CN108734184A - A kind of method and device that sensitive image is analyzed - Google Patents
A kind of method and device that sensitive image is analyzed Download PDFInfo
- Publication number
- CN108734184A CN108734184A CN201710248908.8A CN201710248908A CN108734184A CN 108734184 A CN108734184 A CN 108734184A CN 201710248908 A CN201710248908 A CN 201710248908A CN 108734184 A CN108734184 A CN 108734184A
- Authority
- CN
- China
- Prior art keywords
- sample
- training
- image
- picture
- kinds
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0641—Shopping interfaces
- G06Q30/0643—Graphical representation of items or shoppers
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the invention discloses a kind of method and devices analyzed sensitive image, are related to image identification technical field, can promote the automatization level for advertising pictures recognition detection, reduce manual examination and verification cost.The present invention includes：By being clustered to the sample image in the training sample set, further according to the sample image by cluster, by the corresponding all kinds of identification model of convolutional neural networks training, the identification model obtained later using training identifies corresponding all kinds of sensitization picture from picture library to be detected.The present invention is suitable for identifying the sensitization picture on line platform.
Description
Technical field
The present invention relates to image identification technical field more particularly to a kind of method analyzed sensitive image and dresses
It sets.
Background technology
With the construction of the development and the network platforms such as all kinds of online transaction platforms, on-line marketing platform of Internet technology,
Major operator and size retail shop can all launch the Internet advertising of magnanimity in the network platform all the time.In order to which specification interconnects
The publication behavior of net advertisement, protects the legitimate rights and interests of consumers, in the new Advertising Law promulgated in 2015, clear stipulaties internet
Advertising campaign also has to comply with Advertising Law items regulation.
In current practical application, each network platform monitors the movable means of Internet advertising, mainly passes through detection
Sensitive image differentiates and early warning may illegal advertisement.Existing nude picture detection method refers in particular to pornographic image, phase mostly
The detection means answered and analysis method main development are from foundation《The Law of the P.R.C. on Administrative Penalties for Public Security》With《Criminal law》To obscene pornography
Identification field is monitored, detection mode is mainly based on the sensitive organ of detection.Such as：Hand-designed has fixed color, shape
With the characteristics of image of texture, and doubtful sensitive image is obtained according to the Image Feature Matching manually set.
But the accuracy of identification of existing way is relatively low, often by under normal circumstances birds meat products, underwear, sporting goods,
The advertisements of the classifications commodity such as family planning articles, publicity image are reported by mistake into sensitive image, in the past mainly by being appealed by report side or
The mode of monitoring personnel artificial treatment solves the problems, such as wrong report, this has been difficult to meet the magnanimity to being launched in the network platform at present
The demand that Internet advertising is monitored in real time does not adapt to electric business platform especially and magnanimity demonstration image filtering is supervised
The requirement of control, it is therefore desirable to the higher detection means of the degree of automation is developed, to control cost of labor.
Invention content
The embodiment of the present invention provides a kind of method analyzed sensitive image, can be promoted and advertising pictures are known
The automatization level not detected reduces manual examination and verification cost.
In order to achieve the above objectives, the embodiment of the present invention adopts the following technical scheme that：
In a first aspect, the method that the embodiment of the present invention provides, including：
In extracted training sample set, according to the corresponding sensitive kind of each sample, in the training sample set
Sample image clustered；
According to the sample image by cluster, pass through the corresponding all kinds of identification model of convolutional neural networks training；
The identification model obtained using training identifies corresponding all kinds of sensitization picture from picture library to be detected.
With reference to first aspect, described corresponding according to each sample in the first possible realization method of first aspect
Sensitive kind clusters the sample image in the training sample set, including：
By preset neural network model, the sensitive features of each sample image are extracted from the training sample set,
Wherein, the preset neural network model is trained by imagenet；
By preset clustering algorithm, the sample image that the similarity degree of sensitive features is met to test order is clustered to same
One sample set closes.
The possible realization method of with reference to first aspect the first further includes in second of possible realization method：
In a sample set closes：
According at a distance from cluster centre, by the sample image in the subclass by closely sorting to remote, and sequence is chosen
Preceding specified digit sample image as positive sample；
Utilize obtained positive sample training pattern grader；
By the trained model classifiers, the sample image in being closed to the sample set carries out classification meter
It calculates, and the sample image by the score value being calculated less than pre-determined threshold is rejected.
The possible realization method of with reference to first aspect the first further includes in the third possible realization method：
Utilize extracted pre-training data set, the depth residual error network of the specified number of plies of training, the specified number of plies >=50；
The depth residual error network obtained by training corrects the sample set and closes.
With reference to first aspect, in the 4th kind of possible realization method of first aspect, further include：
All kinds of sensitization pictures is corresponded in the picture library to be detected to be identified, and after obtaining recognition result,
Difficult example sample is extracted from the recognition result；
The parameter of all kinds of identification models is corresponded to according to the difficult example Sample Refreshment.
The 4th kind of possible realization method with reference to first aspect, it is described from institute in the 5th kind of possible realization method
It states and extracts difficult example sample in recognition result, including：
Obtain the score value of each attribute in sensitization picture, wherein the score value of each attribute passes through institute in the sensitization picture
Identification model is stated to be calculated；
According to the sequence of score value from large to small, sort to each attribute of acquired sensitization picture；
The additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence
When spending threshold value, judgement is as the difficult example sample.
With reference to first aspect, in the 6th kind of possible realization method of first aspect, further include：
According to preset business rule, candidate image is acquired from electric business service platform, utilizes acquired candidate image more
The new picture library；
And/or according to preset test order, the trained sample is extracted from the sample database pointed by the test order
This set.
Second aspect, the device that the embodiment of the present invention provides, including：
Cluster module is used in extracted training sample set, according to the corresponding sensitive kind of each sample, to the instruction
The sample image practiced in sample set is clustered；
Training module, for according to the sample image by cluster, passing through the corresponding all kinds of knowledge of convolutional neural networks training
Other model；
Analysis module, the identification model for being obtained using training identify that correspondence is all kinds of from picture library to be detected
Sensitization picture.
In conjunction with second aspect, in the first possible realization method of second aspect, the cluster module is specifically used for
By preset neural network model, the sensitive features of each sample image are extracted from the training sample set；And by pre-
If clustering algorithm, the sample image that the similarity degree of sensitive features is met to test order clusters to the same sample set
It closes；Wherein, the preset neural network model is trained by imagenet.
Further include in second of possible realization method of second aspect in conjunction with second aspect：Filtering module is used for
In a sample set closes：According at a distance from cluster centre, by the sample image in the subclass by closely to remote sequence,
And choose sequence preceding specified digit sample image as positive sample；Obtained positive sample training pattern is recycled to classify
Device；Later by the trained model classifiers, the sample image in being closed to the sample set carries out classified calculating,
And the sample image by the score value being calculated less than pre-determined threshold is rejected；
Further include：Correction module, for utilizing extracted pre-training data set, the depth residual error net of the specified number of plies of training
Network, the specified number of plies >=50；And the depth residual error network obtained by training, it corrects the sample set and closes.
Further include in the third possible realization method of second aspect in conjunction with second aspect：
Update module is identified for corresponding to all kinds of sensitization pictures in the picture library to be detected, and
To after recognition result, the score value of each attribute in sensitization picture is obtained, wherein the score value of each attribute is logical in the sensitization picture
The identification model is crossed to be calculated；And the sequence according to score value from large to small, to each attribute of acquired sensitization picture
Sequence；
The additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence
When spending threshold value, judgement is as the difficult example sample.；And the ginseng of all kinds of identification models is corresponded to according to the difficult example Sample Refreshment
Number.
The method and device provided in an embodiment of the present invention that sensitive image is analyzed, by the training sample set
Sample image in conjunction is clustered, all kinds of by convolutional neural networks training correspondence further according to the sample image by cluster
Identification model corresponding all kinds of Sensitive Graphs are identified from picture library to be detected later using the obtained identification model of training
Piece, to identify whether the picture that the trade company of electric business service platform is uploaded belongs to corresponding all kinds of sensitization picture.It realizes
For trade company upload electric business service platform picture it is automatic detection, scanning, improve for advertising pictures recognition detection oneself
Dynamicization is horizontal, reduces manual examination and verification cost.
Description of the drawings
It to describe the technical solutions in the embodiments of the present invention more clearly, below will be to needed in the embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for ability
For the those of ordinary skill of domain, without creative efforts, it can also be obtained according to these attached drawings other attached
Figure.
Fig. 1 is a kind of possible system architecture schematic diagram provided in an embodiment of the present invention；
Fig. 2 is method flow schematic diagram provided in an embodiment of the present invention；
Fig. 3,4 be specific example provided in an embodiment of the present invention schematic diagram；
Fig. 5,6,7 are the structural schematic diagram of device provided in an embodiment of the present invention.
Specific implementation mode
To make those skilled in the art more fully understand technical scheme of the present invention, below in conjunction with the accompanying drawings and specific embodiment party
Present invention is further described in detail for formula.Embodiments of the present invention are described in more detail below, the embodiment is shown
Example is shown in the accompanying drawings, and in which the same or similar labels are throughly indicated same or similar element or has identical or class
Like the element of function.It is exemplary below with reference to the embodiment of attached drawing description, is only used for explaining the present invention, and cannot
It is construed to limitation of the present invention.Those skilled in the art of the present technique are appreciated that unless expressly stated, odd number shape used herein
Formula " one ", "one", " described " and "the" may also comprise plural form.It is to be further understood that the specification of the present invention
The middle wording " comprising " used refers to that there are the feature, integer, step, operation, element and/or component, but it is not excluded that
Other one or more features of presence or addition, integer, step, operation, element, component and/or their group.It should be understood that
When we say that an element is " connected " or " coupled " to another element, it can be directly connected or coupled to other elements, or
There may also be intermediary elements.In addition, " connection " used herein or " coupling " may include being wirelessly connected or coupling.Here make
Wording "and/or" includes any cell of one or more associated list items and all combines.The art
Technical staff is appreciated that unless otherwise defined all terms (including technical terms and scientific terms) used herein have
Meaning identical with the general understanding of the those of ordinary skill in fields of the present invention.It should also be understood that such as general
Term, which should be understood that, those of defined in dictionary has a meaning that is consistent with the meaning in the context of the prior art, and
Unless being defined as here, will not be explained with the meaning of idealization or too formal.
The embodiment of the present invention specifically may be implemented in a kind of system environments as shown in Figure 1, specifically include：Electric business takes
Business platform, monitoring server and database；
Wherein it is possible to by monitoring server according to preset business rule, candidate image is acquired from electric business service platform, and
Acquired candidate image is utilized to update the picture library in the database.It is taken in electric business specifically, picture library includes publication
The advertising pictures being engaged on platform, or issued on the page that commodity details page, browse page etc. are used to show to consumer
Picture.It can be taken from electric business by monitoring server real-time (such as according to certain update cycle, such as 10 minutes, 1 hour)
All kinds of pictures are acquired as the candidate image on business platform, and import the picture library in the database.
Monitoring service implement body can be the server apparatus being individually made, such as：Rack, blade, tower or machine
Cabinet type server apparatus can also use work station, mainframe computer etc. to have stronger computing capability hardware device；It can also
The server cluster being made of multiple server apparatus.
Database is mainly used for storing picture library, can be specifically a kind of Redis databases or other kinds of distribution
Formula database, relevant database etc., can be specifically include storage device data server and with data server phase
Storage device even, or a kind of server set for database for being made of multiple data servers and storage server
Group's system.
Electric business service platform can be specifically current on-line operation, include all kinds of service sub-systems, for online
The plateform system of transaction, merchandise sales.On hardware view, electric business service platform is also specifically by a series of mutual foundation
The server cluster of communication connection forms, electric business service platform specifically specific construction mode and used architecture standard, can
With the common technology with reference to used in several large-scale net purchase platforms domestic at present, repeat no more in the present embodiment.
The embodiment of the present invention provides a kind of method analyzed sensitive image, as shown in Fig. 2, including：
S1, in extracted training sample set, according to the corresponding sensitive kind of each sample, to the training sample set
In sample image clustered.
Include default in test order library wherein it is possible to also set up test order library in monitoring server in advance
Test order, the test order specifically can by technical staff set and input monitoring server, such as：By technical staff
Design some test templates, wherein a set of test template includes for test order, the correspondence set by specific application scene
Training sample set and test needed for algorithm model, in order to which monitoring server is according to preset test order, from described
The training sample set is extracted in sample database pointed by test order, it specifically can be according to current specific test environment certainly
It is dynamic to transfer (or being operated by technical staff) test order.
In the present embodiment, sensitive kind can be understood as：For the picture of different traffic directions, represented commodity
Either difference can pre-set difference to the type of article for commodity or article of difference in these pictures
The sensitive kind of type, such as：For meat of poultris commodity, interior clothing commodity and family planning articles this 3 kinds of different traffic directions
Commodity are sold the type of merchandize represented by the advertising pictures that the commodity of these commodity are launched and are different, sensitivity can be arranged
Class1,2,3, and the advertising pictures of meat of poultris commodity are clustered to sensitive kind 1, cluster the advertising pictures of interior clothing commodity
It is clustered to sensitive kind 3 to sensitive kind 2, by the advertising pictures of family planning articles.So that being taken from electric business by monitoring server
Collected advertising pictures carry out classification differentiation according to sensitive kind in business platform, such as：Monitoring server is identified in Sensitive Graphs
During filtering, the input of the identification model of sensitization picture be pending image url (Uniform Resoure Locator,
Uniform resource locator), export the attributive classification result for pending image.
S2, the sample image clustered according to process pass through the corresponding all kinds of identification model of convolutional neural networks training.
Wherein, it can be understood as by the corresponding all kinds of identification model of convolutional neural networks training：Training is based on convolution god
Currently used convolutional neural networks technology specifically may be used in identification disaggregated model through network, according to specific practical industry
Business scene can respectively be built corresponding each come the identification model for building sensitization picture identification for different sensitive kinds
The identification model of sensitive kind.
In the present embodiment, in the identification model of the corresponding each sensitive kind of specific training, for different types of sensitivity
Type can configure different recognizers and for identification matched reference data, such as：It can be from electric business service platform
Each service sub-system in extract corresponding professional knowledge rule, professional knowledge rule be used to indicate the traffic direction commodity or
Some exclusive features of person's article, such as：The professional knowledge rule of meat of poultris commodity includes the type of animal limb, internal organ,
And typical legend information, colouring information and the profile information of animal limb, internal organ are used for then for the commodity of sensitive kind 1
Identify that matched reference data may include the typical legend information and profile information of animal limb, internal organ, and in identification decision
During, it is non-sensitive picture by the spectral discrimination for meeting these reference datas；For another example：The professional knowledge of interior clothing commodity
Rule includes the shape, color and Facing material of common underwear (such as can simple process decision chart by color and luster and glossiness
Which region is cloth in piece, which region is the skin of model), then it is matched for identification for the commodity of sensitive kind 2
Reference data may include shape, color, Facing material, and during identification decision, and will meet these reference datas
And sensitive organ is wherein not present, and (traditional discriminant approach may be used in the judgement of sensitive organ, for example public security organ uses
Obscene picture recognition means) spectral discrimination be non-sensitive picture
S3, the identification model obtained using training identify corresponding all kinds of sensitization picture from picture library to be detected.
Wherein, include the picture that is uploaded of trade company of electric business service platform in picture library to be detected, by monitoring service
Device executes the flow of the present embodiment, to identify whether the picture that the trade company of electric business service platform is uploaded belongs to corresponding all kinds of
Sensitization picture.
According to the regulation of new Advertising Law, monitoring and the pipe of the advertising pictures for reinforcing being issued on electric business service platform are needed
Reason.But manually inspection processing is carried out by user feedback and visually at present, that there are efficiency is low, risk is big and heavy workload etc.
Problem, therefore, it is necessary to the monitoring servers in through this embodiment to carry out automatic identification, in order to be able in time to electric business service
The operator of platform and businessman send out warning.
It provides a kind of during filtering is identified for the sensitization picture under electric business business, identifies the calculation of filtering
The training sample of method and the specific method of filtering.Different from the side of conventional method hand-designed color shapes textures characteristics of image
Method, present invention employs convolutional neural networks, reduce the cost of labor of hand-designed feature.It is provided in an embodiment of the present invention right
The method that sensitive image is analyzed, by being clustered to the sample image in the training sample set, further according to process
The sample image of cluster, by the corresponding all kinds of identification model of convolutional neural networks training, the identification obtained later using training
Model identifies corresponding all kinds of sensitization picture, to identify the institute of trade company of electric business service platform from picture library to be detected
Whether the picture of upload belongs to corresponding all kinds of sensitization picture.Realize for trade company upload electric business service platform picture from
Dynamic detection, scanning, improve the automatization level for advertising pictures recognition detection, reduce manual examination and verification cost.
Specifically, described carry out the sample image in the training sample set according to the corresponding sensitive kind of each sample
Cluster, including：
By preset neural network model, the sensitive features of each sample image are extracted from the training sample set.
Again by preset clustering algorithm, the sample image that the similarity degree of sensitive features is met to test order is clustered to the same sample
This subclass.
Wherein, the preset neural network model is by imagenet (in a kind of computer vision system identification project
For image recognition training database) be trained.Such as：The neural network model trained using imagenet, then profit
The sensitive features of candidate image are extracted with the neural network model trained.To be had using preset clustering algorithm later similar
A subset conjunction is merged into the extraction of sensitive features.
Further comprise the concrete mode be filtered, screen, arranged for being closed respectively for each sample set,
It is combined into example with one of sample set, in a sample set closes：
According at a distance from cluster centre, by the sample image in the subclass by closely sorting to remote, and sequence is chosen
Preceding specified digit sample image as positive sample.Utilizing obtained positive sample training pattern grader.Pass through later
The trained model classifiers, the sample image in being closed to the sample set carries out classified calculating, and will calculate
The score value arrived is rejected less than the sample image of pre-determined threshold.Such as:Candidate collection in each subclass, using apart from cluster centre
For 100 nearest samples as positive sample, (for example the classification based on one class svm may be used in training pattern grader
Device)；Grader classified calculating is all used to every figure in subclass, according to classification results, by the lower non-class image of score
It weeds out.
Optionally, one kind is also provided in the present embodiment for the corrected concrete mode of sample set, including：
Utilize extracted pre-training data set, the depth residual error network of the specified number of plies of training, the specified number of plies >=50.
The depth residual error network obtained later by training corrects the sample set and closes.Such as：The public affairs held using ImageNet
The depth residual error network that 1000 class Classification and Identification data sets used in match train 50 layers as pre-training data set is opened, it will
The parameter for the model that pre-training obtains is finely adjusted using the sample set conjunction above with filtering, screening, arrangement processing, from
And it avoids, due to the very few caused over-fitting of Sensitive Graphs training data, also avoiding feature extraction complicated in tional identification algorithm
Step.
For example：As shown in Figure 4, the design for the specific cellular construction of depth residual error network, if depth network
In certain hidden layer be H (x)-x → F (x), if it can be assumed that multiple non-linear layers combination can be similar to a complicated function,
It so similarly assume that the residual error of hidden layer is similar to some complicated function.So hidden layer can be expressed as H (x)
=F (x)+x.So a kind of completely new residual error structural unit is obtained, the output of residual unit is cascade by multiple convolutional layers
It exports and is added between input element (to ensure that convolutional layer output is identical with input element dimension), after being activated using ReLU
It arrives.This structure is cascaded up, depth residual error network has just been obtained.
Further, the present embodiment additionally provides a kind of mode advanced optimizing supervision recognition result, specifically includes：
All kinds of sensitization pictures is corresponded in the picture library to be detected to be identified, and after obtaining recognition result,
Difficult example sample is extracted from the recognition result.And the parameter of all kinds of identification models is corresponded to according to the difficult example Sample Refreshment.
To optimize convolutional neural networks parameter using difficult example, enhance the recognition capability of algorithm model.
Wherein, the difficult example sample of the extraction from the recognition result, including：
Obtain the score value of each attribute in sensitization picture, wherein the score value of each attribute passes through institute in the sensitization picture
Identification model is stated to be calculated.
According to the sequence of score value from large to small, sort to each attribute of acquired sensitization picture.Wherein, sensitization picture
Attribute can be understood as：Information associated with the image data of sensitization picture, such as：Title, source station address, the date,
The information such as resolution ratio, size, tag along sort, these associated information are added usually as the attribute information of image data in image
In data.
The additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence
When spending threshold value, judgement is as the difficult example sample.Such as：The attribute of a sensitization picture includes title, source station address, day
Phase, resolution ratio, size, tag along sort ... wait 10 attribute, calculate this 10 attribute by the identification model, obtain score value
Maximum preceding 3 attribute is：Source station address (score value 0.4), title (score value 0.3) and tag along sort (score value 0.1),
And confidence threshold value is 0.7, then the score value 0.8 of source station address+title+tag along sort is more than 0.7, judges that this is a sensitive
The difficult example sample of picture.
Wherein it is possible to enhance algorithm recognition capability by Progressive CNN, with specific reference to the difficult example of detection as a result,
It adds it in convolutional neural networks training data, enhances these effects of difficult example in the sample, to promote identification model
For the recognition capability of these violation pictures being difficult to differentiate between.
Through this embodiment, the master map, details figure and solarization free hand drawing for uploading electric business service platform for trade company is realized to occur
Obscene violation picture carry out automatic detection scanning, especially improve the intelligent level for advertising pictures management, reduce
Manual examination and verification cost finally also reduces management platform risk.
And the different from the past simple dichotomy for being divided into Sensitive Graphs and non-sensitive figure, the present invention are directed to the possibility field of electric business
If violation image is divided into Ganlei's (sensitive kind) by scape, increases the specific aim for special category picture recognition, improve simultaneously
Recognition accuracy.Such as：In actual test, trade company's management platform in electric business service platform increases newly daily at present to be uploaded
2000000 images, it is huge by manual examination and verification cost, needed for 100 working hours man day.After the present embodiment, need daily artificial
The picture number further verified is reduced within 500, reduces by 4000 times of cost of labor, while reducing artificial participation, is reduced
The risk that maloperation is brought.
The embodiment of the present invention also provides a kind of device analyzed sensitive image as shown in Figure 5, and the device is specific
It may operate on monitoring server as shown in Figure 1, which includes：
Cluster module is used in extracted training sample set, according to the corresponding sensitive kind of each sample, to the instruction
The sample image practiced in sample set is clustered；
Training module, for according to the sample image by cluster, passing through the corresponding all kinds of knowledge of convolutional neural networks training
Other model；
Analysis module, the identification model for being obtained using training identify that correspondence is all kinds of from picture library to be detected
Sensitization picture.
Wherein, the cluster module is specifically used for by preset neural network model, from the training sample set
Extract the sensitive features of each sample image；And by preset clustering algorithm, the similarity degree of sensitive features is met into test rule
Sample image then is clustered to the same sample set and is closed；Wherein, the preset neural network model is instructed by imagenet
Practice.
Further, as shown in fig. 6, further including：Filtering module, used in being closed in a sample set：According to cluster
The distance at center by the sample image in the subclass by closely sorting to remote, and chooses the sample to sort in preceding specified digit
Image is as positive sample；Recycle obtained positive sample training pattern grader；Pass through the trained model later
Grader, the sample image in being closed to the sample set carries out classified calculating, and the score value being calculated is less than pre- gating
The sample image of limit is rejected；
Further include：Correction module, for utilizing extracted pre-training data set, the depth residual error net of the specified number of plies of training
Network, the specified number of plies >=50；And the depth residual error network obtained by training, it corrects the sample set and closes.
Further, all kinds of quick for being corresponded in the picture library to be detected as shown in fig. 7, update module
Sense picture is identified, and after obtaining recognition result, obtains the score value of each attribute in sensitization picture, wherein the Sensitive Graphs
The score value of each attribute is calculated by the identification model in piece；And the sequence according to score value from large to small, to being obtained
Each attribute of the sensitization picture taken sorts；
The additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence
When spending threshold value, judgement is as the difficult example sample.；And the ginseng of all kinds of identification models is corresponded to according to the difficult example Sample Refreshment
Number.
The device provided in an embodiment of the present invention that sensitive image is analyzed, by the training sample set
Sample image is clustered, and further according to the sample image by cluster, passes through the corresponding all kinds of identification of convolutional neural networks training
Model, the identification model obtained later using training identify corresponding all kinds of sensitization picture from picture library to be detected, to
Identify whether the picture that the trade company of electric business service platform is uploaded belongs to corresponding all kinds of sensitization picture.It realizes for trade company
Automatic detection, the scanning for uploading the picture of electric business service platform, improve the automatization level for advertising pictures recognition detection,
Reduce manual examination and verification cost.
Each embodiment in this specification is described in a progressive manner, identical similar portion between each embodiment
Point just to refer each other, and each embodiment focuses on the differences from other embodiments.Especially for equipment reality
For applying example, since it is substantially similar to the method embodiment, so describing fairly simple, related place is referring to embodiment of the method
Part explanation.The above description is merely a specific embodiment, but protection scope of the present invention is not limited to
This, any one skilled in the art in the technical scope disclosed by the present invention, the variation that can readily occur in or replaces
It changes, should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claim
Subject to enclosing.
Claims (11)
1. a kind of method analyzed sensitive image, which is characterized in that including：
In extracted training sample set, according to the corresponding sensitive kind of each sample, to the sample in the training sample set
This image is clustered；
According to the sample image by cluster, pass through the corresponding all kinds of identification model of convolutional neural networks training；
The identification model obtained using training identifies corresponding all kinds of sensitization picture from picture library to be detected.
2. according to the method described in claim 1, it is characterized in that, it is described according to the corresponding sensitive kind of each sample to the instruction
The sample image practiced in sample set is clustered, including：
By preset neural network model, the sensitive features of each sample image are extracted from the training sample set, wherein
The preset neural network model is trained by imagenet；
By preset clustering algorithm, the sample image that the similarity degree of sensitive features is met to test order is clustered to same
Sample set closes.
3. according to the method described in claim 2, it is characterized in that, further including：
In a sample set closes：
According at a distance from cluster centre, by the sample image in the subclass by closely sorting to remote, and sequence is chosen preceding
The sample image of specified digit is as positive sample；
Utilize obtained positive sample training pattern grader；
By the trained model classifiers, the sample image in being closed to the sample set carries out classified calculating, and
Sample image by the score value being calculated less than pre-determined threshold is rejected.
4. according to the method described in claim 2, it is characterized in that, further including：
Utilize extracted pre-training data set, the depth residual error network of the specified number of plies of training, the specified number of plies >=50；
The depth residual error network obtained by training corrects the sample set and closes.
5. according to the method described in claim 1, it is characterized in that, further including：
All kinds of sensitization pictures is corresponded in the picture library to be detected to be identified, and after obtaining recognition result, from institute
It states and extracts difficult example sample in recognition result；
The parameter of all kinds of identification models is corresponded to according to the difficult example Sample Refreshment.
6. according to the method described in claim 5, it is characterized in that, described extract difficult example sample, packet from the recognition result
It includes：
Obtain the score value of each attribute in sensitization picture, wherein the score value of each attribute passes through the knowledge in the sensitization picture
Other model is calculated；According to the sequence of score value from large to small, sort to each attribute of acquired sensitization picture；
The additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence level threshold
When value, judgement is as the difficult example sample.
7. according to the method described in claim 1, it is characterized in that, further including：
According to preset business rule, candidate image is acquired from electric business service platform, acquired candidate image is utilized to update institute
State picture library；
And/or according to preset test order, the training sample set is extracted from the sample database pointed by the test order
It closes.
8. a kind of device analyzed sensitive image, which is characterized in that including：
Cluster module is used in extracted training sample set, according to the corresponding sensitive kind of each sample, to the trained sample
Sample image in this set is clustered；
Training module, for according to the sample image by cluster, passing through the corresponding all kinds of identification mould of convolutional neural networks training
Type；
Analysis module, the identification model for being obtained using training identify corresponding all kinds of sensitivity from picture library to be detected
Picture.
9. according to the method described in claim 8, it is characterized in that, the cluster module, is specifically used for passing through preset nerve
Network model extracts the sensitive features of each sample image from the training sample set；It, will and by preset clustering algorithm
The sample image that the similarity degree of sensitive features meets test order is clustered to the conjunction of the same sample set；Wherein, described default
Neural network model trained by imagenet.
10. according to the method described in claim 8, it is characterized in that, further including：Filtering module, in a sample set
In conjunction：According at a distance from cluster centre, by the sample image in the subclass by closely sorting to remote, and sequence is chosen preceding
The sample image of specified digit is as positive sample；Recycle obtained positive sample training pattern grader；Pass through later by
Trained model classifiers, the sample image in being closed to the sample set carries out classified calculating, and will be calculated
Score value is rejected less than the sample image of pre-determined threshold；
Further include：Correction module, for utilizing extracted pre-training data set, training to specify the depth residual error network of the number of plies,
The specified number of plies >=50；And the depth residual error network obtained by training, it corrects the sample set and closes.
11. according to the method described in claim 8, it is characterized in that, further including：
Update module is identified for corresponding to all kinds of sensitization pictures in the picture library to be detected, and is known
After other result, the score value of each attribute in sensitization picture is obtained, wherein the score value of each attribute passes through institute in the sensitization picture
Identification model is stated to be calculated；And the sequence according to score value from large to small, it sorts to each attribute of acquired sensitization picture；
And the additive value to sort in the score value of the attribute of preceding specified digit is obtained, when the additive value is more than preset confidence threshold value
When, judgement is as the difficult example sample.The parameter of all kinds of identification models is corresponded to further according to the difficult example Sample Refreshment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710248908.8A CN108734184B (en) | 2017-04-17 | 2017-04-17 | Method and device for analyzing sensitive image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710248908.8A CN108734184B (en) | 2017-04-17 | 2017-04-17 | Method and device for analyzing sensitive image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108734184A true CN108734184A (en) | 2018-11-02 |
| CN108734184B CN108734184B (en) | 2022-06-07 |
Family
ID=63923944
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710248908.8A Active CN108734184B (en) | 2017-04-17 | 2017-04-17 | Method and device for analyzing sensitive image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108734184B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109829069A (en) * | 2018-12-28 | 2019-05-31 | 广州华多网络科技有限公司 | Image audit processing method, device, electronic equipment and storage medium |
| CN109831699A (en) * | 2018-12-28 | 2019-05-31 | 广州华多网络科技有限公司 | Image audit processing method, device, electronic equipment and storage medium |
| CN109919170A (en) * | 2018-11-29 | 2019-06-21 | 阿里巴巴集团控股有限公司 | Change evaluation method, apparatus, electronic equipment and computer readable storage medium |
| CN110110982A (en) * | 2019-04-26 | 2019-08-09 | 特赞（上海）信息科技有限公司 | The checking method and device of intention material |
| CN110210356A (en) * | 2019-05-24 | 2019-09-06 | 厦门美柚信息科技有限公司 | A kind of picture discrimination method, apparatus and system |
| CN110222846A (en) * | 2019-05-13 | 2019-09-10 | 中国科学院计算技术研究所 | A kind of the information safety protection method and information security system of Internet terminal |
| CN110456955A (en) * | 2019-08-01 | 2019-11-15 | 腾讯科技（深圳）有限公司 | Exposure dress ornament detection method, device, system, equipment and storage medium |
| CN111144399A (en) * | 2018-11-06 | 2020-05-12 | 富士通株式会社 | Apparatus and method for processing image |
| CN111311316A (en) * | 2020-02-03 | 2020-06-19 | 支付宝(杭州)信息技术有限公司 | Method and device for depicting merchant portrait, electronic equipment, verification method and system |
| CN111626778A (en) * | 2020-05-25 | 2020-09-04 | 陶乐仪 | Advertisement pushing system and method |
| CN111726648A (en) * | 2020-06-28 | 2020-09-29 | 百度在线网络技术（北京）有限公司 | Method, device and equipment for detecting image data and computer readable storage medium |
| CN114004645A (en) * | 2021-10-29 | 2022-02-01 | 浙江省民营经济发展中心(浙江省广告监测中心) | Fuse media advertisement wisdom monitoring platform and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060095202A1 (en) * | 2004-11-01 | 2006-05-04 | Hitachi, Ltd. | Method of delivering difference map data |
| CN103679132A (en) * | 2013-07-15 | 2014-03-26 | 北京工业大学 | A sensitive image identification method and a system |
| CN106228185A (en) * | 2016-07-20 | 2016-12-14 | 武汉盈力科技有限公司 | A kind of general image classifying and identifying system based on neutral net and method |
-
2017
- 2017-04-17 CN CN201710248908.8A patent/CN108734184B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060095202A1 (en) * | 2004-11-01 | 2006-05-04 | Hitachi, Ltd. | Method of delivering difference map data |
| CN103679132A (en) * | 2013-07-15 | 2014-03-26 | 北京工业大学 | A sensitive image identification method and a system |
| CN106228185A (en) * | 2016-07-20 | 2016-12-14 | 武汉盈力科技有限公司 | A kind of general image classifying and identifying system based on neutral net and method |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111144399B (en) * | 2018-11-06 | 2024-03-05 | 富士通株式会社 | Apparatus and method for processing image |
| CN111144399A (en) * | 2018-11-06 | 2020-05-12 | 富士通株式会社 | Apparatus and method for processing image |
| CN109919170A (en) * | 2018-11-29 | 2019-06-21 | 阿里巴巴集团控股有限公司 | Change evaluation method, apparatus, electronic equipment and computer readable storage medium |
| CN109919170B (en) * | 2018-11-29 | 2023-12-05 | 创新先进技术有限公司 | Change evaluation method, change evaluation device, electronic device and computer-readable storage medium |
| CN109831699A (en) * | 2018-12-28 | 2019-05-31 | 广州华多网络科技有限公司 | Image audit processing method, device, electronic equipment and storage medium |
| CN109829069A (en) * | 2018-12-28 | 2019-05-31 | 广州华多网络科技有限公司 | Image audit processing method, device, electronic equipment and storage medium |
| CN109831699B (en) * | 2018-12-28 | 2021-07-20 | 广州华多网络科技有限公司 | Image auditing processing method and device, electronic equipment and storage medium |
| CN110110982A (en) * | 2019-04-26 | 2019-08-09 | 特赞（上海）信息科技有限公司 | The checking method and device of intention material |
| CN110222846B (en) * | 2019-05-13 | 2021-07-20 | 中国科学院计算技术研究所 | Information security method and information security system for internet terminal |
| CN110222846A (en) * | 2019-05-13 | 2019-09-10 | 中国科学院计算技术研究所 | A kind of the information safety protection method and information security system of Internet terminal |
| CN110210356A (en) * | 2019-05-24 | 2019-09-06 | 厦门美柚信息科技有限公司 | A kind of picture discrimination method, apparatus and system |
| CN110456955B (en) * | 2019-08-01 | 2022-03-29 | 腾讯科技（深圳）有限公司 | Exposed clothing detection method, device, system, equipment and storage medium |
| CN110456955A (en) * | 2019-08-01 | 2019-11-15 | 腾讯科技（深圳）有限公司 | Exposure dress ornament detection method, device, system, equipment and storage medium |
| CN111311316B (en) * | 2020-02-03 | 2023-05-23 | 支付宝(杭州)信息技术有限公司 | Method and device for depicting merchant portrait, electronic equipment, verification method and system |
| CN111311316A (en) * | 2020-02-03 | 2020-06-19 | 支付宝(杭州)信息技术有限公司 | Method and device for depicting merchant portrait, electronic equipment, verification method and system |
| CN111626778A (en) * | 2020-05-25 | 2020-09-04 | 陶乐仪 | Advertisement pushing system and method |
| CN111726648A (en) * | 2020-06-28 | 2020-09-29 | 百度在线网络技术（北京）有限公司 | Method, device and equipment for detecting image data and computer readable storage medium |
| CN114004645A (en) * | 2021-10-29 | 2022-02-01 | 浙江省民营经济发展中心(浙江省广告监测中心) | Fuse media advertisement wisdom monitoring platform and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108734184B (en) | 2022-06-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108734184A (en) | A kind of method and device that sensitive image is analyzed | |
| CN105303150B (en) | Realize the method and system of image procossing | |
| Narihira et al. | Learning lightness from human judgement on relative reflectance | |
| CN106408030B (en) | SAR image classification method based on middle layer semantic attribute and convolutional neural networks | |
| CN108171184A (en) | Method for distinguishing is known based on Siamese networks again for pedestrian | |
| Jing et al. | Yarn-dyed fabric defect classification based on convolutional neural network | |
| CN104866831B (en) | The face recognition algorithms of characteristic weighing | |
| CN104915673B (en) | A kind of objective classification method and system of view-based access control model bag of words | |
| CN103119607A (en) | Optimization of human activity determination from video | |
| CN109670546B (en) | Commodity matching and quantity regression recognition algorithm based on preset template | |
| CN108647702B (en) | Large-scale food material image classification method based on transfer learning | |
| CN110287856A (en) | A kind of security personnel's behavior analysis system, method and device | |
| CN110163722A (en) | The big data analysis system and analysis method precisely sold for agricultural product | |
| CN107622071B (en) | Clothes image retrieval system and method under non-source-retrieval condition through indirect correlation feedback | |
| CN109712324A (en) | A kind of automatic vending machine image-recognizing method, good selling method and vending equipment | |
| CN104573699A (en) | Trypetid identification method based on medium field intensity magnetic resonance dissection imaging | |
| Zhang et al. | A review on cluster estimation methods and their application to neural spike data | |
| Siddalingappa et al. | Anomaly detection on medical images using autoencoder and convolutional neural network | |
| CN107944373A (en) | A kind of video anomaly detection method based on deep learning | |
| Wei et al. | (Retracted) Image analysis and pattern recognition method of three-dimensional process in physical education teaching based on big data | |
| CN109214467B (en) | Transformer operator dressing robust identification method considering classifier output sensitivity | |
| CN104573745A (en) | Fruit fly classification method based on magnetic resonance imaging | |
| CN110414792A (en) | Component centralized procurement management system and Related product based on BIM and big data | |
| Zhang et al. | Colour‐patterned fabric‐defect detection using unsupervised and memorial defect‐free features | |
| Shermila et al. | DEEPFIC: Food item classification with calorie calculation using dragonfly deep learning network |
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 | ||
| CB02 | Change of applicant information |
Address after: 210000, 1-5 story, Jinshan building, 8 Shanxi Road, Nanjing, Jiangsu. Applicant after: SUNING.COM Co.,Ltd. Address before: 210042 Suning Headquarters, No. 1 Suning Avenue, Xuanwu District, Nanjing City, Jiangsu Province Applicant before: SUNING COMMERCE GROUP Co.,Ltd. |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |