CN111488875B

CN111488875B - Vehicle insurance claim settlement loss checking method and device based on image recognition and electronic equipment

Info

Publication number: CN111488875B
Application number: CN202010585381.XA
Authority: CN
Inventors: 刘海龙; 苏孝强; 李新科; 王尧; 郭吉鹏
Original assignee: Aibao Technology Co ltd
Current assignee: Aibao Technology Co ltd
Priority date: 2020-06-24
Filing date: 2020-06-24
Publication date: 2021-04-13
Anticipated expiration: 2040-06-24
Also published as: CN111488875A

Abstract

The invention provides a vehicle insurance claim loss checking method and device based on image recognition and electronic equipment, and relates to the technical field of vehicle insurance claims. The method comprises the following steps: obtaining damage assessment case information, wherein the damage assessment case information comprises a damage assessment list and a damage assessment picture set; identifying vehicle appearance piece information in each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position; identifying vehicle damage information in each damage assessment picture, wherein the vehicle damage information comprises damage categories and damage mask positions; matching the component mask and the damage mask to determine a standard component damage list; mapping the standard component damage list into a standard component replacement list according to a preset maintenance logic list; and judging whether the damage assessment list is abnormal or not according to the standard component repair list. By automatically identifying the vehicle appearance piece information and the vehicle damage information in each damage assessment picture, the leakage false increase cases in damage assessment are effectively detected.

Description

Vehicle insurance claim settlement loss checking method and device based on image recognition and electronic equipment

Technical Field

The invention relates to the technical field of vehicle insurance claims, in particular to a vehicle insurance claim damage checking method and device based on image recognition and electronic equipment.

Background

The current vehicle insurance loss checking business of the insurance company is manually processed by a loss checking specialist, wherein the most important link is to check case pictures uploaded to an information system of the insurance company and check whether repair items in a loss assessment list accord with vehicle damage positions and degrees presented in the pictures one by one. The number of pictures in a car insurance case is dozens of pictures and hundreds of pictures, and the items in the loss assessment list are numerous, so that the inspection work is time-consuming and labor-consuming even for experienced loss checking personnel, and in order to ensure the time limit requirement of business processing, the inspection of some pictures is often not required to be given up without full inspection, so that the risk of virtual increase and loss expansion in loss assessment cannot be well controlled, and the loss is brought to insurance companies for settlement.

In addition, the car insurance case picture comprises various contents of an accident scene picture, a car damage picture, a person injury picture, a certificate picture and a document picture, and although various insurance companies have normative requirements on the uploading sequence of the pictures in the link of investigating and determining the damage, the normative of the uploaded pictures can not be ensured due to various reasons in actual production and implementation. This also leads to a core loss, where the various images faced by the staff are still disordered, and in a double-car case or a multi-car case (the number of which accounts for more than 60% of the total number of cases), the images of the cars of each party of the accident are mixed together, which further increases the difficulty of work.

In view of this, a machine-assisted manual mode needs to be introduced to improve efficiency in vehicle insurance loss checking, the machine performs comprehensive inspection and checking on all case pictures and loss assessment repair projects to prompt risks, and workers only need to pay attention to and check cases and specific repair projects with machine prompt risks.

Disclosure of Invention

In order to solve at least one technical problem in the prior art, the invention provides a claim settlement and loss checking method based on image recognition, which automatically classifies case pictures according to image contents and selects pictures of each accident car by detecting and recognizing license plates of pictures containing cars and calculating a characteristic feature vector of each picture containing cars; by automatically identifying the vehicle damage information and the vehicle appearance piece information in each damage assessment picture, risk items which are possibly added in the damage assessment change list are automatically prompted to workers, and corresponding picture evidences are provided for the workers to refer to.

According to one aspect of the invention, an automobile insurance claim damage checking method based on image recognition comprises the following steps: obtaining damage assessment case information of a damage assessment case, wherein the damage assessment case information comprises a damage assessment list and a damage assessment picture set; identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position, and the vehicle damage information comprises a damage type and a damage mask position; matching the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises damaged component names and corresponding damage types; mapping the standard component damage list into a standard component replacement list according to a preset maintenance logic list, wherein the standard component replacement list comprises damaged component names and corresponding replacement information; and judging whether the damage assessment list is abnormal or not according to the standard component repair list.

Further optionally, before the mapping the standard component damage list into a standard component repair list according to a preset repair logic list, the method further includes: determining a component missing list according to the vehicle appearance piece information and a preset component adjacency relation; the mapping of the standard component damage list to a standard component replacement list according to a preset maintenance logic list is as follows: and mapping the standard component damage list and the component missing list into the standard component replacement list according to a preset maintenance logic list, wherein the standard component damage list further comprises a missing component name and a corresponding missing state mark.

Further optionally, the method for identifying the vehicle exterior piece information in each damage assessment picture further includes: determining vehicle direction information in each loss assessment picture; and correcting the vehicle appearance piece information identified in each damage assessment picture according to the vehicle direction information, wherein the vehicle direction information is used for distinguishing similar vehicle appearance pieces at different positions.

Further optionally, the damage assessment case is a double-vehicle or multi-vehicle case, the damage assessment picture set includes a vehicle-containing picture and a vehicle-free picture, and the vehicle-containing picture includes a license plate-containing picture and a license plate-free picture; the vehicle appearance piece information and the vehicle damage information in each damage assessment picture are identified by the following steps: determining a vehicle frame picture set corresponding to each vehicle from the damage assessment picture set; and acquiring the vehicle appearance piece information and the vehicle damage information of each vehicle by using the vehicle frame picture set corresponding to each vehicle.

Further optionally, the determining, from the damage assessment picture set, a vehicle frame picture set corresponding to each vehicle includes: sequentially carrying out license plate detection and recognition and vehicle frame detection on each image containing the vehicle to determine the corresponding relation between each image containing the license plate and license plate information; sequentially determining the corresponding relation between each license plate-free picture and the license plate information through a deep learning network by utilizing the color information and the type information of the vehicle body; and determining a vehicle frame image set corresponding to each vehicle according to the corresponding relation between each image containing the license plate and the license plate information and the corresponding relation between each image without the license plate and the license plate information.

Further optionally, before the license plate detection and recognition and the vehicle frame detection are sequentially performed on each image containing the license plate to determine the corresponding relationship between each image containing the license plate and the license plate information, the method further includes: carrying out picture sorting on the loss assessment picture set to obtain a picture containing a car and a picture without a car; sequentially judging whether each car-containing picture is a medium-long scene picture or not; if yes, marking the picture containing the vehicle as a medium-long-range picture; if not, the picture containing the car is marked as a close-range picture.

Further optionally, the method for sequentially performing license plate detection and recognition and vehicle frame detection on each image containing the vehicle to determine the corresponding relationship between each image containing the license plate and the license plate information includes: sequentially carrying out license plate detection on each picture containing the vehicle, and sorting the pictures containing the vehicle into pictures containing license plates and pictures without license plates; sequentially carrying out vehicle frame detection and license plate identification on each image containing the license plate; the vehicle frame detection is that vehicle image region detection is carried out on the picture containing the license plate marked as the medium-long-range view picture, a vehicle image region is formed by cutting, and a vehicle image region is directly formed on the picture containing the license plate marked as the close-range view picture; performing geometric matching on the vehicle image area and the license plate detection result to obtain at least one first vehicle frame picture and/or second vehicle frame picture; sequentially detecting the vehicle frame of each picture without the license plate; and the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the medium-distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and the second vehicle frame picture is directly formed on the picture containing the license plate marked as the close-range view picture.

Further optionally, the step of sequentially determining the corresponding relationship between the second vehicle frame picture of each license plate-free picture and the license plate information through the deep learning network by using the vehicle body color information and the vehicle type information comprises the steps of: extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network; carrying out similarity clustering on the color vehicle type characterization vectors extracted from all vehicle frame pictures; establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on the first vehicle frame picture containing the license plate picture; and establishing a corresponding relation between the license plate information and each vehicle frame picture.

Further optionally, the sequentially detecting and identifying the license plate of each image containing the license plate further comprises: identifying license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and a license plate number; and comparing and calibrating the license plate number with the real license plate number in the damage assessment case information.

According to another aspect of the invention, an image recognition-based vehicle insurance claim damage checking device comprises: the system comprises an acquisition module, a processing module and a display module, wherein the acquisition module is used for acquiring the damage assessment case information of a damage assessment case, and the damage assessment case information comprises a damage assessment list and a damage assessment picture set; the vehicle appearance piece information identification module is used for identifying vehicle appearance piece information in each damage assessment picture, and the vehicle appearance piece information comprises a component type and a component mask position; the vehicle damage information identification module is used for identifying vehicle damage information in each damage assessment picture, and the vehicle damage information comprises damage categories and damage mask positions; the matching module is used for matching the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises damaged component names and corresponding damage types; the standard component replacement and repair system comprises a mapping module, a maintenance module and a maintenance module, wherein the mapping module is used for mapping a standard component damage list into a standard component replacement and repair list according to a preset maintenance logic list, and the standard component replacement and repair list comprises damaged component names and corresponding replacement and repair information; and the judging module is used for judging whether the damage assessment list is abnormal or not according to the standard component repair list.

Further optionally, the apparatus further comprises: and the part missing identification module is used for determining a part missing list according to the vehicle appearance piece information and the preset part adjacency relation.

Further optionally, the vehicle exterior piece information identification module further includes: the vehicle direction determining submodule is used for determining vehicle direction information in each damage assessment picture; and the correction submodule is used for identifying the vehicle damage information and the vehicle appearance piece information in each damage assessment picture according to the vehicle direction information, and the vehicle direction information is used for distinguishing similar vehicle appearance pieces at different positions.

Further optionally, the apparatus further comprises: the vehicle frame identification module is used for determining a vehicle frame picture set corresponding to each vehicle from the damage assessment picture set; and the information acquisition module is used for acquiring the vehicle appearance piece information and the vehicle damage information of each vehicle by using the vehicle frame picture set corresponding to each vehicle.

Further optionally, the vehicle frame identification module further includes: the vehicle-containing picture sorting submodule is used for sorting pictures of the loss assessment picture set to obtain vehicle-containing pictures and vehicle-free pictures which are used for the vehicle-containing pictures; a medium-short distant view judgment submodule: the corresponding relation determining submodule is used for sequentially judging whether each vehicle-containing picture is a medium-long scene picture or not; if so, marking the picture containing the car as a medium-long-range picture; if not, the picture containing the car is marked as a close-range picture. The vehicle license plate containing picture corresponding relation determining sub-module is used for sequentially carrying out vehicle license plate detection and recognition and vehicle frame detection on each vehicle license plate containing picture so as to determine the corresponding relation between each vehicle license plate containing picture and the vehicle license plate information; the license plate-free picture corresponding relation determining submodule is used for sequentially determining the corresponding relation between each license plate-free picture and the license plate information through a deep learning network by utilizing the color information of the vehicle body and the vehicle type information; and the vehicle frame image set generating module is used for determining a vehicle frame image set corresponding to each vehicle according to the corresponding relation between each image containing the license plate and the license plate information and the corresponding relation between each image without the license plate and the license plate information.

Further optionally, the license plate-containing picture corresponding relation determining sub-module further includes: the license plate detection unit is used for sequentially detecting the license plate of each image containing the vehicle and sorting the images containing the vehicle into the images containing the license plate and the images without the license plate; the vehicle frame and license plate recognition detection unit is used for sequentially carrying out vehicle frame detection and license plate recognition on each image containing the license plate; the vehicle frame detection is that vehicle image areas are detected aiming at the images containing the license plates marked as the middle distant view images, the vehicle image areas are formed by cutting, and the vehicle image areas are directly formed aiming at the images containing the license plates marked as the close view images; geometrically matching the vehicle image area with the license plate detection result to obtain at least one first vehicle frame picture and/or second vehicle frame picture; identifying license plate information in a license plate containing picture, wherein the license plate information comprises license plate region information and a license plate number; the license plate-free vehicle frame detection unit is used for sequentially detecting the vehicle frames of each license plate-free picture; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the middle distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and the second vehicle frame picture is directly formed on the picture containing the license plate marked as the close view picture.

Further optionally, the sub-module for determining the corresponding relationship between the images without license plates further includes: the color vehicle type representation vector calculation unit is used for extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network; the clustering unit is used for carrying out similarity clustering on the color vehicle type characterization vectors extracted from all the vehicle frame pictures; the license plate information corresponding relation establishing unit is used for establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on the first vehicle frame picture containing the license plate picture; and the vehicle frame and license plate corresponding relation establishing unit is used for establishing the corresponding relation between the license plate information and each vehicle frame picture.

Further optionally, the license plate-containing picture corresponding relation determining sub-module further includes: the license plate recognition sub-module is used for recognizing license plate information in the image containing the license plate, and the license plate information comprises license plate region information and license plate numbers; and the comparison and calibration submodule is used for comparing and calibrating the license plate number with the real license plate number in the damage assessment case information.

According to another aspect of the present invention, an electronic device includes: at least one processor; and

a memory coupled to the at least one processor; wherein the memory stores a computer program executable by the at least one processor to implement the image recognition-based vehicle insurance claim reimbursement damage method.

The invention has the beneficial effects that:

1. the vehicle license plate detection and recognition are carried out on the images containing the vehicles, the characteristic feature vector of each image containing the vehicles is calculated, the characteristic vectors embedded with vehicle type and color information are learned through a depth measurement network, the characteristic vectors are clustered according to the similarity, the images of the vehicle frames of different vehicles in the case are automatically separated, the case images are automatically classified according to the image content, and the work efficiency of the nuclear damage and environmental protection is effectively improved.

2. The method comprises the steps of automatically identifying vehicle damage information and vehicle appearance piece information in each damage assessment picture, training two example segmentation models of component identification and damage identification, and performing geometric mapping on masks of the component identification and the damage identification to obtain a damage list. And automatically prompting the risk items which are probably false-increased in the damage assessment and repair list to the staff, and effectively detecting the leakage false-increased cases in the damage assessment.

3. And introducing direction information for correction in the identification process of the vehicle appearance piece information to obtain a correction result. And then the accuracy of vehicle outward appearance piece information identification is promoted.

4. The information of the absence of the vehicle parts in the image is determined by utilizing the inherent adjacency relation of the vehicle parts, so that the accuracy of vehicle damage information identification is further improved.

Drawings

FIG. 1 is a flow chart illustrating a vehicle insurance claim damage checking method based on image recognition according to an embodiment of the present invention;

FIG. 2 is a flow chart illustrating another vehicle insurance claim damage checking method based on image recognition according to an embodiment of the present invention;

fig. 3, 4, and 5 show schematic flow charts of one implementation of step 202, step 2023, and step 2024 in fig. 2, respectively;

FIG. 6 is a flow chart illustrating another vehicle insurance claim damage checking method based on image recognition according to an embodiment of the present invention;

fig. 7 is a functional structure diagram of another vehicle insurance claim settlement and damage checking device based on image recognition.

Detailed Description

The content of the invention will now be discussed with reference to a number of exemplary embodiments. It is to be understood that these examples are discussed only to enable those of ordinary skill in the art to better understand and thus implement the teachings of the present invention, and are not meant to imply any limitations on the scope of the invention.

As used herein, the term "include" and its variants are to be read as open-ended terms meaning "including, but not limited to. The term "based on" is to be read as "based, at least in part, on". The terms "one embodiment" and "an embodiment" are to be read as "at least one embodiment". The term "another embodiment" is to be read as "at least one other embodiment".

Example 1:

as shown in fig. 1, the embodiment provides a vehicle insurance claim damage checking method based on image recognition, which mainly includes the following steps:

101. obtaining the loss assessment case information of the loss assessment case, wherein the loss assessment case information comprises a loss assessment list and a loss assessment picture set;

in the process of acquiring the loss assessment case information, the loss assessment picture sets in the loss assessment case information need to be classified first, so that the core loss process is accelerated, and the core loss complexity is reduced.

102. And identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position, and the vehicle damage information comprises a damage type and a damage mask position.

103. And matching the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises the names of damaged components and corresponding damage types.

104. And mapping the standard component damage list into a standard component replacement list according to the preset maintenance logic list, wherein the standard component replacement list comprises the names of the damaged components and corresponding replacement information.

105. And judging whether the damage assessment list is abnormal or not according to the standard component repair list.

The technical scheme provided by the embodiment has the beneficial effects that: the method comprises the steps of automatically identifying vehicle damage information and vehicle appearance piece information in each damage assessment picture, training two example segmentation models of component identification and damage identification, and performing geometric mapping on masks of the component identification and the damage identification to obtain a damage list. Risk items which are possibly added falsely in the damage assessment and repair list are automatically prompted to the working personnel, and the detection rate of the damage assessment and leakage falsely added cases is effectively improved.

Example 2:

as shown in fig. 2, the embodiment provides a vehicle insurance claim damage checking method based on image recognition, which mainly includes the following steps:

201. obtaining the loss assessment case information of the loss assessment case, wherein the loss assessment case information comprises a loss assessment list and a loss assessment picture set;

the loss assessment picture set in the loss assessment case information comprises an accident scene picture (namely a distant view picture), a vehicle loss picture (namely a medium and close view picture), a vehicle part replacement picture, a certificate picture, a document picture and the like.

202. Determining a vehicle frame picture set corresponding to each vehicle from the loss assessment picture set;

in some embodiments, as shown in fig. 3, step 202 may be implemented by, but is not limited to, the following process:

2021. sorting pictures according to the loss assessment picture set to obtain pictures containing cars and pictures without cars;

2022. sequentially judging whether each picture containing the car is a medium-long scene picture or not; if so, marking the picture containing the car as a medium-long-range picture; if not, the picture containing the car is marked as a close-range picture.

2023. Sequentially carrying out license plate detection and recognition and vehicle frame detection on each image containing the vehicle to determine the corresponding relation between each image containing the license plate and the license plate information;

in some embodiments, as shown in fig. 4, step 2023 may be implemented by, but is not limited to, the following process:

20231. sequentially detecting the license plate of each picture containing the vehicle, and sorting the pictures containing the vehicle into pictures containing the license plate and pictures without the license plate;

the steps of detecting and identifying the license plate of each picture containing the license plate comprise: the license plate detection can use a detection model such as YOLO, SSD, fast RCNN and the like, and once a license plate region is detected on a certain picture, the license plate region can be cut out and rotationally corrected into a horizontal rectangle. The license plate number then continues to be identified using the CRNN model.

20232. Sequentially carrying out vehicle frame detection and license plate identification on each license plate-containing picture; the vehicle frame detection is that vehicle image areas are detected aiming at the images containing the license plates marked as the middle distant view images, the vehicle image areas are formed by cutting, and the vehicle image areas are directly formed aiming at the images containing the license plates marked as the close view images; geometrically matching the vehicle image area with the license plate detection result to obtain at least one first vehicle frame picture and possibly a second vehicle frame picture; the first vehicle frame picture is a vehicle frame picture containing a license plate in the vehicle image area, and the second vehicle frame picture is a vehicle frame picture containing no license plate in the vehicle image area; and identifying license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and license plate number.

In a specific embodiment, two vehicle image areas appear on a picture containing a license plate, one of the two vehicle image areas contains the license plate, and the other vehicle image area does not contain the license plate. In the identification process, a first vehicle frame picture and a second vehicle frame picture can be identified from the license plate containing pictures. In other embodiments, there may be a case where only one first vehicle frame picture is recognized or a plurality of first vehicle frame pictures and a plurality of second vehicle frame pictures are recognized.

The license plate character recognition is not complete and correct due to the fact that the license plate region is possibly incomplete or the image quality is poor, the recognized license plate information can be compared with the real license plate number of each accident vehicle in the damage assessment case information, and the accident vehicle matched with the case is considered to be matched if the editing distance of the character strings of the recognized license plate information and the real license plate number of each accident vehicle in the damage assessment case information is smaller than a threshold value.

Recording an identification result Info _ Plate = (I, Rect, Plate) of each license Plate information, wherein I is an image where a vehicle frame BB is located, and Rect is a rectangular frame position where a vehicle frame picture is located in the to-be-identified vehicle-containing picture. The Plate is the license Plate number of a certain accident vehicle in case information.

The medium and long shot pictures generally contain one or more vehicle frames, and the vehicle frames are detected and cut by using detection models such as YOLO, SSD, Faster RCNN and the like.

Wherein, the whole picture is directly used as a vehicle frame because the shooting distance of the close shot picture is short. Recording a set of information Info _ BB = (I, Rect) for each vehicle frame BB, where I is an image where the vehicle frame BB is located, and Rect is a rectangular frame position of the vehicle frame on the to-be-identified vehicle-containing picture.

20233. Sequentially detecting the vehicle frame of each picture without the license plate; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the middle distant view picture, a second vehicle frame picture is formed in the vehicle image area through cutting, and the second vehicle frame picture is directly formed on the picture containing the license plate marked as the close view picture;

20234. identifying license plate information in a license plate containing picture, wherein the license plate information comprises license plate region information and a license plate number;

20235. and comparing and calibrating the license plate number with the real license plate number in the damage assessment case information.

2024. Sequentially determining the corresponding relation between each license plate-free picture and the license plate information through a deep learning network by utilizing the color information and the type information of the vehicle body;

in some embodiments, as shown in fig. 5, step 2024 may be implemented by, but is not limited to, the following process:

20241. extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network;

in this embodiment, the pre-trained neural network may be trained through the insurance company's historical damage scenario information. A large number of pictures of the same vehicle at different shooting distances and angles are obtained through historical damage assessment case information, triple training data of 'target vehicle pictures', 'same vehicle pictures' and 'different vehicle pictures' can be constructed, a depth measurement network is trained through a triple loss function, and characteristic feature vectors of the vehicle in a semantic space are obtained. In the inference stage, a color vehicle type characterization vector V is extracted from the image of each vehicle frame BB_BB。

20242. Carrying out similarity clustering on the color vehicle type characterization vectors extracted from all vehicle frame pictures;

in this embodiment, the similarity may be calculated by using the magnitude of the cosine included angle between vectors or the distance in the euclidean distance. In this embodiment, K-means clustering is used as n groups (G)₁,G₂,...G_n) Wherein n corresponds to the number of involved vehicles in the damage assessment case information. If n =1, the case is single-vehicle, meaning that clustering is not needed, and if n =2, the case is double-vehicle accident, and so on. Characterizing the color vehicle type by a vector V_BBAnd the cluster attribution G of the vector are added into the vehicle frame information, Info _ BB = (I, Rect, V)_BB，G)。

20243. Establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on a first vehicle frame picture containing a license plate picture;

and according to the license Plate information obtained by license Plate detection and recognition, corresponding the result obtained in the clustering process to the accident vehicle corresponding to the license Plate information, and checking the geometric inclusion relationship between the license Plate information and the picture of the vehicle frame where the license Plate information is located, so that a cluster group G for each group of vehicle frame pictures can be obtained, and the cluster group G corresponds to the license Plate information Plate of one accident vehicle. Namely, a vehicle frame picture set of each accident vehicle in the case is obtained.

20244. And establishing a corresponding relation between the license plate information and each vehicle frame picture.

2025. And determining a vehicle frame image set corresponding to each vehicle according to the corresponding relation between each image containing the license plate and the license plate information and the corresponding relation between each image without the license plate and the license plate information. And executing the next operation by using the vehicle frame picture set corresponding to each vehicle to acquire the vehicle damage information and the vehicle appearance piece information of each vehicle.

203. And acquiring the vehicle appearance piece information and the vehicle damage information of each vehicle by using the vehicle frame picture set corresponding to each vehicle. The vehicle appearance piece information comprises a component type and a component mask position, and the vehicle damage information comprises a damage type and a damage mask position;

and performing image segmentation on all the vehicle frame pictures to obtain one or more pieces of vehicle damage information, wherein the result comprises each damage type, corresponding probability and damage mask positions. Predefined identifiable categories of damage include, but are not limited to, scratch, deformation (with deformation being differentiated by mild, moderate, or severe), tear, puncture, glass breakage, and other common categories of damage. The damage mask position is described by a binary image with the same width and height as the image, wherein the pixels which are positioned inside the outline of the damage mask position are 1, and other pixels are 0.

And (4) carrying out vehicle appearance piece segmentation on all the vehicle frame pictures to obtain component category information, corresponding probability and component mask positions on the pictures. It should be noted that the vehicle exterior parts mentioned in the present embodiment refer to exterior parts including, but not limited to, a door shell, a fender, a hood, a trunk lid, headlights, tail lights, a windshield, and the like. The appearance elements of the same name and different orientations are considered in the present invention as different vehicle appearance elements, such as a headlight (left) and a headlight (right), a front door shell (left) and a front door shell (right), etc.

204. Determining vehicle direction information in each loss assessment picture;

and judging the shooting direction of the vehicle according to the picture content. In this embodiment, it is preferable to perform direction vectorization around the vehicle, divided into k direction directions. That is, when k =12, the corresponding orientation includes front, left front middle, left rear, right rear middle, right front middle, and right front. Each orientation corresponds to a certain combination of vehicle exterior parts, e.g. a "left front" orientation picture typically corresponds to a combination of front bumper, front fender (left), headlights, etc.

205. And correcting the information of the vehicle appearance pieces identified in the damage assessment picture according to the vehicle direction information, wherein the vehicle direction information is used for distinguishing similar vehicle appearance pieces at different positions.

206. Matching the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises damaged component names and corresponding damage types;

207. determining a component missing list according to the vehicle appearance piece information and the preset component adjacency relation;

208. mapping the standard component damage list into a standard component replacement list according to a preset maintenance logic list; the standard component damage list and the component missing list are mapped into a standard component replacement list according to a preset maintenance logic list, and the standard component damage list further comprises missing component names and corresponding missing state marks.

209. And judging whether the damage assessment list is abnormal or not according to the standard component repair list.

It should be noted that, the order and the flow relation of the above steps are not limited in the embodiment of the present invention, and this embodiment only shows one possible implementation scheme, and in engineering application, specifically, may be adjusted by a person skilled in the art according to engineering needs, and are not described herein again.

The scheme provided by the embodiment has the beneficial effects that:

1. by detecting and identifying the license plate of the picture containing the car and calculating the characteristic feature vector of each picture containing the car, learning the characteristic vector embedded with the model and color information of the car system by adopting a depth measurement network, clustering the characteristic vectors according to the similarity, automatically separating the picture blocks of different cars in the case, automatically classifying the case pictures according to the content of the pictures, and effectively improving the work efficiency of the nuclear damage loop.

2. The method comprises the steps of automatically identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, training two example segmentation models of component identification and damage identification, and performing geometric mapping on masks of the component identification and the damage identification to obtain a damage list. And automatically prompting the risk items which are probably false-increased in the damage assessment and repair list to the staff, and effectively detecting the leakage false-increased cases in the damage assessment.

4. The vehicle part missing information in the image is determined by utilizing the inherent adjacency relation of the vehicle parts, so that the accuracy of vehicle damage information identification is further improved.

Example 3:

the embodiment provides a vehicle insurance claim settlement and loss checking method based on image recognition, which mainly comprises the following steps:

301. obtaining damage assessment case information, wherein the damage assessment case information comprises a damage assessment list and a damage assessment picture set;

the loss assessment picture set in the loss assessment case information comprises an accident scene picture (namely a distant view picture), a vehicle loss picture (namely a medium and close view picture), a vehicle part replacement picture, a certificate picture, a document picture and the like. Training is required prior to the specific application. Collecting the loss assessment picture set in the historical loss assessment case, labeling according to the categories and training a multi-classification network, wherein the network structure can be ResNet, VGG, GoogleNet and the like.

302. Determining a vehicle frame picture set corresponding to each vehicle from the loss assessment picture set;

3021. sorting pictures according to the loss assessment picture set to obtain pictures containing cars and pictures without cars;

3022. sequentially judging whether each picture containing the car is a medium-long scene picture or not; if so, marking the picture containing the car as a medium-long-range picture; if not, the picture containing the car is marked as a close-range picture.

3023. Sequentially carrying out license plate detection and recognition and vehicle frame detection on each image containing the vehicle to determine the corresponding relation between each image containing the license plate and the license plate information;

30231. sequentially detecting the license plate of each picture containing the vehicle, and sorting the pictures containing the vehicle into pictures containing the license plate and pictures without the license plate;

30232. Sequentially carrying out vehicle frame detection and license plate identification on each license plate-containing picture; the vehicle frame detection is that vehicle image areas are detected aiming at the images containing the license plates marked as the middle distant view images, the vehicle image areas are formed by cutting, and the vehicle image areas are directly formed aiming at the images containing the license plates marked as the close view images; geometrically matching the vehicle image area with the license plate detection result to obtain at least one first vehicle frame picture and/or second vehicle frame picture; identifying license plate information in a license plate containing picture, wherein the license plate information comprises license plate region information and a license plate number;

30233. Sequentially detecting the vehicle frame of each picture without the license plate; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the middle distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and the second vehicle frame picture is directly formed on the picture containing the license plate marked as the close view picture.

30234. Identifying license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and a license plate number;

30235. and comparing and calibrating the license plate number with the real license plate number in the damage assessment case information.

3024. Sequentially determining the corresponding relation between each license plate-free picture and the license plate information through a deep learning network by utilizing the color information and the type information of the vehicle body;

30241. extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network;

in this embodiment, the pre-trained neural network may be trained through the insurance company's historical damage scenario information. A large number of pictures of the same vehicle at different shooting distances and angles are obtained through historical damage assessment case information, so that triple training data of 'target vehicle pictures', 'same vehicle pictures' and 'different vehicle pictures' can be constructed, a depth measurement network is trained through a triple loss function, characteristic feature vectors of the vehicle in a semantic space are obtained, the goal is to enable the characteristic vectors of the same vehicle picture to be as close as possible in the space, the characteristic vectors of different vehicle pictures are separated as far as possible in the space, and semantic expressions of information such as vehicle type structures and vehicle body colors of the vehicle are embedded into the characteristic vectors learned through the constructed triple training data. In the inference stage, a color vehicle type characterization vector V is extracted from the image of each vehicle frame BB_BB。

30242. Carrying out similarity clustering on the color vehicle type characterization vectors extracted from all vehicle frame pictures;

30243. Establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on a first vehicle frame picture containing a license plate picture;

30244. And establishing a corresponding relation between the license plate information and each vehicle frame picture.

3025. And determining a vehicle frame image set corresponding to each vehicle according to the corresponding relation between each vehicle frame image containing the license plate image and the license plate information and the corresponding relation between each vehicle frame image without the license plate image and the license plate information. And executing the next operation by using the vehicle frame picture set corresponding to each vehicle to acquire the vehicle damage information and the vehicle appearance piece information of each vehicle.

303. Identifying vehicle appearance piece information in the vehicle frame picture of each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position;

In this embodiment, the position of a part mask containing a certain appearance on a car picture is described by a binary image with the same width and height as the image, wherein the pixels located inside the outline of the part mask take 1 and the other pixels take 0.

In a typical embodiment, pictures in historical damage assessment case information can be collected, component categories and component mask positions in all vehicle appearance information on the pictures can be labeled, and an example segmentation network such as MaskRCNN is trained. In a specific application process, the component category, the corresponding component probability and the component mask position of the vehicle appearance piece information can be accurately segmented by the example segmentation network. In order to ensure the reliability of the output result, the probability of the appearance identification can be subjected to threshold interception, and the output is carried out when the probability exceeds a certain threshold.

304. Identifying vehicle damage information in the vehicle frame picture of each damage assessment picture, wherein the vehicle damage information comprises damage categories and damage mask positions;

In a typical embodiment, pictures in historical damage assessment case information can be collected, and vehicle damage types and mask positions in the pictures are labeled so as to train an instance segmentation network such as MaskRCNN. In a specific application process, all damage categories, corresponding damage probabilities and damage mask positions in the vehicle damage information in the damage assessment picture set can be accurately segmented by the example segmentation network. In order to ensure the reliability of the output result, threshold interception can be carried out on the identified damage probability, and the damage probability is output only when a certain threshold is exceeded.

305. Determining vehicle direction information in each loss assessment picture;

and judging the shooting direction of the vehicle according to the picture content. In this embodiment, it is preferable to perform direction vectorization around the vehicle, divided into k direction directions. That is, when k =12, the corresponding orientation includes front, left front middle, left rear, right rear middle, right front middle, and right front. Each orientation corresponds to a certain combination of vehicle exterior parts, e.g. a "left front" orientation picture typically corresponds to a combination of front bumper, front fender (left), headlights, etc. In a typical embodiment, vehicle pictures in historical damage assessment case information can be collected, labeled according to the above orientation categories, and classification networks such as ResNet, VGG, GoogleNet and the like are trained. In a specific application process, the probability that each vehicle frame picture belongs to each predefined direction category is obtained.

306. And correcting the vehicle appearance piece information in each damage assessment picture according to the vehicle direction information, wherein the vehicle direction information is used for distinguishing similar vehicle appearances at different positions.

Due to the structural characteristics of the vehicle, in the output result of step 308, it is easy to have the mask position information basically overlapped, but the vehicle appearance piece pairs with different categories, such as front fender (left) and rear fender (right), front door shell (left) and rear door shell (right), front windshield glass and rear windshield glass, and so on. These component pairs need to be accurately distinguished by a combination of consideration of the overall orientation of the vehicle and the surrounding components.

In the present embodiment, the set of such confusable component pairs is preset to CS. Referring to a mode of carrying out non-maximum suppression (NMS) processing on the segmentation output results of the same category in the MaskRCNN model, and carrying out segmentation output results of the confusable category components; non-maximum suppression (NMS) merging may also be performed with reference to the orientation information.

One specific implementation is as follows: in a car-containing picture, if the cross-over position of the part MASK is greater than the IOU _ MASK by a certain threshold value and the part pair (Pi, Pj) belonging to the confusable pair set CS exists in the car-associated appearance information output in step 308, then:

(1) obtaining the vehicle direction information of Pi and Pj in step 309 and the corresponding probabilities wi and wj of the vehicle direction information

(2) The probabilities SCOREi corresponding to Pi and the probabilities SCOREj corresponding to Pj are weighted by the probabilities corresponding to the vehicle direction information. In the present embodiment, the component type of the vehicle exterior information having a large value after the weighting process is selected. In this embodiment, the specific weighting method is as follows: if wi SCOREi > wj SCOREi, then part Pi is left and part Pj is discarded, otherwise part Pj is left and part Pi is discarded.

307. And matching the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises the names of damaged components and corresponding damage types.

For each injury category D_iDamage Mask position Mask _ D of_iThe picture is combined with each appearance piece P on the picture containing the car_jMask position Mask _ P of the component_jFind the intersection and calculate the intersection Mask _ D_iP_jThe area of (a). When the area of the intersection area occupying the damaged area is larger than a certain threshold value, the appearance piece P is considered to be_jHas the damage category D_iAfter all the damaged areas of the whole picture are processed, a list of (part name, damage type) tuples can be obtained, for example:

[ front windshield, glass breakage ], (hood, scuff), (hood, slight deformation) ]

308. Determining a component missing list according to the vehicle appearance piece information and the preset component adjacency relation;

in addition to scratches, deformation, and the like, a vehicle may actually be damaged by a special damage, that is, a complete loss of a component, and such a damage is often found in components such as a rearview mirror, a headlight, and a tail lamp. Determining such vehicle damage does not apply to the method of matching the component mask and the damage mask in step 311 because the damage model has difficulty in efficiently describing the image features where the component is missing.

In this embodiment, the absence is determined by inference using the preset component adjacency relationship. In one embodiment, if there are m adjacent appearance pieces Q1, Q2, … Q around the appearance piece P_mP may be completely spatially surrounded. And all the appearance pieces Q1, Q2 and … Q are recognized to be present in the image through step 311_mIf the appearance piece P is not recognized, the more reliable estimation P is lost.

One typical set of embodiments is: p-rearview mirror (left), Q- { engine cover, front windshield, front door glass (left), front fender (left) }. In this step, a list of (part name, damage type) duplets is output upon detection of a part missing, for example: [ BACK-UP MIRROR (LEFT, LOSS) ].

309. And mapping the standard component damage list and the component missing list into a standard component replacement list according to a preset maintenance logic list.

310. And judging whether the damage assessment list is abnormal or not according to the standard component repair list.

Comparing the standard component repair and replacement list output in the step 309 and the repair and replacement list in the damage assessment list item by item according to the component name, if the maintenance grade of a certain component in the manual damage assessment list is higher than the maintenance grade determined by the machine; or a component exists on the manual damage assessment list but does not exist on the machine determined repair assessment list, the vehicle may be prompted to run a risk of leakage and false increase in the component damage.

Meanwhile, all vehicle frame images containing the part in the part segmentation result can be made into visual evidence and provided for the loss-checking personnel for further checking.

The technical scheme provided by the embodiment has the beneficial effects that:

2. The method comprises the steps of automatically identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, training two example segmentation models of component identification and damage identification, and performing geometric mapping on masks of the component identification and the damage identification to obtain a damage list. Risk items which are possibly added falsely in the damage assessment and repair list are automatically prompted to the working personnel, and the detection rate of the damage assessment and leakage falsely added cases is effectively improved.

4. The information of the missing vehicle parts in the image is judged by utilizing the inherent adjacency relation of the vehicle parts, so that the accuracy of identifying the vehicle damage information is further improved.

Example 4:

as shown in fig. 3, the embodiment provides a vehicle insurance claim settlement device based on image recognition, including:

the obtaining module 401 is configured to obtain damage assessment case information, where the damage assessment case information includes a damage assessment list and a damage assessment picture set.

A vehicle frame identification module 402, configured to determine, from the damage assessment picture set, a vehicle frame picture set corresponding to each vehicle; the vehicle frame recognition module further includes:

the car-containing picture sorting submodule 4021 is used for carrying out picture sorting on the damage assessment picture set to obtain a car-containing picture and a car-free picture;

a medium and near perspective judgment sub-module 4022, which is used for sequentially judging whether each car-containing picture is a medium and far perspective picture; if so, marking the picture containing the car as a medium-long-range picture; if not, the picture containing the car is marked as a close-range picture.

The license plate containing picture corresponding relation determining sub-module 4023 is used for sequentially carrying out license plate detection and recognition and vehicle frame detection on each license plate containing picture so as to determine the corresponding relation between each license plate containing picture and license plate information; the license plate picture corresponding relation determining sub-module 4023 further includes:

the license plate detection unit 40231 is used for sequentially detecting the license plates of each image containing the vehicle and sorting the images containing the vehicle into images containing the license plates and images without the license plates;

the vehicle frame and license plate recognition detection unit 40232 is used for sequentially carrying out vehicle frame detection and license plate recognition on each image containing the license plate; the vehicle frame detection is that vehicle image areas are detected aiming at the images containing the license plates marked as the middle distant view images, the vehicle image areas are formed by cutting, and the vehicle image areas are directly formed aiming at the images containing the license plates marked as the close view images; geometrically matching the vehicle image area with the license plate detection result to obtain at least one first vehicle frame picture and/or second vehicle frame picture;

the license plate-free vehicle frame detection unit 40233 is used for sequentially detecting the vehicle frames of each license plate-free picture; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the middle distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and the second vehicle frame picture is directly formed on the picture containing the license plate marked as the close view picture.

The license plate recognition unit 40234 is used for recognizing license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and a license plate number;

and the comparison and calibration unit 40235 is used for comparing and calibrating the license plate number with the real license plate number in the damage assessment case information.

The license plate-free picture corresponding relation determining sub-module 4024 is used for sequentially determining the corresponding relation between each license plate-free picture and the license plate information through a deep learning network by using the color information of the vehicle body and the information of the vehicle type; the corresponding relation determining submodule for the license plate-free picture further comprises:

the color vehicle type characterization vector calculation unit 40241 is used for extracting a color vehicle type characterization vector of each vehicle frame picture by using a pre-trained neural network;

the clustering unit 40242 is used for performing similarity clustering on the color vehicle type characterization vectors extracted from all the vehicle frame pictures;

the license plate information corresponding relation establishing unit 40243 is used for establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on a first vehicle frame picture containing a license plate picture;

a vehicle frame and license plate corresponding relation establishing unit 40244, configured to establish a corresponding relation between license plate information and each vehicle frame picture.

And the vehicle frame image set generating sub-module 4025 is used for determining a vehicle frame image set corresponding to each vehicle according to the corresponding relationship between each image containing the license plate and the license plate information and the corresponding relationship between each image without the license plate and the license plate information.

A vehicle appearance piece information identification module 403, configured to identify vehicle appearance piece information in the vehicle frame picture of each damage assessment picture, where the vehicle appearance piece information includes a component category and a component mask position; the vehicle exterior piece information identification module 403 further includes:

a vehicle direction determination submodule 4031 for determining vehicle direction information in the vehicle frame picture of each damage assessment picture;

and a modification submodule 4032, configured to modify vehicle exterior piece information identified in the damage assessment picture according to vehicle direction information, where the vehicle direction information is used to distinguish similar vehicle exterior pieces at different positions.

A vehicle damage information identification module 404, configured to identify vehicle damage information in each damage assessment picture, where the vehicle damage information includes a damage category and a damage mask position;

a matching module 405, configured to match the component mask and the damage mask of each damage assessment picture to determine a standard component damage list of the vehicle, where the standard component damage list includes a damaged component name and a corresponding damage type;

and a component missing identification module 406 for determining a component missing list according to the vehicle exterior piece information and the preset component adjacency relation.

The mapping module 407 is configured to map the standard component damage list into a standard component replacement list according to a preset maintenance logic list, where the standard component replacement list includes names of damaged components and corresponding replacement information; and the judging module is used for judging whether the damage assessment list is abnormal or not according to the standard component repair list.

The determining module 408 is configured to determine whether the damage assessment list is abnormal according to the standard component repair list.

The beneficial effect of this embodiment lies in:

Example 5:

according to another aspect of the present invention, an electronic device includes: at least one processor; and a memory coupled to the at least one processor; wherein the memory stores a computer program that can be executed by the at least one processor to implement the method of the invention.

Those of ordinary skill in the art will appreciate that the various illustrative modules and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described apparatuses and devices may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the modules is merely a logical division, and in actual implementation, there may be other divisions, for example, multiple modules or components may be combined or integrated into another system, or some features may be omitted, or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or modules, and may be in an electrical, mechanical or other form.

The modules described as separate parts may or may not be physically separate, and parts displayed as modules may or may not be physical modules, may be located in one place, or may be distributed on a plurality of network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the embodiment of the present invention.

In addition, functional modules in the embodiments of the present invention may be integrated into one processing module, or each of the modules may exist alone physically, or two or more modules are integrated into one module.

The functions, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method for transmitting/receiving the power saving signal according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a U disk, a removable hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

The above description is only a preferred embodiment of the application and is illustrative of the principles of the technology employed. It will be appreciated by a person skilled in the art that the scope of the invention as referred to in the present application is not limited to the embodiments with a specific combination of the above-mentioned features, but also covers other embodiments with any combination of the above-mentioned features or their equivalents without departing from the inventive concept. For example, the above features may be replaced with (but not limited to) features having similar functions disclosed in the present application.

It should be understood that the order of execution of the steps in the summary of the invention and the embodiments of the present invention does not absolutely imply any order of execution, and the order of execution of the steps should be determined by their functions and inherent logic, and should not be construed as limiting the process of the embodiments of the present invention.

Claims

1. A vehicle insurance claim settlement loss method based on image recognition is characterized by comprising the following steps:

obtaining damage assessment case information of a damage assessment case, wherein the damage assessment case information comprises a damage assessment list and a damage assessment picture set;

the loss assessment case is a double-vehicle or multi-vehicle case, the loss assessment picture set comprises a vehicle-containing picture and a vehicle-free picture, and the vehicle-containing picture comprises a license plate-containing picture and a license plate-free picture; carrying out picture sorting on the loss assessment picture set to obtain a picture containing a car and a picture without a car; sequentially judging whether each car-containing picture is a medium-long scene picture or not; if yes, marking the picture containing the vehicle as a medium-long-range picture; if not, marking the picture containing the car as a close-range picture;

sequentially carrying out license plate detection and recognition and vehicle frame detection on each image containing the vehicle so as to determine the corresponding relation between each image containing the license plate and the license plate information, and the method comprises the following steps: sequentially carrying out license plate detection on each picture containing the vehicle, and sorting the pictures containing the vehicle into pictures containing license plates and pictures without license plates; sequentially carrying out vehicle frame detection and license plate identification on each image containing the license plate; the vehicle frame detection is that vehicle image region detection is carried out on the picture containing the license plate marked as the medium-long-range view picture, a vehicle image region is formed by cutting, and a vehicle image region is directly formed on the picture containing the license plate marked as the close-range view picture; performing geometric matching on the vehicle image area and the license plate recognition result to obtain at least one first vehicle frame picture and/or second vehicle frame picture; sequentially detecting the vehicle frame of each license plate-free picture; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the medium-distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and a second vehicle frame picture is directly formed on the picture containing the license plate marked as the close-range view picture;

sequentially carrying out license plate detection and identification on each license plate-containing picture further comprises the following steps:

identifying license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and a license plate number; comparing and calibrating the license plate number with the real license plate number in the damage assessment case information;

the method for sequentially determining the corresponding relation between each license plate-free picture and the license plate information through the deep learning network by utilizing the color information and the model information of the vehicle body comprises the following steps: extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network; carrying out similarity clustering on the color vehicle type characterization vectors extracted from all vehicle frame pictures; establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on the first vehicle frame picture containing the license plate picture; establishing a corresponding relation between license plate information and each vehicle frame picture;

determining a vehicle frame picture set corresponding to each vehicle according to the corresponding relation between each picture containing the license plate and the license plate information and the corresponding relation between each picture without the license plate and the license plate information;

acquiring vehicle appearance piece information and vehicle damage information of each vehicle by using a vehicle frame picture set corresponding to each vehicle; identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position, and the vehicle damage information comprises a damage type and a damage mask position; determining vehicle direction information in each loss assessment picture; correcting the vehicle appearance piece information identified in each damage assessment picture according to the vehicle direction information, wherein the vehicle direction information is used for distinguishing similar vehicle appearance pieces at different positions;

matching the component mask position and the damage mask position of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises damaged component names and corresponding damage types; determining a component missing list according to the vehicle appearance piece information and a preset component adjacency relation;

mapping the standard component damage list and the component missing list into a standard component replacement list according to a preset maintenance logic list, wherein the standard component replacement list comprises missing component names and corresponding missing state marks, and the standard component replacement list also comprises damaged component names and corresponding replacement information;

and judging whether the damage assessment list is abnormal or not according to the standard component repair list.

2. An automobile insurance claim settlement damage checking device based on image recognition is characterized by comprising:

the system comprises an acquisition module, a processing module and a display module, wherein the acquisition module is used for acquiring the damage assessment case information of a damage assessment case, and the damage assessment case information comprises a damage assessment list and a damage assessment picture set;

the in-vehicle long-range scene recognition module is used for the case of double vehicles or multiple vehicles, the damage assessment picture set comprises a vehicle-containing picture and a vehicle-free picture, and the vehicle-containing picture comprises a license plate-containing picture and a license plate-free picture; carrying out picture sorting on the loss assessment picture set to obtain a picture containing a car and a picture without a car; sequentially judging whether each car-containing picture is a medium-long scene picture or not; if yes, marking the picture containing the vehicle as a medium-long-range picture; if not, marking the picture containing the car as a close-range picture;

the vehicle appearance piece information identification module is used for sequentially carrying out license plate detection and identification and vehicle frame detection on each image containing the vehicle so as to determine the corresponding relation between each image containing the license plate and the license plate information, and comprises the following steps: sequentially carrying out license plate detection on each picture containing the vehicle, and sorting the pictures containing the vehicle into pictures containing license plates and pictures without license plates; sequentially carrying out vehicle frame detection and license plate identification on each image containing the license plate; the vehicle frame detection is that vehicle image region detection is carried out on the picture containing the license plate marked as the medium-long-range view picture, a vehicle image region is formed by cutting, and a vehicle image region is directly formed on the picture containing the license plate marked as the close-range view picture; performing geometric matching on the vehicle image area and the license plate recognition result to obtain at least one first vehicle frame picture and/or second vehicle frame picture; sequentially detecting the vehicle frame of each license plate-free picture; the vehicle frame detection is that vehicle image area detection is carried out on the picture containing the license plate marked as the medium-distant view picture, a second vehicle frame picture is formed in the vehicle image area through clipping, and a second vehicle frame picture is directly formed on the picture containing the license plate marked as the close-range view picture; sequentially carrying out license plate detection and identification on each license plate-containing picture further comprises the following steps: identifying license plate information in the image containing the license plate, wherein the license plate information comprises license plate region information and a license plate number; comparing and calibrating the license plate number with the real license plate number in the damage assessment case information; the method for sequentially determining the corresponding relation between each license plate-free picture and the license plate information through the deep learning network by utilizing the color information and the model information of the vehicle body comprises the following steps: extracting a color vehicle type representation vector of each vehicle frame picture by using a pre-trained neural network; carrying out similarity clustering on the color vehicle type characterization vectors extracted from all vehicle frame pictures; establishing a corresponding relation between license plate information and a color vehicle type characterization vector clustering group according to a license plate recognition result on the first vehicle frame picture containing the license plate picture; establishing a corresponding relation between license plate information and each vehicle frame picture; determining a vehicle frame picture set corresponding to each vehicle according to the corresponding relation between each picture containing the license plate and the license plate information and the corresponding relation between each picture without the license plate and the license plate information; acquiring vehicle appearance piece information and vehicle damage information of each vehicle by using a vehicle frame picture set corresponding to each vehicle; identifying vehicle appearance piece information and vehicle damage information in each damage assessment picture, wherein the vehicle appearance piece information comprises a component type and a component mask position, and the vehicle damage information comprises a damage type and a damage mask position; determining vehicle direction information in each loss assessment picture; correcting the vehicle appearance piece information identified in each damage assessment picture according to the vehicle direction information, wherein the vehicle direction information is used for distinguishing similar vehicle appearance pieces at different positions;

the matching module is used for matching the component mask position and the damage mask position of each damage assessment picture to determine a standard component damage list of the vehicle, wherein the standard component damage list comprises damaged component names and corresponding damage types; determining a component missing list according to the vehicle appearance piece information and a preset component adjacency relation;

the standard component repair system comprises a mapping module, a maintenance module and a maintenance module, wherein the mapping module is used for mapping the standard component damage list and the component missing list into a standard component repair list according to a preset maintenance logic list, the standard component repair list comprises missing component names and corresponding missing state marks, and the standard component repair list also comprises damaged component names and corresponding repair information;

and the judging module is used for judging whether the damage assessment list is abnormal or not according to the standard component repair list.

3. An electronic device, comprising:

at least one processor; and

a memory coupled to the at least one processor; wherein,

the memory stores a computer program executable by the at least one processor to implement the image recognition-based vehicle insurance claim damage review method of claim 1.