CN106372651A - Picture quality detection method and device - Google Patents
Picture quality detection method and device Download PDFInfo
- Publication number
- CN106372651A CN106372651A CN201610704799.1A CN201610704799A CN106372651A CN 106372651 A CN106372651 A CN 106372651A CN 201610704799 A CN201610704799 A CN 201610704799A CN 106372651 A CN106372651 A CN 106372651A
- Authority
- CN
- China
- Prior art keywords
- photo
- training
- resolution
- convolutional neural
- neural networks
- 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 38
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 93
- 238000012549 training Methods 0.000 claims description 156
- 239000013598 vector Substances 0.000 claims description 50
- 238000000605 extraction Methods 0.000 claims description 33
- 238000000034 method Methods 0.000 claims description 27
- 239000000284 extract Substances 0.000 claims description 14
- 238000003062 neural network model Methods 0.000 claims description 5
- 230000003760 hair shine Effects 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 description 8
- 210000002569 neuron Anatomy 0.000 description 6
- 230000000694 effects Effects 0.000 description 4
- 230000004913 activation Effects 0.000 description 2
- 239000011521 glass Substances 0.000 description 2
- 230000007935 neutral effect Effects 0.000 description 2
- 230000000149 penetrating effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 210000004247 hand Anatomy 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000029058 respiratory gaseous exchange Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Abstract
The invention relates to a picture quality detection method and device. The picture quality detection method comprises the following steps of after a car insurance claims settlement server receives a claims settlement photo uploaded by a user terminal, performing definition identification on the received claims settlement photo by using a pretrained and generated deep convolutional neural network model so as to determine the definition grade of the claims settlement photo; and if the definition grade of the claims settlement photo is lower than the preset definition grade, sending first reminding information to the user terminal so as to remind the user to upload the claims settlement photo again. The claims settlement photo is subjected to the definition identification through the pretrained and generated deep convolutional neural network model; each claims settlement photo uploaded by the user is enabled to be the claims settlement photo capable of being accurately analyzed to obtain the car insurance site information; therefore the work efficiency improvement of a self-help claims settlement system is facilitated; and the experience of the user is improved.
Description
Technical field
The present invention relates to technical field of image processing, more particularly, to a kind of detection method of picture quality and device.
Background technology
At present, in vehicle insurance intelligent self-service Claims Resolution system, the quality of picture ceases manner of breathing with the accurate rate of vehicle image identification
Close.Specifically, if user uploads clearly vehicle image, this Claims Resolution system can analyze the feelings at vehicle insurance scene very accurately
Condition;If conversely, the vehicle image that user is uploaded is not clear, Claims Resolution system cannot draw according to the analysis of described vehicle image
Vehicle insurance field data is it is impossible to normal work.Therefore, whether how to identify the definition of the vehicle image that user uploads exactly
Meet the requirements, meet the needs of analysis, become a problem demanding prompt solution.
Content of the invention
The technical problem to be solved is to provide a kind of detection method of picture quality and device.
The technical scheme is that a kind of detection method of picture quality, described picture
The detection method of quality includes:
S1, vehicle insurance Claims Resolution server, after receiving the Claims Resolution photo of user terminal uploads, is generated using training in advance
Depth convolutional neural networks model enters line definition identification to the Claims Resolution photo receiving, to determine the clear of described Claims Resolution photo
Degree grade;
S2, if the levels of sharpness of described Claims Resolution photo is less than default levels of sharpness, sends the first information extremely
Described user terminal, to remind user again to upload Claims Resolution photo.
Preferably, before described step s1, the method also includes:
S01, the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and is extracted under each classification
In Claims Resolution photo, as training photo, under each classification of extraction, remaining Claims Resolution photo is as checking for the Claims Resolution photo of preset ratio
Photo;
S02, carries out feature extraction to each training photo under each classification, to obtain depth convolution to be entered to described
The first pixel vectors in neural network model, and will be defeated for corresponding first pixel vectors of each training photo under each classification
Enter to described depth convolutional neural networks model, to train the depth convolutional neural networks model generating for identification;
S03, carries out feature extraction to each checking photo under each classification, to obtain the depth that input generates to training
The second pixel vectors in convolutional neural networks model, and by each classification under corresponding second pixel of each checking photo to
In the depth convolutional neural networks model that amount input generates to training, to verify the depth convolutional neural networks model that training generates
Accuracy rate;
S04, if the accuracy rate of the depth convolutional neural networks model of training generation is more than or equal to predetermined threshold value, training knot
Bundle.
Preferably, after described step s03, the method also includes: s05, if the depth convolutional neural networks that training generates
The accuracy rate of model is less than predetermined threshold value, then send the second prompting message to described user terminal, to remind user to increase Claims Resolution
The sample size of photo.
Preferably, the described step bag that each training photo under each classification or checking photo are carried out with feature extraction
Include:
For the lower each training photo of each classification or checking photo, using different convolution kernels from each train photo or
First block of pixels of checking photo begins stepping through and carries out convolution algorithm to last block of pixels, is shone with extracting each training
Piece or checking photo corresponding different characteristic figure;
Pond and rasterization process are carried out to the characteristic pattern of each training photo extracting or checking photo, will extract
The each training photo eigen figure going out is processed into the first consistent pixel vectors of dimension, by each checking photo eigen extracting
Figure is processed into the second consistent pixel vectors of dimension.
Preferably, in described step s02, the parameter of described depth convolutional neural networks model is by using back kick
Broadcast what the estimation of bp method obtained.
The technical scheme that the present invention solves above-mentioned technical problem is also as follows: a kind of detection means of picture quality, described figure
The detection means of piece quality includes:
Identification module, for the depth after receiving the Claims Resolution photo of user terminal uploads, being generated using training in advance
Convolutional neural networks model enters line definition identification to the Claims Resolution photo receiving, to determine definition of described Claims Resolution photo etc.
Level;
Prompting module, if the levels of sharpness for described Claims Resolution photo is less than default levels of sharpness, sends first
Information extremely described user terminal, to remind user again to upload Claims Resolution photo.
Preferably, the detection means of described picture quality also includes:
Sort module, for the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and extracts every
In Claims Resolution photo under one classification, the Claims Resolution photo of preset ratio, as training photo, extracts remaining Claims Resolution under each classification and shines
Piece is as checking photo;
Training module, for each classification under each training photo carry out feature extraction, with obtain to be entered to institute
State the first pixel vectors in depth convolutional neural networks model, and by each training photo corresponding first under each classification
Pixel vectors input to described depth convolutional neural networks model, to train the depth convolutional neural networks generating for identification
Model;
Authentication module, for carrying out feature extraction to each checking photo under each classification, to obtain input to training
The second pixel vectors in the depth convolutional neural networks model generating, and will be corresponding for each checking photo under each classification
Second pixel vectors input in the depth convolutional neural networks model generating to training, to verify the depth convolution god that training generates
Accuracy rate through network model;
Terminate module, if the accuracy rate of the depth convolutional neural networks model generating for training is more than or equal to default threshold
Value, then training terminates.
Preferably, the detection means of described picture quality also includes:
Loop module, if the accuracy rate of the depth convolutional neural networks model generating for training is less than predetermined threshold value,
Send the second prompting message to described user terminal, to remind user to increase the sample size of Claims Resolution photo.
Preferably, described training module is specifically for for each training photo under each classification, using different convolution
Core begins stepping through from first block of pixels of each training photo and carries out convolution algorithm to last block of pixels, every to extract
One training photo corresponding different characteristic figure;The characteristic pattern of each training photo extracting is carried out at pond and rasterisation
Reason, each training photo eigen figure extracting is processed into the first consistent pixel vectors of dimension;
Described authentication module is specifically for for each checking photo under each classification, using different convolution kernels from every
First block of pixels of one checking photo begins stepping through and carries out convolution algorithm to last block of pixels, to extract each checking
Photo corresponding different characteristic figure;Pond and rasterization process are carried out to the characteristic pattern of each checking photo extracting, will
The each checking photo eigen figure extracting is processed into the second consistent pixel vectors of dimension.
Preferably, described training module is specifically for estimating to obtain described depth convolution by using back-propagation bp method
The parameter of neural network model.
The invention has the beneficial effects as follows: compared to prior art, the detection method of picture quality of the present invention and dress
Put when processing vehicle insurance Claims Resolution photo, Claims Resolution photo is uploaded to vehicle insurance Claims Resolution server from user terminal, is given birth to by training in advance
The depth convolutional neural networks model becoming enters line definition analysis to Claims Resolution photo, and whether the levels of sharpness of determination Claims Resolution photo
Satisfaction is actually needed, if the levels of sharpness of described Claims Resolution photo is unsatisfactory for being actually needed, sends to user terminal and reminds
Information, to remind it again to upload Claims Resolution photo.The depth convolutional neural networks model pair that the present invention is generated by training in advance
Claims Resolution photo enters line definition identification it is ensured that the Claims Resolution photo that user is uploaded is all can to analyze exactly to draw vehicle insurance scene
The Claims Resolution photo of information, so, is favorably improved the work efficiency of self-service Claims Resolution system, improves Consumer's Experience.
Brief description
Fig. 1 is the schematic flow sheet of the detection method first embodiment of picture quality of the present invention;
Fig. 2 is the schematic flow sheet of the detection method second embodiment of picture quality of the present invention;
Fig. 3 is the schematic flow sheet of the detection method 3rd embodiment of picture quality of the present invention;
Fig. 4 is to carry out carrying out convolution during feature extraction to each training photo under each classification or checking photo in Fig. 2
Schematic diagram;
Fig. 5 is the structural representation of detection means one embodiment of picture quality of the present invention.
Specific embodiment
Below in conjunction with accompanying drawing, the principle of the present invention and feature are described, example is served only for explaining the present invention, and
Non- for limiting the scope of the present invention.
As shown in figure 1, Fig. 1 is the schematic flow sheet of detection method one embodiment of picture quality of the present invention, this picture product
The detection method of matter comprises the following steps:
Step s1, vehicle insurance Claims Resolution server, after receiving the Claims Resolution photo of user terminal uploads, is given birth to using training in advance
The depth convolutional neural networks model becoming enters line definition identification to the Claims Resolution photo receiving, to determine described Claims Resolution photo
Levels of sharpness;
In the present embodiment, the Claims Resolution needing into line definition identification of server receive user terminal upload of being settled a claim by vehicle insurance
Photo.User terminal can be mobile phone, panel computer, personal digital assistant (personal digital assistant,
Pda), the intelligent terminal such as wearable device (for example, intelligent watch, intelligent glasses etc.) or other any suitable electronics set
Standby.
In the present embodiment, training in advance generates depth convolutional neural networks model, the depth being generated using this training in advance
Convolutional neural networks model enters line definition identification to the Claims Resolution photo uploading, and specifically, training in advance generates depth convolution god
It is the neutral net of a multilamellar through network model, including feature extraction layer (c layer), Feature Mapping layer (s layer), every layer by multiple
Two dimensional surface forms, and each plane is made up of multiple independent neurons, each feature extraction layer (c layer) followed by
It is used for seeking the computation layer (s layer) of local average and second extraction, this distinctive structure of feature extraction twice makes it in identification
There is higher distortion tolerance to input sample.Claims Resolution photo inputs the depth convolutional neural networks generating to this training in advance
During Model Identification, in c layer, local experiences region (i.e. the sub-fraction of the photo) phase of the input of each neuron and preceding layer
Even, extract the feature in this local experiences region, after extraction, the position relationship between itself and other feature also determines therewith;In s layer,
It is made up of multiple Feature Mapping, and each Feature Mapping is a plane, and in plane, the weights of all neurons are equal.Feature is reflected
Penetrating structure can adopt the little sigmoid function of influence function core as the activation primitive of convolutional network so that Feature Mapping has
There is shift invariant.Claims Resolution photo enters after line definition identification through the depth convolutional neural networks model that this training in advance generates,
Photo of settling a claim exports respectively according to levels of sharpness, and wherein, levels of sharpness can be divided into fine definition grade, middle definition
Grade, low definition grade etc., it is of course also possible to distinguish levels of sharpness in other ways.
Step s2, if the levels of sharpness of described Claims Resolution photo is less than default levels of sharpness, sends the first prompting letter
Cease to described user terminal, to remind user again to upload Claims Resolution photo.
In the present embodiment, if the levels of sharpness of the Claims Resolution photo going out through depth convolutional neural networks Model Identification is less than
When the Claims Resolution photo of default levels of sharpness, such as upload is low definition grade, then the Claims Resolution of explanation user terminal uploads is shone
Piece is undesirable photo it is impossible to for the live situation of analysis vehicle insurance exactly, at this moment send first to user terminal
Prompting message, to remind user again to upload Claims Resolution photo;Certainly, if having higher requirement to the definition of Claims Resolution photo
When, need the Claims Resolution photo of user terminal uploads more fine definition grade, for example, upload the Claims Resolution photo of fine definition grade, and
The Claims Resolution photo of user terminal uploads is if middle levels of sharpness or low definition grade, all undesirable, needs
Send the first prompting message to user terminal, to remind user again to upload Claims Resolution photo.
Compared with prior art, the present embodiment uploads Claims Resolution photo from user terminal to vehicle insurance Claims Resolution server, by pre-
The depth convolutional neural networks model first training generation enters line definition analysis to Claims Resolution photo, determines the definition of Claims Resolution photo
Whether grade meets is actually needed, if the levels of sharpness of described Claims Resolution photo is unsatisfactory for being actually needed, to user terminal
Send prompting message, to remind it again to upload Claims Resolution photo.The depth convolutional Neural that the present embodiment is generated by training in advance
Network model enters line definition to Claims Resolution photo and identifies it is ensured that the Claims Resolution photo that user is uploaded is all to analyze exactly
Go out the Claims Resolution photo of vehicle insurance field data, so, be favorably improved the work efficiency of self-service Claims Resolution system, improve Consumer's Experience.
In a preferred embodiment, as shown in Fig. 2 on the basis of the embodiment of above-mentioned Fig. 1, above-mentioned step s1 it
Front inclusion:
S01, the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and is extracted under each classification
In Claims Resolution photo, as training photo, under each classification of extraction, remaining Claims Resolution photo is as checking for the Claims Resolution photo of preset ratio
Photo;
S02, carries out feature extraction to each training photo under each classification, to obtain depth convolution to be entered to described
The first pixel vectors in neural network model, and will be defeated for corresponding first pixel vectors of each training photo under each classification
Enter to described depth convolutional neural networks model, to train the depth convolutional neural networks model generating for identification;
S03, carries out feature extraction to each checking photo under each classification, to obtain the depth that input generates to training
The second pixel vectors in convolutional neural networks model, and by each classification under corresponding second pixel of each checking photo to
In the depth convolutional neural networks model that amount input generates to training, to verify the depth convolutional neural networks model that training generates
Accuracy rate;
S04, if the accuracy rate of the depth convolutional neural networks model of training generation is more than or equal to predetermined threshold value, training knot
Bundle.
In the present embodiment, when training generates depth convolutional neural networks model, can be in advance according to predetermined definition
Grade is classified to the Claims Resolution photo of predetermined number, for example, carry out levels of sharpness classification to 500,000 Claims Resolution photos, permissible
Classified according to predetermined levels of sharpness according to the resolution of Claims Resolution photo in advance, high-resolution Claims Resolution photo is high definition
Clear degree grade, the Claims Resolution photo of intermediate resolution are middle levels of sharpness, the Claims Resolution photo of low resolution is low definition grade.
After the completion of classification, the Claims Resolution photo that the Claims Resolution photo under each classification is respectively extracted with preset ratio, as training photo, for example will
Predetermined number Claims Resolution photo in 70% Claims Resolution photo as training photo, using each classification remaining Claims Resolution photo as
Checking photo, such as using remaining 30% Claims Resolution photo as training photo.
Then, feature extraction is carried out to each training photo under each classification, obtain different characteristic patterns to extract, right
Characteristic pattern carries out process of convolution, to finally give the first pixel vectors to depth convolutional neural networks model for the input, utilizes
The training of this first pixel vectors generates the depth convolutional neural networks model for identification;Each checking under each classification is shone
Piece carries out feature extraction, obtains different characteristic patterns to extract, carries out process of convolution to characteristic pattern, to finally give input to depth
The second pixel vectors in degree convolutional neural networks model, the depth convolution god being generated using the checking training of this second pixel vectors
Accuracy rate through network model.If checking obtains training the accuracy rate of the depth convolutional neural networks model generating to be more than or equal to
Preset value, is greater than equal to 0.95, then the depth convolutional neural networks model that explanation training generates can reach expected clear
Clear degree recognition effect, training terminates, and the depth convolutional neural networks model that subsequently can be generated using this training is entered to Claims Resolution photo
Line definition identifies.
In a preferred embodiment, as shown in figure 3, on the basis of the embodiment of above-mentioned Fig. 2, in above-mentioned steps s03
Also include afterwards:
S05, if the accuracy rate of the depth convolutional neural networks model of training generation is less than predetermined threshold value, to described user
Terminal sends the second prompting message, to remind user to increase the sample size of Claims Resolution photo, is back to described step s01 and follows
Ring.
In the present embodiment, if the accuracy rate of the depth convolutional neural networks model of training generation is less than predetermined threshold value, example
Such as less than 0.95, then the depth convolutional neural networks model that explanation training generates can not reach expected definition identification effect
Really, send the second prompting message to user terminal, to remind user to increase the sample size of Claims Resolution photo, based on increased Claims Resolution
Photo continues depth convolutional neural networks model is trained, and specifically, is receiving being increased of user terminal uploads
After Claims Resolution photo, may return in above-mentioned step s01, increased Claims Resolution photo is carried out point by predetermined levels of sharpness
Class, till the accuracy rate of the depth convolutional neural networks model that training generates is more than or equal to predetermined threshold value.
In a preferred embodiment, on the basis of the embodiment of above-mentioned Fig. 2, above-mentioned steps s02 are under each classification
Each training photo step of carrying out feature extraction include:
For each training photo under each classification, using different convolution kernels from each first pixel training photo
BOB(beginning of block) travels through and carries out convolution algorithm to last block of pixels, to extract each training photo corresponding different characteristic figure;
Pond and rasterization process are carried out to the characteristic pattern of each training photo extracting, by each instruction extracting
Practice photo eigen figure and be processed into the first consistent pixel vectors of dimension;
The step carrying out feature extraction to each checking photo under each classification in step s03 includes:
For each checking photo under each classification, using different convolution kernels from each first pixel verifying photo
BOB(beginning of block) travels through and carries out convolution algorithm to last block of pixels, to extract each checking photo corresponding different characteristic figure;
Pond and rasterization process are carried out to the characteristic pattern of each checking photo extracting, each tests extract
License piece characteristic pattern is processed into the second consistent pixel vectors of dimension.
In the present embodiment, for each training photo under each classification or checking photo, using different convolution kernels from every
First block of pixels of one training photo or checking photo begins stepping through and carries out convolution algorithm to last block of pixels, such as Fig. 4
Shown, for each block of pixels to be entered [(0,0), (1,0), (2,0), (0,1), (1,1), (2,1), (0,2), (1,2),
(2,2)], using convolution kernel (i, h, g, f, e, d, c, b, a) carry out convolution algorithm, from the first block of pixels carry out convolution algorithm and time
Go through to last block of pixels, each block of pixels correspondence obtains output vector (1,1), and the set of all output vectors can be instructed
Practice photo or checking photo corresponding different characteristic figure.Then, pond and rasterisation are carried out to the characteristic pattern of each training photo
Process, each training photo eigen figure extracting is processed into the first consistent pixel vectors of dimension, and, to each checking
The characteristic pattern of photo carries out pond and rasterization process, each checking photo eigen figure extracting is processed into dimension consistent
Second pixel vectors.
In a preferred embodiment, on the basis of the embodiment of above-mentioned Fig. 2, in above-mentioned steps s02, depth convolution god
Parameter through network model is estimated to obtain by using back-propagation bp method.Wherein, in training, for each training photo
Sequentially carry out convolution, Chi Hua, raster manipulation, the residual error between the actual value of Claims Resolution photo and estimated value can be obtained, using every
The residual error of secondary generation is inversely updated to parameter by rasterisation, Chi Hua, convolution, repeatedly forward direction inversely carry out above-mentioned
Operation is till global error restrains.Obtain the parameter of depth convolutional neural networks model when global error restrains.Wherein,
When training for the first time, the parameter of this depth convolutional neural networks model uses default parameterss, will not be described here.
As shown in figure 5, Fig. 5 is the structural representation of detection means one embodiment of picture quality of the present invention, this picture product
The detection means of matter includes:
Identification module 101, for the depth after receiving the Claims Resolution photo of user terminal uploads, being generated using training in advance
Degree convolutional neural networks model enters line definition identification to the Claims Resolution photo receiving, to determine the definition of described Claims Resolution photo
Grade;
In the present embodiment, the detection means of picture quality is integrated in vehicle insurance Claims Resolution server.User terminal can be handss
Machine, panel computer, personal digital assistant (personal digital assistant, pda), wearable device (for example, intelligence
Wrist-watch, intelligent glasses etc.) etc. intelligent terminal or other any suitable electronic equipment.
In the present embodiment, training in advance generates depth convolutional neural networks model, the depth being generated using this training in advance
Convolutional neural networks model enters line definition identification to the Claims Resolution photo uploading, and specifically, training in advance generates depth convolution god
It is the neutral net of a multilamellar through network model, including feature extraction layer (c layer), Feature Mapping layer (s layer), every layer by multiple
Two dimensional surface forms, and each plane is made up of multiple independent neurons, each feature extraction layer (c layer) followed by
It is used for seeking the computation layer (s layer) of local average and second extraction, this distinctive structure of feature extraction twice makes it in identification
There is higher distortion tolerance to input sample.Claims Resolution photo inputs the depth convolutional neural networks generating to this training in advance
During Model Identification, in c layer, local experiences region (i.e. the sub-fraction of the photo) phase of the input of each neuron and preceding layer
Even, extract the feature in this local experiences region, after extraction, the position relationship between itself and other feature also determines therewith;In s layer,
It is made up of multiple Feature Mapping, and each Feature Mapping is a plane, and in plane, the weights of all neurons are equal.Feature is reflected
Penetrating structure can adopt the little sigmoid function of influence function core as the activation primitive of convolutional network so that Feature Mapping has
There is shift invariant.Claims Resolution photo enters after line definition identification through the depth convolutional neural networks model that this training in advance generates,
Photo of settling a claim exports respectively according to levels of sharpness, and wherein, levels of sharpness can be divided into fine definition grade, middle definition
Grade, low definition grade etc., it is of course also possible to distinguish levels of sharpness in other ways.
Prompting module 102, if the levels of sharpness for described Claims Resolution photo is less than default levels of sharpness, sends the
One information extremely described user terminal, to remind user again to upload Claims Resolution photo.
In the present embodiment, if the levels of sharpness of the Claims Resolution photo going out through depth convolutional neural networks Model Identification is less than
When the Claims Resolution photo of default levels of sharpness, such as upload is low definition grade, then the Claims Resolution of explanation user terminal uploads is shone
Piece is undesirable photo it is impossible to for the live situation of analysis vehicle insurance exactly, at this moment send first to user terminal
Prompting message, to remind user again to upload Claims Resolution photo;Certainly, if having higher requirement to the definition of Claims Resolution photo
When, need the Claims Resolution photo of user terminal uploads more fine definition grade, for example, upload the Claims Resolution photo of fine definition grade, and
The Claims Resolution photo of user terminal uploads is if middle levels of sharpness or low definition grade, all undesirable, needs
Send the first prompting message to user terminal, to remind user again to upload Claims Resolution photo.
In a preferred embodiment, on the basis of the embodiment of above-mentioned Fig. 5, the detection means of above-mentioned picture quality is also
Including:
Sort module, for the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and extracts every
In Claims Resolution photo under one classification, the Claims Resolution photo of preset ratio, as training photo, extracts remaining Claims Resolution under each classification and shines
Piece is as checking photo;
Training module, for each classification under each training photo carry out feature extraction, with obtain to be entered to institute
State the first pixel vectors in depth convolutional neural networks model, and by each training photo corresponding first under each classification
Pixel vectors input to described depth convolutional neural networks model, to train the depth convolutional neural networks generating for identification
Model;
Authentication module, for carrying out feature extraction to each checking photo under each classification, to obtain input to training
The second pixel vectors in the depth convolutional neural networks model generating, and will be corresponding for each checking photo under each classification
Second pixel vectors input in the depth convolutional neural networks model generating to training, to verify the depth convolution god that training generates
Accuracy rate through network model;
Terminate module, if the accuracy rate of the depth convolutional neural networks model generating for training is more than or equal to default threshold
Value, then training terminates.
In the present embodiment, when training generates depth convolutional neural networks model, can be in advance according to predetermined definition
Grade is classified to the Claims Resolution photo of predetermined number, for example, carry out levels of sharpness classification to 500,000 Claims Resolution photos, permissible
Classified according to predetermined levels of sharpness according to the resolution of Claims Resolution photo in advance, high-resolution Claims Resolution photo is high definition
Clear degree grade, the Claims Resolution photo of intermediate resolution are middle levels of sharpness, the Claims Resolution photo of low resolution is low definition grade.
After the completion of classification, the Claims Resolution photo that the Claims Resolution photo under each classification is respectively extracted with preset ratio, as training photo, for example will
Predetermined number Claims Resolution photo in 70% Claims Resolution photo as training photo, using each classification remaining Claims Resolution photo as
Checking photo, such as using remaining 30% Claims Resolution photo as training photo.
Then, feature extraction is carried out to each training photo under each classification, obtain different characteristic patterns to extract, right
Characteristic pattern carries out process of convolution, to finally give the first pixel vectors to depth convolutional neural networks model for the input, utilizes
The training of this first pixel vectors generates the depth convolutional neural networks model for identification;Each checking under each classification is shone
Piece carries out feature extraction, obtains different characteristic patterns to extract, carries out process of convolution to characteristic pattern, to finally give input to depth
The second pixel vectors in degree convolutional neural networks model, the depth convolution god being generated using the checking training of this second pixel vectors
Accuracy rate through network model.If checking obtains training the accuracy rate of the depth convolutional neural networks model generating to be more than or equal to
Preset value, is greater than equal to 0.95, then the depth convolutional neural networks model that explanation training generates can reach expected clear
Clear degree recognition effect, training terminates, and the depth convolutional neural networks model that subsequently can be generated using this training is entered to Claims Resolution photo
Line definition identifies.
In a preferred embodiment, on the basis of the above embodiments, the detection means of above-mentioned picture quality is also wrapped
Include: loop module, if the accuracy rate of the depth convolutional neural networks model generating for training is less than predetermined threshold value, to described
User terminal sends the second prompting message, to remind user to increase the sample size of Claims Resolution photo, triggers described identification module simultaneously
Circulation.
In the present embodiment, if the accuracy rate of the depth convolutional neural networks model of training generation is less than predetermined threshold value, example
Such as less than 0.95, then the depth convolutional neural networks model that explanation training generates can not reach expected definition identification effect
Really, send the second prompting message to user terminal, to remind user to increase the sample size of Claims Resolution photo, based on increased Claims Resolution
Photo continues depth convolutional neural networks model is trained, and specifically, is receiving being increased of user terminal uploads
After Claims Resolution photo, above-mentioned identification module can be triggered and circulate, and increased Claims Resolution photo is entered by predetermined levels of sharpness
Row classification, till the accuracy rate of the depth convolutional neural networks model that training generates is more than or equal to predetermined threshold value.
In a preferred embodiment, on the basis of the above embodiments, described training module is specifically for for every
Each training photo under one classification, is begun stepping through to from first block of pixels of each training photo using different convolution kernels
A block of pixels carries out convolution algorithm afterwards, to extract each training photo corresponding different characteristic figure;Each to extract
The characteristic pattern of training photo carries out pond and rasterization process, and each training photo eigen figure extracting is processed into dimension
The first consistent pixel vectors;
Described authentication module is specifically for for each checking photo under each classification, using different convolution kernels from every
First block of pixels of one checking photo begins stepping through and carries out convolution algorithm to last block of pixels, to extract each checking
Photo corresponding different characteristic figure;Pond and rasterization process are carried out to the characteristic pattern of each checking photo extracting, will
The each checking photo eigen figure extracting is processed into the second consistent pixel vectors of dimension.
In the present embodiment, for each training photo under each classification or checking photo, using different convolution kernels from every
First block of pixels of one training photo or checking photo begins stepping through and carries out convolution algorithm to last block of pixels, such as Fig. 4
Shown, for each block of pixels to be entered [(0,0), (1,0), (2,0), (0,1), (1,1), (2,1), (0,2), (1,2),
(2,2)], using convolution kernel (i, h, g, f, e, d, c, b, a) carry out convolution algorithm, from the first block of pixels carry out convolution algorithm and time
Go through to last block of pixels, each block of pixels correspondence obtains output vector (1,1), and the set of all output vectors can be instructed
Practice photo or checking photo corresponding different characteristic figure.Then, pond and rasterisation are carried out to the characteristic pattern of each training photo
Process, each training photo eigen figure extracting is processed into the first consistent pixel vectors of dimension, and, to each checking
The characteristic pattern of photo carries out pond and rasterization process, each checking photo eigen figure extracting is processed into dimension consistent
Second pixel vectors.
In a preferred embodiment, on the basis of the above embodiments, described training module is specifically for by profit
Estimate to obtain the parameter of described depth convolutional neural networks model with back-propagation bp method.Wherein, in training, for each
Training photo sequentially carries out convolution, Chi Hua, raster manipulation, can obtain residual between the actual value of Claims Resolution photo and estimated value
Difference, is inversely updated to parameter by rasterisation, Chi Hua, convolution using each residual error producing, repeatedly positive reverse
Ground carries out aforesaid operations till global error convergence.Obtain depth convolutional neural networks model when global error restrains
Parameter.Wherein, when training for the first time, the parameter of this depth convolutional neural networks model uses default parameterss, and here is not done
Repeat.
The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all spirit in the present invention and
Within principle, any modification, equivalent substitution and improvement made etc., should be included within the scope of the present invention.
Claims (10)
1. a kind of detection method of picture quality is it is characterised in that the detection method of described picture quality includes:
S1, vehicle insurance settles a claim server after receiving the Claims Resolution photo of user terminal uploads, the depth being generated using training in advance
Convolutional neural networks model enters line definition identification to the Claims Resolution photo receiving, to determine definition of described Claims Resolution photo etc.
Level;
S2, if the levels of sharpness of described Claims Resolution photo is less than default levels of sharpness, sends the first information extremely described
User terminal, to remind user again to upload Claims Resolution photo.
2. according to claim 1 the detection method of picture quality it is characterised in that before described step s1, the method is also
Including:
S01, the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and is extracted the Claims Resolution under each classification
In photo, the Claims Resolution photo of preset ratio, as training photo, extracts remaining Claims Resolution photo under each classification and shines as checking
Piece;
S02, carries out feature extraction to each training photo under each classification, to obtain depth convolutional Neural to be entered to described
The first pixel vectors in network model, and by each classification under corresponding first pixel vectors of each training photo input to
In described depth convolutional neural networks model, to train the depth convolutional neural networks model generating for identification;
S03, carries out feature extraction to each checking photo under each classification, to obtain the depth convolution that input generates to training
The second pixel vectors in neural network model, and will be defeated for corresponding second pixel vectors of each checking photo under each classification
Enter in the depth convolutional neural networks model generating to training, to verify the standard of the depth convolutional neural networks model of training generation
Really rate;
S04, if the accuracy rate of the depth convolutional neural networks model of training generation is more than or equal to predetermined threshold value, training terminates.
3. according to claim 2 the detection method of picture quality it is characterised in that after described step s03, the method is also
Including:
S05, if the accuracy rate of the depth convolutional neural networks model of training generation is less than predetermined threshold value, to described user terminal
Send the second prompting message, to remind user to increase the sample size of Claims Resolution photo.
4. according to claim 2 picture quality detection method it is characterised in that described to each classification under each instruction
Practice photo or checking photo carry out the step of feature extraction and includes:
For each training photo under each classification or checking photo, using different convolution kernels from each training photo or checking
First block of pixels of photo begins stepping through and carries out convolution algorithm to last block of pixels, with extract each training photo or
Checking photo corresponding different characteristic figure;
Pond and rasterization process are carried out to the characteristic pattern of each training photo extracting or checking photo, by extract
Each training photo eigen figure is processed into the first consistent pixel vectors of dimension, at each checking photo eigen figure extracting
Manage into the second consistent pixel vectors of dimension.
5. according to claim 2 the detection method of picture quality it is characterised in that in described step s02, described depth
The parameter of convolutional neural networks model is estimated to obtain by using back-propagation bp method.
6. a kind of detection means of picture quality is it is characterised in that the detection means of described picture quality includes:
Identification module, for the depth convolution after receiving the Claims Resolution photo of user terminal uploads, being generated using training in advance
Neural network model enters line definition identification to the Claims Resolution photo receiving, to determine the levels of sharpness of described Claims Resolution photo;
Prompting module, if the levels of sharpness for described Claims Resolution photo is less than default levels of sharpness, sends the first prompting
Information extremely described user terminal, to remind user again to upload Claims Resolution photo.
7. according to claim 6 picture quality detection means it is characterised in that the detection means of described picture quality also
Including:
Sort module, for the Claims Resolution photo of predetermined number is classified by predetermined levels of sharpness, and extracts each point
In Claims Resolution photo under class, the Claims Resolution photo of preset ratio, as training photo, extracts remaining Claims Resolution photo under each classification and makees
For verifying photo;
Training module, for carrying out feature extraction to each training photo under each classification, to obtain depth to be entered to described
Degree convolutional neural networks model in the first pixel vectors, and by each classification under corresponding first pixel of each training photo
Vector inputs to described depth convolutional neural networks model, to train the depth convolutional neural networks mould generating for identification
Type;
Authentication module, for carrying out feature extraction to each checking photo under each classification, is generated to training with obtaining input
Depth convolutional neural networks model in the second pixel vectors, and by each classification under each checking photo corresponding second
Pixel vectors input in the depth convolutional neural networks model generating to training, to verify the depth convolutional Neural net that training generates
The accuracy rate of network model;
Terminate module, if the accuracy rate of the depth convolutional neural networks model generating for training is more than or equal to predetermined threshold value,
Training terminates.
8. according to claim 7 picture quality detection means it is characterised in that the detection means of described picture quality also
Including:
Loop module, if the accuracy rate of the depth convolutional neural networks model generating for training is less than predetermined threshold value, to institute
State user terminal and send the second prompting message, to remind user to increase the sample size of Claims Resolution photo.
9. according to claim 7 the detection means of picture quality it is characterised in that described training module is specifically for right
Each training photo under each classification, is begun stepping through from first block of pixels of each training photo using different convolution kernels
Carry out convolution algorithm to last block of pixels, to extract each training photo corresponding different characteristic figure;To extract
The characteristic pattern of each training photo carries out pond and rasterization process, and each training photo eigen figure extracting is processed into
The first consistent pixel vectors of dimension;
Described authentication module is specifically for for each checking photo under each classification, being tested from each using different convolution kernels
First block of pixels of license piece begins stepping through and carries out convolution algorithm to last block of pixels, to extract each checking photo
Corresponding different characteristic figure;Pond and rasterization process are carried out to the characteristic pattern of each checking photo extracting, will extract
The each checking photo eigen figure going out is processed into the second consistent pixel vectors of dimension.
10. according to claim 7 the detection means of picture quality it is characterised in that described training module is specifically for logical
Cross and estimate to obtain the parameter of described depth convolutional neural networks model using back-propagation bp method.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610704799.1A CN106372651B (en) | 2016-08-22 | 2016-08-22 | The detection method and device of picture quality |
PCT/CN2017/091306 WO2018036276A1 (en) | 2016-08-22 | 2017-06-30 | Image quality detection method, device, server and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610704799.1A CN106372651B (en) | 2016-08-22 | 2016-08-22 | The detection method and device of picture quality |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106372651A true CN106372651A (en) | 2017-02-01 |
CN106372651B CN106372651B (en) | 2018-03-06 |
Family
ID=57878027
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610704799.1A Active CN106372651B (en) | 2016-08-22 | 2016-08-22 | The detection method and device of picture quality |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN106372651B (en) |
WO (1) | WO2018036276A1 (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107342962A (en) * | 2017-07-03 | 2017-11-10 | 北京邮电大学 | Deep learning intelligence Analysis On Constellation Map method based on convolutional neural networks |
WO2018036276A1 (en) * | 2016-08-22 | 2018-03-01 | 平安科技(深圳)有限公司 | Image quality detection method, device, server and storage medium |
CN107918916A (en) * | 2017-09-13 | 2018-04-17 | 平安科技(深圳)有限公司 | Self-service Claims Resolution application processing method, device, computer equipment and storage medium |
WO2018191435A1 (en) * | 2017-04-11 | 2018-10-18 | Alibaba Group Holding Limited | Picture-based vehicle loss assessment method and apparatus, and electronic device |
CN108764261A (en) * | 2018-05-31 | 2018-11-06 | 努比亚技术有限公司 | A kind of image processing method, mobile terminal and storage medium |
CN109785312A (en) * | 2019-01-16 | 2019-05-21 | 创新奇智(广州)科技有限公司 | A kind of image fuzzy detection method, system and electronic equipment |
CN110689322A (en) * | 2019-09-27 | 2020-01-14 | 成都知识视觉科技有限公司 | Artificial intelligence auxiliary claims checking system suitable for insurance claims settlement process |
CN110766033A (en) * | 2019-05-21 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
US10789786B2 (en) | 2017-04-11 | 2020-09-29 | Alibaba Group Holding Limited | Picture-based vehicle loss assessment |
US10817956B2 (en) | 2017-04-11 | 2020-10-27 | Alibaba Group Holding Limited | Image-based vehicle damage determining method and apparatus, and electronic device |
WO2021012891A1 (en) * | 2019-07-23 | 2021-01-28 | 平安科技(深圳)有限公司 | Vehicle loss assessment method, device, apparatus, and storage medium |
CN112788131A (en) * | 2020-12-31 | 2021-05-11 | 平安科技(深圳)有限公司 | Method, system and storage medium for generating early warning picture based on artificial intelligence |
CN114241180A (en) * | 2021-12-15 | 2022-03-25 | 平安科技(深圳)有限公司 | Image detection method and device for vehicle damage claims, computer equipment and storage medium |
US11544914B2 (en) | 2021-02-18 | 2023-01-03 | Inait Sa | Annotation of 3D models with signs of use visible in 2D images |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112911393B (en) * | 2018-07-24 | 2023-08-01 | 广州虎牙信息科技有限公司 | Method, device, terminal and storage medium for identifying part |
CN109145903A (en) * | 2018-08-22 | 2019-01-04 | 阿里巴巴集团控股有限公司 | A kind of image processing method and device |
CN109948625A (en) * | 2019-03-07 | 2019-06-28 | 上海汽车集团股份有限公司 | Definition of text images appraisal procedure and system, computer readable storage medium |
CN109949323B (en) * | 2019-03-19 | 2022-12-20 | 广东省农业科学院农业生物基因研究中心 | Crop seed cleanliness judgment method based on deep learning convolutional neural network |
CN110705847A (en) * | 2019-09-18 | 2020-01-17 | 中国南方电网有限责任公司超高压输电公司广州局 | Intelligent substation inspection method and system based on image recognition technology |
CN110795579B (en) * | 2019-10-29 | 2022-11-18 | Oppo广东移动通信有限公司 | Picture cleaning method and device, terminal and storage medium |
CN111327831B (en) * | 2020-03-30 | 2021-09-10 | 北京智美智学科技有限公司 | Image acquisition method and device for UGC, electronic equipment and system |
CN111709906A (en) * | 2020-04-13 | 2020-09-25 | 北京深睿博联科技有限责任公司 | Medical image quality evaluation method and device |
CN111563663B (en) * | 2020-04-16 | 2023-03-21 | 五邑大学 | Robot, service quality evaluation method and system |
CN112365451A (en) * | 2020-10-23 | 2021-02-12 | 微民保险代理有限公司 | Method, device and equipment for determining image quality grade and computer readable medium |
CN112803341A (en) * | 2020-12-31 | 2021-05-14 | 国网浙江省电力有限公司嘉兴供电公司 | Non-invasive cable anti-breaking monitoring device and method |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103544705A (en) * | 2013-10-25 | 2014-01-29 | 华南理工大学 | Image quality testing method based on deep convolutional neural network |
CN104866524A (en) * | 2015-04-10 | 2015-08-26 | 大连交通大学 | Fine classification method for commodity images |
CN105160678A (en) * | 2015-09-02 | 2015-12-16 | 山东大学 | Convolutional-neural-network-based reference-free three-dimensional image quality evaluation method |
CN105512624A (en) * | 2015-12-01 | 2016-04-20 | 天津中科智能识别产业技术研究院有限公司 | Smile face recognition method and device for human face image |
CN105719188A (en) * | 2016-01-22 | 2016-06-29 | 平安科技(深圳)有限公司 | Method and server for achieving insurance claim anti-fraud based on consistency of multiple pictures |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101609500B (en) * | 2008-12-01 | 2012-07-25 | 公安部第一研究所 | Quality estimation method of exit-entry digital portrait photos |
KR101906827B1 (en) * | 2012-04-10 | 2018-12-05 | 삼성전자주식회사 | Apparatus and method for taking a picture continously |
CN105405054A (en) * | 2015-12-11 | 2016-03-16 | 平安科技(深圳)有限公司 | Insurance claim antifraud implementation method based on claim photo deep learning and server |
CN106372651B (en) * | 2016-08-22 | 2018-03-06 | 平安科技(深圳)有限公司 | The detection method and device of picture quality |
-
2016
- 2016-08-22 CN CN201610704799.1A patent/CN106372651B/en active Active
-
2017
- 2017-06-30 WO PCT/CN2017/091306 patent/WO2018036276A1/en active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103544705A (en) * | 2013-10-25 | 2014-01-29 | 华南理工大学 | Image quality testing method based on deep convolutional neural network |
CN104866524A (en) * | 2015-04-10 | 2015-08-26 | 大连交通大学 | Fine classification method for commodity images |
CN105160678A (en) * | 2015-09-02 | 2015-12-16 | 山东大学 | Convolutional-neural-network-based reference-free three-dimensional image quality evaluation method |
CN105512624A (en) * | 2015-12-01 | 2016-04-20 | 天津中科智能识别产业技术研究院有限公司 | Smile face recognition method and device for human face image |
CN105719188A (en) * | 2016-01-22 | 2016-06-29 | 平安科技(深圳)有限公司 | Method and server for achieving insurance claim anti-fraud based on consistency of multiple pictures |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018036276A1 (en) * | 2016-08-22 | 2018-03-01 | 平安科技(深圳)有限公司 | Image quality detection method, device, server and storage medium |
US10817956B2 (en) | 2017-04-11 | 2020-10-27 | Alibaba Group Holding Limited | Image-based vehicle damage determining method and apparatus, and electronic device |
WO2018191435A1 (en) * | 2017-04-11 | 2018-10-18 | Alibaba Group Holding Limited | Picture-based vehicle loss assessment method and apparatus, and electronic device |
US11049334B2 (en) | 2017-04-11 | 2021-06-29 | Advanced New Technologies Co., Ltd. | Picture-based vehicle loss assessment |
WO2018191437A1 (en) * | 2017-04-11 | 2018-10-18 | Alibaba Group Holding Limited | Image-based vehicle loss assessment method, apparatus, and system, and electronic device |
US10789786B2 (en) | 2017-04-11 | 2020-09-29 | Alibaba Group Holding Limited | Picture-based vehicle loss assessment |
CN107342962A (en) * | 2017-07-03 | 2017-11-10 | 北京邮电大学 | Deep learning intelligence Analysis On Constellation Map method based on convolutional neural networks |
WO2019052226A1 (en) * | 2017-09-13 | 2019-03-21 | 平安科技(深圳)有限公司 | Processing method and apparatus for self-service claim settlement application, computer device and storage medium |
CN107918916A (en) * | 2017-09-13 | 2018-04-17 | 平安科技(深圳)有限公司 | Self-service Claims Resolution application processing method, device, computer equipment and storage medium |
CN108764261A (en) * | 2018-05-31 | 2018-11-06 | 努比亚技术有限公司 | A kind of image processing method, mobile terminal and storage medium |
CN109785312B (en) * | 2019-01-16 | 2020-10-09 | 创新奇智(广州)科技有限公司 | Image blur detection method and system and electronic equipment |
CN109785312A (en) * | 2019-01-16 | 2019-05-21 | 创新奇智(广州)科技有限公司 | A kind of image fuzzy detection method, system and electronic equipment |
CN110766033A (en) * | 2019-05-21 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
WO2021012891A1 (en) * | 2019-07-23 | 2021-01-28 | 平安科技(深圳)有限公司 | Vehicle loss assessment method, device, apparatus, and storage medium |
CN110689322A (en) * | 2019-09-27 | 2020-01-14 | 成都知识视觉科技有限公司 | Artificial intelligence auxiliary claims checking system suitable for insurance claims settlement process |
CN112788131A (en) * | 2020-12-31 | 2021-05-11 | 平安科技(深圳)有限公司 | Method, system and storage medium for generating early warning picture based on artificial intelligence |
US11544914B2 (en) | 2021-02-18 | 2023-01-03 | Inait Sa | Annotation of 3D models with signs of use visible in 2D images |
CN114241180A (en) * | 2021-12-15 | 2022-03-25 | 平安科技(深圳)有限公司 | Image detection method and device for vehicle damage claims, computer equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN106372651B (en) | 2018-03-06 |
WO2018036276A1 (en) | 2018-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106372651B (en) | The detection method and device of picture quality | |
CN110674688B (en) | Face recognition model acquisition method, system and medium for video monitoring scene | |
CN111612807B (en) | Small target image segmentation method based on scale and edge information | |
CN108229479A (en) | The training method and device of semantic segmentation model, electronic equipment, storage medium | |
CN108805016B (en) | Head and shoulder area detection method and device | |
CN106415594A (en) | A method and a system for face verification | |
CN115063573A (en) | Multi-scale target detection method based on attention mechanism | |
CN107563274A (en) | A kind of vehicle checking method and method of counting of the video based on confrontation e-learning | |
CN104408728A (en) | Method for detecting forged images based on noise estimation | |
CN109325435B (en) | Video action recognition and positioning method based on cascade neural network | |
CN110956080A (en) | Image processing method and device, electronic equipment and storage medium | |
CN113361567B (en) | Image processing method, device, electronic equipment and storage medium | |
CN114529687A (en) | Image reconstruction method and device, electronic equipment and computer readable storage medium | |
Guo et al. | Haze visibility enhancement for promoting traffic situational awareness in vision-enabled intelligent transportation | |
CN111275070B (en) | Signature verification method and device based on local feature matching | |
CN113807237B (en) | Training of in vivo detection model, in vivo detection method, computer device, and medium | |
CN116630917A (en) | Lane line detection method | |
CN115661611A (en) | Infrared small target detection method based on improved Yolov5 network | |
CN115661803A (en) | Image definition detection method, electronic device, and computer-readable storage medium | |
CN114743045A (en) | Small sample target detection method based on double-branch area suggestion network | |
CN114596609A (en) | Audio-visual counterfeit detection method and device | |
CN113780241A (en) | Acceleration method and device for detecting salient object | |
CN112733686A (en) | Target object identification method and device used in image of cloud federation | |
Varkentin et al. | Development of an application for vehicle classification using neural networks technologies | |
CN112699928B (en) | Non-motor vehicle detection and identification method based on deep convolutional network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |