to the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample.
  • the vehicle damage shape label type corresponding to the vehicle damage sample image and the vehicle damage shape result corresponding to the vehicle damage sample partial image corresponding to the vehicle damage sample image are input into the second loss function.
  • determine the second loss value the second loss function can be set according to requirements, and input the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample into the third loss function wherein, the third loss value is determined, the third loss function is preferably a binary classification loss function, and the vehicle damage shape result represents the damage type of the vehicle damage shape identified in the vehicle damage sample image.
  • a second loss value is determined, including:
  • L 2 is the second loss value
  • n is the total number of all the partial images of the vehicle damage samples corresponding to the vehicle damage sample images
  • S is the original local sample image in the vehicle damage sample local image
  • F is the destructive sample image in the partial image of the vehicle damage sample
  • d is the true weight corresponding to the original local sample image
  • (1-d) is the pseudo weight corresponding to the destructive sample image.
  • the true weight is a weight parameter preset to the original local sample image
  • the pseudo weight is a weight parameter preset to the destructive sample image
  • S70 Determine a total loss value according to the first loss value, the second loss value and the third loss value.
  • step S70 that is, determining the total loss value according to the first loss value, the second loss value and the third loss value, including:
  • the attention mechanism technique is to continuously enhance the weights useful for identification among the first loss weight, the second loss weight and the third loss weight in the iterative training process, so that the enhancement is useful.
  • the result of feature recognition is to continuously enhance the weights useful for identification among the first loss weight, the second loss weight and the third loss weight in the iterative training process, so that the enhancement is useful.
  • L is the total loss value
  • L 1 is the first loss value
  • L 2 is the second loss value
  • L 3 is the third loss value
  • ⁇ 1 is the first loss weight
  • ⁇ 2 is the second loss weight
  • ⁇ 3 is the third loss weight.
  • the present application realizes that through the attention mechanism technology, the first loss weight, the second loss weight and the third loss weight are adjusted according to the target area result, the vehicle damage shape result and the authenticity result; A loss value, the second loss value, the third loss value, the first loss weight, the second loss weight and the third loss weight are input into the total loss function in the initial detection model , and the total loss value is calculated through the total loss function, so that the results of useful feature recognition are automatically enhanced, and the efficiency of recognition and training is improved.
  • the convergence condition may be the condition that the total loss value is small and will not decrease after 10,000 calculations, that is, the total loss value is small and will not decrease after 10,000 calculations.
  • the convergence condition can also be the condition that the total loss value is less than the set threshold, that is, when When the total loss value is less than the set threshold, the training is stopped, and the initial detection model after convergence is recorded as the vehicle damage detection model after the training is completed.
  • the step of determining the first loss value can continuously approach the accurate result, so that the accuracy of the recognition is getting higher and higher. In this way, the identification of vehicle damage shape types can be optimized, and the accuracy and reliability of lung feature identification can be improved.
  • the present application realizes obtaining a vehicle damage sample set;
  • the vehicle damage sample set includes vehicle damage sample images, and one vehicle damage sample image is associated with a vehicle damage label group;
  • the vehicle damage label group includes a vehicle damage rectangle.
  • area and vehicle damage morphological label type input the vehicle damage sample image into an initial detection model containing initial parameters; extract the global vehicle damage morphological feature in the vehicle damage sample image, identify the target area result, and determine the target area according to the target The first loss value is determined based on the area result and the vehicle damage rectangular area; according to the target area in the target area result, the vehicle damage sample image is extracted and rearranged to obtain a plurality of vehicle damage sample images.
  • the corresponding partial image of the vehicle damage sample; one of the partial image of the vehicle damage sample is associated with a authenticity label; according to the result of the target area, the partial image of the vehicle damage sample corresponding to the vehicle damage sample image is performed.
  • damage shape recognition and identify the sample results corresponding to the partial images of the vehicle damage samples; the sample results include the vehicle damage shape results and the true and false results; according to the vehicle damage shape label type corresponding to the vehicle damage sample image and the vehicle damage morphological result corresponding to the vehicle damage sample partial image corresponding to the vehicle damage sample image, a second loss value is determined, and at the same time, according to the authenticity label and the vehicle damage sample partial image corresponding to the According to the true and false results, a third loss value is determined; according to the first loss value, the second loss value and the third loss value, a total loss value is determined; when the total loss value does not reach a preset value When the convergence condition is met, the initial parameters of the initial detection model are iteratively updated, and when the total loss value reaches the preset
  • the present application provides a vehicle damage detection model training method, which realizes the identification of the target area result in the vehicle damage sample image through the global vehicle damage morphological feature recognition, and then destructively performs destructive testing on the target area in the target area result.
  • the partial image of the vehicle damage sample containing the associated authenticity label is obtained by rearranging the local image of the vehicle damage sample, and the local vehicle damage form recognition is performed on the local image of the vehicle damage sample, and the vehicle damage form result combined with the result of the target area and the authenticity result of the authenticity identification are obtained, so as to determine
  • the total loss value is used for training, which realizes the identification of the target area results of the large-grained damage dimension through the learning of the global vehicle damage morphological features, and combines the target area results to learn the local vehicle damage of the local area of the fine-grained damage dimension under the large-grained damage dimension.
  • Morphological recognition and by rearranging the local images of destructive vehicle damage samples to prevent over-fitting in recognition, and learning noise authenticity features to identify the authenticity of noise, it can reduce the interference of noise in the recognition process and improve
  • the present application can improve the accuracy and quality of vehicle damage shape recognition.
  • Local vehicle damage shape recognition, and learning the true and false features of noise to reduce the influence of interference noise realize the accurate and rapid identification of vehicle damage shape types in the image, and improve the recognition accuracy. Reduce costs and improve training efficiency.
  • the vehicle damage detection method provided by the present application can be applied in the application environment as shown in FIG. 1 , wherein the client (computer device) communicates with the server through the network.
  • the client computer equipment
  • the server includes but is not limited to various personal computers, notebook computers, smart phones, tablet computers, cameras and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a vehicle damage detection method is provided, and the technical solution mainly includes the following steps S100-S200:
  • the vehicle will leave traces of damage.
  • the staff of the insurance company will take relevant photos of the traffic accident, including the photos of the vehicle damage, and the staff will upload the photos of the vehicle damage to the server.
  • to trigger the vehicle damage detection instruction to acquire the vehicle damage image contained in the vehicle damage detection instruction, where the vehicle damage image is a photographed photograph of vehicle damage.
  • S200 Input the vehicle damage image into a vehicle damage detection model trained by the above-mentioned vehicle damage detection model training method, and perform vehicle damage shape recognition on the vehicle damage image by using the vehicle damage detection model to obtain a final result;
  • the final result characterizes the type of vehicle damage morphology in the vehicle damage image.
  • the final result can be obtained by simply inputting the vehicle damage image into the trained vehicle damage detection model, and performing the vehicle damage shape recognition through the vehicle damage detection model.
  • the identification process of the vehicle damage shape is accelerated, and the recognition accuracy of the vehicle damage shape is improved. rate and quality.
  • the present application obtains a vehicle damage image; inputs the vehicle damage image into a vehicle damage detection model trained by the above-mentioned vehicle damage detection model training method, and performs vehicle damage shape recognition on the vehicle damage image through the vehicle damage detection model, A final result is obtained; the final result characterizes the type of vehicle damage in the vehicle damage image.
  • the present application reduces the impact of noise, improves the speed of vehicle damage form recognition, and thus improves the accuracy of vehicle damage form type recognition. rate and quality, improve the efficiency of loss assessment, reduce costs, and improve customer satisfaction.
  • a vehicle damage detection model training device is provided, and the vehicle damage detection model training device corresponds one-to-one with the vehicle damage detection model training method in the above embodiment.
  • the vehicle damage detection model training device includes an acquisition module 11 , an input module 12 , an extraction module 13 , a rearrangement module 14 , a first identification module 15 , a determination module 16 , a loss module 17 and a training module 18 .
  • the detailed description of each functional module is as follows:
  • the obtaining module 11 is configured to obtain a vehicle damage sample set; the vehicle damage sample set includes vehicle damage sample images, and one vehicle damage sample image is associated with a vehicle damage label group; the vehicle damage label group includes a vehicle damage rectangular area and vehicle damage shape label type;
  • an input module 12 for inputting the vehicle damage sample image into an initial detection model containing initial parameters
  • the extraction module 13 is configured to extract the global vehicle damage morphological feature in the vehicle damage sample image, identify the target area result, and determine the first loss value according to the target area result and the vehicle damage rectangular area;
  • the rearrangement module 14 is configured to extract and rearrange the vehicle damage sample images according to the target area in the target area result, to obtain a plurality of vehicle damage sample partial images corresponding to the vehicle damage sample images;
  • the local image of the vehicle damage sample is associated with a authenticity label;
  • the first recognition module 15 is configured to perform local vehicle damage shape recognition on all the vehicle damage sample partial images corresponding to the vehicle damage sample images according to the result of the target area, and identify the corresponding vehicle damage sample partial images.
  • the determining module 16 is configured to determine the type of vehicle damage form label corresponding to the vehicle damage sample image and the vehicle damage form result corresponding to the partial image of the vehicle damage sample corresponding to the vehicle damage sample image. the second loss value, and simultaneously determine the third loss value according to the authenticity label and the authenticity result corresponding to the partial image of the vehicle damage sample;
  • a loss module 17 configured to determine a total loss value according to the first loss value, the second loss value and the third loss value
  • the training module 18 is configured to iteratively update the initial parameters of the initial detection model when the total loss value does not reach the preset convergence condition, until the total loss value reaches the preset convergence condition, the convergence
  • the subsequent initial detection model is recorded as a trained vehicle damage detection model.
  • each module in the above-mentioned vehicle damage detection model training device may be implemented in whole or in part by software, hardware and combinations thereof.
  • the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a vehicle damage detection device is provided, and the vehicle damage detection device corresponds one-to-one with the vehicle damage detection method in the above embodiment.
  • the vehicle damage detection device includes a receiving module 101 and a second identification module 102 .
  • the detailed description of each functional module is as follows:
  • the receiving module 101 is configured to receive a vehicle damage detection instruction and obtain a vehicle damage image
  • the second recognition module 102 is configured to input the vehicle damage image into a vehicle damage detection model trained by the above-mentioned vehicle damage detection model training method, and perform vehicle damage shape recognition on the vehicle damage image through the vehicle damage detection model, A final result is obtained; the final result characterizes the morphological type of vehicle damage in the vehicle damage image.
  • All or part of the modules in the above vehicle damage detection device can be implemented by software, hardware and combinations thereof.
  • the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a computer device in one embodiment, the computer device may be a server, and its internal structure diagram may be as shown in FIG. 6 .
  • the computer device includes a processor, memory, a network interface, and a database connected by 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 readable storage medium, an internal memory.
  • the readable storage medium stores an operating system, computer readable instructions and a database.
  • the internal memory provides an environment for the execution of the operating system and computer-readable instructions in the readable storage medium.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions when executed by the processor, implement a vehicle damage detection model training method, or a vehicle damage detection method.
  • the readable storage medium provided by this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.
  • a computer device which includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor.
  • the processor executes the computer-readable instructions
  • the vehicle in the above embodiment is implemented
  • the damage detection model training method, or the vehicle damage detection method in the above embodiment is implemented when the processor executes the computer-readable instruction.
  • one or more readable storage media storing computer-readable instructions are provided, and the readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage media medium; computer-readable instructions are stored on the readable storage medium, and when the computer-readable instructions are executed by one or more processors, make one or more processors implement the vehicle damage detection model training method in the above-mentioned embodiment, or When the computer-readable instructions are executed by the processor, the vehicle damage detection method in the above embodiment is implemented.
  • Nonvolatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • 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 chain Road (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

一种车损检测模型训练、车损检测方法、装置、设备及介质，涉及人工智能的分类模型领域，该方法包括：通过获取车损样本集；输入含有初始参数的初始检测模型；提取车损样本图像中的全局车损形态特征，获取目标区域结果，并确定出第一损失值；根据目标区域结果中的目标区域，得到多个车损样本局部图像；通过局部车损形态特征提取，识别出车损形态结果，同时进行真伪识别，得到真伪结果；确定出第二损失值，同时确定出第三损失值；从而得到总损失值；通过总损失值训练模型，得到训练完成的车损检测模型。该方法实现减少干扰噪声的影响，实现了准确地、快速地识别出图像中的车损形态类型。

Description

车损检测模型训练、车损检测方法、装置、设备及介质
本申请要求于2020年12月25日提交中国专利局、申请号为202011566971.4，发明名称为“车损检测模型训练、车损检测方法、装置、设备及介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能的分类模型领域，尤其涉及一种车损检测模型训练、车损检测方法、装置、计算机设备及存储介质。
背景技术
发明人发现在车辆发生交通事故后，车辆的某些部位会留下破损、刮伤等损伤的痕迹，目前，保险公司一般是通过人工识别由车主或业务人员拍摄的交通事故的车辆损伤图像，进而对图像中车辆的损伤部位的损伤类型进行人工识别并判定；该方案的不足之处在于：由于存在标准理解不一、观察经验不足等影响，可能会导致人工识别的损伤类型不符；例如：由于凹陷和刮擦难以通过目测图像加以分辨，定损人员很容易就将凹陷的损伤类型确定为刮擦的损伤类型，上述情况下导致的定损失误，会大大降低了定损的准确性；在可能导致保险公司的成本损失的同时，也会降低车主或客户的满意度；此外，人工定损的工作量巨大，定损效率低下，在需要满足一定的定损准确度的情况下，会进一步增加工作量，降低工作效率。
发明内容
本申请提供一种车损检测模型训练、车损检测方法、装置、计算机设备及存储介质，实现了提取出目标区域和确定出第一损失值，再对目标区域进行提取及重排，识别出结合大粒度损伤结果的车损形态结果及真伪识别的真伪结果，并确定第二损失值和第三损失值，从而得到总损失值，进而根据总损失值进行训练，直至收敛，以防止识别中过拟合，并聚焦学习于车损形态特征的细粒程度，和学习减少干扰噪声的特征的影响，提高了识别的准确率，减少了成本，提高了训练效率。
一种车损检测模型训练方法，包括：
获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
将所述车损样本图像输入含有初始参数的初始检测模型；
提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车 损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
一种车损检测方法，包括：
接收到车损检测指令，获取车损图像；
将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
一种车损检测模型训练装置，包括：
获取模块，用于获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
输入模块，用于将所述车损样本图像输入含有初始参数的初始检测模型；
提取模块，用于提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
重排模块，用于根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
第一识别模块，用于根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
确定模块，用于根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
损失模块，用于根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
训练模块，用于在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
一种车损检测装置，包括：
接收模块，用于接收到车损检测指令，获取车损图像；
第二识别模块，用于将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
将所述车损样本图像输入含有初始参数的初始检测模型；
提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时还实现如下步骤：
接收到车损检测指令，获取车损图像；
将所述车损图像输入通过车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
本申请提供的车损检测模型训练方法、装置、计算机设备及存储介质，通过获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；将所述车损样本图像输入含有初始参数的初始检测模型；提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型，因此，本申请提供了一种车损检测模型训练方法，实现了通过全局车损形态特征识别，识别出车损样本图像中的目标区域结果，再对目标区域结果中的目标区域进行破坏性的重排得到含有关联真伪标签的车损样本局部图像，对车损样本局部图像进行局部车损形态识别，得到结合目标区域结果的车损形态结果及真伪识别的真伪结果，从而确定出总损失值进行训练，实现了通过全局车损形态特征的学习中识别大粒度损伤维度的目标区域结果，结合目标区域结果学习大粒度损伤维度下的细粒度损伤维度的局部区域的局部车损形态识别，并通过重排的具有破坏性的车损样本局部图像以防止识别中过拟合，以及学习噪声真伪特征以识别噪声的真伪，能够将减少噪声在识别过程中的干扰，提高了车损形态检测的准确性和可靠性，本申请能够提升了车损形态识别的准确率和质量，通过车损形态的针对大粒度维度的全局车损形态特征的识别和针对细粒度维度的局部车损形态识别，以及学习噪声真伪特征减少干扰噪声的影响，实现了准确地、快速地识别出图像中的车损形态类型，提高了识别准确率，减少了成本，提高了训练效率。
本申请提供的车损检测方法、装置、计算机设备及存储介质，通过获取车损图像；将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损 检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型，如此，本申请提高了识别速度，从而提升了对车损形态类型识别的准确率及可靠性，提高了定损效率，减少了成本，提高了客户满意度。
本申请的一个或多个实施例的细节在下面的附图和描述中提出，本申请的其他特征和优点将从说明书、附图以及权利要求变得明显。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对本申请实施例的描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一实施例中车损检测模型训练方法或车损检测方法的应用环境示意图；
图2是本申请一实施例中车损检测模型训练方法的流程图；
图3是本申请一实施例中车损检测方法的流程图；
图4是本申请一实施例中车损检测模型训练装置的原理框图；
图5是本申请一实施例中车损检测装置的原理框图；
图6是本申请一实施例中计算机设备的示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请提供的车损检测模型训练方法，可应用在如图1的应用环境中，其中，客户端(计算机设备)通过网络与服务器进行通信。其中，客户端(计算机设备)包括但不限于为各种个人计算机、笔记本电脑、智能手机、平板电脑、摄像头和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中，如图2所示，提供一种车损检测模型训练方法，其技术方案主要包括以下步骤S10-S80：
S10，获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型。
可理解地，所述车损样本集为所述车损样本图像的集合，所述车损样本图像为历史收集的车损的图像，将所述车损样本图像作为训练车损检测模型的样本，一个所述车损样本图像与一个所述车损标签组关联，所述车损标签组包括所述车损矩形区域和所述车损形态标签类型，所述车损标签组为与所述车损样本图像关联的一组标签的集合，所述车损标签组定义了与其对应的所述车损样本图像中的车损区域的坐标范围和属于哪一种车损形态标签的类型，所述车损矩形区域为车损区域的坐标范围的区域，所述车损形态标签类型为对车损形态定义的各大粒度损伤维度的大粒度损伤类型下的细粒度损伤维度的细粒度损伤类型的标签类型，所述大粒度损伤类型为定义的损伤类型分类的树状结构中第一层的根节点，即所有损伤类型分类中划分的大类的损伤类型，所述细粒度损伤类型为定义的在所述大粒度损伤类型下细分的损伤类型，比如大粒度损伤类型包括擦伤，凹陷和开裂等三种大粒度损伤类型，细粒度损伤类型有为在大粒度损伤类型为凹陷下的下凹，褶皱，死褶等类型，所述车损形态标签类型包括大粒度损伤类型下的细分度损伤类型，例如：车损形态标签类型包含有“凹陷-褶皱”的标签。
S20，将所述车损样本图像输入含有初始参数的初始检测模型。
可理解地，所述初始检测模型包含有所述初始参数，所述初始参数可以为预设的初始的参数值。
S30，提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值。
可理解地，对输入的图像进行所述全局车损形态特征的提取以及识别出该图像中的含有损伤的目标区域和大粒度损伤维度的损伤类型，所述全局车损形态特征为与所述大粒度损伤类型相关的损伤区域的特征，比如下凹，褶皱，死褶等类型的特征，提取所述车损样本图像中的全局车损形态特征，即提取出所述车损样本图像中大粒度损伤类型下具有全局性的和共性损伤的特征，通过对提取的所述全局车损形态特征进行目标识别，将具有所述全局车损形态特征的局域进行框定，从而能够识别出所述目标区域结果，所述目标区域结果包括所述目标区域和大粒度损伤结果，所述目标区域结果为识别出所述车损样本图像中存在车损形态的矩形区域范围和大粒度损伤维度下的损伤类型，所述目标区域为一个识别出的具有车损形态的矩形区域范围，所述大粒度损伤结果为识别出的具有大粒度损伤维度下的损伤类型的结果。
其中，将所述目标区域结果和所述车损矩形区域输入第一损失函数中，通过所述第一损失函数计算出所述第一损失值，所述第一损失值表明了识别出的所述目标区域结果与所述车损矩形区域的差距。
作为优选，所述初始检测模型包括基于SSD与FPN模型架构的目标区域检测模型，所述目标区域检测模型用于提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值的，所述目标区域检测模型的网络结构为基于所述SSD与FPN模型架构构建的网络结构，所述SSD与FPN模型架构为在SSD的神经网络结构中应用FPN模型架构对SSD架构中的各层进行融合的架构，所述FPN即为特征金字塔，所述FPN模型架构为先压缩(下采样)、再增大(上采样)、然后同层级的进行融合的架构，所述SSD模型架构为目标区域生成以及多特征信息融合的目标检测模型。
S40，根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联。
可理解地，从所述车损样本图像中提取及重排出所述车损样本局部图像，所述提取及重排的过程为对识别出的所述目标区域结果中的目标区域进行提取，得到原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联，以及对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联。
其中，所述车损样本局部图像包括所述原始局部样本图像和所述破坏性样本图像，所述原始局部样本图像为直接从所述车损样本图像中根据所述目标区域结果中的目标区域提取出的图像，所述破坏性样本图像为对所述原始局部样本图像进行重排过的图像，所述真伪标签包括真标签和伪标签，所述真伪标签为区分真伪的标签，所述真标签为未加工的原始的图像赋予的标签，所述伪标签为重排过的图像赋予的标签。
在一实施例中，所述步骤S40中，即所述根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像，包括：
S401，根据所述目标区域结果中的目标区域，对所述车损样本图像进行提取，得到与所述车损样本图像对应的原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联。
可理解地，所述提取过程为从所述车损样本图像中按照所述目标区域的范围复制出的 过程，从而得到与所述车损样本图像对应的所述原始局部样本图像，以及建立所述原始局部样本图像与所述真伪标签中的真标签的关联关系。
S402，对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联。
可理解地，所述分割重排的处理过程为对所述原始局部样本图像按照预设区域进行分割，再重新排列后得到相同大小的图像的过程，所述预设区域可以根据需求设定，所述预设区域可以根据识别出的所述目标区域结果中的目标区域的大小确定其尺寸大小，也可以根据识别出的所述目标区域结果中的大粒度损伤结果确定其尺寸大小，所述预设区域为M×N的区域，其中，M×N的区域小于所述目标区域的尺寸大小，优先地，所述预设区域的默认尺寸大小为5×4的区域，将分割后的多个与预设区域相同大小的区域确定为待排列区域，对所有所述待排列区域进行随机排列，重新排列成与所述原始局部样本图像一样大小的图像，从而经过分割重排处理后的所述原始局部样本图像，可以得到多个与所述原始局部样本图像对应的所述破坏性样本图像，建立所述破坏性样本图像与所述真伪标签中的伪标签的关联关系，所述破坏性样本图像为具有视觉上的噪声的图像，所述破坏性样本图像能够防止局部车损形态识别的过拟合，以及为噪声真伪特征学习提供了数据。
本申请实现了根据所述目标区域结果中的目标区域，对所述车损样本图像进行提取，得到与所述车损样本图像对应的原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联，以及对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联，如此，能够得到具有真标签的原始局部样本图像和具有伪标签的破坏性样本图像，为后续的车损形态类型的识别提高了准确性，以及防止了局部车损形态识别出现过拟合，并为后续的局部车损形态识别提供了数据来源。
S50，根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果。
可理解地，所述目标区域结果还包括与所述车损样本局部图形对应的所述大粒度损伤结果，所述大粒度损伤结果为识别出的具有大粒度损伤维度下的损伤类型的结果，所述局部车损形态识别为选择与该大粒度损伤结果对应的细粒度卷积层，通过选择的所述细粒度卷积层对所述车损样本局部图像进行所述局部车损形态特征提取和所述噪声真伪特征提取，分别对提取所述局部车损形态特征和所述噪声真伪特征进行识别的过程，所述局部车损形态识别通过分开两个分支进行卷积提取，所述局部车损形态特征为大粒度损伤类型下的细粒度损伤类型的相关特征，所述噪声真伪特征为造成损伤形状不规则或者不流畅的噪声的特征，所述局部车损形态识别包括细粒度车损识别和真伪识别，在识别过程中结合提取的所述局部车损形态特征提取和所述噪声真伪特征之间的关联关系，彼此相互影响细粒度车损识别和真伪识别，调整所述细粒度卷积层中的参数，将细粒度车损识别和真伪识别的精度不断提高，所述细粒度车损识别为针对于细粒度损伤维度的局部特征的识别，所述真伪识别为针对视觉噪音的真伪特征进行识别，所述噪声真伪特征所述车损形态结果为经过所述细粒度车损识别之后确定的细粒度损伤维度的损伤类型的结果，所述真伪结果为经过所述真伪识别之后确定的是否存在噪声的结果，所述真伪结果表明了所述车损样本局部图像的是否存在噪声的真伪情况。
作为优选，所述初始检测模型还包括基于ResNet50模型架构的车损形态检测模型，所述车损形态检测模型的网络结构为ResNet50模型架构的网络结构，所述车损形态检测模型用于实现根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果的模型。
在一实施例中，所述步骤S50中，即所述根据所述目标区域结果对所有与所述车损样 本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果，包括：
S501，选择与大粒度损伤结果对应的细粒度卷积层；所述目标区域结果包括所述大粒度损伤结果。
可理解地，所述大粒度损伤结果为识别出的具有大粒度损伤维度下的损伤类型的结果，所述局部车损形态特征为大粒度损伤类型下的细粒度损伤类型的相关特征，所述初始检测模型中包含有多个与各大粒度损伤类型一一对应的所述细粒度卷积层，各所述细粒度卷积层的层级结构可以一样，也可以不一样。
S502，通过选择的所述细粒度卷积层对所述车损样本局部图像进行局部车损形态特征提取，同时进行噪声真伪特征提取。
可理解地，所述细粒度卷积层的层级结构可以为ResNet50模型的卷积层的层级结构，通过该细粒度卷积层对所述局部特征向量图进行卷积，从而提取出所述局部车损形态特征，以及提取出所诉噪声真伪特征。
S503，根据提取的所述局部车损形态特征进行细粒度车损识别，得到所述车损形态结果，以及根据提取的所述噪声真伪特征进行真伪识别，得到所述真伪结果。
可理解地，所述细粒度车损识别为针对于细粒度损伤维度的局部特征的识别，所述真伪识别为针对视觉噪音的真伪特征进行识别，所述车损形态结果为经过所述细粒度车损识别之后确定的细粒度损伤维度的损伤类型的结果，所述细粒度车损识别可理解为对卷积后的所述局部车损形态特征进行全连接分类识别，识别出具有大粒度损伤类型下的细粒度损伤类型的所述车损形态结果的过程，所述真伪结果为经过所述真伪识别之后确定的是否存在噪声的结果，所述真伪结果表明了所述车损样本局部图像的是否存在噪声的真伪情况。
S504，将所述车损形态结果和所述真伪结果确定为所述样本结果。可理解地，将所述车损形态结果和所述真伪结果标记为所述样本结果。
本申请实现了选择与大粒度损伤结果对应的细粒度卷积层；所述目标区域结果包括所述大粒度损伤结果；通过选择的所述细粒度卷积层对所述车损样本局部图像进行局部车损形态特征提取，同时进行噪声真伪特征提取；根据提取的所述局部车损形态特征进行细粒度车损识别，得到所述车损形态结果，以及根据提取的所述噪声真伪特征进行真伪识别，得到所述真伪结果；将所述车损形态结果和所述真伪结果确定为所述样本结果，如此，实现了通过选择与大粒度损伤结果对应的细粒度卷积层，对车损样本局部图像进行局部车损形态特征和噪声真伪特征的提取，并且进行细粒度车损识别和噪声真伪识别，能够识别出含有车损形态结果和真伪结果的样本结果，实现了自动识别出具有大粒度损伤类型下的细粒度损伤类型的车损形态结果，以及是否存在噪声的真伪结果，去除噪声对识别过程的影响，提高了识别精度，提升了识别准确率和质量。
S60，根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值。
可理解地，将与所述车损样本图像对应的所述车损形态标签类型和与所述车损样本图像对应的所述车损样本局部图像对应的车损形态结果输入第二损失函数中，确定出所述第二损失值，所述第二损失函数可以根据需求设定，以及将与所述车损样本局部图像对应的所述真伪标签和所述真伪结果输入第三损失函数中，确定出所述第三损失值，所述第三损失函数优选为二分类损失函数，所述车损形态结果表征了所述车损样本图像的识别出的车损形态的损伤类型。
在一实施例中，所述步骤S60中，即所述根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，包括：
S601，将所述车损形态标签类型和所述车损形态结果输入第二损失函数中，获取所述第二损失函数输出的所述第二损失值；所述第二损失函数为：
Figure PCTCN2021083782-appb-000001
其中：
L 2为第二损失值；
n为与所述车损样本图像对应的所有所述车损样本局部图像的总数；
S为所述车损样本局部图像中的所述原始局部样本图像；
F为所述车损样本局部图像中的所述破坏性样本图像；
d为与所述原始局部样本图像对应的真权重；
(1-d)为与所述破坏性样本图像对应的伪权重。
可理解地，所述真权重为预设给所述原始局部样本图像的权重参数，所述伪权重为预设给所述破坏性样本图像的权重参数。
S70，根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值。
在一实施例中，所述步骤S70中，即所述根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值，包括：
S701，运用注意力机制技术，根据所述目标区域结果、所述车损形态结果和所述真伪结果，调整第一损失权重、第二损失权重和第三损失权重,。
可理解地，所述注意力机制技术为在迭代训练过程中不断增强所述第一损失权重、所述第二损失权重和所述第三损失权重之中对识别有用的权重，以将增强有用的特征识别的结果。
S702，将所述第一损失值、所述第二损失值、所述第三损失值、所述第一损失权重、所述第二损失权重和所述第三损失权重输入所述初始检测模型中的总损失函数中，通过所述总损失函数计算出所述总损失值；所述总损失函数为：
L＝α 1L 1+α 2L 2+α 3L 3
其中：
L为总损失值；
L 1为所述第一损失值；
L 2为所述第二损失值；
L 3为所述第三损失值；
α 1为所述第一损失权重；
α 2为所述第二损失权重；
α 3为所述第三损失权重。
本申请实现了通过注意力机制技术，根据所述目标区域结果、所述车损形态结果和所述真伪结果，调整第一损失权重、第二损失权重和第三损失权重；将所述第一损失值、所述第二损失值、所述第三损失值、所述第一损失权重、所述第二损失权重和所述第三损失权重输入所述初始检测模型中的总损失函数中，通过所述总损失函数计算出所述总损失值，如此，实现了自动增强有用的特征识别的结果，提高了识别的效率及训练的效率。
S80，在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
可理解地，所述收敛条件可以为所述总损失值经过了10000次计算后值为很小且不会再下降的条件，即在所述总损失值经过10000次计算后值为很小且不会再下降时，停止训练，并将收敛之后的所述初始检测模型记录为训练完成的车损检测模型；所述收敛条件也 可以为所述总损失值小于设定阈值的条件，即在所述总损失值小于设定阈值时，停止训练，并将收敛之后的所述初始检测模型记录为训练完成的车损检测模型，如此，在所述总损失值未达到预设的收敛条件时，不断调整所述初始检测模型的初始参数，并触发通过所述目标区域检测模型提取所述车损样本图像中的全局车损形态特征，获取所述目标区域检测模型输出的目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值的步骤，可以不断向准确的结果靠拢，让识别的准确率越来越高。如此，能够优化车损形态类型识别，提高了肺部特征识别的准确性和可靠性。
如此，本申请实现了通过获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；将所述车损样本图像输入含有初始参数的初始检测模型；提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型，因此，本申请提供了一种车损检测模型训练方法，实现了通过全局车损形态特征识别，识别出车损样本图像中的目标区域结果，再对目标区域结果中的目标区域进行破坏性的重排得到含有关联真伪标签的车损样本局部图像，对车损样本局部图像进行局部车损形态识别，得到结合目标区域结果的车损形态结果及真伪识别的真伪结果，从而确定出总损失值进行训练，实现了通过全局车损形态特征的学习中识别大粒度损伤维度的目标区域结果，结合目标区域结果学习大粒度损伤维度下的细粒度损伤维度的局部区域的局部车损形态识别，并通过重排的具有破坏性的车损样本局部图像以防止识别中过拟合，以及学习噪声真伪特征以识别噪声的真伪，能够将减少噪声在识别过程中的干扰，提高了车损形态检测的准确性和可靠性，本申请能够提升了车损形态识别的准确率和质量，通过车损形态的针对大粒度维度的全局车损形态特征的识别和针对细粒度维度的局部车损形态识别，以及学习噪声真伪特征减少干扰噪声的影响，实现了准确地、快速地识别出图像中的车损形态类型，提高了识别准确率，减少了成本，提高了训练效率。
本申请提供的车损检测方法，可应用在如图1的应用环境中，其中，客户端(计算机设备)通过网络与服务器进行通信。其中，客户端(计算机设备)包括但不限于为各种个人计算机、笔记本电脑、智能手机、平板电脑、摄像头和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中，如图3所示，提供一种车损检测方法，其技术方案主要包括以下步骤S100-S200：
S100，接收到车损检测指令，获取车损图像。
可理解地，在车辆发生交通事故后，车辆会留下损伤的痕迹，保险公司的工作人员会拍摄交通事故的相关照片，这些照片包括车辆损伤的照片，工作人员将车辆损伤的照片上传至服务器，以触发所述车损检测指令，获取所述车损检测指令中含有的所述车损图像，所述车损图像为拍摄的车辆损伤的照片。
S200，将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
可理解地，只需将所述车损图像输入训练完成的车损检测模型，通过所述车损检测模型进行所述车损形态识别，就可以得到所述最终结果，所述车损形态识别为通过提取全局车损形态特征识别出区域，然后在具有大粒度损伤类型的区域下识别出细粒度损伤类型的识别过程，加快了车损形态的识别速度，从而提升了车损形态的识别准确率和质量。
本申请通过获取车损图像；将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型，如此，本申请减少了噪声影响，提高了车损形态识别速度，从而提升了对车损形态类型识别的准确率及质量，提高了定损效率，减少了成本，提高了客户满意度。
在一实施例中，提供一种车损检测模型训练装置，该车损检测模型训练装置与上述实施例中车损检测模型训练方法一一对应。如图4所示，该车损检测模型训练装置包括获取模块11、输入模块12、提取模块13、重排模块14、第一识别模块15、确定模块16、损失模块17和训练模块18。各功能模块详细说明如下：
获取模块11，用于获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
输入模块12，用于将所述车损样本图像输入含有初始参数的初始检测模型；
提取模块13，用于提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
重排模块14，用于根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
第一识别模块15，用于根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
确定模块16，用于根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
损失模块17，用于根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
训练模块18，用于在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
关于车损检测模型训练装置的具体限定可以参见上文中对于车损检测模型训练方法的限定，在此不再赘述。上述车损检测模型训练装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一实施例中，提供一种车损检测装置，该车损检测装置与上述实施例中车损检测方法一一对应。如图5所示，该车损检测装置包括接收模块101和第二识别模块102。各功能模块详细说明如下：
接收模块101，用于接收到车损检测指令，获取车损图像；
第二识别模块102，用于将所述车损图像输入如上述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
关于车损检测装置的具体限定可以参见上文中对于车损检测方法的限定，在此不再赘述。上述车损检测装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图6所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括可读存储介质、内存储器。该可读存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为可读存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种车损检测模型训练方法，或者车损检测方法。本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
在一个实施例中，提供了一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令，处理器执行计算机可读指令时实现上述实施例中车损检测模型训练方法，或者处理器执行计算机可读指令时实现上述实施例中车损检测方法。
在一个实施例中，提供了一个或多个存储有计算机可读指令的可读存储介质，本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质；该可读存储介质上存储有计算机可读指令，该计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器实现上述实施例中车损检测模型训练方法，或者计算机可读指令被处理器执行时实现上述实施例中车损检测方法。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质或易失性可读存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种车损检测模型训练方法，其中，包括：
    获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
    将所述车损样本图像输入含有初始参数的初始检测模型；
    提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
    根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
    根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
    根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
    根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
  2. 如权利要求1所述的车损检测模型训练方法，其中，所述根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像，包括：
    根据所述目标区域结果中的目标区域，对所述车损样本图像进行提取，得到与所述车损样本图像对应的原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联；
    对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联。
  3. 如权利要求1所述的车损检测模型训练方法，其中，所述根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果，包括：
    选择与大粒度损伤结果对应的细粒度卷积层；所述目标区域结果包括所述大粒度损伤结果；
    通过选择的所述细粒度卷积层对所述车损样本局部图像进行局部车损形态特征提取，同时进行噪声真伪特征提取；
    根据提取的所述局部车损形态特征进行细粒度车损识别，得到所述车损形态结果，以及根据提取的所述噪声真伪特征进行真伪识别，得到所述真伪结果；
    将所述车损形态结果和所述真伪结果确定为所述样本结果。
  4. 如权利要求2所述的车损检测模型训练方法，其中，所述根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，包括：
    将所述车损形态标签类型和所述车损形态结果输入第二损失函数中，获取所述第二损失函数输出的所述第二损失值；所述第二损失函数为：
    Figure PCTCN2021083782-appb-100001
    其中：
    L 2为第二损失值；
    n为与所述车损样本图像对应的所有所述车损样本局部图像的总数；
    S为所述车损样本局部图像中的所述原始局部样本图像；
    F为所述车损样本局部图像中的所述破坏性样本图像；
    d为与所述原始局部样本图像对应的真权重；
    (1-d)为与所述破坏性样本图像对应的伪权重。
  5. 如权利要求1所述的车损检测模型训练方法，其中，所述根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值，包括：
    运用注意力机制技术，根据所述目标区域结果、所述车损形态结果和所述真伪结果，调整第一损失权重、第二损失权重和第三损失权重；
    将所述第一损失值、所述第二损失值、所述第三损失值、所述第一损失权重、所述第二损失权重和所述第三损失权重输入所述初始检测模型中的总损失函数中，通过所述总损失函数计算出所述总损失值；所述总损失函数为：
    L＝α 1L 1+α 2L 2+α 3L 3
    其中：
    L为总损失值；
    L 1为所述第一损失值；
    L 2为所述第二损失值；
    L 3为所述第三损失值；
    α 1为所述第一损失权重；
    α 2为所述第二损失权重；
    α 3为所述第三损失权重。
  6. 一种车损检测方法，其中，包括：
    接收到车损检测指令，获取车损图像；
    将所述车损图像输入如权利要求1至5任一项所述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
  7. 一种车损检测模型训练装置，其中，包括：
    获取模块，用于获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
    输入模块，用于将所述车损样本图像输入含有初始参数的初始检测模型；
    提取模块，用于提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
    重排模块，用于根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
    第一识别模块，用于根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
    确定模块，用于根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时 根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
    损失模块，用于根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
    训练模块，用于在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
  8. 一种车损检测装置，其中，包括：
    接收模块，用于接收到车损检测指令，获取车损图像；
    第二识别模块，用于将所述车损图像输入如权利要求1至5任一项所述车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
  9. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其中，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
    将所述车损样本图像输入含有初始参数的初始检测模型；
    提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
    根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
    根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
    根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
    根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
  10. 如权利要求9所述的计算机设备，其中，所述根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像，包括：
    根据所述目标区域结果中的目标区域，对所述车损样本图像进行提取，得到与所述车损样本图像对应的原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联；
    对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联。
  11. 如权利要求9所述的计算机设备，其中，所述根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果，包括：
    选择与大粒度损伤结果对应的细粒度卷积层；所述目标区域结果包括所述大粒度损伤结果；
    通过选择的所述细粒度卷积层对所述车损样本局部图像进行局部车损形态特征提取， 同时进行噪声真伪特征提取；
    根据提取的所述局部车损形态特征进行细粒度车损识别，得到所述车损形态结果，以及根据提取的所述噪声真伪特征进行真伪识别，得到所述真伪结果；
    将所述车损形态结果和所述真伪结果确定为所述样本结果。
  12. 如权利要求10所述的计算机设备，其中，所述根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，包括：
    将所述车损形态标签类型和所述车损形态结果输入第二损失函数中，获取所述第二损失函数输出的所述第二损失值；所述第二损失函数为：
    Figure PCTCN2021083782-appb-100002
    其中：
    L 2为第二损失值；
    n为与所述车损样本图像对应的所有所述车损样本局部图像的总数；
    S为所述车损样本局部图像中的所述原始局部样本图像；
    F为所述车损样本局部图像中的所述破坏性样本图像；
    d为与所述原始局部样本图像对应的真权重；
    (1-d)为与所述破坏性样本图像对应的伪权重。
  13. 如权利要求9所述的计算机设备，其中，所述根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值，包括：
    运用注意力机制技术，根据所述目标区域结果、所述车损形态结果和所述真伪结果，调整第一损失权重、第二损失权重和第三损失权重；
    将所述第一损失值、所述第二损失值、所述第三损失值、所述第一损失权重、所述第二损失权重和所述第三损失权重输入所述初始检测模型中的总损失函数中，通过所述总损失函数计算出所述总损失值；所述总损失函数为：
    L＝α 1L 1+α 2L 2+α 3L 3
    其中：
    L为总损失值；
    L 1为所述第一损失值；
    L 2为所述第二损失值；
    L 3为所述第三损失值；
    α 1为所述第一损失权重；
    α 2为所述第二损失权重；
    α 3为所述第三损失权重。
  14. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其中，所述处理器执行所述计算机可读指令时实现如下步骤：
    接收到车损检测指令，获取车损图像；
    将所述车损图像输入通过车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
  15. 一个或多个存储有计算机可读指令的可读存储介质，其中，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
    获取车损样本集；所述车损样本集包括车损样本图像，一个所述车损样本图像与一个车损标签组关联；所述车损标签组包括车损矩形区域和车损形态标签类型；
    将所述车损样本图像输入含有初始参数的初始检测模型；
    提取所述车损样本图像中的全局车损形态特征，识别出目标区域结果，并根据所述目标区域结果和所述车损矩形区域，确定出第一损失值；
    根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像；一个所述车损样本局部图像与一个真伪标签关联；
    根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果；所述样本结果包括车损形态结果和真伪结果；
    根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，同时根据与所述车损样本局部图像对应的所述真伪标签和所述真伪结果，确定出第三损失值；
    根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述初始检测模型的初始参数，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述初始检测模型记录为训练完成的车损检测模型。
  16. 如权利要求15所述的可读存储介质，其中，所述根据所述目标区域结果中的目标区域，对所述车损样本图像提取及重排，得到多个与所述车损样本图像对应的车损样本局部图像，包括：
    根据所述目标区域结果中的目标区域，对所述车损样本图像进行提取，得到与所述车损样本图像对应的原始局部样本图像，并将所述原始局部样本图像与所述真伪标签中的真标签关联；
    对所述原始局部样本图像进行分割重排，得到破坏性样本图像，并将所述破坏性样本图像与所述真伪标签中的伪标签关联。
  17. 如权利要求15所述的可读存储介质，其中，所述根据所述目标区域结果对所有与所述车损样本图像对应的所述车损样本局部图像进行局部车损形态识别，识别出与所述车损样本局部图像对应的样本结果，包括：
    选择与大粒度损伤结果对应的细粒度卷积层；所述目标区域结果包括所述大粒度损伤结果；
    通过选择的所述细粒度卷积层对所述车损样本局部图像进行局部车损形态特征提取，同时进行噪声真伪特征提取；
    根据提取的所述局部车损形态特征进行细粒度车损识别，得到所述车损形态结果，以及根据提取的所述噪声真伪特征进行真伪识别，得到所述真伪结果；
    将所述车损形态结果和所述真伪结果确定为所述样本结果。
  18. 如权利要求16所述的可读存储介质，其中，所述根据与所述车损样本图像对应的所述车损形态标签类型和与该车损样本图像对应的所述车损样本局部图像对应的所述车损形态结果，确定出第二损失值，包括：
    将所述车损形态标签类型和所述车损形态结果输入第二损失函数中，获取所述第二损失函数输出的所述第二损失值；所述第二损失函数为：
    Figure PCTCN2021083782-appb-100003
    其中：
    L 2为第二损失值；
    n为与所述车损样本图像对应的所有所述车损样本局部图像的总数；
    S为所述车损样本局部图像中的所述原始局部样本图像；
    F为所述车损样本局部图像中的所述破坏性样本图像；
    d为与所述原始局部样本图像对应的真权重；
    (1-d)为与所述破坏性样本图像对应的伪权重。
  19. 如权利要求15所述的可读存储介质，其中，所述根据所述第一损失值、所述第二损失值和所述第三损失值，确定总损失值，包括：
    运用注意力机制技术，根据所述目标区域结果、所述车损形态结果和所述真伪结果，调整第一损失权重、第二损失权重和第三损失权重；
    将所述第一损失值、所述第二损失值、所述第三损失值、所述第一损失权重、所述第二损失权重和所述第三损失权重输入所述初始检测模型中的总损失函数中，通过所述总损失函数计算出所述总损失值；所述总损失函数为：
    L＝α 1L 1+α 2L 2+α 3L 3
    其中：
    L为总损失值；
    L 1为所述第一损失值；
    L 2为所述第二损失值；
    L 3为所述第三损失值；
    α 1为所述第一损失权重；
    α 2为所述第二损失权重；
    α 3为所述第三损失权重。
  20. 一个或多个存储有计算机可读指令的可读存储介质，其中，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
    接收到车损检测指令，获取车损图像；
    将所述车损图像输入通过车损检测模型训练方法训练完成的车损检测模型，通过所述车损检测模型对所述车损图像进行车损形态识别，得到最终结果；所述最终结果表征了所述车损图像中的车损形态类型。
PCT/CN2021/083782 2020-12-25 2021-03-30 车损检测模型训练、车损检测方法、装置、设备及介质 WO2022134354A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011566971.4A CN112668462B (zh) 2020-12-25 2020-12-25 车损检测模型训练、车损检测方法、装置、设备及介质
CN202011566971.4 2020-12-25

Publications (1)

Publication Number Publication Date
WO2022134354A1 true WO2022134354A1 (zh) 2022-06-30

Family

ID=75409670

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/083782 WO2022134354A1 (zh) 2020-12-25 2021-03-30 车损检测模型训练、车损检测方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN112668462B (zh)
WO (1) WO2022134354A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117372791A (zh) * 2023-12-08 2024-01-09 齐鲁空天信息研究院 细粒度定向能毁伤区域检测方法、装置及存储介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113920565A (zh) * 2021-10-29 2022-01-11 上海商汤智能科技有限公司 真伪识别方法、装置、电子设备以及存储介质
CN114155417B (zh) * 2021-12-13 2022-07-19 中国科学院空间应用工程与技术中心 图像目标的识别方法、装置、电子设备及计算机存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569864A (zh) * 2018-09-04 2019-12-13 阿里巴巴集团控股有限公司 基于gan网络的车损图像生成方法和装置
US20200111203A1 (en) * 2018-10-08 2020-04-09 Baidu Online Network Technology (Beijing) Co., Ltd. Method and apparatus for generating vehicle damage information
CN111667011A (zh) * 2020-06-08 2020-09-15 平安科技（深圳）有限公司 损伤检测模型训练、车损检测方法、装置、设备及介质
CN111680746A (zh) * 2020-06-08 2020-09-18 平安科技（深圳）有限公司 车损检测模型训练、车损检测方法、装置、设备及介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695605B (zh) * 2020-05-20 2024-05-10 平安科技（深圳）有限公司 基于oct图像的图像识别方法、服务器及存储介质
CN111598190B (zh) * 2020-07-21 2020-10-30 腾讯科技（深圳）有限公司 图像目标识别模型的训练方法、图像识别方法及装置
CN111709408B (zh) * 2020-08-18 2020-11-20 腾讯科技（深圳）有限公司 图像真伪检测方法和装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569864A (zh) * 2018-09-04 2019-12-13 阿里巴巴集团控股有限公司 基于gan网络的车损图像生成方法和装置
US20200111203A1 (en) * 2018-10-08 2020-04-09 Baidu Online Network Technology (Beijing) Co., Ltd. Method and apparatus for generating vehicle damage information
CN111667011A (zh) * 2020-06-08 2020-09-15 平安科技（深圳）有限公司 损伤检测模型训练、车损检测方法、装置、设备及介质
CN111680746A (zh) * 2020-06-08 2020-09-18 平安科技（深圳）有限公司 车损检测模型训练、车损检测方法、装置、设备及介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117372791A (zh) * 2023-12-08 2024-01-09 齐鲁空天信息研究院 细粒度定向能毁伤区域检测方法、装置及存储介质
CN117372791B (zh) * 2023-12-08 2024-03-22 齐鲁空天信息研究院 细粒度定向能毁伤区域检测方法、装置及存储介质

Also Published As

Publication number Publication date
CN112668462B (zh) 2024-05-07
CN112668462A (zh) 2021-04-16

Similar Documents

Publication Publication Date Title
WO2021017261A1 (zh) 识别模型训练方法、图像识别方法、装置、设备及介质
WO2021135499A1 (zh) 损伤检测模型训练、车损检测方法、装置、设备及介质
WO2022134354A1 (zh) 车损检测模型训练、车损检测方法、装置、设备及介质
WO2020253629A1 (zh) 检测模型训练方法、装置、计算机设备和存储介质
CN111860670B (zh) 域自适应模型训练、图像检测方法、装置、设备及介质
CN109241903B (zh) 样本数据清洗方法、装置、计算机设备及存储介质
WO2019233166A1 (zh) 一种表面缺陷检测方法、装置及电子设备
WO2021114809A1 (zh) 车辆损伤特征检测方法、装置、计算机设备及存储介质
WO2020252917A1 (zh) 一种模糊人脸图像识别方法、装置、终端设备及介质
CN114897779B (zh) 基于融合注意力的宫颈细胞学图像异常区域定位方法及装置
WO2021175006A1 (zh) 车辆图像检测方法、装置、计算机设备及存储介质
CN107784288A (zh) 一种基于深度神经网络的迭代定位式人脸检测方法
CN111626295B (zh) 车牌检测模型的训练方法和装置
WO2022057309A1 (zh) 肺部特征识别方法、装置、计算机设备及存储介质
CN112884782B (zh) 生物对象分割方法、装置、计算机设备和存储介质
CN116596875B (zh) 晶圆缺陷检测方法、装置、电子设备及存储介质
CN114639102B (zh) 基于关键点与尺寸回归的细胞分割方法和装置
CN111860582A (zh) 图像分类模型构建方法、装置、计算机设备和存储介质
CN114897764A (zh) 基于标准化通道注意力的肺结节假阳性排除方法及装置
CN112418196A (zh) 人群数量预测方法、装置、计算机设备及存储介质
US11756319B2 (en) Shift invariant loss for deep learning based image segmentation
CN112330652A (zh) 基于深度学习的染色体识别方法、装置和计算机设备
CN112926610A (zh) 车牌图像筛选模型的构建方法与车牌图像筛选方法
CN113780131B (zh) 文本图像朝向识别方法和文本内容识别方法、装置、设备
CN116612474B (zh) 对象检测方法、装置、计算机设备及计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21908366

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21908366

Country of ref document: EP

Kind code of ref document: A1