CN108446330A - Promotion object processing method and device and computer-readable storage medium - Google Patents
Promotion object processing method and device and computer-readable storage medium Download PDFInfo
- Publication number
- CN108446330A CN108446330A CN201810150833.4A CN201810150833A CN108446330A CN 108446330 A CN108446330 A CN 108446330A CN 201810150833 A CN201810150833 A CN 201810150833A CN 108446330 A CN108446330 A CN 108446330A
- Authority
- CN
- China
- Prior art keywords
- popularization
- library
- promoted
- characteristic information
- promoting
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling 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
- G06Q30/0242—Determining effectiveness of 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
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted 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
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0277—Online advertisement
Abstract
The invention discloses a popularization object processing method and device and a computer readable storage medium, which can identify popularization objects. The promotion object processing method comprises the following steps: crawling promotion objects regularly, generating a first promotion object library, and generating an attribute tag for each promotion object; the crawled popularization objects are subjected to duplicate removal processing regularly to obtain a second popularization object library; and identifying characteristic information in promotion objects in a second promotion object library by adopting a target detection technology based on deep learning, and adding the identified characteristic information into attribute labels of all promotion objects which are the same as the promotion objects in the first promotion object library. Through batch processing, the image processing speed is improved, and researchers can analyze and research massive popularization objects according to the characteristic information of each popularization object to obtain a popularization effect monitoring report.
Description
Technical field
The present invention relates to image processing techniques, espespecially a kind of popularization object processing method, device and one kind are computer-readable
Storage medium.
Background technology
With the continuous development of internet+(Internet Plus), the especially extensive use of smart mobile phone, increasingly
More Internet advertisings using be not added with detection code by the way of occur, for advertising research, these advertisements be difficult to science,
Completely, efficient mode is studied and is assessed.Meanwhile with the maturation of the development of artificial intelligence technology and computation vision technology,
Make it possible the content recognition of Internet advertising picture, is asked for magnanimity internet the undetectable of advertisement of not raising the price
Topic is badly in need of a solution.
Invention content
In order to solve the above technical problem, the present invention provides a kind of popularization object processing method, device and a kind of calculating
Machine readable storage medium storing program for executing can identify popularization object.
In order to reach the object of the invention, the present invention provides a kind of popularization object processing methods, including：
Popularization object is periodically crawled, first is generated and promotes library of object, attribute tags are generated for each object of promoting；
Duplicate removal processing periodically is carried out to the popularization object crawled, obtains the second popularization library of object；
The feature letter promoted in library of object in object is promoted using the target detection technique identification second based on deep learning
Breath, the characteristic information that will identify that are added to described first and promote all popularizations pair identical with the popularization object in library of object
In the attribute tags of elephant.
Further, described periodically to crawl popularization object, it generates first and promotes library of object, belong to each to promote object and generating
Property label, including：
A variety of user properties are preset, interval prefixed time interval crawls the popularization object of predeterminated position in the same period,
The popularization object is downloaded, first is generated and promotes library of object, generates attribute tags for each object of promoting, the attribute tags are extremely
Include less：Promote the mark of object.
Further, described that duplicate removal processing periodically is carried out to the popularization object crawled, the second popularization library of object is obtained, is wrapped
It includes：
The perceptual hash value for each promoting object in the first popularization library of object is calculated, compares the Hamming distance for promoting object two-by-two
From the popularization object that Hamming distance is less than to preset value is set as one group, generates second and promotes library of object, second popularization pair
As the information in library includes that group identifies, the mark of object is respectively promoted in group.
Further, described that popularization pair in library of object is promoted using the target detection technique identification second based on deep learning
Characteristic information as in, the characteristic information that will identify that are added to identical as the popularization object in the first popularization library of object
All popularization objects attribute tags in, including：
One in any one group is selected to promote object from the second popularization library of object, using the target based on deep learning
Detection technique identifies that the characteristic information in the popularization object, the characteristic information that will identify that are added to described first and promote object
It is promoted described in library in the attribute tags of object and belongs to its of same group with the popularization object in the first popularization library of object
In his all attribute tags for promoting object；All groups of above-mentioned processing of carry out in library of object are promoted to second.
Further, the feature letter identified using the target detection technique based on deep learning in the popularization object
Breath, including：
Use CNN layers of extraction of one group of convolutional neural networks characteristic pattern feature map promoted in object；
Region suggests that RPN layers of network judges that anchor anchor belongs to foreground or background by grader, and frame is recycled to return
Bounding box regression correct anchor and obtain accurate suggestion proposal；
Characteristic pattern and proposal are collected in ROI Pooling layers of area-of-interest pond, extract and build after these comprehensive information
Discuss characteristic pattern proposal feature map；
The classification of full connect layers of judgement suggestion feature figures of connection full, the classification is the feature promoted in object
Information.
Further, entirely after the classification of connect layers of judgement suggestion feature figures of connection full, the method further includes：Profit
The position offset bbox_pred that bounding box regression obtain each proposal is returned with frame.
Further, the method further includes：All attribute tags for promoting object in library of object are promoted to first to carry out
Statistical analysis, the promotion effect for the object that puts it over.
Further, the popularization object is that internet is not raised the price advertisement, and the characteristic information includes brand message.
Further, the characteristic information further includes：The displaying area of brand message accounts for the percentage of the picture gross area.
In order to reach the object of the invention, the present invention also provides a kind of popularization object handles devices, including crawl module, go
Molality block and processing module, wherein：
It is described to crawl module, for periodically crawling popularization object, generates first and promote library of object, object life is promoted to be each
At attribute tags；
The deduplication module obtains the second popularization library of object for periodically carrying out duplicate removal processing to the popularization object crawled；
The processing module is pushed away for being promoted in library of object using the target detection technique identification second based on deep learning
Characteristic information in wide object, the characteristic information that will identify that are added in the first popularization library of object and the popularization object
In identical all attribute tags for promoting object.
Further, the module that crawls periodically crawls popularization object, generates first and promotes library of object, for each popularization pair
As generation attribute tags, including：
The module that crawls presets a variety of user properties, and interval prefixed time interval crawls predeterminated position in the same period
Popularization object, download the popularization object, generate first and promote library of object, attribute tags, institute are generated for each object of promoting
Attribute tags are stated to include at least：Promote the mark of object.
Further, the deduplication module periodically carries out duplicate removal processing to the popularization object crawled, obtains the second popularization pair
As library, including：
The deduplication module calculates the perceptual hash value that object is each promoted in the first popularization library of object, compares popularization two-by-two
The Hamming distance of object, the popularization object that Hamming distance is less than to preset value are set as one group, generate second and promote library of object, institute
It includes that group identifies to state the information in the second popularization library of object, and the mark of object is respectively promoted in group.
Further, the processing module promotes library of object using the target detection technique identification second based on deep learning
Characteristic information in middle popularization object, the characteristic information that will identify that are added in the first popularization library of object and the popularization
In the identical all attribute tags for promoting object of object, including：
The processing module selects one in any one group to promote object from the second popularization library of object, using based on depth
The target detection technique of degree study identifies the characteristic information in the popularization object, and the characteristic information that will identify that is added to described
First promote library of object described in promote object attribute tags in and first promote library of object in the popularization object category
In same group other all attribute tags for promoting object；All groups of above-mentioned processing of carry out in library of object are promoted to second.
Further, described device further includes analysis module, is used to promote all popularization objects in library of object to first
Attribute tags it is for statistical analysis, the promotion effect for the object that puts it over.
In order to reach the object of the invention, the present invention also provides a kind of computer readable storage mediums, are stored thereon with meter
Calculation machine instructs, when which is executed by processor the step of the realization above method.
The embodiment of the present invention by modes such as web crawlers, obtains first and promotes object (picture materials) in internet, so
After carry out duplicate removal processing, finally use Faster R-CNN deep learning systems, identify picture materials in characteristic information, and will
Characteristic information is added in original popularization library of object, so that researcher can be according to each characteristic information for promoting object to sea
The popularization object of amount is analyzed and researched, and effect monitoring report is promoted.The embodiment of the present invention passes through duplicate removal processing so that
It is not necessary that each popularization object is identified when identification feature information, same or analogous popularization object is only identified once
, the characteristic information of a certain popularization object that will identify that is added to all categories for promoting objects identical with the popularization object
Property label in, by the processing of batch, substantially increase image processing speed, in particular for magnanimity promote object scene.
Other features and advantages of the present invention will be illustrated in the following description, also, partly becomes from specification
It obtains it is clear that understand through the implementation of the invention.The purpose of the present invention and other advantages can be by specification, rights
Specifically noted structure is realized and is obtained in claim and attached drawing.
Description of the drawings
Attached drawing is used for providing further understanding technical solution of the present invention, and a part for constitution instruction, with this
The embodiment of application technical solution for explaining the present invention together, does not constitute the limitation to technical solution of the present invention.
Fig. 1 is one method flow diagram of the embodiment of the present invention；
Fig. 2 is two devices structural schematic diagram of the embodiment of the present invention；
Fig. 3 is the present invention using the detection of Faster R-CNN brands and brand recognition flow chart in example；
Fig. 4 is the present invention using Faster R-CNN neural network structure figures in example.
Specific implementation mode
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention
Embodiment be described in detail.It should be noted that in the absence of conflict, in the embodiment and embodiment in the application
Feature mutually can arbitrarily combine.
Step shown in the flowchart of the accompanying drawings can be in the computer system of such as a group of computer-executable instructions
It executes.Also, although logical order is shown in flow charts, and it in some cases, can be with suitable different from herein
Sequence executes shown or described step.
Embodiment one
The present embodiment describes a kind of processing method for promoting object, as shown in Figure 1, including the following steps：
Step 11, popularization object is periodically crawled, first is generated and promotes library of object, attribute mark is generated for each object of promoting
Label；
Above-mentioned popularization object for example can be the ad material obtained based on internet, including be not limited to obtain from following channel
：Website, APP, internet television etc..Since Internet advertising material quantity is more, the period crawled can be with minute
For unit, such as it can be 30 minutes, 60 minutes, 90 minutes, 120 minutes etc..
Step 12, duplicate removal processing periodically is carried out to the popularization object crawled, obtains the second popularization library of object；
Since duplicate removal processing operand is larger, can not real-time implementation, therefore be arranged duplicate removal processing period as unit of day,
Such as it could be provided as the popularization object that daily night or morning crawl the same day and carry out duplicate removal.
Duplicate removal described in this step can will judge that identical picture is classified as one group, and second promotes library of object to organize as list
The information of object is promoted in position storage.
Step 13, it is promoted in library of object and is promoted in object using the target detection technique identification second based on deep learning
Characteristic information, the characteristic information that will identify that are added to identical with the popularization object all in the first popularization library of object
In the attribute tags for promoting object.
Step 13 is repeated until all popularization objects in the second popularization library of object are disposed.
Used in step 13 in the present embodiment based on the target detection technique of deep learning for Faster R-CNN.If
The popularization object crawled in step 11 is that internet is not raised the price advertisement, and the characteristic information may include brand message, optionally,
Can also include：The displaying area of brand message accounts for the percentage of the picture gross area.
Using the embodiment of the present invention, by duplicate removal processing, the picture of identical (or similar) is divided in same group so that
It is not necessary that each popularization object is identified when identification feature information, same or analogous popularization object is only identified once
, the characteristic information of a certain popularization object that will identify that is added to all categories for promoting objects identical with the popularization object
In property label, by the processing of batch, image processing speed is substantially increased.The processing of 11-13 through the above steps, can be fast
It is fast to obtain all characteristic informations for promoting objects, it carries out promoting the tactile analysis up to effect of object to promoting object convenient for researcher
Research.
Optionally, after step 13, further include：
Step 14, the attribute tags that all popularization objects in library of object are promoted to first are for statistical analysis, put it over
The promotion effect of object.
Wherein, described periodically to crawl popularization object in above-mentioned steps 11, it generates first and promotes library of object, each to push away
Wide object generates attribute tags, including：
A variety of user properties are preset, interval prefixed time interval crawls the popularization object of predeterminated position in the same period,
The popularization object is downloaded, first is generated and promotes library of object, generates attribute tags for each object of promoting, the attribute tags are extremely
Include less：Promote the mark of object.Crawling operation can utilize web crawlers to realize.
Wherein, described that duplicate removal processing periodically is carried out to the popularization object crawled in above-mentioned steps 12, obtain the second popularization
Library of object, including：
The perceptual hash value for each promoting object in the first popularization library of object is calculated, compares the Hamming distance for promoting object two-by-two
From the popularization object that Hamming distance is less than to preset value is set as one group, and the popularization object that Hamming distance is less than preset value is thought
It is same or analogous, generates second and promote library of object, the information recorded in the second popularization library of object includes that group identifies, group
The middle mark for promoting object, can also include the corresponding perceptual hash value of each popularization object.
Wherein, described using the second popularization pair of target detection technique identification based on deep learning in above-mentioned steps 13
As in library promote object in characteristic information, the characteristic information that will identify that be added to it is described first promote library of object in it is described
It promotes in the identical all attribute tags for promoting object of object, including：
One in any one group is selected to promote object (can be any one) from the second popularization library of object, using base
Identify that the characteristic information in the popularization object, the characteristic information that will identify that are added in the target detection technique of deep learning
It is described first promote library of object described in promote object attribute tags in and first promote library of object in it is described promote pair
As belong to same group other it is all promote objects attribute tags in；All groups of above-mentioned places of carry out in library of object are promoted to second
Reason finishes until the characteristic information of any popularization object in each group identifies, first promotes all popularizations in library of object at this time
Characteristic information is added in the attribute tags of object.
Specifically, the feature letter identified using the target detection technique based on deep learning in the popularization object
Breath, including：
Use CNN layers of extraction of one group of convolutional neural networks characteristic pattern feature map promoted in object；
Region suggests that RPN layers of network judges that anchor anchor belongs to foreground or background by grader, and frame is recycled to return
Bounding box regression correct anchor and obtain accurate suggestion proposal；
Characteristic pattern and proposal are collected in ROI Pooling layers of area-of-interest pond, extract and build after these comprehensive information
Discuss characteristic pattern proposal feature map；
The classification of full connect layers of judgement suggestion feature figures of connection full, the classification is the feature promoted in object
Information.
Optionally, after the classification of connect layers of judgement suggestion feature figures of connection full entirely, the method further includes：Profit
The position offset bbox_pred that bounding box regression obtain each proposal, i.e. feature are returned with frame
The displaying area of information accounts for the percentage of the picture gross area.
Embodiment two
The present embodiment describes a kind of processing unit for promoting object, this implementation is also applied for described in above method embodiment
Example, this embodiment is not repeated.As shown in Fig. 2, described device includes crawling module 21, deduplication module 22 and processing module 23,
Wherein：
It is described to crawl module 21, for periodically crawling popularization object, generates first and promote library of object, object is promoted to be each
Generate attribute tags；
The deduplication module 22 obtains the second popularization object for periodically carrying out duplicate removal processing to the popularization object crawled
Library；
The processing module 23, for being promoted in library of object using the target detection technique identification second based on deep learning
The characteristic information in object is promoted, the characteristic information that will identify that is added in the first popularization library of object and the popularization pair
As in identical all attribute tags for promoting object.
In one alternate embodiment, the module 21 that crawls periodically crawls popularization object, generates first and promotes library of object,
Attribute tags are generated for each popularization object, including：
The module 21 that crawls presets a variety of user properties, and interval prefixed time interval, which crawls, presets position in the same period
The popularization object set downloads the popularization object, generates first and promotes library of object, and attribute tags are generated for each object of promoting,
The attribute tags include at least：Promote the mark of object.
In one alternate embodiment, the deduplication module 22 periodically carries out duplicate removal processing to the popularization object crawled, obtains
Library of object is promoted to second, including：
The deduplication module 22 calculates the perceptual hash value that object is each promoted in the first popularization library of object, compares push away two-by-two
The Hamming distance of wide object, the popularization object that Hamming distance is less than to preset value are set as one group, generate second and promote library of object,
Described second information promoted in library of object includes that group identifies, and the mark of object is respectively promoted in group.
In one alternate embodiment, the processing module 23 is using the target detection technique identification the based on deep learning
Two promote the characteristic information promoted in library of object in object, and the characteristic information that will identify that is added to described first and promotes library of object
In with it is described popularization object it is identical it is all promote objects attribute tags in, including：
The processing module 23 selects one in any one group to promote object from the second popularization library of object, using based on
The target detection technique of deep learning identifies that the characteristic information in the popularization object, the characteristic information that will identify that are added to institute
State described in the first popularization library of object promote object attribute tags in and first promote library of object in the popularization object
Belong in same group other all attribute tags for promoting object；All groups of above-mentioned places of carry out in library of object are promoted to second
Reason.
The processing module 23 identifies the feature in the popularization object using the target detection technique based on deep learning
Information, including：
The processing module 23 uses CNN layers of extraction of one group of convolutional neural networks characteristic pattern promoted in object
feature map；
The processing module 23 makes region suggest that RPN layers of network judges that anchor anchor belongs to foreground or the back of the body by grader
Scape recycles frame to return bounding box regression and correct anchor and obtain and accurately suggests proposal；
The processing module 23 makes ROI Pooling layers of area-of-interest pond collect characteristic pattern and proposal, comprehensive
Suggestion feature figure proposal feature map are extracted after these information；
The processing module 23 makes to connect connect layers of classifications for judging suggestion feature figure of full entirely, and the classification is
Promote the characteristic information in object.
Optionally, entirely after the classification of connect layers of judgement suggestion feature figures of connection full, the also profit of the processing module 23
The position offset bbox_pred that bounding box regression obtain each proposal is returned with frame.
In one alternate embodiment, described device further includes analysis module, is used to promote institute in library of object to first
There are the attribute tags for promoting object for statistical analysis, the promotion effect for the object that puts it over.
Using example
This example is that internet is not raised the price and is specifically described for ad material to promote object, is included the following steps：
Step 1：It does not raise the price web advertisement material collection：In this example by disposing distributed network crawler system, mould
Intend a variety of user properties, crawls web advertisement material of not raising the price, and the material labeling to crawling, specifically include：
(1) a variety of user properties are simulated and crawls ad material；
By presetting a variety of UA (User Agent) in crawler system, such as age, gender, region, mobile phone model, consumption
Custom, media custom etc., crawl the ad material of the targeted sites of same period；Downloads ad material simultaneously adds attribute tags
Spider_AD_Label, Spider_AD_Label include but not limited to：Spider_AD_ID, Original_URL, AD_
Path, Site_ID, Media_Type, Unix_Time, Area_ID, AD_Info, AD_UA, wherein Spider_AD_ID is indicated
The unique identifier of ad material, Original_URL indicate original URL (the Uniform Resource of ad material
Locator, uniform resource locator), AD_Path indicates that the server storage path of ad material, Site_ID indicate material
Source (website or APP), Media_Tpye indicate the media types (website, APP, internet television etc.) in ad material source,
Unix_Time indicates that material crawls the time, and Area_ID indicates the dispensing urban information of ad material, and it is wide that AD_UA expressions crawl this
Accuse the UA information used when material.Depending on the needs researched and analysed, the content of above-mentioned attribute tags can be increased and decreased.
(2) ad material is crawled at times；
According to demand, can take 30 minutes, 60 minutes, 120 minutes constant durations start reptile and crawl ad material.
Step 2：Material data pre-processes：Using day as time interval, perception hash algorithm (Perceptual hash are used
Algorithm duplicate removal) is carried out to the picture materials that reptile crawls, generates the libraries AD_Img, includes but not limited to following parameter in library：
AD_Unique_ID, Spider_AD_ID, Img_Phash, the wherein consistent figure of AD_Unique_ID field references image content
Piece group serial number, Spider_AD_ID fields are derived from the Spider_AD_ID in Spider_AD_Label labels, indicate ad material
Unique identifier, Img_pHash fields indicate material pHash values；Process of data preprocessing specifically includes：
(1) the perceptual hash value pHash_Value of all ad materials is calculated, and generates AD_Img labels, wherein AD_
Img_ID fields are generated by self-propagation mode；
(2) all materials are traversed successively, and calculate the Hamming distance of ad material pHash_Value two-by-two, if Hamming
Distance is less than or equal to preset value (being, for example, 0), then it is assumed that two images are same or similar, and same or analogous image is classified as one
Group, group number are identified using AD_Unique_ID, and AD_Unique_ID is globally unique in system；
(3) AD_Unique_ID fields in AD_Img are extracted, AD_Img_List lists are obtained, which includes only AD_
Mono- field of Unique_ID.
Step 3：Ad material intelligent identifying system identifies the brand message in material：It is extracted at random in the libraries AD_Img
An ad material in AD_Img_List corresponding to AD_Unique_ID, and it is input to Faster R-CNN deep learnings system
System, system will export one group of brand message AD_Img_Brand, including but not limited to：AD_Unique_ID, Brand_ID and
Proportion, wherein AD_Unique_ID identify one group of same or analogous material, and an AD_Unique_ID can be corresponded to
Multigroup AD_Img_Brand (i.e. including multiple brand messages in a material), Brand_ID indicates branded content or class in material
Not, Proportion indicates that the displaying area of the branded content accounts for the percentage of the picture gross area.System core algorithm is
Faster R-CNN brands position detections and branded content identify neural network, the detection of Faster R-CNN brands and brand recognition
Flow chart is as shown in figure 3, Faster R-CNN deep learning neural networks form structure as shown in figure 4, ROI is projected as in Fig. 3
ROI Projection indicate that area-of-interest projection, the ponds ROI layer are RoI Pooling layer, Fc Full
Connect layer indicate that full articulamentum, RoI characteristic vectors are RoI Feature Vector, Deep ConvNet, Conv
Feature Map, Softmax, Bbox regressor are without universal Chinese technical term.Conv layers, 13 of 13 in Fig. 4
Relu layers are that following manner arranges with 4 pond (pooling) layers：Conv layers, relu layers, conv layers, relu layers, pooling
Layer, conv layers, relu layers, conv layers, relu layers, pooling layers, conv layers, relu layers, conv layers, relu layers, conv layers,
Relu layers, pooling layers, conv layers, relu layers, conv layers, relu layers, conv layers, relu layers, pooling layers, conv layers,
Relu layers, conv layers, relu layers, conv layers, relu layers.2 relu layers in Fig. 4 and 2 full articulamentums are following manner row
Row：Full articulamentum, relu layers, full articulamentum, relu layers.Conv, relu, Reshape, Softmax in Fig. 4 is without universal Chinese
Technical term.
Feature recognition flow can be divided into following four part：
(1) Faster RCNN use one group of CNN (Convolutional Neural Network, convolutional neural networks)
Feature maps (characteristic pattern) in layer extraction material, the feature maps will be shared for follow-up RPN (Region
Network is suggested in Proposal Networks, region) layer and full articulamentum (fully connected layers, abbreviation FC)；
(2) RPN (Region Proposal Networks) network is for generating region proposals；The layer passes through
Softmax graders judge that anchors (anchor) belongs to foreground (foreground) or background (background), recycle
Bounding box regression (frame recurrence) correct anchors and obtain accurate proposals；
(3) ROI (region of interest, area-of-interest) Pooling (pond) layer collects the feature of input
Maps and proposals extracts proposal feature maps after these comprehensive information, is sent into follow-up full articulamentum judgement
Target category；
(4) Classification (classification)；The parts Classification utilize the proposal obtained
Feature maps reuse softmax graders and obtain each proposal tools by full connect (full connection) layer
Body belongs to that classification (i.e. brand), exports cls_prob probability vectors, i.e. material content type, and expression belongs to some brand
Probability；The position offset of each proposal is obtained using bounding box regression (frame recurrence) simultaneously
Bbox_pred, for returning more accurate target detection frame, that is, the displaying area for obtaining branded content accounts for total figure piece area
Percentage.
Step 4：Export analysis of advertising results result of not raising the price：
(1) inverse network image preprocessing, the processes such as identification, recalls, finds the original corresponding to AD_Unique_ID step by step
The reptile material of beginning, and add in brand message label to the AD_Info fields of Spider_AD_Label；
(2) Spider_AD_Label after identification is input in existing sample study system, which will
According to demand, advertising results monitoring report of not raising the price is exported, it may include PV (page view, page browsing amount), UV (user
View, independent visitor), Reach (covering surface), frequency (frequency), medium property, user's portrait, ROI (investment repayments
Rate), advertisement is touched up to key indexes such as recruitment evaluation, SOV (Share of voice, advertisement occupation rate).
This exemplary advertising pictures of not raising the price launches effect monitoring method and is divided into the progress of four stages, is first initial ad element
Material data acquisition phase, the stage simulate a variety of user properties by distributed network reptile and crawl the nets such as electric business, video website
The advertising pictures material of not raising the price of page and the ends APP, and will be after material labeling；Subsequently into data preprocessing phase, the stage
The material crawled is subjected to the operations such as duplicate removal；Enter back into ad material cognitive phase；The stage is by pretreated ad material
It is input to image detection and identifying system based on Frast R-CNN, which will identify and mark the brand in ad material
Information and scene information；It is finally to calculate not raising the price the advertisement delivery effect stage, the stage is wide by the URL of ad material and correlation
Accuse attribute be input in sample data research system, output do not raise the price advertisement touch up to recruitment evaluation, SOV (Share of voice,
Advertisement occupation rate) etc. advertisement delivery effects testing result.This example passes through Frast R-CNN deep learning systems, intelligent recognition
The picture materials of not raising the price of internet, efficiently solve internet and do not raise the price the effect monitoring demand of advertisement, and this method can be applied
The advertisement delivery effect research that do not raise the price in internet.
It will appreciated by the skilled person that whole or certain steps in method disclosed hereinabove, system, dress
Function module/unit in setting may be implemented as software, firmware, hardware and its combination appropriate.In hardware embodiment,
Division between the function module/unit referred in the above description not necessarily corresponds to the division of physical unit；For example, one
Physical assemblies can have multiple functions or a function or step that can be executed by several physical assemblies cooperations.Certain groups
Part or all components may be implemented as by processor, such as the software that digital signal processor or microprocessor execute, or by
It is embodied as hardware, or is implemented as integrated circuit, such as application-specific integrated circuit.Such software can be distributed in computer-readable
On medium, computer-readable medium may include computer storage media (or non-transitory medium) and communication media (or temporarily
Property medium).As known to a person of ordinary skill in the art, term computer storage medium is included in for storing information (such as
Computer-readable instruction, data structure, program module or other data) any method or technique in the volatibility implemented and non-
Volatibility, removable and nonremovable medium.Computer storage media include but not limited to RAM, ROM, EEPROM, flash memory or its
His memory technology, CD-ROM, digital versatile disc (DVD) or other optical disc storages, magnetic holder, tape, disk storage or other
Magnetic memory apparatus or any other medium that can be used for storing desired information and can be accessed by a computer.This
Outside, known to a person of ordinary skill in the art to be, communication media generally comprises computer-readable instruction, data structure, program mould
Other data in the modulated data signal of block or such as carrier wave or other transmission mechanisms etc, and may include any information
Delivery media.
Although disclosed herein embodiment it is as above, the content only for ease of understanding the present invention and use
Embodiment is not limited to the present invention.Technical staff in any fields of the present invention is taken off not departing from the present invention
Under the premise of the spirit and scope of dew, any modification and variation, but the present invention can be carried out in the form and details of implementation
Scope of patent protection, still should be subject to the scope of the claims as defined in the appended claims.
Claims (15)
1. a kind of popularization object processing method, which is characterized in that including：
Popularization object is periodically crawled, first is generated and promotes library of object, attribute tags are generated for each object of promoting；
Duplicate removal processing periodically is carried out to the popularization object crawled, obtains the second popularization library of object；
The characteristic information promoted in library of object in object is promoted using the target detection technique identification second based on deep learning, it will
The characteristic information identified is added to described first and promotes all popularization objects identical with the popularization object in library of object
In attribute tags.
2. according to the method described in claim 1, it is characterized in that,
It is described periodically to crawl popularization object, it generates first and promotes library of object, attribute tags are generated for each object of promoting, including：
A variety of user properties are preset, interval prefixed time interval crawls the popularization object of predeterminated position in the same period, downloads
The popularization object generates first and promotes library of object, generates attribute tags for each object of promoting, the attribute tags are at least wrapped
It includes：Promote the mark of object.
3. according to the method described in claim 1, it is characterized in that,
It is described that duplicate removal processing periodically is carried out to the popularization object crawled, the second popularization library of object is obtained, including：
The perceptual hash value for each promoting object in the first popularization library of object is calculated, compares the Hamming distance for promoting object two-by-two,
The popularization object that Hamming distance is less than to preset value is set as one group, generates second and promotes library of object, described second promotes object
Information in library includes that group identifies, and the mark of object is respectively promoted in group.
4. according to the method described in claim 1, it is characterized in that,
It is described that the feature letter promoted in library of object in object is promoted using the target detection technique identification second based on deep learning
Breath, the characteristic information that will identify that are added to described first and promote all popularizations pair identical with the popularization object in library of object
In the attribute tags of elephant, including：
One in any one group is selected to promote object from the second popularization library of object, using the target detection based on deep learning
Technology identifies the characteristic information in the popularization object, and the characteristic information that will identify that, which is added to described first, to be promoted in library of object
Belong to same group other institutes in the attribute tags for promoting object and in the first popularization library of object with the popularization object
Have in the attribute tags for promoting object；
All groups of above-mentioned processing of carry out in library of object are promoted to second.
5. according to the method described in claim 4, it is characterized in that,
The characteristic information identified using the target detection technique based on deep learning in the popularization object, including：
Use CNN layers of extraction of one group of convolutional neural networks characteristic pattern feature map promoted in object；
Region suggests that RPN layers of network judges that anchor anchor belongs to foreground or background by grader, and frame is recycled to return
Bounding box regression correct anchor and obtain accurate suggestion proposal；
Characteristic pattern and proposal are collected in ROI Pooling layers of area-of-interest pond, and extraction suggestion is special after these comprehensive information
Sign figure proposal feature map；
The classification of full connect layers of judgement suggestion feature figures of connection full, the classification are the feature letter promoted in object
Breath.
6. according to the method described in claim 5, it is characterized in that,
After the classification of full connect layers of judgement suggestion feature figures of connection full, the method further includes：It is returned using frame
Bounding box regression obtain the position offset bbox_pred of each proposal.
7. according to the method described in claim 1, it is characterized in that,
The method further includes：The attribute tags that all popularization objects in library of object are promoted to first are for statistical analysis, obtain
Promote the promotion effect of object.
8. according to the method described in any claim in claim 1-7, which is characterized in that
The popularization object is that internet is not raised the price advertisement, and the characteristic information includes brand message.
9. according to the method described in claim 8, it is characterized in that,
The characteristic information further includes：The displaying area of brand message accounts for the percentage of the picture gross area.
10. a kind of popularization object handles device, which is characterized in that including crawling module, deduplication module and processing module, wherein：
It is described to crawl module, for periodically crawling popularization object, generates first and promote library of object, generate and belong to for each popularization object
Property label；
The deduplication module obtains the second popularization library of object for periodically carrying out duplicate removal processing to the popularization object crawled；
The processing module, for promoting popularization pair in library of object using the target detection technique identification second based on deep learning
Characteristic information as in, the characteristic information that will identify that are added to identical as the popularization object in the first popularization library of object
All popularization objects attribute tags in.
11. device according to claim 10, which is characterized in that
The module that crawls periodically crawls popularization object, generates first and promotes library of object, and attribute mark is generated for each object of promoting
Label, including：
The module that crawls presets a variety of user properties, and interval prefixed time interval crawls pushing away for predeterminated position in the same period
Wide object downloads the popularization object, generates first and promotes library of object, and attribute tags, the category are generated for each object of promoting
Property label includes at least：Promote the mark of object.
12. device according to claim 10, which is characterized in that
The deduplication module periodically carries out duplicate removal processing to the popularization object crawled, obtains the second popularization library of object, including：
The deduplication module calculates the perceptual hash value that object is each promoted in the first popularization library of object, compares popularization object two-by-two
Hamming distance, the popularization object that Hamming distance is less than to preset value is set as one group, generates second and promotes library of object, described the
Two information promoted in library of object include that group identifies, and the mark of object is respectively promoted in group.
13. device according to claim 10, which is characterized in that
The processing module is promoted in library of object using the target detection technique identification second based on deep learning and is promoted in object
Characteristic information, the characteristic information that will identify that be added to it is described first promote library of object in institute identical with the popularization object
Have in the attribute tags for promoting object, including：
The processing module selects one in any one group to promote object from the second popularization library of object, using based on depth
The target detection technique of habit identifies that the characteristic information in the popularization object, the characteristic information that will identify that are added to described first
It promotes in the attribute tags for promoting object described in library of object and belongs to same in the first popularization library of object with the popularization object
In one group other all attribute tags for promoting object；
All groups of above-mentioned processing of carry out in library of object are promoted to second.
14. device according to claim 10, which is characterized in that
Described device further includes analysis module, is used to promote all attribute tags for promoting object in library of object to first carrying out
Statistical analysis, the promotion effect for the object that puts it over.
15. a kind of computer readable storage medium, is stored thereon with computer instruction, which is characterized in that the instruction is by processor
The step of any claim the method in claim 1-9 is realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810150833.4A CN108446330B (en) | 2018-02-13 | 2018-02-13 | Promotion object processing method and device and computer-readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810150833.4A CN108446330B (en) | 2018-02-13 | 2018-02-13 | Promotion object processing method and device and computer-readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108446330A true CN108446330A (en) | 2018-08-24 |
| CN108446330B CN108446330B (en) | 2022-05-13 |
Family
ID=63192496
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810150833.4A Active CN108446330B (en) | 2018-02-13 | 2018-02-13 | Promotion object processing method and device and computer-readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108446330B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109447719A (en) * | 2018-12-17 | 2019-03-08 | 厦门美柚信息科技有限公司 | Targeted promotion commodity automatic determination method, device, medium and electronic equipment |
| CN109740729A (en) * | 2018-12-14 | 2019-05-10 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN113971592A (en) * | 2021-12-23 | 2022-01-25 | 成都易播科技有限公司 | Supervision evaluation method, system and device for promotion information release main body |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100082427A1 (en) * | 2008-09-30 | 2010-04-01 | Yahoo! Inc. | System and Method for Context Enhanced Ad Creation |
| US20150139485A1 (en) * | 2013-11-15 | 2015-05-21 | Facebook, Inc. | Pose-aligned networks for deep attribute modeling |
| AU2013254921A1 (en) * | 2013-11-07 | 2015-05-21 | Canon Kabushiki Kaisha | Method, apparatus and system for determining a label for a group of individuals represented in images |
| KR20160059403A (en) * | 2014-11-18 | 2016-05-26 | 주식회사 솔루엠 | Apparatus for generating advertisement image using display of a plurality of electronic information labels and advertisement method by thereof |
| CN105677844A (en) * | 2016-01-06 | 2016-06-15 | 北京摩比万思科技有限公司 | Mobile advertisement big data directional pushing and user cross-screen recognition method |
| CN105825396A (en) * | 2016-03-11 | 2016-08-03 | 合网络技术（北京）有限公司 | Co-occurrence-based advertisement label clustering method and system |
| CN105913275A (en) * | 2016-03-25 | 2016-08-31 | 哈尔滨工业大学深圳研究生院 | Clothes advertisement putting method and system based on video leading role identification |
| WO2017019643A1 (en) * | 2015-07-24 | 2017-02-02 | Videoamp, Inc. | Targeting tv advertising slots based on consumer online behavior |
| WO2017019646A1 (en) * | 2015-07-24 | 2017-02-02 | Videoamp, Inc. | Sequential delivery of advertising content across media devices |
| CN106383887A (en) * | 2016-09-22 | 2017-02-08 | 深圳市博安达信息技术股份有限公司 | Environment-friendly news data acquisition and recommendation display method and system |
| CN107203598A (en) * | 2017-05-08 | 2017-09-26 | 广州智慧城市发展研究院 | A kind of method and system for realizing image switch labels |
| CN107358264A (en) * | 2017-07-14 | 2017-11-17 | 深圳市唯特视科技有限公司 | A kind of method that graphical analysis is carried out based on machine learning algorithm |
| CN107545271A (en) * | 2016-06-29 | 2018-01-05 | 阿里巴巴集团控股有限公司 | Image-recognizing method, device and system |
| CN107562742A (en) * | 2016-06-30 | 2018-01-09 | 苏宁云商集团股份有限公司 | A kind of image processing method and device |
| CN107636646A (en) * | 2015-08-03 | 2018-01-26 | 谷歌有限责任公司 | Facility grappling is carried out using the imaging of geo-location |
-
2018
- 2018-02-13 CN CN201810150833.4A patent/CN108446330B/en active Active
Patent Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100082427A1 (en) * | 2008-09-30 | 2010-04-01 | Yahoo! Inc. | System and Method for Context Enhanced Ad Creation |
| AU2013254921A1 (en) * | 2013-11-07 | 2015-05-21 | Canon Kabushiki Kaisha | Method, apparatus and system for determining a label for a group of individuals represented in images |
| US20150139485A1 (en) * | 2013-11-15 | 2015-05-21 | Facebook, Inc. | Pose-aligned networks for deep attribute modeling |
| KR20160059403A (en) * | 2014-11-18 | 2016-05-26 | 주식회사 솔루엠 | Apparatus for generating advertisement image using display of a plurality of electronic information labels and advertisement method by thereof |
| WO2017019643A1 (en) * | 2015-07-24 | 2017-02-02 | Videoamp, Inc. | Targeting tv advertising slots based on consumer online behavior |
| WO2017019646A1 (en) * | 2015-07-24 | 2017-02-02 | Videoamp, Inc. | Sequential delivery of advertising content across media devices |
| CN107636646A (en) * | 2015-08-03 | 2018-01-26 | 谷歌有限责任公司 | Facility grappling is carried out using the imaging of geo-location |
| CN105677844A (en) * | 2016-01-06 | 2016-06-15 | 北京摩比万思科技有限公司 | Mobile advertisement big data directional pushing and user cross-screen recognition method |
| CN105825396A (en) * | 2016-03-11 | 2016-08-03 | 合网络技术（北京）有限公司 | Co-occurrence-based advertisement label clustering method and system |
| CN105913275A (en) * | 2016-03-25 | 2016-08-31 | 哈尔滨工业大学深圳研究生院 | Clothes advertisement putting method and system based on video leading role identification |
| CN107545271A (en) * | 2016-06-29 | 2018-01-05 | 阿里巴巴集团控股有限公司 | Image-recognizing method, device and system |
| CN107562742A (en) * | 2016-06-30 | 2018-01-09 | 苏宁云商集团股份有限公司 | A kind of image processing method and device |
| CN106383887A (en) * | 2016-09-22 | 2017-02-08 | 深圳市博安达信息技术股份有限公司 | Environment-friendly news data acquisition and recommendation display method and system |
| CN107203598A (en) * | 2017-05-08 | 2017-09-26 | 广州智慧城市发展研究院 | A kind of method and system for realizing image switch labels |
| CN107358264A (en) * | 2017-07-14 | 2017-11-17 | 深圳市唯特视科技有限公司 | A kind of method that graphical analysis is carried out based on machine learning algorithm |
Non-Patent Citations (4)
| Title |
|---|
| DONG MINGZHI等: "TRANSFERRING CNNS TO MULTI-INSTANCE MULTI-LABEL CLASSIFICATION ON SMALL DATASETS", 《2017 24TH IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP)》 * |
| 吕云翔等: "基于机器学习的监控视频行人检测与追踪系统的设计与实现", 《工业和信息化教育》 * |
| 张国燕: "基于标签的个性化广告精准营销系统设计与实现", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
| 张广: "基于贝叶斯方法的图像标注研究与系统实现", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740729A (en) * | 2018-12-14 | 2019-05-10 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN109447719A (en) * | 2018-12-17 | 2019-03-08 | 厦门美柚信息科技有限公司 | Targeted promotion commodity automatic determination method, device, medium and electronic equipment |
| CN113971592A (en) * | 2021-12-23 | 2022-01-25 | 成都易播科技有限公司 | Supervision evaluation method, system and device for promotion information release main body |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108446330B (en) | 2022-05-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110796204B (en) | Video tag determining method, device and server | |
| CN108205766A (en) | Information-pushing method, apparatus and system | |
| CN105893407A (en) | Individual user portraying method and system | |
| CN106204165A (en) | A kind of advertisement placement method and device | |
| CN109933699A (en) | A kind of construction method and device of academic portrait model | |
| CN111882361A (en) | Audience accurate advertisement pushing method and system based on artificial intelligence and readable storage medium | |
| CN111078940B (en) | Image processing method, device, computer storage medium and electronic equipment | |
| CN108446330A (en) | Promotion object processing method and device and computer-readable storage medium | |
| CN115002200B (en) | Message pushing method, device, equipment and storage medium based on user portrait | |
| CN110598095B (en) | Method, device and storage medium for identifying article containing specified information | |
| CN108734516A (en) | Advertisement placement method and device | |
| CN113592605B (en) | Product recommendation method, device, equipment and storage medium based on similar products | |
| CN114936301B (en) | Intelligent household building material data management method, device, equipment and storage medium | |
| CN111881901A (en) | Screenshot content detection method and device and computer-readable storage medium | |
| CN113269232B (en) | Model training method, vectorization recall method, related equipment and storage medium | |
| US20230316106A1 (en) | Method and apparatus for training content recommendation model, device, and storage medium | |
| KR102559950B1 (en) | An AI-based optimal advertising recommendation system | |
| CN110765352B (en) | User interest identification method and device | |
| CN108734498B (en) | Advertisement pushing method and device | |
| CN116957035A (en) | Model training method and method for recommending media content | |
| CN110942056A (en) | Clothing key point positioning method and device, electronic equipment and medium | |
| CN114372202A (en) | Content recommendation method and device, storage medium and electronic equipment | |
| CN112765444A (en) | Method, device and equipment for extracting target text segment and storage medium | |
| CN114372580A (en) | Model training method, storage medium, electronic device, and computer program product | |
| CN109902531B (en) | User management method, device, medium and electronic equipment |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20210122 Address after: Room 2020, 2 / F, building 27, No. 25, North Third Ring Road West, Haidian District, Beijing 100089 Applicant after: Beijing minglue Zhaohui Technology Co.,Ltd. Address before: 100012 Room 302, block B, Ruipu building, No.15, Hongjunying South Road, Beiyuan, Chaoyang District, Beijing Applicant before: BEIJING XINSIGHT TECHNOLOGY Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |