WO2020073510A1 - 一种基于神经网络的车辆定损方法、服务器及介质 - Google Patents

一种基于神经网络的车辆定损方法、服务器及介质 Download PDF

Info

Publication number
WO2020073510A1
WO2020073510A1 PCT/CN2018/124307 CN2018124307W WO2020073510A1 WO 2020073510 A1 WO2020073510 A1 WO 2020073510A1 CN 2018124307 W CN2018124307 W CN 2018124307W WO 2020073510 A1 WO2020073510 A1 WO 2020073510A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
preset
fixed
network model
damage
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2018/124307
Other languages
English (en)
French (fr)
Inventor
马进
王健宗
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020073510A1 publication Critical patent/WO2020073510A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08—Insurance
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30248—Vehicle exterior or interior

Definitions

  • the present application belongs to the field of artificial intelligence technology, and particularly relates to a neural network-based vehicle loss determination method, a server, and a non-volatile readable storage medium.
  • the embodiments of the present application provide a neural network-based vehicle loss determination method, a server, and a non-volatile readable storage medium to solve the problems of existing vehicle loss determination methods that have low loss accuracy and high labor costs problem.
  • a first aspect of the embodiments of the present application provides a method for vehicle loss determination based on a neural network, including:
  • the fixed-loss evaluation image sequence includes a fixed-loss evaluation image obtained by photographing the accident vehicle from each preset orientation of the accident vehicle;
  • a damage level probability vector of the accident vehicle is determined based on the feature vectors of all the fixed-loss evaluation images; the value of each element in the damage level probability vector is used for Identify the probability that the accident vehicle belongs to the preset damage level corresponding to the element;
  • the preset damage level corresponding to the element with the largest value in the damage level probability vector is determined as the damage level of the accident vehicle.
  • This application obtains the fixed-loss evaluation images captured from the various preset orientations of the accident vehicle, determines the feature vector of each fixed-loss evaluation image through the preset neural network model's feature extraction layer, and uses the preset neural network model's probability calculation layer Based on the feature vectors of all fixed damage assessment images, the damage level probability vector of the accident vehicle is determined, and the preset damage level corresponding to the element with the largest median value in the damage level probability vector is determined as the damage level of the accident vehicle, thereby achieving Intelligent, saves labor costs, and at the same time, because the fixed damage assessment images taken from the various preset orientations of the accident vehicle can reflect the overall damage situation of the accident vehicle, the embodiments of the present application are based on the preset orientation from the accident vehicle The feature vectors of the captured fixed-loss assessment images comprehensively determine the damage level of the accident vehicle, which improves the accuracy of the fixed-damage vehicle.
  • FIG. 1 is an implementation flowchart of a neural network-based vehicle loss determination method provided by the first embodiment of the present application
  • FIG. 2 is a specific implementation flowchart of S13 in a neural network-based vehicle loss determination method provided by a second embodiment of the present application;
  • FIG. 3 is a flow chart of an implementation of a neural network-based vehicle loss determination method provided by a third embodiment of the present application.
  • FIG. 4 is a specific implementation flowchart of S03 in a neural network-based vehicle loss determination method provided by a fourth embodiment of the present application;
  • FIG. 5 is a structural block diagram of a server provided by an embodiment of the present application.
  • FIG. 6 is a structural block diagram of a server according to another embodiment of the present application.
  • FIG. 1 is an implementation flowchart of a neural network-based vehicle loss determination method provided by the first embodiment of the present application.
  • the execution subject of the vehicle fixed loss method based on the neural network is a server.
  • the neural network-based vehicle loss determination method shown in Figure 1 includes the following steps:
  • S11 Acquire a fixed-loss assessment image sequence of the accident vehicle to be determined; the fixed-loss assessment image sequence includes a fixed-loss assessment image obtained by photographing the accident vehicle from each preset orientation of the accident vehicle.
  • the process of determining damage to an accident vehicle is the process of determining the damage level of the accident vehicle.
  • the damage level is used to describe the degree of damage to the vehicle.
  • the damage level can be defined based on the damage degree of the vehicle, and the defined damage level is the preset damage level. For example, an injury level with an injury level of 10% to 30% can be defined as mild injury, an injury level with an injury level of 30% to 50% can be defined as moderate injury, and an injury level with an injury level greater than 50% can be defined as For severe injuries, then the preset injury levels include: mild injury, moderate injury and severe injury.
  • the owner or the person with fixed damage to the vehicle can send a fixed loss request to the server through a designated application (APP) installed on the terminal device such as a mobile phone or tablet computer.
  • APP application
  • the vehicle damage determination request carries the vehicle identification of the accident vehicle to be damaged and the damage assessment image sequence.
  • the vehicle identification can be a license plate number, a vehicle identification number (Vehicle Identification Number, VIN), or an engine number, etc.
  • the vehicle identification can be manually entered into the terminal device by the owner or the loss-determining person.
  • the fixed-loss evaluation image sequence includes fixed-loss evaluation images obtained by photographing the accident vehicle from various preset orientations of the accident vehicle.
  • the preset orientation can be set according to actual needs, without limitation here, for example, the preset orientation may include but not limited to: front front, left front, left side, left rear, front rear, right rear, right side and right Forward.
  • the owner or the person with fixed damage can take pictures of the accident vehicle from the front, left front, left side, left rear, right rear, right rear, right side, and right front of the accident vehicle, and then obtain the vehicle used for the accident
  • Multiple fixed-loss evaluation images for fixed-loss evaluation are the front image, left front image, left side image, left rear image, right rear image, right rear image, right side image, and right front image of the accident vehicle
  • These fixed-loss evaluation images constitute a sequence of fixed-loss evaluation images.
  • the server may obtain the vehicle identification and the fixed loss evaluation image sequence of the accident vehicle to be damaged from the vehicle fixed loss request.
  • S12 Perform feature extraction on each of the fixed-loss evaluation images in the fixed-loss evaluation image sequence through a feature extraction layer of a preset neural network model to obtain feature vectors of the fixed-loss evaluation images.
  • the preset neural network model is obtained by training a pre-built original neural network model through a machine learning algorithm based on a preset number of sample data.
  • Each piece of data in the sample data is composed of a sequence of fixed damage assessment images of an accident vehicle and the damage vector probability vector of the accident vehicle.
  • the original neural network includes a feature extraction layer and a probability calculation layer connected in sequence. among them:
  • the feature extraction layer is used to extract the feature vector of the fixed-loss evaluation image.
  • the feature extraction layer is composed of at least one convolutional layer, and each convolutional layer corresponds to a preset convolution kernel, and the preset convolution kernel is used to interact with the fixed-loss evaluation image.
  • the corresponding image matrix is subjected to a convolution operation to extract the feature vector of the fixed-loss evaluation image.
  • the convolution kernel parameters of the preset convolution kernel need to be learned in the training of the original neural network model.
  • the probability calculation layer is used to calculate the damage level probability vector of each accident vehicle based on the feature vectors of all the fixed damage assessment images of each accident vehicle output by the feature extraction layer. It should be noted that the number of elements contained in the damage level probability vector is the same as the number of preset damage levels, and the value of each element in the damage level probability vector is used to identify that the accident vehicle belongs to the preset damage level corresponding to the element The probability.
  • a probability calculation function is pre-built in the probability calculation layer, the independent variable of the probability calculation function is the feature vector of each fixed-loss evaluation image, and the dependent variable of the probability calculation function is that the accident vehicle belongs to the probability calculation The probability of the preset damage level corresponding to the function.
  • Each independent variable in each probability calculation function corresponds to a weight coefficient. The weight coefficient corresponding to each independent variable needs to be learned in the training of the original neural network model.
  • the original neural network model When training the original neural network model, use the fixed-loss evaluation image sequence of the accident vehicle included in each sample data as the input of the original neural network model, and the damage level probability vector of the accident vehicle included in each sample data as the The output of the original neural network model learns the convolution kernel parameters of each preset convolution kernel included in the feature extraction layer and the weight coefficients corresponding to the independent variables in each probability calculation function included in the probability calculation layer, and then completes the original neural network For the training of the network model, the original neural network model after completing the training is the preset neural network model in the embodiment of the present application.
  • the server imports all the fixed-loss evaluation images in the fixed-loss evaluation image sequence into the preset neural network model.
  • the server performs feature extraction on each fixed-loss evaluation image in the fixed-loss evaluation image sequence through the feature extraction layer of the preset neural network model to obtain the feature vector of each fixed-loss evaluation image.
  • S12 may specifically include the following steps:
  • the feature extraction layer determine the image matrix corresponding to the fixed-loss evaluation image based on the correspondence between the position information of each pixel in the fixed-loss evaluation image and the pixel value, and check the image matrix by preset convolution Perform convolution processing to obtain the feature vector of the fixed-loss evaluation image.
  • the fixed-loss evaluation image is a two-dimensional matrix formed by arranging a plurality of pixels, and the position information of each pixel in the fixed-loss evaluation image is used to describe the rank order of each pixel in the two-dimensional matrix.
  • Each pixel in the fixed-loss evaluation image corresponds to a three-dimensional pixel value.
  • the three-dimensional pixel value includes the pixel values on the three color channels R, G, and B.
  • the server extracts the pixel value of each pixel in the fixed-loss evaluation image as an image based on the correspondence between the position information of each pixel in the fixed-loss evaluation image and the pixel value in the feature extraction layer of the preset neural network The value of the element at the same position as the pixel in the matrix, and then the image matrix corresponding to the fixed-loss evaluation image is obtained. It should be noted that each element in the image matrix corresponding to the fixed-loss evaluation image is represented by a three-dimensional pixel value.
  • the server After obtaining the image matrix corresponding to the fixed-loss evaluation image, the server performs a convolution operation on the image matrix corresponding to the fixed-loss evaluation image and a preset convolution kernel to obtain the feature vector of the fixed-loss evaluation image.
  • the specific process of the server performing the convolution operation on the image matrix corresponding to the fixed-loss evaluation image and the preset convolution kernel can be: using the preset convolution kernel on the image matrix in a preset step size from left to right and from top to bottom Perform sliding, and at each sliding position, multiply the preset convolution kernel with the sub-matrix composed of the elements at the corresponding position in the image matrix, and use the multiplied result as the corresponding position in the feature vector of the fixed-loss evaluation image For the value of the element, after the preset convolution kernel slides on the image matrix corresponding to the fixed-loss evaluation image, the value of each element in the feature vector of the fixed-loss evaluation image has been determined.
  • the server uses the feature vectors of all the fixed-loss assessment images of the accident vehicle as the preset neural network model Based on the input of the probability calculation layer, the probability calculation layer of the preset neural network model is based on the feature vectors of all the fixed damage assessment images of the accident vehicle to determine the damage level probability vector of the accident vehicle. It should be noted that the value of each element in the damage level probability vector is used to identify the probability that the accident vehicle belongs to the preset damage level corresponding to the element, and the sum of the values of all elements in the damage level probability vector is 1.
  • S13 may be implemented through S131 to S132 shown in FIG. 2:
  • S131 Determine, in the probability calculation layer, based on the weight coefficients of the feature vectors of the pre-learned preset damage assessment images with respect to the preset damage levels, determine the The weight coefficient of the feature vector relative to each of the preset damage levels.
  • S132 Based on the weight coefficients of the feature vectors of the fixed damage assessment images of the accident vehicle relative to each of the preset damage levels, perform the feature vectors of the fixed loss assessment images of the accident vehicle Weighted summation to obtain the damage level probability vector of the accident vehicle.
  • the server when the preset neural network model is trained through a preset number of sample data, the server will learn the characteristics of each loss assessment image included in the sample data at the probability calculation layer of the preset neural network model The weight of the vector relative to each preset damage level, when the amount of sample data is large enough, the server can learn the weight of the feature vectors of all possible fixed-loss evaluation images relative to each preset damage level, this application implements The example recognizes all possible fixed-loss evaluation images as pre-set loss evaluation images, that is, the pre-set loss evaluation images contain all possible fixed-loss evaluation images.
  • the probability calculation layer of the preset neural network model is based on the features of each preset loss assessment image learned in advance
  • the vector is relative to the weight coefficient of each preset damage level
  • the weight vector of the feature vector of each fixed damage assessment image of the accident vehicle to be determined is determined relative to each preset damage level, and is based on the respective fixed damage assessment image of the accident vehicle
  • the feature vectors of the various fixed-loss evaluation images of the accident vehicle are weighted and summed to obtain the damage level probability vector of the accident vehicle.
  • the feature vectors of each fixed-damage assessment image of the accident vehicle can be expressed as: x 1 , x 2 , x 3 , ..., x 8
  • the predetermined damage level includes mild damage
  • moderate damage and Severe injuries are expressed as y 1 , y 2 and y 3 respectively .
  • weight coefficients of x 1 , x 2 , x 3 , ..., x 8 relative to y 1 that are learned by the server in advance are a 11 , a 21 , a 31 , ..., a 81 respectively ; x 1 learned in advance , X 2 , x 3 ,..., X 8 weight coefficients relative to y 2 are respectively a 12 , a 22 , a 32 , ..., a 82 ; x 1 , x 2 , x 3 , ...
  • the weight coefficients of x 8 relative to y 3 are respectively a 13 , a 23 , a 33 , ..., a 83 ; then, based on the weight coefficients of the feature vectors of each fixed-loss evaluation image relative to each damage level, the The damage level probability vector obtained by weighting and summing the feature vectors of the various fixed-loss assessment images of the accident vehicle is:
  • x 1 a 11 + x 2 a 21 + x 3 a 31 + ... + x 8 a 81 represents the probability that the accident vehicle belongs to the damage level y1;
  • x 1 a 12 + x 2 a 22 + x 3 a 32 + ... + x 8 a 82 represents the probability that the accident vehicle belongs to the damage level y2;
  • x 1 a 13 + x 2 a 23 + x 3 a 33 + ... + x 8 a 83 represents the probability that the accident vehicle belongs to the damage level y3.
  • S14 Determine the preset damage level corresponding to the element with the largest value in the damage level probability vector as the damage level of the accident vehicle.
  • the server determines the preset damage level corresponding to the element with the largest median value in the damage level probability vector as the damage level of the accident vehicle.
  • a neural network-based vehicle damage determination method obtaineds fixed loss assessment images taken from various preset orientations of the accident vehicle, and determines each of the features through the preset neural network model's feature extraction layer.
  • the feature vector of the fixed-loss evaluation image based on the feature vectors of all the fixed-loss evaluation images through the preset neural network model's probability calculation layer, determine the damage level probability vector of the accident vehicle, and map the element with the largest value of the damage level probability vector
  • the damage level is determined to be the damage level of the accident vehicle, thereby realizing the intelligent vehicle damage determination and saving labor costs.
  • the fixed damage assessment images taken from the various preset orientations of the accident vehicle can reflect the whole of the accident vehicle
  • the embodiment of the present application comprehensively determines the damage level of the accidental vehicle based on the feature vectors of the fixed-loss evaluation images captured from the various preset orientations of the accidental vehicle, and improves the accuracy of the fixed-vehicle damage.
  • FIG. 3 is an implementation flowchart of a neural network-based vehicle loss determination method provided by a third embodiment of the present application.
  • a neural network-based vehicle loss determination method provided in this embodiment may include S01 to S104 before S11, as described in detail as follows:
  • S01 Obtain a preset sample data set, and divide the sample data set into a training set and a test set; each sample data in the sample data set is composed of an image sequence of a fixed damage assessment of the accident vehicle and the accident vehicle
  • the damage level probability vector constitutes.
  • the original neural network model Before determining the damage level of the accident vehicle to be determined, the original neural network model needs to be constructed first.
  • the original neural network includes a feature extraction layer and a probability calculation layer connected in sequence.
  • a feature extraction layer and a probability calculation layer connected in sequence.
  • each piece of sample data in the sample data set is composed of a sequence of fixed damage assessment images of an accident vehicle and the damage level probability vector of the accident vehicle. It can be understood that the probability vector of the damage level of the accident vehicle included in each sample data may be obtained by manually performing fixed damage assessment on the accident vehicle.
  • the server may divide the sample data set into a training set and a test set based on the preset allocation ratio.
  • the training set is used to train the original neural network model
  • the test set is used to verify the accuracy of the trained original neural network model.
  • the preset distribution ratio can be set according to actual needs, and is not limited here.
  • S02 Train a pre-built original neural network model based on the training set, determine convolution kernel parameters of preset convolution kernels included in the feature extraction layer of the original neural network model, and determine the original neural network The weight coefficients of the feature vectors of each preset damage assessment image included in the probability calculation layer of the model relative to each of the preset damage levels.
  • the server trains the pre-built original neural network model based on the training set.
  • the fixed loss assessment image sequence of the accident vehicle included in each sample data in the training set is used as the original nerve
  • the damage level probability vector of the accident vehicle included in each sample data in the training set is used as the output of the original neural network model to determine the convolution kernel of the preset convolution kernel included in the feature extraction layer of the original neural network model Parameters, and the weight coefficients of the feature vectors of each preset loss assessment image included in the probability calculation layer of the original neural network model with respect to each preset damage level, that is, each preset included in the feature extraction layer by the server based on the training set
  • the convolution kernel parameters of the convolution kernel and the feature vectors of each preset loss evaluation image included in the probability calculation layer are learned with respect to the weight coefficients of each preset damage level, and then the training of the original neural network model is completed.
  • the server After the server completes the training of the original neural network model based on the training set, it validates the original neural network model that has been trained based on the test set.
  • S03 can be implemented through S031 to S033 as shown in FIG. 4, which are described in detail as follows:
  • the server verifies the original neural network model that has been trained based on the test set, it guides the sequence of the fixed-loss assessment image of the accident vehicle included in each sample data in the test set as the original neural network model that has been trained Input to determine the predicted value of the damage level probability vector corresponding to each sample data in the test set through the original neural network model that has been trained.
  • Error value predictive , value actual
  • n is the number of elements contained in the damage level probability vector
  • the server determines the predicted value of the damage level probability vector corresponding to each sample data in the test set, and then the damage level probability vector of the accident vehicle included in each sample data in the test set and the damage corresponding to each sample data
  • the predicted value of the level probability vector is substituted into the above formula to calculate the predicted error of the trained original neural network model.
  • the prediction error of the trained original neural network model is used to identify the vehicle fixed loss accuracy of the trained original neural network model.
  • S033 Compare the prediction error of the original neural network model with a preset error threshold, and determine the verification result of the original neural network model based on the comparison result; wherein, if the comparison result is that of the original neural network model If the prediction error is less than or equal to the preset error threshold, the verification result is determined to be verified; if the comparison result is that the prediction error of the original neural network model is greater than the preset error threshold, the verification is determined The result is that the verification failed.
  • the server compares the prediction error of the original neural network model that has completed training with a preset error threshold, and determines the value of the completed training based on the comparison result. Verification results of the original neural network model.
  • the preset error threshold is the allowable error value of the vehicle's fixed loss accuracy in practical applications.
  • the server determines the verification result of the original neural network model that has been trained to be verified; if the comparison result is that the prediction error of the original neural network model that has been trained is greater than the preset error threshold, it means that the original training has been completed. The accuracy of the vehicle's fixed loss of the neural network model exceeds the allowable error range. At this time, the server determines the verification result of the original neural network model that has completed the training as a verification failure.
  • the server detects that the original neural network model that has completed the training passes the verification, the original neural network model that has completed the training is determined as the preset neural network model.
  • the server detects that the verification of the trained original neural network model fails, it uses a back propagation algorithm to the volume of the preset convolution kernel included in the feature extraction layer of the original neural network model Accumulate kernel parameters and / or adjust the weight vectors of each preset damage assessment image contained in the probability calculation layer with respect to the weight coefficients of each preset damage level, and adjust the original neural network model after parameter adjustment based on the test set Verify again until the verification is passed, and the original neural network model with adjusted parameters is determined as the preset neural network model.
  • a neural network-based vehicle loss determination method trains a pre-built original neural network model through a training set containing a certain number of sample data, and passes a test containing a certain number of sample data. Set to verify the accuracy of the fixed neural network model of the trained original neural network model. After the verification is passed, the trained original neural network model is used as the subsequent preset neural network model for determining the damage level of the accident vehicle , Thereby improving the accuracy of the vehicle's fixed loss.
  • FIG. 5 is a structural block diagram of a server according to an embodiment of the present application.
  • the server in this embodiment is a server.
  • Each unit included in the server is used to execute each step in the embodiments corresponding to FIGS. 1 to 4.
  • the server 500 includes a first acquisition unit 51, a feature extraction unit 52, a first determination unit 53, and a second determination unit 54. among them:
  • the first acquisition unit 51 is used to acquire a fixed-loss evaluation image sequence of the accident vehicle to be determined for damage; the fixed-loss evaluation image sequence includes a fixed-loss evaluation obtained by photographing the accident vehicle from each preset orientation of the accident vehicle image.
  • the feature extraction unit 52 is configured to perform feature extraction on each of the fixed-loss evaluation images in the fixed-loss evaluation image sequence through a feature extraction layer of a preset neural network model to obtain feature vectors of the fixed-loss evaluation images.
  • the first determining unit 53 is used to determine the probability vector of the damage level of the accident vehicle based on the feature vectors of all the fixed-loss evaluation images in the probability calculation layer of the preset neural network model; The value of each element is used to identify the probability that the accident vehicle belongs to the preset damage level corresponding to the element.
  • the second determining unit 54 is configured to determine the preset damage level corresponding to the element with the largest value in the damage level probability vector as the damage level of the accident vehicle.
  • the feature extraction unit 52 is specifically used to:
  • the first determining unit 53 includes: a weight determining unit and a probability determining unit. among them:
  • the weight determination unit is used to determine, in the probability calculation layer, each of the fixed losses of the accident vehicle based on the weight coefficients of the feature vectors of each of the preset damage assessment images learned in advance with respect to each of the preset damage levels Weighting coefficients of the feature vector of the evaluation image relative to each of the preset damage levels.
  • the probability determining unit is configured to compare the weight coefficients of the feature vectors of the fixed-damage evaluation images of the accident vehicle with respect to each of the preset damage levels, and compare the The feature vectors are weighted and summed to obtain the damage level probability vector of the accident vehicle.
  • the server 500 further includes: a second acquisition unit, a training unit, a verification unit, and a third determination unit. among them:
  • the second obtaining unit is used to obtain a preset sample data set, and divide the sample data set into a training set and a test set; each sample data in the sample data set is determined by an image sequence of the damage assessment of the accident vehicle And the damage vector probability vector of the accident vehicle.
  • the training unit is used to train the pre-built original neural network model based on the training set, determine the convolution kernel parameters of the preset convolution kernel included in the feature extraction layer of the original neural network model, and determine the original The weight coefficients of the feature vectors of each preset damage assessment image included in the probability calculation layer of the neural network model with respect to each of the preset damage levels.
  • the verification unit is used to verify the trained original neural network model based on the test set.
  • the third determining unit is configured to determine the original neural network model that has completed training as the preset neural network model if the verification is passed.
  • the verification unit includes: a prediction unit, an error calculation unit, and a fourth determination unit. among them:
  • the prediction unit is used to import the fixed-loss evaluation image sequence of the accident vehicle included in each sample data of the test set into the original neural network model that has been trained to obtain the corresponding The predicted value of the damage level probability vector.
  • the error calculation unit is used to calculate the trained value based on the predicted value of the damage level probability vector of the accident vehicle included in each sample data in the test set and the damage level probability vector corresponding to each sample data
  • the prediction error of the original neural network model :
  • Error value predictive , value actual
  • n is the number of elements contained in the damage level probability vector
  • the fourth determining unit is used to compare the prediction error of the original neural network model with a preset error threshold, and determine the verification result of the original neural network model based on the comparison result; wherein, if the comparison result is the original If the prediction error of the neural network model is less than or equal to the preset error threshold, the verification result is determined to be verified; if the comparison result is that the prediction error of the original neural network model is greater than the preset error threshold, then It is determined that the verification result is that the verification fails.
  • the server obtains the fixed-loss evaluation images captured from each preset orientation of the accident vehicle, and determines the feature vector of each fixed-loss evaluation image through the feature extraction layer of the preset neural network model , Through the probability calculation layer of the preset neural network model, based on the feature vectors of all the fixed-loss evaluation images, determine the damage level probability vector of the accident vehicle, and determine the preset damage level corresponding to the element with the largest value in the damage level probability vector as the accident vehicle The damage level of the vehicle, thereby realizing the intelligent vehicle damage determination and saving labor costs.
  • this application comprehensively determines the damage level of the accidental vehicle based on the feature vectors of the fixed damage assessment images captured from the various preset orientations of the accidental vehicle, thereby improving the accuracy of the vehicle's fixed loss.
  • the server 6 of this embodiment includes: a processor 60, a memory 61, and computer-readable instructions 62 stored in the memory 61 and executable on the processor 60, for example, a neural network-based Procedures for vehicle damage determination methods.
  • the processor 60 executes the computer-readable instructions 62
  • the steps in the embodiments of the above neural network-based vehicle loss determination methods are implemented, for example, S11 to S14 shown in FIG. 1.
  • the processor 60 executes the computer-readable instructions 62
  • the computer-readable instructions 62 may be divided into one or more units, and the one or more units are stored in the memory 61 and executed by the processor 60 to complete the application .
  • the one or more units may be a series of computer-readable instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer-readable instructions 62 in the server 6.
  • the computer-readable instructions 62 may be divided into a first acquisition unit, a feature extraction unit, a first determination unit, and a second determination unit, and the specific functions of each unit are as described above.
  • the server may include, but is not limited to, the processor 60 and the memory 61. Those skilled in the art may understand that FIG. 6 is only an example of the server 6 and does not constitute a limitation on the server 6, and may include more or less components than shown, or combine certain components, or different components, for example
  • the server may also include input and output devices, network access devices, buses, and the like.
  • the so-called processor 60 can be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the storage 61 may be an internal storage unit of the server 6, such as a hard disk or a memory of the server 6.
  • the memory 61 may also be an external storage device of the server 6, such as a plug-in hard disk equipped on the server 6, a smart memory card (Smart) Media (SMC), a secure digital (SD) card, Flash card (Flash Card), etc. Further, the memory 61 may also include both an internal storage unit of the server 6 and an external storage device.
  • the memory 61 is used to store the computer-readable instructions and other programs and data required by the server.
  • the memory 61 can also be used to temporarily store data that has been or will be output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Development Economics (AREA)
  • Technology Law (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

一种基于神经网络的车辆定损方法、服务器(500)及非易失性可读存储介质，适用于人工智能技术领域，该方法包括：获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像(S11)；通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量(S12)；在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率(S13)；将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别(S14)，从而实现了车辆定损的智能化，节省了人工成本，提高了车辆定损的准确性。

Description

一种基于神经网络的车辆定损方法、服务器及介质
本申请申明享有2018年10月11日递交的申请号为201811182147.1、名称为“一种基于神经网络的车辆定损方法、服务器及介质”中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请属于人工智能技术领域，尤其涉及一种基于神经网络的车辆定损方法、服务器及非易失性可读存储介质。
背景技术
在车险理赔过程中，通常需要先对事故车辆进行定损，再基于定损结果确定事故车辆的赔偿金额。现有技术通常是由车辆定损人员依据自己的既往经验对事故车辆进行人工定损，而不同定损人员的定损标准及经验丰富程度不一，导致最终得到的定损结果的准确率较低，且人工进行车辆定损的进度较慢，人工成本较高。
技术问题
本申请实施例提供了一种基于神经网络的车辆定损方法、服务器及非易失性可读存储介质，以解决现有的车辆定损方法存在的定损准确率较低且人工成本高的问题。
技术解决方案
本申请实施例的第一方面提供了一种基于神经网络的车辆定损方法，包括：
获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像；
通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量；
在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率；
将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
有益效果
本申请通过获取从事故车辆的各个预设方位拍摄得到的定损评估图像，通过预设神经网络模型的特征提取层确定各个定损评估图像的特征向量，通过预设神经网络模型的概率计算层基于所有定损评估图像的特征向量，确定事故车辆的损伤级别概率向量，将损伤级别概率 向量中值最大的元素对应的预设损伤级别确定为事故车辆的损伤级别，从而实现了车辆定损的智能化，节省了人工成本，同时，由于从事故车辆的各个预设方位拍摄得到的定损评估图像能够反映事故车辆的整体受损情况，因此本申请实施例基于从事故车辆的各个预设方位拍摄得到的定损评估图像的特征向量综合确定事故车辆的损伤级别，提高了车辆定损的准确性。
附图说明
图1是本申请第一实施例提供的一种基于神经网络的车辆定损方法的实现流程图；
图2是本申请第二实施例提供的一种基于神经网络的车辆定损方法中S13的具体实现流程图；
图3是本申请第三实施例提供的一种基于神经网络的车辆定损方法的实现流程图；
图4是本申请第四实施例提供的一种基于神经网络的车辆定损方法中S03的具体实现流程图；
图5是本申请实施例提供的一种服务器的结构框图；
图6是本申请另一实施例提供的一种服务器的结构框图。
本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
请参阅图1，图1是本申请第一实施例提供的一种基于神经网络的车辆定损方法的实现流程图。本实施例中，基于神经网络的车辆定损方法的执行主体为服务器。如图1所示的基于神经网络的车辆定损方法包括以下步骤：
S11：获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像。
在车险理赔过程中，需要先对发生事故的事故车辆进行定损，再基于定损结果确定事故车辆的赔偿金额。对事故车辆进行定损的过程即为确定事故车辆的损伤级别的过程。其中，损伤级别用于描述车辆的损伤程度。在实际应用中，可以基于车辆的损伤程度来定义损伤级别，定义的损伤级别即为预设损伤级别。例如，可以将损伤程度为10％～30％的损伤级别定义为轻度损伤，将损伤程度为30％～50％的损伤级别定义为中度损伤，将损伤程度大于50％的损伤级别定义为重度损伤，那么，预设损伤级别即包括：轻度损伤、中度损伤及重度损伤。
在发生车辆事故后，车主或车辆定损人员可以通过手机、平板电脑等终端设备上安装的指定应用程序(application，APP)向服务器发送车辆定损请求。该车辆定损请求中携带待定损的事故车辆的车辆标识和定损评估图像序列。
车辆标识可以是车牌号码、车辆识别号码(Vehicle Identification Number，VIN)或发动机号码等，车辆标识可由车主或定损人员手动输入至终端设备中。
定损评估图像序列包括从事故车辆的各个预设方位对事故车辆进行拍摄得到的定损评估图像。预设方位可以根据实际需求设置，此处不做限制，例如，预设方位可以包括但不限于：正前方、左前方、左侧方、左后方、正后方、右后方、右侧方及右前方。即车主或定损人员可以从事故车辆的正前方、左前方、左侧方、左后方、正后方、右后方、右侧方及右前方分别对事故车辆进行拍照，进而得到用于对事故车辆进行定损评估的多张定损评估图像，分别为事故车辆的正前方图像、左前方图像、左侧方图像、左后方图像、正后方图像、右后方图像、右侧方图像及右前方图像，这些定损评估图像即构成定损评估图像序列。
本申请实施例中，服务器接收到终端设备发送的车辆定损请求后，可以从该车辆定损请求中获取待定损的事故车辆的车辆标识和定损评估图像序列。
S12：通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量。
预设神经网络模型是基于预设数量的样本数据，通过机器学习算法对预先构建的原始神经网络模型进行训练得到的。样本数据中的每条数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成。
原始神经网络包括依次连接的特征提取层和概率计算层。其中：
特征提取层用于提取定损评估图像的特征向量，特征提取层由至少一个卷积层构成，每个卷积层对应一个预设卷积核，预设卷积核用于与定损评估图像对应的图像矩阵进行卷积操作，以提取定损评估图像的特征向量。预设卷积核的卷积核参数需要在对原始神经网络模型的训练中学习得到。
概率计算层用于基于特征提取层输出的每一事故车辆的所有定损评估图像的特征向量，计算每一事故车辆的损伤级别概率向量。需要说明的是，损伤级别概率向量中包含的元素个数与预设损伤级别的个数相同，损伤级别概率向量中的每个元素的值用于标识事故车辆属于该元素对应的预设损伤级别的概率。对于每一预设损伤级别，在概率计算层均预先构建有一个概率计算函数，概率计算函数的自变量为各个定损评估图像的特征向量，概率计算函数的因变量为事故车辆属于该概率计算函数对应的预设损伤级别的概率，每一概率计算函数中的每一自变量均对应一个权重系数，每一自变量对应的权重系数需要在对原始神经网络模型的训练中学习得到。
在对原始神经网络模型进行训练时，将每条样本数据中包括的事故车辆的定损评估图像序列作为原始神经网络模型的输入，将每条样本数据中包括的事故车辆的损伤级别概率向量 作为原始神经网络模型的输出，对特征提取层包含的各个预设卷积核的卷积核参数以及概率计算层包含的各个概率计算函数中的自变量对应的权重系数进行学习，进而完成对原始神经网络模型的训练，完成训练的原始神经网络模型即为本申请实施例中的预设神经网络模型。
本申请实施例中，服务器在获取到待定损的事故车辆的定损评估图像序列后，将定损评估图像序列中的所有定损评估图像导入预设神经网络模型中。服务器通过预设神经网络模型的特征提取层对定损评估图像序列中的各个定损评估图像进行特征提取，得到各个定损评估图像的特征向量。
作为本申请一实施例，S12具体可以包括以下步骤：
在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系，确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
定损评估图像是由多个像素点排列而成的二维矩阵，定损评估图像中各像素点的位置信息用于描述各像素点在该二维矩阵中所处的行列次序。定损评估图像中的每个像素点均对应一个三维像素值，该三维像素值包含像素点在R、G、B三个颜色通道上的值。本申请实施例中，服务器在预设神经网络的特征提取层基于定损评估图像中个像素点的位置信息与像素值的对应关系，将定损评估图像中每个像素点的像素值作为图像矩阵中与该像素点处于同一位置的元素的值，进而得到定损评估图像对应的图像矩阵。需要说明的是，定损评估图像对应的图像矩阵中的每一元素均通过一个三维像素值表示。
服务器得到定损评估图像对应的图像矩阵后，将定损评估图像对应的图像矩阵与预设卷积核进行卷积运算，进而得到定损评估图像的特征向量。
服务器将定损评估图像对应的图像矩阵与预设卷积核进行卷积运算的具体过程可以为：采用预设卷积核在图像矩阵上以预设步长从左至右、从上至下进行滑动，在每个滑动到的位置上，将预设卷积核与图像矩阵中相应位置的元素构成的子矩阵进行相乘，将相乘结果作为定损评估图像的特征向量中相应位置的元素的值，预设卷积核在定损评估图像对应的图像矩阵上滑动完成后，定损评估图像的特征向量中各元素的值则均已确定。
S13：在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率。
本申请实施例中，服务器通过预设神经网络模型的特征提取层提取出事故车辆的各个定损评估图像的特征向量后，将事故车辆的所有定损评估图像的特征向量作为预设神经网络模型的概率计算层的输入，在预设神经网络模型的概率计算层基于事故车辆的所有定损评估图 像的特征向量，确定事故车辆的损伤级别概率向量。需要说明的是，损伤级别概率向量中的每个元素的值用于标识事故车辆属于该元素对应的预设损伤级别的概率，损伤级别概率向量中所有元素的值之和为1。
作为本申请一实施例，S13可以通过如图2所示的S131～S132实现：
S131：在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数。
S132：分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
本申请实施例中，在通过预设数量的样本数据对预设神经网络模型进行训练时，服务器会在预设神经网络模型的概率计算层学习到样本数据中包含的每一定损评估图像的特征向量相对于各个预设损伤级别的权重，在样本数据的数据量足够大的情况下，服务器可以学习到所有可能的定损评估图像的特征向量相对于各个预设损伤级别的权重，本申请实施例将所有可能的定损评估图像识别为预设定损评估图像，即预设定损评估图像包含了所有可能的定损评估图像。
服务器通过预设神经网络模型的特征提取层提取出事故车辆的各个定损评估图像的特征向量后，在预设神经网络模型的概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个预设损伤级别的权重系数，确定待定损的事故车辆的各个定损评估图像的特征向量相对于各个预设损伤级别的权重系数，并分别基于事故车辆的各个定损评估图像的特征向量相对于每一损伤级别的权重系数，将事故车辆的各个定损评估图像的特征向量进行加权求和，得到事故车辆的损伤级别概率向量。
示例性的，可以将事故车辆的各个定损评估图像的特征向量分别表示为：x 1、x 2、x 3、……、x 8，预设损伤级别包含的轻度损伤、中度损伤及重度损伤分别表示为y 1、y 2及y 3。若服务器预先学习到的x 1、x 2、x 3、……、x 8相对于y 1的权重系数分别为a 11、a 21、a 31、……、a 81；预先学习到的x 1、x 2、x 3、……、x 8相对于y 2的权重系数分别为a 12、a 22、a 32、……、a 82；预先学习到的x 1、x 2、x 3、……、x 8相对于y 3的权重系数分别为a 13、a 23、a 33、……、a 83；那么，基于各个定损评估图像的特征向量相对于每一损伤级别的权重系数，将事故车辆的各个定损评估图像的特征向量进行加权求和得到的损伤级别概率向量为：
Figure PCTCN2018124307-appb-000001
其中，x 1a 11+x 2a 21+x 3a 31+...+x 8a 81表示事故车辆属于损伤级别y1的概率；
x 1a 12+x 2a 22+x 3a 32+...+x 8a 82表示事故车辆属于损伤级别y2的概率；
x 1a 13+x 2a 23+x 3a 33+...+x 8a 83表示事故车辆属于损伤级别y3的概率。
S14：将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
损伤级别概率向量中某一元素的值越大，表明事故车辆属于该元素对应的损伤级别的概率越大。因此，本申请实施例中，服务器得到事故车辆的损伤级别概率向量后，将损伤级别概率向量中值最大的元素对应的预设损伤级别确定为事故车辆的损伤级别。
以上可以看出，本实施例提供的一种基于神经网络的车辆定损方法通过获取从事故车辆的各个预设方位拍摄得到的定损评估图像，通过预设神经网络模型的特征提取层确定各个定损评估图像的特征向量，通过预设神经网络模型的概率计算层基于所有定损评估图像的特征向量，确定事故车辆的损伤级别概率向量，将损伤级别概率向量中值最大的元素对应的预设损伤级别确定为事故车辆的损伤级别，从而实现了车辆定损的智能化，节省了人工成本，同时，由于从事故车辆的各个预设方位拍摄得到的定损评估图像能够反映事故车辆的整体受损情况，因此本申请实施例基于从事故车辆的各个预设方位拍摄得到的定损评估图像的特征向量综合确定事故车辆的损伤级别，提高了车辆定损的准确性。
请参阅图3，图3是本申请第三实施例提供的一种基于神经网络的车辆定损方法的实现流程图。相对于图1对应的实施例，本实施例提供的一种基于神经网络的车辆定损方法在S11之前，还可以包括S01～S104，详述如下：
S01：获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成。
在确定待定损的事故车辆的损伤级别之前，需要先构建原始神经网络模型。原始神经网络包括依次连接的特征提取层和概率计算层。特征提取层和概率计算层的具体结构及原理请参照第一实施例S12中的相关描述，此处不再赘述。
在构建好原始神经网络模型后，服务器获取预设的样本数据集。其中，样本数据集中的 每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成。可以理解的是，每条样本数据中包含的事故车辆的损伤级别概率向量可以是由人工对事故车辆进行定损评估得到的。
服务器获取到预设的样本数据集后，可以基于预设分配比例将样本数据集分为训练集和测试集。训练集用于对原始神经网络模型进行训练，测试集用于对已完成训练的原始神经网络模型的准确度进行校验。预设分配比例可以根据实际需求设置，此处不做限制，例如，预设分配比例可以为：训练集：测试集＝3:1。即样本数据集中3/4的样本数据用于训练原始神经网络模型，1/4的样本数据用于对已完成训练的原始神经网络模型的准确度进行校验。
S02：基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数。
本实施例中，服务器基于训练集对预先构建的原始神经网络模型进行训练，在对原始神经网络模型进行训练时，将训练集中每条样本数据包含的事故车辆的定损评估图像序列作为原始神经网络模型的输入，将训练集中每条样本数据包含的事故车辆的损伤级别概率向量作为原始神经网络模型的输出，确定原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个预设损伤级别的权重系数，即服务器基于训练集对特征提取层包含的各个预设卷积核的卷积核参数以及概率计算层包含的各个预设定损评估图像的特征向量相对于各个预设损伤级别的权重系数进行学习，进而完成对原始神经网络模型的训练。
S03：基于所述测试集对已完成训练的所述原始神经网络模型进行验证。
服务器基于训练集完成对原始神经网络模型的训练后，基于测试集对已完成训练的原始神经网络模型进行验证。
具体的，S03可以通过如图4所示的S031～S033实现，详述如下：
S031：将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值。
本实施例中，服务器在基于测试集对已完成训练的原始神经网络模型进行验证时，将测试集中每条样本数据包含的事故车辆的定损评估图像序列导作为已完成训练的原始神经网络模型的输入，以通过已完成训练的原始神经网络模型确定测试集中每条样本数据各自对应的损伤级别概率向量的预测值。
S032：基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
Figure PCTCN2018124307-appb-000002
其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
Figure PCTCN2018124307-appb-000003
为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
Figure PCTCN2018124307-appb-000004
为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值。
本实施例中，服务器确定出测试集中每条样本数据各自对应的损伤级别概率向量的预测值后，将测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条样本数据对应的损伤级别概率向量的预测值代入上述公式中，计算已完成训练的原始神经网络模型的预测误差。
已完成训练的原始神经网络模型的预测误差用于标识已完成训练的原始神经网络模型的车辆定损准确度。其中，已完成训练的原始神经网络模型的预测误差值越大，表明已完成训练的原始神经网络模型的车辆定损准确度越低。
S033：将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
本实施例中，服务器得到已完成训练的原始神经网络模型的预测误差后，将已完成训练的原始神经网络模型的预测误差与预设误差阈值进行比较，并基于比较结果确定对已完成训练的原始神经网络模型的验证结果。其中，预设误差阈值为实际应用中可允许的车辆定损准确度误差值。
其中，若比较结果为已完成训练的原始神经网络模型的预测误差小于或等于所述预设误差阈值，则说明已完成训练的原始神经网络模型的车辆定损准确度在可允许的误差范围内，此时服务器将对已完成训练的原始神经网络模型的验证结果确定为验证通过；若比较结果为已完成训练的原始神经网络模型的预测误差大于预设误差阈值，则说明已完成训练的原始神经网络模型的车辆定损准确度超过了可允许的误差范围，此时服务器将对已完成训练的原始 神经网络模型的验证结果确定为验证未通过。
S04：若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
本实施例中，服务器若检测到对已完成训练的原始神经网络模型的验证通过，则将已完成训练的原始神经网络模型确定为预设神经网络模型。
作为本申请另一实施例，服务器若检测到对已完成训练的原始神经网络模型的验证未通过，则通过反向传播算法对原始神经网络模型的特征提取层包含的预设卷积核的卷积核参数和/或对概率计算层包含的各个预设定损评估图像的特征向量相对于各个预设损伤级别的权重系数进行调整，并基于测试集对参数调整后的原始神将网络模型进行再次验证，直至验证通过，将调整参数后的原始神经网络模型确定为预设神经网络模型。
以上可以看出，本实施提供的一种基于神经网络的车辆定损方法通过包含一定数量的样本数据的训练集对预先构建的原始神经网络模型进行训练，并通过包含一定数量的样本数据的测试集对已完成训练的原始神经网络模型的车辆定损准确度进行验证，在验证通过后，才将已完成训练的原始神经网络模型作为后续用于确定事故车辆的损伤级别的预设神经网络模型，从而提高了车辆定损的准确度。
请参阅图5，图5是本申请实施例提供的一种服务器的结构框图。本实施例中的服务器为服务器。该服务器包括的各单元用于执行图1至图4对应的实施例中的各步骤。具体请参阅图1至图4以及图1至图4所对应的实施例中的相关描述。为了便于说明，仅示出了与本实施例相关的部分。参见图5，服务器500包括：第一获取单元51、特征提取单元52、第一确定单元53及第二确定单元54。其中：
第一获取单元51用于获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像。
特征提取单元52用于通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量。
第一确定单元53用于在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率。
第二确定单元54用于将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
作为本申请一实施例，特征提取单元52具体用于：
在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系， 确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
作为本申请一实施例，第一确定单元53包括：权重确定单元及概率确定单元。其中：
权重确定单元用于在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数。
概率确定单元用于分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
作为本申请一实施，服务器500还包括：第二获取单元、训练单元、验证单元及第三确定单元。其中：
第二获取单元用于获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成。
训练单元用于基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数。
验证单元用于基于所述测试集对已完成训练的所述原始神经网络模型进行验证。
第三确定单元用于若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
作为本申请一实施，验证单元包括：预测单元、误差计算单元及第四确定单元。其中：
预测单元用于将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值。
误差计算单元用于基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
Figure PCTCN2018124307-appb-000005
其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
Figure PCTCN2018124307-appb-000006
为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
Figure PCTCN2018124307-appb-000007
为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值。
第四确定单元用于将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
以上可以看出，本实施例提供的一种服务器通过获取从事故车辆的各个预设方位拍摄得到的定损评估图像，通过预设神经网络模型的特征提取层确定各个定损评估图像的特征向量，通过预设神经网络模型的概率计算层基于所有定损评估图像的特征向量，确定事故车辆的损伤级别概率向量，将损伤级别概率向量中值最大的元素对应的预设损伤级别确定为事故车辆的损伤级别，从而实现了车辆定损的智能化，节省了人工成本，同时，由于从事故车辆的各个预设方位拍摄得到的定损评估图像能够反映事故车辆的整体受损情况，因此本申请实施例基于从事故车辆的各个预设方位拍摄得到的定损评估图像的特征向量综合确定事故车辆的损伤级别，提高了车辆定损的准确性。
图6是本申请另一实施例提供的一种服务器的结构框图。如图6所示，该实施例的服务器6包括：处理器60、存储器61以及存储在所述存储器61中并可在所述处理器60上运行的计算机可读指令62，例如基于神经网络的车辆定损方法的程序。处理器60执行所述计算机可读指令62时实现上述各个基于神经网络的车辆定损方法各实施例中的步骤，例如图1所示的S11至S14。或者，所述处理器60执行所述计算机可读指令62时实现上述图5对应的实施例中各单元的功能，例如，图5所示的单元51至54的功能，具体请参阅图5对应的实施例中的相关描述，此处不赘述。
示例性的，所述计算机可读指令62可以被分割成一个或多个单元，所述一个或者多个单元被存储在所述存储器61中，并由所述处理器60执行，以完成本申请。所述一个或多个单元可以是能够完成特定功能的一系列计算机可读指令段，该指令段用于描述所述计算机可读指令62在所述服务器6中的执行过程。例如，所述计算机可读指令62可以被分割成第一获取单元、特征提取单元、第一确定单元及第二确定单元，各单元具体功能如上所述。
所述服务器可包括，但不仅限于，处理器60、存储器61。本领域技术人员可以理解，图 6仅仅是服务器6的示例，并不构成对服务器6的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述服务器还可以包括输入输出设备、网络接入设备、总线等。
所称处理器60可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器61可以是所述服务器6的内部存储单元，例如服务器6的硬盘或内存。所述存储器61也可以是所述服务器6的外部存储设备，例如所述服务器6上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器61还可以既包括所述服务器6的内部存储单元也包括外部存储设备。所述存储器61用于存储所述计算机可读指令以及所述服务器所需的其他程序和数据。所述存储器61还可以用于暂时地存储已经输出或者将要输出的数据。
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. 一种基于神经网络的车辆定损方法，其特征在于，包括：
    获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像；
    通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量；
    在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率；
    将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
  2. 根据权利要求1所述的基于神经网络的车辆定损方法，其特征在于，所述通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量，包括：
    在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系，确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
  3. 根据权利要求1所述的基于神经网络的车辆定损方法，其特征在于，所述在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量，包括：
    在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
  4. 根据权利要求1-3任一项所述的基于神经网络的车辆定损方法，其特征在于，所述获取待定损的事故车辆的定损评估图像序列之前，还包括：
    获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成；
    基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    基于所述测试集对已完成训练的所述原始神经网络模型进行验证；
    若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
  5. 根据权利要求4所述的基于神经网络的车辆定损方法，其特征在于，所述基于所述测试集对已完成训练的所述原始神经网络模型进行验证，包括：
    将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值；
    基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
    Figure PCTCN2018124307-appb-100001
    其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
    Figure PCTCN2018124307-appb-100002
    为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
    Figure PCTCN2018124307-appb-100003
    为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值；
    将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果 为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
  6. 一种服务器，其特征在于，包括：
    第一获取单元，用于获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像；
    特征提取单元，用于通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量；
    第一确定单元，用于在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率；
    第二确定单元，用于将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
  7. 根据权利要求6所述的服务器，其特征在于，所述特征提取单元具体用于：
    在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系，确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
  8. 根据权利要求6所述的服务器，其特征在于，所述第一确定单元包括：
    权重确定单元，用于在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    概率确定单元，用于分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
  9. 根据权利要求6-8任一项所述的服务器，其特征在于，还包括：
    第二获取单元，用于获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成；
    训练单元，用于基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述 原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    验证单元，用于基于所述测试集对已完成训练的所述原始神经网络模型进行验证；
    第三确定单元，用于若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
  10. 根据权利要求9所述的服务器，其特征在于，所述验证单元包括：
    预测单元，用于将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值；
    误差计算单元，用于基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
    Figure PCTCN2018124307-appb-100004
    其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
    Figure PCTCN2018124307-appb-100005
    为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
    Figure PCTCN2018124307-appb-100006
    为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值；
    第四确定单元，用于将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
  11. 一种服务器，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，其特征在于，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故 车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像；
    通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量；
    在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率；
    将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
  12. 根据权利要求11所述的服务器，其特征在于，所述通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量，包括：
    在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系，确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
  13. 根据权利要求11所述的服务器，其特征在于，所述在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量，包括：
    在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
  14. 根据权利要求13所述的服务器，其特征在于，所述获取待定损的事故车辆的定损评估图像序列之前，还包括：
    获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向 量构成；
    基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    基于所述测试集对已完成训练的所述原始神经网络模型进行验证；
    若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
  15. 根据权利要求11-14任一项所述的服务器，其特征在于，所述基于所述测试集对已完成训练的所述原始神经网络模型进行验证，包括：
    将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值；
    基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
    Figure PCTCN2018124307-appb-100007
    其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
    Figure PCTCN2018124307-appb-100008
    为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
    Figure PCTCN2018124307-appb-100009
    为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值；
    将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
  16. 一种非易失性可读存储介质，所述非易失性可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如下步骤：
    获取待定损的事故车辆的定损评估图像序列；所述定损评估图像序列包括从所述事故车辆的各个预设方位对所述事故车辆进行拍摄得到的定损评估图像；
    通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量；
    在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量；所述损伤级别概率向量中的每个元素的值用于标识所述事故车辆属于该元素对应的预设损伤级别的概率；
    将所述损伤级别概率向量中值最大的元素对应的预设损伤级别确定为所述事故车辆的损伤级别。
  17. 根据权利要求16所述的非易失性可读存储介质，其特征在于，所述通过预设神经网络模型的特征提取层对所述定损评估图像序列中的各个所述定损评估图像进行特征提取，得到各个所述定损评估图像的特征向量，包括：
    在所述特征提取层基于所述定损评估图像中各像素点的位置信息与像素值的对应关系，确定所述定损评估图像对应的图像矩阵，并通过预设卷积核对所述图像矩阵进行卷积处理，得到所述定损评估图像的特征向量。
  18. 根据权利要求16所述的非易失性可读存储介质，其特征在于，所述在所述预设神经网络模型的概率计算层基于所有所述定损评估图像的特征向量，确定所述事故车辆的损伤级别概率向量，包括：
    在所述概率计算层基于预先学习到的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数，确定所述事故车辆的各个所述定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    分别基于所述事故车辆的各个所述定损评估图像的特征向量相对于每一所述预设损伤级别的权重系数，将所述事故车辆的各个所述定损评估图像的特征向量进行加权求和，得到所述事故车辆的损伤级别概率向量。
  19. 根据权利要求16-18任一项所述的非易失性可读存储介质，其特征在于，所述获取待定损的事故车辆的定损评估图像序列之前，还包括：
    获取预设的样本数据集，并将所述样本数据集划分为训练集和测试集；所述样本数据 集中的每条样本数据均由一事故车辆的定损评估图像序列及该事故车辆的损伤级别概率向量构成；
    基于所述训练集对预先构建的原始神经网络模型进行训练，确定所述原始神经网络模型的特征提取层所包含的预设卷积核的卷积核参数，以及确定所述原始神经网络模型的概率计算层所包含的各个预设定损评估图像的特征向量相对于各个所述预设损伤级别的权重系数；
    基于所述测试集对已完成训练的所述原始神经网络模型进行验证；
    若验证通过，则将已完成训练的所述原始神经网络模型确定为所述预设神经网络模型。
  20. 根据权利要求19所述的非易失性可读存储介质，其特征在于，所述基于所述测试集对已完成训练的所述原始神经网络模型进行验证，包括：
    将所述测试集中每条样本数据包含的事故车辆的定损评估图像序列导入已完成训练的所述原始神经网络模型中，得到所述测试集中每条所述样本数据各自对应的损伤级别概率向量的预测值；
    基于所述测试集中每条样本数据包含的事故车辆的损伤级别概率向量及每条所述样本数据对应的损伤级别概率向量的预测值，通过以下公式计算已训练完成的所述原始神经网络模型的预测误差：
    Figure PCTCN2018124307-appb-100010
    其中，Error(value predictive,value actual)为已训练完成的所述原始神经网络模型的预测误差，n为所述损伤级别概率向量中所包含的元素个数，
    Figure PCTCN2018124307-appb-100011
    为所述样本数据包含的事故车辆的损伤级别概率向量中第i个元素的值，
    Figure PCTCN2018124307-appb-100012
    为所述样本数据对应的损伤级别概率向量的预测值中第i个元素的值；
    将所述原始神经网络模型的预测误差与预设误差阈值进行比较，基于比较结果确定对所述原始神经网络模型的验证结果；其中，若所述比较结果为所述原始神经网络模型的预测误差小于或等于所述预设误差阈值，则确定所述验证结果为验证通过；若所述比较结果为所述原始神经网络模型的预测误差大于所述预设误差阈值，则确定所述验证结果为验证未通过。
PCT/CN2018/124307 2018-10-11 2018-12-27 一种基于神经网络的车辆定损方法、服务器及介质 Ceased WO2020073510A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811182147.1A CN109215027B (zh) 2018-10-11 2018-10-11 一种基于神经网络的车辆定损方法、服务器及介质
CN201811182147.1 2018-10-11

Publications (1)

Publication Number Publication Date
WO2020073510A1 true WO2020073510A1 (zh) 2020-04-16

Family

ID=64979646

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/124307 Ceased WO2020073510A1 (zh) 2018-10-11 2018-12-27 一种基于神经网络的车辆定损方法、服务器及介质

Country Status (2)

Country Link
CN (1) CN109215027B (zh)
WO (1) WO2020073510A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114037320A (zh) * 2021-11-18 2022-02-11 北京城市系统工程研究中心 一种基于神经网络模型的地下管网运行能力评估方法
CN114186586A (zh) * 2021-12-08 2022-03-15 华中科技大学 基于二维卷积神经网络的损伤识别方法及设备
CN116434047A (zh) * 2023-03-29 2023-07-14 邦邦汽车销售服务(北京)有限公司 基于数据处理的车辆损伤范围确定方法及系统

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948811A (zh) * 2019-01-31 2019-06-28 德联易控科技(北京)有限公司 车辆定损的处理方法、装置及电子设备
CN110569865B (zh) * 2019-02-01 2020-07-17 阿里巴巴集团控股有限公司 识别车体方向的方法和装置
CN111582009B (zh) * 2019-02-19 2023-09-15 富士通株式会社 训练分类模型的装置和方法及利用分类模型分类的装置
CN110135437B (zh) 2019-05-06 2022-04-05 北京百度网讯科技有限公司 用于车辆的定损方法、装置、电子设备和计算机存储介质
CN111275121B (zh) * 2020-01-23 2023-07-18 北京康夫子健康技术有限公司 一种医学影像处理方法、装置和电子设备
CN111680746B (zh) * 2020-06-08 2023-08-04 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
CN112085610B (zh) * 2020-09-07 2023-08-22 中国平安财产保险股份有限公司 目标物定损方法、装置、电子设备及计算机可读存储介质
CN112329596B (zh) * 2020-11-02 2021-08-24 中国平安财产保险股份有限公司 目标物定损方法、装置、电子设备及计算机可读存储介质
CN112906139A (zh) * 2021-04-08 2021-06-04 平安科技（深圳）有限公司 车辆故障风险评估方法、装置、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106127747A (zh) * 2016-06-17 2016-11-16 史方 基于深度学习的汽车表面损伤分类方法及装置
CN107358596A (zh) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置、电子设备及系统
CN108256720A (zh) * 2017-11-07 2018-07-06 中国平安财产保险股份有限公司 一种保险理赔风险评估方法及终端设备
CN108399382A (zh) * 2018-02-13 2018-08-14 阿里巴巴集团控股有限公司 车险图像处理方法和装置
CN108446618A (zh) * 2018-03-09 2018-08-24 平安科技（深圳）有限公司 车辆定损方法、装置、电子设备及存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106021548A (zh) * 2016-05-27 2016-10-12 大连楼兰科技股份有限公司 基于分布式人工智能图像识别的远程定损方法及系统
CN107403424B (zh) * 2017-04-11 2020-09-18 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置及电子设备
CN107730485B (zh) * 2017-08-03 2020-04-10 深圳壹账通智能科技有限公司 车辆定损方法、电子设备及计算机可读存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106127747A (zh) * 2016-06-17 2016-11-16 史方 基于深度学习的汽车表面损伤分类方法及装置
CN107358596A (zh) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置、电子设备及系统
CN108256720A (zh) * 2017-11-07 2018-07-06 中国平安财产保险股份有限公司 一种保险理赔风险评估方法及终端设备
CN108399382A (zh) * 2018-02-13 2018-08-14 阿里巴巴集团控股有限公司 车险图像处理方法和装置
CN108446618A (zh) * 2018-03-09 2018-08-24 平安科技（深圳）有限公司 车辆定损方法、装置、电子设备及存储介质

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114037320A (zh) * 2021-11-18 2022-02-11 北京城市系统工程研究中心 一种基于神经网络模型的地下管网运行能力评估方法
CN114186586A (zh) * 2021-12-08 2022-03-15 华中科技大学 基于二维卷积神经网络的损伤识别方法及设备
CN116434047A (zh) * 2023-03-29 2023-07-14 邦邦汽车销售服务(北京)有限公司 基于数据处理的车辆损伤范围确定方法及系统
CN116434047B (zh) * 2023-03-29 2024-01-09 邦邦汽车销售服务(北京)有限公司 基于数据处理的车辆损伤范围确定方法及系统

Also Published As

Publication number Publication date
CN109215027A (zh) 2019-01-15
CN109215027B (zh) 2024-05-24

Similar Documents

Publication Publication Date Title
WO2020073510A1 (zh) 一种基于神经网络的车辆定损方法、服务器及介质
CN111340077B (zh) 基于注意力机制的视差图获取方法和装置
US12530738B2 (en) Neural network training method, image processing method, and apparatus
US9390475B2 (en) Backlight detection method and device
CN109840477B (zh) 基于特征变换的受遮挡人脸识别方法及装置
CN110765860A (zh) 摔倒判定方法、装置、计算机设备及存储介质
CN114387353B (zh) 一种相机标定方法、标定装置及计算机可读存储介质
CN111079764A (zh) 一种基于深度学习的低照度车牌图像识别方法及装置
CN110956149B (zh) 宠物身份核验方法、装置、设备及计算机可读存储介质
CN109948439B (zh) 一种活体检测方法、系统及终端设备
CN109359542B (zh) 基于神经网络的车辆损伤级别的确定方法及终端设备
US12482059B2 (en) Method and apparatus for generating panoramic image based on deep learning network
CN113658097A (zh) 一种眼底图像质量增强模型的训练方法及装置
CN112488053A (zh) 一种人脸识别方法、装置、机器人及存储介质
CN116524206B (zh) 目标图像的识别方法及装置
CN110458754A (zh) 图像生成方法及终端设备
CN114820779B (zh) 位姿估计方法、装置、终端设备及存储介质
CN111340722A (zh) 图像处理方法、处理装置、终端设备及可读存储介质
CN113269812B (zh) 图像预测模型的训练及应用方法、装置、设备、存储介质
CN107564013B (zh) 融合局部信息的场景分割修正方法与系统
CN114119377B (zh) 一种图像处理方法及装置
CN120125455A (zh) 图像处理方法、芯片、电子设备及存储介质
CN116958870B (zh) 一种视频特征提取方法、装置、可读存储介质及终端设备
CN112801134A (zh) 基于区块链和图像的手势识别模型训练与分发方法与装置
CN117579944A (zh) 基于饱和校正与细节增强的过曝光图像处理方法及系统

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: 18936570

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: 18936570

Country of ref document: EP

Kind code of ref document: A1