CN109784171A - Car damage identification method for screening images, device, readable storage medium storing program for executing and server - Google Patents
Car damage identification method for screening images, device, readable storage medium storing program for executing and server Download PDFInfo
- Publication number
- CN109784171A CN109784171A CN201811529042.9A CN201811529042A CN109784171A CN 109784171 A CN109784171 A CN 109784171A CN 201811529042 A CN201811529042 A CN 201811529042A CN 109784171 A CN109784171 A CN 109784171A
- Authority
- CN
- China
- Prior art keywords
- image
- hidden layer
- frame image
- layer node
- neural network
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000012216 screening Methods 0.000 title claims abstract description 18
- 238000003860 storage Methods 0.000 title claims abstract description 17
- 238000003062 neural network model Methods 0.000 claims abstract description 48
- 230000003287 optical effect Effects 0.000 claims abstract description 13
- 238000007873 sieving Methods 0.000 claims abstract description 12
- 230000006870 function Effects 0.000 claims description 25
- 230000008569 process Effects 0.000 claims description 21
- 238000012549 training Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 7
- 210000005036 nerve Anatomy 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000005516 engineering process Methods 0.000 abstract description 4
- 238000000605 extraction Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 239000000284 extract Substances 0.000 description 4
- 235000021167 banquet Nutrition 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000012015 optical character recognition Methods 0.000 description 2
- 239000000725 suspension Substances 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 241000283070 Equus zebra Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000033228 biological regulation Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000010485 coping Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000000151 deposition Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006748 scratching Methods 0.000 description 1
- 230000002393 scratching effect Effects 0.000 description 1
Abstract
The invention belongs to field of computer technology more particularly to a kind of car damage identification method for screening images, device, computer readable storage medium and servers based on image recognition.The car damage identification video that the method receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to handle in preset neural network model, obtains the assessed value of each frame image;According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and each frame image is added into respectively in the image sequence at corresponding vehicle position;Choose preferred setting loss image of the image as each vehicle position of the smallest preset number of assessment errors respectively from the image sequence at each vehicle position.Through the embodiment of the present invention, artificial progress optical sieving is replaced using neural network model, reduces the consumption to human resources, greatly improves working efficiency.
Description
Technical field
The invention belongs to field of computer technology more particularly to a kind of car damage identification method for screening images, device, computer
Readable storage medium storing program for executing and server.
Background technique
With the development of vehicle technology and the sharp increase of vehicle fleet size, occur between vehicle some such as scratching, knock into the back
The probability of accident is also greatly increasing.When these accidents occur, car damage identification is usually carried out by traffic police or insurance company.
Before carrying out car damage identification, the video of staff's floor first is generally required, then manually from the video
In filter out the image at each position of vehicle, the foundation as car damage identification.But such car damage identification optical sieving mode
Need to expend a large amount of human resources, efficiency is very low.
Summary of the invention
In view of this, the embodiment of the invention provides a kind of car damage identification method for screening images, device, computer-readable depositing
Storage media and server need to expend a large amount of human resources, efficiency ten in a manner of existing car damage identification optical sieving to solve
Divide low problem.
The first aspect of the embodiment of the present invention provides a kind of car damage identification method for screening images, may include:
The car damage identification video that receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;
The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to preset nerve
It is handled in network model, obtains the assessed value of each frame image;
According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and by each frame image distinguish
It is added into the image sequence at corresponding vehicle position;
The image of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position as each
The preferred setting loss image at a vehicle position.
The second aspect of the embodiment of the present invention provides a kind of car damage identification optical sieving device, may include:
Image zooming-out module, for the car damage identification video that receiving terminal apparatus is sent, and from the car damage identification video
It is middle to extract each frame image;
Feature vector computing module, for calculating separately the feature vector of each frame image;
Assessed value computing module, for the feature vector of each frame image to be separately input in preset neural network model
It is handled, obtains the assessed value of each frame image;
Image sequence constructing module, for according to the assessed value of each frame image determine each frame image respectively corresponding to vehicle
Position, and each frame image is added into respectively in the image sequence at corresponding vehicle position;
It is preferred that setting loss image chooses module, for choosing assessment errors respectively most from the image sequence at each vehicle position
Preferred setting loss image of the image of small preset number as each vehicle position.
The third aspect of the embodiment of the present invention provides a kind of computer readable storage medium, the computer-readable storage
Media storage has computer-readable instruction, and the computer-readable instruction realizes following steps when being executed by processor:
The car damage identification video that receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;
The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to preset nerve
It is handled in network model, obtains the assessed value of each frame image;
According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and by each frame image distinguish
It is added into the image sequence at corresponding vehicle position;
The image of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position as each
The preferred setting loss image at a vehicle position.
The fourth aspect of the embodiment of the present invention provides a kind of server, including memory, processor and is stored in institute
The computer-readable instruction that can be run in memory and on the processor is stated, the processor executes described computer-readable
Following steps are realized when instruction:
The car damage identification video that receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;
The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to preset nerve
It is handled in network model, obtains the assessed value of each frame image;
According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and by each frame image distinguish
It is added into the image sequence at corresponding vehicle position;
The image of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position as each
The preferred setting loss image at a vehicle position.
Existing beneficial effect is the embodiment of the present invention compared with prior art: the embodiment of the present invention is set receiving terminal
The car damage identification video that preparation is sent extracts each frame image from the car damage identification video, calculates separately the feature of each frame image
Vector, and the feature vector of each frame image is separately input to handle in preset neural network model, obtain each frame figure
The assessed value of picture, then according to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and by each frame
Image is added into respectively in the image sequence at corresponding vehicle position, is finally selected respectively from the image sequence at each vehicle position
Take preferred setting loss image of the image of the smallest preset number of assessment errors as each vehicle position.Implement through the invention
Example, image is directly extracted according to car damage identification video and to image classification, is screened according to position, and also uses nerve net
Network model replaces artificial progress optical sieving, reduces the consumption to human resources, greatly improves working efficiency.
Detailed description of the invention
It to describe the technical solutions in the embodiments of the present invention more clearly, below will be to embodiment or description of the prior art
Needed in attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is only of the invention some
Embodiment for those of ordinary skill in the art without any creative labor, can also be according to these
Attached drawing obtains other attached drawings.
Fig. 1 is a kind of one embodiment flow chart of car damage identification method for screening images in the embodiment of the present invention;
Fig. 2 is schematic flow diagram of the neural network model to the treatment process of feature vector;
Fig. 3 is the schematic flow diagram of the training process of neural network model;
Fig. 4 is a kind of one embodiment structure chart of car damage identification optical sieving device in the embodiment of the present invention;
Fig. 5 is a kind of schematic block diagram of server in the embodiment of the present invention.
Specific embodiment
In order to make the invention's purpose, features and advantages of the invention more obvious and easy to understand, below in conjunction with the present invention
Attached drawing in embodiment, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that disclosed below
Embodiment be only a part of the embodiment of the present invention, and not all embodiment.Based on the embodiments of the present invention, this field
Those of ordinary skill's all other embodiment obtained without making creative work, belongs to protection of the present invention
Range.
In the present embodiment, after car accident occurs, car owner can connect backstage with video and attend a banquet, and carry out video and report a case to the security authorities,
Backstage is attended a banquet insurmountable problem, can turn to task dam site investigation person, is carried out car owner, is attended a banquet, tripartite's view between the person of surveying
Frequently, the person of surveying can carry out video interception to the video taken at the scene, and the picture for removing user terminal from uploads.
Further, it is contemplated that the person of surveying carries out also expending biggish workload when video interception, in the present embodiment also
It can be used a kind of car damage identification method for screening images as shown in Figure 1, by background server automatically to being taken on site
Video carries out optical sieving:
Step S101, the car damage identification video that receiving terminal apparatus is sent, and extracted respectively from the car damage identification video
Frame image.
After car accident occurs, car owner or the person of surveying can be by installing on the terminal devices such as mobile phone, tablet computer
Specified application program (APP) to server send car damage identification request.Accident vehicle is carried in car damage identification request
Vehicles identifications and live shooting car damage identification video, the vehicles identifications can be license plate number, vehicle identification number
Perhaps other marks by car owner or can survey for (Vehicle Identification Number, VIN), engine number
Member is input in terminal device, can also carry out optical character identification to the image comprising the vehicles identifications by terminal device
(Optical Character Recognition, OCR) identification obtains.It should include the accident in the car damage identification video
Image of each vehicle position of vehicle after accident, for example, should include bumper, car door, vehicle in the car damage identification video
Image of each vehicle position such as wheel, suspension, chassis, engine, cylinder after accident, the car damage identification video can be by end
The photographic device of end equipment shoots to obtain.
Server can therefrom extract the car damage identification video, go forward side by side one after receiving car damage identification request
Step ground extracts each frame image from the car damage identification video.
Step S102, the feature vector of each frame image is calculated separately.
In the present embodiment, inherently a kind of image based on local shape factor of the calculating of feature vector indicate with
The method of measuring similarity, the extraction of local feature are divided into two steps: it extracts target critical point and key point is described, it is crucial
The positioning of point is the basis of target identification, the usually Local Extremum of variation of image grayscale, containing significant structural information,
Even these points can also be rich in easy without actual intuitive visual meaning, but on certain angle, some scale
In matched information.Feature point description establishes feature vector, and the selection of feature space determines which characteristic of image participates in
Matching.The feature description of characteristic point should be invariant to various change, with ensure bottom line by position, visual angle, scale and
The influence of the factors such as illumination.Select reasonable feature space that can reduce all kinds of image change factors to matching algorithm speed, steady
The influence of strong property.
Firstly, carrying out the critical point detection based on multiscale.In order to guarantee that the feature extracted becomes scale
Holding stability is changed, the present embodiment carries out the detection of image key points in scale space, and Scale-space theory occurs earliest
In computer vision field, the purpose is to the Analysis On Multi-scale Features of simulated image data.The main thought of Scale-space theory is benefit
Change of scale is carried out to original image with Gaussian kernel, with obtain it is multiple dimensioned under image indicate.
Critical point detection is mainly generated by graphical rule space, and difference of Gaussian pyramid is established, and candidate key point obtains, and is closed
Key point fine positioning and its screening and key point principal direction determine several part compositions.
Graphical rule space, which generates, mainly generates the image sequence under different scale space to given two dimensional image
Figure.
Difference of Gaussian pyramid, which is established, mainly carries out difference of Gaussian (Difference of to scale space images sequence
Gaussian, DOG) operation, i.e., the difference of adjacent gaussian filtering image is mainly to find the pass with scale feature is stablized
Key point.
The acquisition of candidate key point is defined on adjacent scale space mainly in the difference of Gaussian spatial pyramid of foundation
The interior candidate put as image key points with local maximum or local minimum.The middle layer of Gaussian difference scale space
Each pixel and same layer adjacent 8 pixels, upper one layer of adjacent 9 pixels and next layer of adjacent 9 pictures
26 neighbor pixels are compared vegetarian refreshments in total.If pixel is all bigger than the Gauss difference value of 26 adjacent pixels or all small,
Then the point can be used as candidate key point.
The pixel value that key point fine positioning and its screening are primarily due to difference of Gaussian image is more sensitive to noise and edge,
Therefore, it further to be screened in the Local Extremum that difference of Gaussian space detects, and be reoriented to sub-pixel and accuracy rule
Spend position.Also to remove simultaneously low contrast characteristic point and unstable skirt response point, stability and mentioned with enhancing matching
High noise resisting ability.
The determination main purpose of key point principal direction is to guarantee rotational invariance, the gradient side based on characteristic point neighborhood territory pixel
It is each characteristic point assigned direction parameter to distribution character.It is sampled in the neighborhood window centered on characteristic point, and uses gradient
The gradient direction of direction histogram statistics neighborhood territory pixel.
By above step, the extraction of image characteristic point is completed, there are three information for each characteristic point: position, scale
The direction and.It is then possible to the key point feature extraction based on gradient orientation histogram statistics.
In image key points expression, it is not enough to be formed merely with the position of key point, scale and directional information and sentences enough
The certainly feature of property then needs to extract gray-scale statistical characteristics to the region around key point with scale size.Feature extraction it
Before, reference axis is rotated to be to the direction of key point first, to ensure rotational invariance.Then 8 × 8 are taken centered on key point
Window calculates the gradient orientation histogram in 8 directions on every 4 × 4 fritter, draws the accumulated value of each gradient direction.This
The kind united thought of neighborhood directivity information enhances the antimierophonic ability of algorithm, simultaneously for the feature containing position error
With also providing preferable fault-tolerance.To enhance matched robustness in practical calculating process, the region of feature extraction will be enlarged by
Range, to each key point using 4 × 4 totally 16 seed points describe, in this way 128 can be generated for a key point
Data ultimately form the feature vector of 128 dimensions.It, can be by feature vector for the influence for further removing the variation of illumination contrast
Length normalization method.
Step S103, the feature vector of each frame image is separately input to handle in preset neural network model,
Obtain the assessed value of each frame image.
It is needed during car damage identification according to multiple scene image informations come definitive result, for example, each part of vehicle bitmap
As information, bumper, car door, wheel, suspension, chassis, engine, cylinder etc. position are specifically included, wherein each position is equal
Need an at least corresponding scene image information, further, it is also possible to have the image information of driver, the image information of driving license,
Site environment image information (road, zebra stripes, traffic lights, traffic mark etc.).
Live image number required in total during car damage identification is denoted as N herein, then is needed from car damage identification video
Each frame image in identify this N kind live image respectively.
In order to reach this purpose, in the present embodiment first using a large amount of image pattern for carrying label to the mind
Training is carried out through network model, each label value corresponds to a kind of live image, for example, if the label value of certain sample is 1,
It represents the sample then as the 1st kind of live image, if the label value of certain sample is 2, represents the sample as the 2nd kind of live image,
And so on, after training is completed, image to be analyzed is handled using the neural network model, obtains its assessment
Value, the assessed value can reflect classification belonging to image to be analyzed.
As shown in Fig. 2, the neural network model may include steps of the treatment process of the feature vector of input:
Step S1031, component of the described eigenvector in each dimension is identified as the neural network model
Input layer data.
Neural network model in the present embodiment may include input layer, hidden layer and output layer.The input layer is used for
Input data, including more than two input layers are received from outside, the hidden layer is used to handle data, including
More than two hidden layer nodes, the output layer is for exporting processing result, including an output node layer.
The component of the input layer and described eigenvector in each dimension corresponds.For example, if feature to
Amount shares 3 dimensions, respectively dimension 1, dimension 2 and dimension 3, then of the input layer of corresponding neural network model
Number also should be 3, respectively input layer 1, input layer 2 and input layer 3, wherein input layer 1 and dimension
1 is corresponding, and input layer 2 is corresponding with dimension 2, and input layer 3 is corresponding with dimension 3.
Step S1032, in the hidden layer node of the neural network model respectively using fuzzy Gauss subordinating degree function pair
The input layer data are handled, and hidden layer node data are obtained.
Treatment process may particularly denote are as follows:
Wherein, i is the label of input layer, and value range is [1, n], and n is the number of input layer, and j is hidden
Label containing node layer, value range are [1, h], and h is the number of hidden layer node, and x is input layer data, xiFor x
In i-th of input layer input layer data, ΦjIt (x) is the hidden layer node data of j-th of hidden layer node,
Gij(xi) be j-th of hidden layer node i-th of fuzzy Gauss subordinating degree function, μijFor i-th of mould of j-th of hidden layer node
Paste the mathematic expectaion of Gauss subordinating degree function, σijFor the standard of i-th of fuzzy Gauss subordinating degree function of j-th of hidden layer node
Difference, exp are natural exponential function.
Preferably, the hidden layer node data can also be normalized, to reduce the hidden layer node
The difference of data, specifically, maximum value and minimum value in the available hidden layer node data, then according to most
The hidden layer node data are normalized in big value and the minimum value, obtain normalized node in hidden layer
According to.
For example, the hidden layer node data can be normalized by following formula:
Wherein, ΨjIt (x) is the normalized hidden layer node data of j-th of hidden layer node, ΦmaxIt (x) is Φj(x) in
Maximum value, ΦminIt (x) is Φj(x) minimum value in.
Step S1033, summation is weighted to the hidden layer node data respectively using preset weight, obtained described
Assessed value.
For the hidden layer node data not being normalized, the calculation formula of the assessed value can be with are as follows:
For normalized hidden layer node data, the calculation formula of the assessed value can be with are as follows:
Wherein: ωjFor weight corresponding with the hidden layer node data of j-th of hidden layer node, R (x) is output layer section
Point data namely the assessed value.
In the present embodiment, the training process of the neural network model includes the steps that as shown in Figure 3:
Step S301, calculate using the neural network model to each sample in preset history image sample database into
Whole degree of deviation when row processing.
It include the feature of the image at each vehicle position acquired in history setting loss record in the history image sample database
The feature vector of vector and the image at non-vehicle position, and the sample at each vehicle position in the history image sample database
Number should be greater than preset quantity threshold, and the quantity threshold can be configured according to the actual situation, in order to guarantee to train
As a result accuracy rate generally requires as each vehicle position increase number of samples as much as possible, for example, can be by the number
Threshold value is set as 10000,20000,50000 etc., and the present embodiment is not specifically limited in this embodiment.
In the present embodiment, the whole degree of deviation can be calculated according to the following formula:
Wherein, t is the label of sample, and 1≤t≤T, T are the total sample number in the history image sample database, EtFor t
The training error of a sample, yt' be t-th of sample assessed value, ytFor the desired value of t-th of sample, it is contemplated that it is non-negative whole for being worth
Number, it is assumed that vehicle position shares N number of, then the value range of desired value is [0, N], and value 0 represents present image as non-vehicle portion
The image of position, value 1 represent present image and represent present image as the image at vehicle position 1, value 2 as the figure at vehicle position 2
As ..., value n represent present image as vehicle position n image ..., value N represent present image as the figure of vehicle position N
Picture, E are the whole degree of deviation.
Step S302, judge whether the whole degree of deviation is greater than preset threshold value.
If the entirety degree of deviation is greater than the threshold value, S303 is thened follow the steps, S301 is then returned to step, until
The entirety degree of deviation is less than or equal to the threshold value.
If the entirety degree of deviation is less than or equal to the threshold value, S304 is thened follow the steps.
Step S303, the parameter of the neural network model is adjusted.
In the present embodiment, the parameter that can be adjusted includes: μij、σijAnd/or ωj。
To μijFor being adjusted, the formula that is adjusted to it can be with are as follows: μ 'ij=μij+kΔμ.Wherein, μ 'ijFor
Value adjusted, Δ μ are adjusting step, can be preset according to the actual situation, and k is regulation coefficient, and value can be to appoint
Meaning integer.
To σijFor being adjusted, the formula that is adjusted to it can be with are as follows: σ 'ij=σij+kΔσ.Wherein, σ 'ijFor
Value adjusted, Δ σ are adjusting step, can be preset according to the actual situation.
To ωjFor being adjusted, the formula that is adjusted to it can be with are as follows: ω 'j=ωj+kΔω.Wherein, ω 'j
For value adjusted, Δ ω is adjusting step, can be preset according to the actual situation.
It is especially noted that in actual parameter tuning process, it can be just for any one parameter therein
It is adjusted, can also be adjusted, these parameters can also be adjusted simultaneously for any two parameter therein, this
Embodiment is not especially limited this.
Step S304, terminate the training process to the neural network model.
Neural network model at this time have passed through a large amount of sample training, and its whole degree of deviation be maintained at one it is lesser
It in range, is handled using feature vector of the neural network model to image, a more accurate assessed value can be obtained.
Step S104, according to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and will be each
Frame image is added into respectively in the image sequence at corresponding vehicle position.
In the present embodiment, vehicle position corresponding to any frame image can be determined according to the following formula:
CarPartSq=argmin (| R (x) -0 |, | R (x) -1 |, | R (x) -2 | ..., | R (x)-n | ..., | R (x)-N
|)
Wherein, R (x) is the assessed value of the frame image, and argmin is minimum independent variable function, and CarPartSq is the frame figure
As the serial number at corresponding vehicle position.
Step S105, the figure of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position
As the preferred setting loss image as each vehicle position.
By the above process, multiple image might have in the image sequence at each vehicle position, it can also be further
It is screened out from it optimal several frame images namely the preferred setting loss image.For example, can be according to the following formula from each vehicle portion
Preferred setting loss image of the frame image as the vehicle position is only chosen in the image sequence of position:
SelImgSqn=argmin (| R1(x)-n|,|R2(x)-n|,|R3(x)-n|,...,|Rq(x)-n|,...,|RQ(n)
(x)-n|)
Wherein, q is the picture numbers in image sequence, is schemed in the image sequence that 1≤q≤Q (n), Q (n) are vehicle position n
As sum, RqIt (x) is the assessed value of q frame image in image sequence, | Rq(x)-n | q frame image comments as in image sequence
Estimate error, SelImgSqnFor the serial number of the preferred setting loss image of vehicle position n.
The above process is identified using image of the neural network model to each vehicle position, such side
Factor in need of consideration is excessive simultaneously for formula, will lead to that neural network model is extremely complex, and accuracy rate can also decline.Therefore, at this
In the alternatively possible realization of embodiment, it is preferred to use the mode of neural network cluster respectively identifies each frame image, i.e.,
A neural network model, respectively neural network model 1, neural network model is separately provided for each vehicle position
2 ..., neural network model n ..., neural network model N, this N number of neural network model forms a neural network model collection
Group, wherein each neural network model desired value is 0 or 1, it is contemplated that is worth and represents present image for 0 and be not belonging to the neural network
The corresponding vehicle position of model, it is contemplated that be worth and represent present image for 1 and belong to the corresponding vehicle position of the neural network model.Wherein
The processing of each neural network model and training process are similar with aforementioned process, and details are not described herein again.
In conclusion the embodiment of the present invention receive terminal device transmission car damage identification video, from the car damage identification
Each frame image is extracted in video, calculates separately the feature vector of each frame image, and the feature vector of each frame image difference is defeated
Enter into preset neural network model and handled, obtain the assessed value of each frame image, then according to the assessment of each frame image
Value determines vehicle position corresponding to each frame image respectively, and each frame image is added into the image at corresponding vehicle position respectively
In sequence, the image conduct of the smallest preset number of assessment errors is finally chosen respectively from the image sequence at each vehicle position
The preferred setting loss image at each vehicle position.Through the embodiment of the present invention, image and right is directly extracted according to car damage identification video
Image classification is screened according to position, and also using neural network model replace it is artificial carry out optical sieving, reduce pair
The consumption of human resources, greatly improves working efficiency.
It should be understood that the size of the serial number of each step is not meant that the order of the execution order in above-described embodiment, each process
Execution sequence should be determined by its function and internal logic, the implementation process without coping with the embodiment of the present invention constitutes any limit
It is fixed.
The embodiment of the present invention is shown corresponding to a kind of car damage identification method for screening images, Fig. 4 described in foregoing embodiments
A kind of one embodiment structure chart of the car damage identification optical sieving device provided.
In the present embodiment, a kind of car damage identification optical sieving device may include:
Image zooming-out module 401 is regarded for the car damage identification video that receiving terminal apparatus is sent, and from the car damage identification
Each frame image is extracted in frequency;
Feature vector computing module 402, for calculating separately the feature vector of each frame image;
Assessed value computing module 403, for the feature vector of each frame image to be separately input to preset neural network mould
It is handled in type, obtains the assessed value of each frame image;
Image sequence constructing module 404, corresponding to determining each frame image respectively according to the assessed value of each frame image
Vehicle position, and each frame image is added into respectively in the image sequence at corresponding vehicle position;
It is preferred that setting loss image chooses module 405, missed for choosing assessment respectively from the image sequence at each vehicle position
Preferred setting loss image of the image of the smallest preset number of difference as each vehicle position.
Further, the assessed value computing module may include:
Input layer data determination unit, for component of the described eigenvector in each dimension to be identified as
Point of the input layer data of the neural network model, the input layer and described eigenvector in each dimension
Amount corresponds;
Hidden layer node data computing unit, for distinguishing according to the following formula in the hidden layer node of the neural network model
The input layer data are handled using fuzzy Gauss subordinating degree function, obtain hidden layer node data:
Wherein, i is the label of input layer, and value range is [1, n], and n is the number of input layer, and j is hidden
Label containing node layer, value range are [1, h], and h is the number of hidden layer node, and x is input layer data, xiFor x
In i-th of input layer input layer data, ΦjIt (x) is the hidden layer node data of j-th of hidden layer node,
Gij(xi) be j-th of hidden layer node i-th of fuzzy Gauss subordinating degree function, μijFor i-th of mould of j-th of hidden layer node
Paste the mathematic expectaion of Gauss subordinating degree function, σijFor the standard of i-th of fuzzy Gauss subordinating degree function of j-th of hidden layer node
Difference, exp are natural exponential function;
Assessed value computing unit is asked for being weighted respectively to the hidden layer node data using preset weight
With obtain the assessed value.
Further, the assessed value computing module can also include:
Most it is worth acquiring unit, for obtaining maximum value and minimum value in the hidden layer node data;
Normalized unit is returned for the hidden layer node data to be normalized according to the following formula
The one hidden layer node data changed:
Wherein, ΨjIt (x) is the normalized hidden layer node data of j-th of hidden layer node, ΦmaxIt (x) is Φj(x) in
Maximum value, ΦminIt (x) is Φj(x) minimum value in.
Further, the car damage identification optical sieving device can also include:
Whole degree of deviation computing module, for calculating using the neural network model to preset history image sample database
In the whole degree of deviation of each sample when being handled;
Threshold value judgment module, for judging whether the whole degree of deviation is greater than preset threshold value;
Parameter adjustment module, if being greater than the threshold value for the whole degree of deviation, to the neural network model
Parameter is adjusted;
Training ending module terminates if being less than or equal to the threshold value for the whole degree of deviation to the nerve
The training process of network model.
Further, the whole degree of deviation computing module can specifically include:
The whole degree of deviation is calculated according to the following formula:
Wherein, t is the label of sample, and 1≤t≤T, T are the total sample number in the history image sample database, EtFor t
The training error of a sample, y 'tFor the assessed value of t-th of sample, ytFor the desired value of t-th of sample, E is the whole deviation
Degree.
It is apparent to those skilled in the art that for convenience and simplicity of description, the device of foregoing description,
The specific work process of module and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, is not described in detail or remembers in some embodiment
The part of load may refer to the associated description of other embodiments.
The schematic block diagram that Fig. 5 shows a kind of server provided in an embodiment of the present invention illustrates only for ease of description
Part related to the embodiment of the present invention.
In the present embodiment, the server 5 may include: processor 50, memory 51 and be stored in the storage
In device 51 and the computer-readable instruction 52 that can run on the processor 50, such as execute above-mentioned car damage identification image sieve
The computer-readable instruction of choosing method.The processor 50 realizes above-mentioned each vehicle when executing the computer-readable instruction 52
Step in setting loss method for screening images embodiment, such as step S101 to S105 shown in FIG. 1.Alternatively, the processor 50
The function of each module/unit in above-mentioned each Installation practice is realized when executing the computer-readable instruction 52, such as shown in Fig. 4
The function of module 401 to 405.
Illustratively, the computer-readable instruction 52 can be divided into one or more module/units, one
Or multiple module/units are stored in the memory 51, and are executed by the processor 50, to complete the present invention.Institute
Stating one or more module/units can be the series of computation machine readable instruction section that can complete specific function, the instruction segment
For describing implementation procedure of the computer-readable instruction 52 in the server 5.
The processor 50 can be central processing unit (Central Processing Unit, CPU), can also be
Other general processors, digital signal processor (Digital Signal Processor, DSP), specific integrated circuit
(Application Specific Integrated Circuit, ASIC), field programmable gate array (Field-
Programmable Gate Array, FPGA) either other programmable logic device, discrete gate or transistor logic,
Discrete hardware components etc..General processor can be microprocessor or the processor is also possible to any conventional processor
Deng.
The memory 51 can be the internal storage unit of the server 5, such as the hard disk or memory of server 5.
The memory 51 is also possible to the External memory equipment of the server 5, such as the plug-in type being equipped on the server 5 is hard
Disk, intelligent memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card, flash card
(Flash Card) etc..Further, the memory 51 can also both include the internal storage unit of the server 5 or wrap
Include External memory equipment.The memory 51 is for storing needed for the computer-readable instruction and the server 5 it
Its instruction and data.The memory 51 can be also used for temporarily storing the data that has exported or will export.
The functional units in various embodiments of the present invention may be integrated into one processing unit, is also possible to each
Unit physically exists alone, and can also be integrated in one unit with two or more units.Above-mentioned integrated unit both may be used
To use formal implementation of hardware, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention substantially or
Person says that all or part of the part that contributes to existing technology or the technical solution can body in the form of software products
Reveal and, which is stored in a storage medium, including several computer-readable instructions are used so that one
Platform computer equipment (can be personal computer, server or the network equipment etc.) executes described in each embodiment of the present invention
The all or part of the steps of method.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-
Only Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. are various can be with
Store the medium of computer-readable instruction.
Embodiment described above is merely illustrative of the technical solution of the present invention, rather than its limitations;Although referring to aforementioned reality
Applying example, invention is explained in detail, those skilled in the art should understand that: it still can be to aforementioned each
Technical solution documented by embodiment is modified or equivalent replacement of some of the technical features;And these are modified
Or replacement, the spirit and scope for technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution.
Claims (10)
1. a kind of car damage identification method for screening images characterized by comprising
The car damage identification video that receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;
The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to preset neural network
It is handled in model, obtains the assessed value of each frame image;
According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and each frame image is added respectively
In the image sequence for entering corresponding vehicle position;
The image of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position as each vehicle
The preferred setting loss image at position.
2. car damage identification method for screening images according to claim 1, which is characterized in that the neural network model is to defeated
The treatment process of the feature vector entered includes:
Component of the described eigenvector in each dimension is identified as to the input layer number of the neural network model
According to the component of the input layer and described eigenvector in each dimension corresponds;
According to the following formula in the hidden layer node of the neural network model respectively using fuzzy Gauss subordinating degree function to described defeated
Enter node layer data to be handled, obtain hidden layer node data:
Wherein, i is the label of input layer, and value range is [1, n], and n is the number of input layer, and j is hidden layer
The label of node, value range are [1, h], and h is the number of hidden layer node, and x is input layer data, xiFor in x
The input layer data of i-th of input layer, ΦjIt (x) is the hidden layer node data of j-th of hidden layer node, Gij
(xi) be j-th of hidden layer node i-th of fuzzy Gauss subordinating degree function, μijI-th for j-th of hidden layer node is fuzzy
The mathematic expectaion of Gauss subordinating degree function, σijFor the standard of i-th of fuzzy Gauss subordinating degree function of j-th of hidden layer node
Difference, exp are natural exponential function;
Summation is weighted to the hidden layer node data respectively using preset weight, obtains the assessed value.
3. car damage identification method for screening images according to claim 2, which is characterized in that distinguish using preset weight
The hidden layer node data are weighted before summation, further includes:
Obtain the maximum value and minimum value in the hidden layer node data;
The hidden layer node data are normalized according to the following formula, obtain normalized hidden layer node data:
Wherein, ΨjIt (x) is the normalized hidden layer node data of j-th of hidden layer node, ΦmaxIt (x) is Φj(x) in most
Big value, ΦminIt (x) is Φj(x) minimum value in.
4. car damage identification method for screening images according to any one of claim 1 to 3, which is characterized in that the nerve
The training process of network model includes:
It calculates whole when being handled using the neural network model each sample in preset history image sample database
The body degree of deviation;
Judge whether the whole degree of deviation is greater than preset threshold value;
If the entirety degree of deviation is greater than the threshold value, the parameter of the neural network model is adjusted, and return and hold
When the row calculating is handled each sample in preset history image sample database using the neural network model
The step of whole degree of deviation, until the whole degree of deviation is less than or equal to the threshold value;
If the entirety degree of deviation is less than or equal to the threshold value, terminate the training process to the neural network model.
5. car damage identification method for screening images according to claim 4, which is characterized in that the calculating uses the nerve
Whole degree of deviation when network model handles each sample in preset history image sample database includes:
The whole degree of deviation is calculated according to the following formula:
Wherein, t is the label of sample, and 1≤t≤T, T are the total sample number in the history image sample database, EtFor t-th of sample
Training error, y 'tFor the assessed value of t-th of sample, ytFor the desired value of t-th of sample, E is the whole degree of deviation.
6. a kind of car damage identification optical sieving device characterized by comprising
Image zooming-out module for the car damage identification video that receiving terminal apparatus is sent, and is mentioned from the car damage identification video
Take each frame image;
Feature vector computing module, for calculating separately the feature vector of each frame image;
Assessed value computing module, for the feature vector of each frame image to be separately input to carry out in preset neural network model
Processing, obtains the assessed value of each frame image;
Image sequence constructing module, for according to the assessed value of each frame image determine each frame image respectively corresponding to vehicle portion
Position, and each frame image is added into respectively in the image sequence at corresponding vehicle position;
It is preferred that setting loss image chooses module, it is the smallest for choosing assessment errors respectively from the image sequence at each vehicle position
Preferred setting loss image of the image of preset number as each vehicle position.
7. a kind of computer readable storage medium, the computer-readable recording medium storage has computer-readable instruction, special
Sign is, realizes that the vehicle as described in any one of claims 1 to 5 is fixed when the computer-readable instruction is executed by processor
The step of damaging method for screening images.
8. a kind of server, including memory, processor and storage can transport in the memory and on the processor
Capable computer-readable instruction, which is characterized in that the processor realizes following steps when executing the computer-readable instruction:
The car damage identification video that receiving terminal apparatus is sent, and each frame image is extracted from the car damage identification video;
The feature vector of each frame image is calculated separately, and the feature vector of each frame image is separately input to preset neural network
It is handled in model, obtains the assessed value of each frame image;
According to the assessed value of each frame image determine each frame image respectively corresponding to vehicle position, and each frame image is added respectively
In the image sequence for entering corresponding vehicle position;
The image of the smallest preset number of assessment errors is chosen respectively from the image sequence at each vehicle position as each vehicle
The preferred setting loss image at position.
9. server according to claim 8, which is characterized in that the neural network model is to the feature vector of input
Treatment process includes:
Component of the described eigenvector in each dimension is identified as to the input layer number of the neural network model
According to the component of the input layer and described eigenvector in each dimension corresponds;
According to the following formula in the hidden layer node of the neural network model respectively using fuzzy Gauss subordinating degree function to described defeated
Enter node layer data to be handled, obtain hidden layer node data:
Wherein, i is the label of input layer, and value range is [1, n], and n is the number of input layer, and j is hidden layer
The label of node, value range are [1, h], and h is the number of hidden layer node, and x is input layer data, xiFor in x
The input layer data of i-th of input layer, ΦjIt (x) is the hidden layer node data of j-th of hidden layer node, Gij
(xi) be j-th of hidden layer node i-th of fuzzy Gauss subordinating degree function, μijI-th for j-th of hidden layer node is fuzzy
The mathematic expectaion of Gauss subordinating degree function, σijFor the standard of i-th of fuzzy Gauss subordinating degree function of j-th of hidden layer node
Difference, exp are natural exponential function;
Summation is weighted to the hidden layer node data respectively using preset weight, obtains the assessed value.
10. server according to claim 9, which is characterized in that in the preset weight of use respectively to the hidden layer
Node data is weighted before summation, further includes:
Obtain the maximum value and minimum value in the hidden layer node data;
The hidden layer node data are normalized according to the following formula, obtain normalized hidden layer node data:
Wherein, ΨjIt (x) is the normalized hidden layer node data of j-th of hidden layer node, ΦmaxIt (x) is Φj(x) in most
Big value, ΦminIt (x) is Φj(x) minimum value in.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811529042.9A CN109784171A (en) | 2018-12-14 | 2018-12-14 | Car damage identification method for screening images, device, readable storage medium storing program for executing and server |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811529042.9A CN109784171A (en) | 2018-12-14 | 2018-12-14 | Car damage identification method for screening images, device, readable storage medium storing program for executing and server |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109784171A true CN109784171A (en) | 2019-05-21 |
Family
ID=66495751
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811529042.9A Pending CN109784171A (en) | 2018-12-14 | 2018-12-14 | Car damage identification method for screening images, device, readable storage medium storing program for executing and server |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109784171A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110723145A (en) * | 2019-09-12 | 2020-01-24 | 精电有限公司 | Vehicle control method and device, computer-readable storage medium and wearable device |
CN112381589A (en) * | 2020-12-04 | 2021-02-19 | 南京浪脆电子商务有限公司 | Intelligent commodity evaluation management system of commodity transaction platform based on cloud computing |
WO2021057069A1 (en) * | 2019-09-27 | 2021-04-01 | 支付宝(杭州)信息技术有限公司 | Computer-executable vehicle loss assessment method and apparatus |
WO2023273296A1 (en) * | 2021-06-30 | 2023-01-05 | 平安科技(深圳)有限公司 | Vehicle image segmentation quality evaluation method and apparatus, device, and storage medium |
US11710158B2 (en) | 2020-01-23 | 2023-07-25 | Ford Global Technologies, Llc | Vehicle damage estimation |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107194323A (en) * | 2017-04-28 | 2017-09-22 | 阿里巴巴集团控股有限公司 | Car damage identification image acquiring method, device, server and terminal device |
CN107358596A (en) * | 2017-04-11 | 2017-11-17 | 阿里巴巴集团控股有限公司 | A kind of car damage identification method based on image, device, electronic equipment and system |
CN107368776A (en) * | 2017-04-28 | 2017-11-21 | 阿里巴巴集团控股有限公司 | Car damage identification image acquiring method, device, server and terminal device |
CN107730485A (en) * | 2017-08-03 | 2018-02-23 | 上海壹账通金融科技有限公司 | Car damage identification method, electronic equipment and computer-readable recording medium |
CN108364253A (en) * | 2018-03-15 | 2018-08-03 | 北京威远图易数字科技有限公司 | Car damage identification method, system and electronic equipment |
CN108388900A (en) * | 2018-02-05 | 2018-08-10 | 华南理工大学 | The video presentation method being combined based on multiple features fusion and space-time attention mechanism |
CN108446618A (en) * | 2018-03-09 | 2018-08-24 | 平安科技(深圳)有限公司 | Car damage identification method, device, electronic equipment and storage medium |
CN108763277A (en) * | 2018-04-10 | 2018-11-06 | 平安科技(深圳)有限公司 | A kind of data analysing method, computer readable storage medium and terminal device |
WO2018205467A1 (en) * | 2017-05-10 | 2018-11-15 | 平安科技(深圳)有限公司 | Automobile damage part recognition method, system and electronic device and storage medium |
-
2018
- 2018-12-14 CN CN201811529042.9A patent/CN109784171A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107358596A (en) * | 2017-04-11 | 2017-11-17 | 阿里巴巴集团控股有限公司 | A kind of car damage identification method based on image, device, electronic equipment and system |
CN107194323A (en) * | 2017-04-28 | 2017-09-22 | 阿里巴巴集团控股有限公司 | Car damage identification image acquiring method, device, server and terminal device |
CN107368776A (en) * | 2017-04-28 | 2017-11-21 | 阿里巴巴集团控股有限公司 | Car damage identification image acquiring method, device, server and terminal device |
WO2018205467A1 (en) * | 2017-05-10 | 2018-11-15 | 平安科技(深圳)有限公司 | Automobile damage part recognition method, system and electronic device and storage medium |
CN107730485A (en) * | 2017-08-03 | 2018-02-23 | 上海壹账通金融科技有限公司 | Car damage identification method, electronic equipment and computer-readable recording medium |
CN108388900A (en) * | 2018-02-05 | 2018-08-10 | 华南理工大学 | The video presentation method being combined based on multiple features fusion and space-time attention mechanism |
CN108446618A (en) * | 2018-03-09 | 2018-08-24 | 平安科技(深圳)有限公司 | Car damage identification method, device, electronic equipment and storage medium |
CN108364253A (en) * | 2018-03-15 | 2018-08-03 | 北京威远图易数字科技有限公司 | Car damage identification method, system and electronic equipment |
CN108763277A (en) * | 2018-04-10 | 2018-11-06 | 平安科技(深圳)有限公司 | A kind of data analysing method, computer readable storage medium and terminal device |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110723145A (en) * | 2019-09-12 | 2020-01-24 | 精电有限公司 | Vehicle control method and device, computer-readable storage medium and wearable device |
WO2021057069A1 (en) * | 2019-09-27 | 2021-04-01 | 支付宝(杭州)信息技术有限公司 | Computer-executable vehicle loss assessment method and apparatus |
US11710158B2 (en) | 2020-01-23 | 2023-07-25 | Ford Global Technologies, Llc | Vehicle damage estimation |
CN112381589A (en) * | 2020-12-04 | 2021-02-19 | 南京浪脆电子商务有限公司 | Intelligent commodity evaluation management system of commodity transaction platform based on cloud computing |
WO2023273296A1 (en) * | 2021-06-30 | 2023-01-05 | 平安科技(深圳)有限公司 | Vehicle image segmentation quality evaluation method and apparatus, device, and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112308092B (en) | Light-weight license plate detection and identification method based on multi-scale attention mechanism | |
CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
WO2019169816A1 (en) | Deep neural network for fine recognition of vehicle attributes, and training method thereof | |
CN109784171A (en) | Car damage identification method for screening images, device, readable storage medium storing program for executing and server | |
CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
CN109284733B (en) | Shopping guide negative behavior monitoring method based on yolo and multitask convolutional neural network | |
CN113065558A (en) | Lightweight small target detection method combined with attention mechanism | |
CN109960742B (en) | Local information searching method and device | |
CN109800682B (en) | Driver attribute identification method and related product | |
CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
CN106250838A (en) | vehicle identification method and system | |
CN104751147A (en) | Image recognition method | |
CN104850857B (en) | Across the video camera pedestrian target matching process of view-based access control model spatial saliency constraint | |
Do et al. | Automatic license plate recognition using mobile device | |
CN110689043A (en) | Vehicle fine granularity identification method and device based on multiple attention mechanism | |
CN110825900A (en) | Training method of feature reconstruction layer, reconstruction method of image features and related device | |
CN113159300A (en) | Image detection neural network model, training method thereof and image detection method | |
CN111160169A (en) | Face detection method, device, equipment and computer readable storage medium | |
CN110363238A (en) | Intelligent vehicle damage identification method, system, electronic equipment and storage medium | |
CN111951283A (en) | Medical image identification method and system based on deep learning | |
CN114049572A (en) | Detection method for identifying small target | |
CN112101195A (en) | Crowd density estimation method and device, computer equipment and storage medium | |
CN110619280B (en) | Vehicle re-identification method and device based on deep joint discrimination learning | |
CN111382638B (en) | Image detection method, device, equipment and storage medium | |
Feng et al. | A novel saliency detection method for wild animal monitoring images with WMSN |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |