CN107590741A - A kind of method and system of predicted pictures popularity - Google Patents
A kind of method and system of predicted pictures popularity Download PDFInfo
- Publication number
- CN107590741A CN107590741A CN201710848290.9A CN201710848290A CN107590741A CN 107590741 A CN107590741 A CN 107590741A CN 201710848290 A CN201710848290 A CN 201710848290A CN 107590741 A CN107590741 A CN 107590741A
- Authority
- CN
- China
- Prior art keywords
- models
- feature
- feature vector
- input
- input feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
This application discloses a kind of method of predicted pictures popularity, method includes：The source data of reception is pre-processed to obtain feature samples；Feature samples carry out random inactivation processing according to pre-set ratio, obtain the first input feature vector；Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast result, and medium range forecast result and the first input feature vector are combined to obtain the second input feature vector；Judge that depth stacks whether regression model restrains according to the second input feature vector；If so, corresponding picture popularity is then generated according to medium range forecast result；If it is not, then combine medium range forecast result with the second input feature vector to obtain the 3rd input feature vector, and using the 3rd input results as new feature samples；This method is capable of the popularity of Accurate Prediction picture；Disclosed herein as well is a kind of system of predicted pictures popularity, a kind of computer-readable recording medium and server, has above beneficial effect.
Description
Technical field
The present invention relates to data analysis field, more particularly to a kind of method of predicted pictures popularity, system and a kind of meter
Calculation machine readable storage medium storing program for executing and server.
Background technology
The rapid development of information technology has promoted the prevalence of social media, and social media changes the mode of people's interaction.
User mainly by way of sending picture, shares oneself life and experience in social media platform.Therefore, social media is accumulated
The image data of magnanimity is tired out.However, the popularity of these pictures is not quite similar.The picture that the user of different popularity is sent out
Popularity differs greatly, and the popularity of the picture of same user's hair is also different.The application in many fields, such as news personalization push away
The design, the dispensing of online advertisement etc. of system is recommended, benefits from social media picture Popularity prediction this subject study.
In the prior art, the social network message outburst detection based on Recognition with Recurrent Neural Network is that user in social networks is sent out
The classification of cloth and the history message of forwarding is predicted, judges whether message breaks out.The prior art is related only to in social media
The prediction of the text message in face, the accurate prediction to social picture popularity can not be realized.
Therefore, how to realize and social picture popularity is precisely predicted, be that those skilled in the art need to solve at present
Technical problem certainly.
The content of the invention
The purpose of the application is to provide a kind of method of predicted pictures popularity, system and a kind of computer-readable storage medium
Matter and server, it can realize and social picture popularity is precisely predicted.
In order to solve the above technical problems, the application provides a kind of method of predicted pictures popularity, this method includes：
Step 1：The source data of reception is pre-processed to obtain feature samples；Wherein, the feature samples include vision
Feature and social characteristics；
Step 2：Random inactivation processing is carried out according to pre-set ratio to the feature samples, obtains the first input feature vector；
Step 3：Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast result,
And the medium range forecast result and first input feature vector are combined to obtain the second input feature vector；
Step 4：Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then by described in
Medium range forecast result combines to obtain the 3rd input feature vector with second input feature vector, and using the 3rd input results as new
The feature samples enter step 2；If so, then enter step 5；
Step 5：The corresponding picture popularity is generated according to the medium range forecast result.
Optionally, the source data of described pair of reception is pre-processed to obtain feature samples and included：
The source data is divided into image data and social data；
Feature extraction is carried out to the image data using artificial neural network and obtains the visual signature；
The social data is converted to obtain using multi-time Scales scale and Z-score (criterion score) standardization described
Social characteristics；
The visual signature and the social characteristics are spliced by preset rules to obtain the feature samples.
Optionally, it is described that the visual signature is obtained to image data progress feature extraction using artificial neural network
Including：
Two-stage feature extraction is carried out to the image data using artificial neural network and obtains low-level features and advanced features；
The low-level features and the advanced features are combined to obtain the visual signature.
Optionally, it is described that low-level features are obtained to image data progress two-stage feature extraction using artificial neural network
Include with advanced features：
Training obtains ResNeXt models on ImageNet data sets and Place365 data sets, Xception models and
DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet models are artificial neural network；
The image data is input in the ResNeXt models, Xception models and DenseNet models, and will
The value of characteristic pattern before the ResNeXt models, Xception models and DenseNet models last layer is extracted and passed through
Feature Compression and feature selecting obtain the low-level features；
Scene information and class by the ResNeXt models, Xception models and DenseNet models to picture prediction
Other information connects to obtain the advanced features.
Present invention also provides a kind of system of predicted pictures popularity, the system includes：
Pretreatment module, for being pre-processed to obtain feature samples to the source data of reception；Wherein, the feature samples
Including visual signature and social characteristics；
Dropout modules, for carrying out random inactivation processing according to pre-set ratio to the feature samples, it is defeated to obtain first
Enter feature；
Block modules, obtained for being predicted operation to first input feature vector using default several regression models
Medium range forecast result, and the medium range forecast result and the input feature vector are combined to obtain the second input feature vector；
Detector modules, for judging that depth stacks whether regression model restrains according to second input feature vector；If
It is no, then the medium range forecast result is combined with second input feature vector to obtain the 3rd input feature vector, and it is defeated by the described 3rd
Enter result and enter next layer of stacking regression model as the new feature samples；If so, then according to the medium range forecast result
Generate the corresponding picture popularity.
Optionally, the pretreatment module includes：
Classification submodule, for the source data to be divided into image data and social data；
Visual Feature Retrieval Process submodule, obtained for carrying out feature extraction to the image data using artificial neural network
The visual signature；
Social characteristics extracting sub-module, for being entered using multi-time Scales scale and Z-score standardization to the social data
Row conversion obtains the social characteristics；
Splice submodule, for splicing the visual signature and the social characteristics by preset rules to obtain the feature
Sample.
Optionally, the Visual Feature Retrieval Process submodule includes：
Two-stage feature extraction unit, obtained for carrying out two-stage feature extraction to the image data using artificial neural network
To low-level features and advanced features；
Assembled unit, for being combined to obtain the visual signature by the low-level features and the advanced features.
Optionally, the two-stage feature extraction includes：
Model training subelement, ResNeXt is obtained for being trained on ImageNet data sets and Place365 data sets
Model, Xception models and DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet moulds
Type is artificial neural network；
Low level feature extraction subelement, for the image data to be input into the ResNeXt models, Xception moulds
In type and DenseNet models, and by before the ResNeXt models, Xception models and DenseNet models last layers
The value of characteristic pattern extract and pass through Feature Compression and feature selecting obtains the low-level features；
Advanced features extract subelement, for by the ResNeXt models, Xception models and DenseNet models pair
The scene information of picture prediction connects to obtain the advanced features with classification information.
Present invention also provides a kind of computer-readable recording medium, is stored thereon with computer program, the computer
Program realizes following steps when performing：
Step 1：The source data of reception is pre-processed to obtain feature samples；Wherein, the feature samples include vision
Feature and social characteristics；
Step 2：Random inactivation processing is carried out according to pre-set ratio to the feature samples, obtains the first input feature vector；
Step 3：Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast result,
And the medium range forecast result and first input feature vector are combined to obtain the second input feature vector；
Step 4：Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then by described in
Medium range forecast result combines to obtain the 3rd input feature vector with second input feature vector, and using the 3rd input results as new
The feature samples enter step 2；If so, then enter step 5；
Step 5：The corresponding picture popularity is generated according to the medium range forecast result.
Present invention also provides a kind of server, including memory and processor, computer is stored with the memory
Program, the processor realize following steps when calling the computer program in the memory：
Step 1：The source data of reception is pre-processed to obtain feature samples；Wherein, the feature samples include vision
Feature and social characteristics；
Step 2：Random inactivation processing is carried out according to pre-set ratio to the feature samples, obtains the first input feature vector；
Step 3：Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast result,
And the medium range forecast result and first input feature vector are combined to obtain the second input feature vector；
Step 4：Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then by described in
Medium range forecast result combines to obtain the 3rd input feature vector with second input feature vector, and using the 3rd input results as new
The feature samples enter step 2；If so, then enter step 5；
Step 5：The corresponding picture popularity is generated according to the medium range forecast result.
The invention provides a kind of method of predicted pictures popularity, the source data of reception is pre-processed to obtain feature
Sample；Wherein, the feature samples include visual signature and social characteristics；To the feature samples according to pre-set ratio carry out with
The processing of machine inactivation, obtains the first input feature vector；The input feature vector is predicted using default several regression models and operated
To medium range forecast result, and the medium range forecast result and first input feature vector be combined to obtain the second input spy
Sign；Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then by the medium range forecast result
Combine to obtain the 3rd input feature vector with second input feature vector, and using the 3rd input results as the new feature sample
This；If so, the corresponding picture popularity is then generated according to the medium range forecast result.
Source data is pre-processed in this method to obtain the feature samples that can be identified, to feature samples according to default ratio
Example, which carries out random inactivation processing, can increase the diversity of feature samples, make picture Popularity prediction more accurate.Utilize recurrence
Model is predicted to obtain predicted value to the first input feature vector of input, and judges that depth stacks whether regression model restrains, if
Do not restrain, repeat random inactivation processing, forecast of regression model, convergence judge operation until depth regression model restrain,
Obtain picture popularity.Due to multiple regression models be present, so multiple predicted values can be obtained, above-mentioned predicted value is averaged
The picture popularity predicted.This method is capable of the popularity of Accurate Prediction picture by the stacking of multiple regression model, has
Beneficial to the development of new media；The application additionally provides a kind of system of predicted pictures popularity, a kind of computer-readable deposited simultaneously
Storage media and server, there is above-mentioned beneficial effect, will not be repeated here.
Brief description of the drawings
In order to illustrate more clearly of the embodiment of the present application, the required accompanying drawing used in embodiment will be done simply below
Introduce, it should be apparent that, drawings in the following description are only some embodiments of the present application, for ordinary skill people
For member, on the premise of not paying creative work, other accompanying drawings can also be obtained according to these accompanying drawings.
A kind of flow chart of the method for predicted pictures popularity that Fig. 1 is provided by the embodiment of the present application；
The flow chart of the method for another predicted pictures popularity that Fig. 2 is provided by the embodiment of the present application；
The flow chart of the method for another predicted pictures popularity that Fig. 3 is provided by the embodiment of the present application；
A kind of structural representation of the system for predicted pictures popularity that Fig. 4 is provided by the embodiment of the present application.
Embodiment
To make the purpose, technical scheme and advantage of the embodiment of the present application clearer, below in conjunction with the embodiment of the present application
In accompanying drawing, the technical scheme in the embodiment of the present application is clearly and completely described, it is clear that described embodiment is
Some embodiments of the present application, rather than whole embodiments.Based on the embodiment in the application, those of ordinary skill in the art
The every other embodiment obtained under the premise of creative work is not made, belong to the scope of the application protection.
Fig. 1, a kind of flow of the method for predicted pictures popularity that Fig. 1 is provided by the embodiment of the present application are referred to below
Figure；
Specific steps can include：
Step S101：The source data of reception is pre-processed to obtain feature samples；Wherein, the feature samples include regarding
Feel feature and social characteristics；
Wherein, the purpose of this programme is the popularity of predicted pictures, therefore the source data received in this step is picture
Relevant information, the source of source data can be the social platforms such as Flickr, microblogging, QQ space or picture sharing website, herein
Acquisition source not to source data is defined, as long as website disclosed, related to picture.It is appreciated that
It is that because the purpose of the present invention is predicted pictures popularity, therefore the source data obtained can be the very high picture of concerned degree,
It can also be the picture having lost focus substantially.But the requirement of legal ethics is in order at, for some carrying flame (such as colors
Feelings, violence, political reaction) source data can rejected when being pre-processed without Popularity prediction.
It is understood that do not only have the visual information (information that image content carries) of picture, also society in source data
Handing over information, (whether comment number, the picture that issuing time, picture such as picture are obtained have label, the number of label, picture header
Length, issue picture user average pageview, user adds number etc. of group).Therefore, enter accordingly to source data
, it is necessary to be pre-processed to the visual information of picture during row pretreatment, it is also desirable to which the information of the social media aspect of picture is entered
Row pretreatment.Due to different from the object of processing, thus it is also different to the mode of its pretreatment carried out.For the vision of picture
Information, it can be believed using depth learning technology (such as ResNeXt, Xception or DenseNet) to extract the vision in source data
Breath obtains visual signature.Further, self-encoding encoder can be used and random forest carries out dimension-reduction treatment and feature selecting obtains
The visual signature less to dimension.Preferably, it can utilize the temporal information in social information can be by the issue date of picture
The division of various dimensions is carried out, for example season, month, hour etc. can be divided into.
Certainly, source data is pre-processed to obtain feature samples and there can be many operations, except correlation mentioned above
Operation is outer, can include the screening to repeating source data (i.e. multiple same photographs)；It can also include by picture/mb-type to source number
According to being classified, such as figure map, landscape figure, animal figure, i.e., first picture/mb-type is made a distinction and carry out follow-up prevalence again
Spend predicted operation.Above-mentioned classification is carried out to picture to be intended only as a kind of preferred embodiment and exist, those skilled in the art can
Comprehensive selection is made with the concrete scene applied according to scheme.
Step S102：Random inactivation processing is carried out according to pre-set ratio to the feature samples, obtains the first input feature vector；
Wherein, the purpose of this step is the diversity of lifting feature sample.Because this programme is returned by stacked multilayer
Prediction of the model realization to picture popularity, therefore the increase of the complexity with regression model, the prediction for popularity are past
It is past to be limited by overfitting.In order to reduce the limitation that overfitting is brought, this programme carry out picture Popularity prediction it
It is preceding that random inactivation processing is carried out to all feature samples.For example, it is assumed that input feature value is 4-D, and the ratio inactivated
Then it is that each basic model obtains 2-D characteristic vectors for 0.5；Assuming that there are 4 basic models in block module, ratio is inactivated herein
On the basis of, then the characteristic size of the input of next piece of module is 8-D.
It is understood that the pre-set ratio in this step, which is those skilled in the art, is carrying out many experiments and demonstration
Obtain afterwards, relatively accurate selection can be carried out according to the concrete condition of source data, to lift the degree of accuracy of Popularity prediction
And efficiency.Generally, can be properly increased at random inactivation in the case of feature samples (i.e. source data) diversity deficiency
The pre-set ratio of reason；And the default of random inactivation processing can be suitably reduced in the case where feature samples diversity is of a relatively high
Ratio.Sum it up, be not defined to the concrete numerical value of pre-set ratio herein, it is directed to different application environments and exists not
Same pre-set ratio.
Step S103：Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast knot
Fruit, and the medium range forecast result and first input feature vector are combined to obtain the second input feature vector；
Wherein, the purpose of this step is that the input feature vector obtained in step S102 is predicted to obtain preliminary prediction
Value.The regression model used in this step can have many kinds, including：Random forest (RF), extreme random tree (EXRT),
XGBoost or Lasso etc., there can also be other regression models certainly, those skilled in the art can voluntarily select,
Therefore the particular type of the regression model to being used in this step is not defined herein.
It is understood that the predetermined number mentioned in this step can 1 can be 2 or more than 2 numeral, but go out
In the multifarious consideration of increase output characteristic, the relatively low harmful effect of diversity only can be brought using a regression model, because
This increases the diversity of output characteristic using different types of regression model more as far as possible.It is, for example, possible to use random forest
(RF), extreme random tree (EXRT), XGBoost and these four regression models of Lasso are next simultaneously defeated to being obtained in step S102
Enter feature to be predicted to obtain preliminary predicted value.Wherein, because structure, the principle of the regression model of each type are different,
Therefore the predicted value that each regression model is calculated is also different, and all predicted values can be averaged to obtain centre
Prediction result.
By discussion above, this programme is to realize picture prediction by stacked multilayer regression model, therefore originally
The step for medium range forecast result and first input feature vector are combined to obtain the second input feature vector in step, be in order to
Judge whether to need to predict again into next layer of regression model.
Step S104：Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then enter
Step S105；If so, then enter step S106；
Wherein, the purpose of this step is that the second input feature vector obtained by step S103 judges that depth stacks regression model
Whether restrain.Depth (number that i.e. regression model stacks) is most important for stacking model, and depth is bigger within the specific limits
Prediction for picture popularity is more accurate, but the depth for increasing stacking simply when reaching some critical value will not increase
Add the degree of accuracy of prediction, the waste of resource can be caused on the contrary.Therefore, this step stacks whether regression model is received by judging depth
Hold back to analyse whether to need to increase the depth stacked.
Do not restrained if depth stacks model, need to continue to stack；If depth model is restrained, illustrate for picture
The prediction of popularity is over.Wherein continue to stack is exactly to utilize the 3rd input feature vector repeat step obtained in step S105
S102, step S103, step S104 flow are until judge that depth stacks model convergence.
Step S105：The medium range forecast result is combined with second input feature vector to obtain the 3rd input feature vector, and
Enter step S102 using the 3rd input results as the new feature samples.
Step S106：The corresponding picture popularity is generated according to the medium range forecast result.
Fig. 2, the stream of the method for another predicted pictures popularity that Fig. 2 is provided by the embodiment of the present application are referred to below
Cheng Tu.
The present embodiment how is directed in a upper embodiment in step S101 to being pre-processed to obtain feature in source data
The specific restriction that sample is made, other steps are substantially the same with a upper embodiment, and it is real that same section can be found in one
A relevant portion is applied, will not be repeated here.
Specific steps can include：
Step S201：The source data is divided into image data and social data；
Wherein, the image data mentioned in this step is exactly the visual information (image content carry information) of picture, society
Intersection number is according to being exactly that (whether comment number that issuing time, picture such as picture are obtained, picture have mark for the information of social media aspect
Label, the number of label, picture header length, issue picture user average pageview, user adds number etc. of group).
Step S202：Feature extraction is carried out to the image data using artificial neural network and obtains the visual signature；
Step S203：The social data is converted to obtain using multi-time Scales scale and Z-score standardization described
Social characteristics；
Step S204：The visual signature and the social characteristics are spliced by preset rules to obtain the feature samples.
Fig. 3, the stream of the method for another predicted pictures popularity that Fig. 3 is provided by the embodiment of the present application are referred to below
Cheng Tu.
The present embodiment is to be directed in a upper embodiment in step S202 how to utilize artificial neural network to the picture number
The specific restriction for obtaining the visual signature according to feature extraction is carried out and being made, other steps and a upper embodiment are substantially
Identical, same section can be found in an embodiment relevant portion, will not be repeated here.
Step S301：Training obtains ResNeXt models on ImageNet data sets and Place365 data sets,
Xception models and DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet models are equal
For artificial neural network；
Step S302：The image data is input to the ResNeXt models, Xception models and DenseNet moulds
In type, and the value of the characteristic pattern before the ResNeXt models, Xception models and DenseNet models last layer is carried
Take and pass through Feature Compression and feature selecting obtains the low-level features；
Step S303：Scene by the ResNeXt models, Xception models and DenseNet models to picture prediction
Information connects to obtain the advanced features with classification information.
Step S304：The low-level features and the advanced features are combined to obtain the visual signature.
Refer to Fig. 4, a kind of structural representation of the system for predicted pictures popularity that Fig. 4 is provided by the embodiment of the present application
Figure；
The system can include：
Pretreatment module 100, for being pre-processed to obtain feature samples to the source data of reception；Wherein, the feature
Sample includes visual signature and social characteristics；
Dropout modules 200, for carrying out random inactivation processing according to pre-set ratio to the feature samples, obtain the
One input feature vector；
Block modules 300, operated for being predicted using default several regression models to first input feature vector
It is combined to obtain the second input feature vector to medium range forecast result, and by the medium range forecast result and the input feature vector；
Detector modules 400, for judging that depth stacks whether regression model restrains according to second input feature vector；
If it is not, the medium range forecast result then to be combined to obtain the 3rd input feature vector with second input feature vector, and by the described 3rd
Input results enter next layer as the new feature samples and stack regression model if so, then according to the medium range forecast result
Generate the corresponding picture popularity.
Wherein, each module that first layer is only stacked to regression model in Fig. 4 is marked, each layer of stacking regression model
There are Dropout modules, Block modules and Detector modules.
In the embodiment of the system for another predicted pictures popularity that the application provides, the pretreatment module 100
Including：
Classification submodule, for the source data to be divided into image data and social data；
Visual Feature Retrieval Process submodule, obtained for carrying out feature extraction to the image data using artificial neural network
The visual signature；
Social characteristics extracting sub-module, for being entered using multi-time Scales scale and Z-score standardization to the social data
Row conversion obtains the social characteristics；
Splice submodule, for splicing the visual signature and the social characteristics by preset rules to obtain the feature
Sample.
Further, the Visual Feature Retrieval Process submodule includes：
Two-stage feature extraction unit, obtained for carrying out two-stage feature extraction to the image data using artificial neural network
To low-level features and advanced features；
Assembled unit, for being combined to obtain the visual signature by the low-level features and the advanced features.
Further, the two-stage feature extraction includes：
Model training subelement, ResNeXt is obtained for being trained on ImageNet data sets and Place365 data sets
Model, Xception models and DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet moulds
Type is artificial neural network；
Low level feature extraction subelement, for the image data to be input into the ResNeXt models, Xception moulds
In type and DenseNet models, and by before the ResNeXt models, Xception models and DenseNet models last layers
The value of characteristic pattern extract and pass through Feature Compression and feature selecting obtains the low-level features；
Advanced features extract subelement, for by the ResNeXt models, Xception models and DenseNet models pair
The scene information of picture prediction connects to obtain the advanced features with classification information.
Because the embodiment of components of system as directed and the embodiment of method part are mutually corresponding, therefore the embodiment of components of system as directed please
Referring to the description of the embodiment of method part, wouldn't repeat here.
Present invention also provides a kind of computer-readable recording medium, there is computer program thereon, the computer program
The step of above-described embodiment provides can be realized when being performed.The storage medium can include：USB flash disk, mobile hard disk, read-only deposit
Reservoir (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disc or
CD etc. is various can be with the medium of store program codes.
Present invention also provides a kind of server, can include memory and processor, have calculating in the memory
Machine program, the processor call the computer sechron in the memory, it is possible to achieve the step that above-described embodiment is provided
Suddenly.Certain server can also include the component such as various network interfaces, power supply.
Each embodiment is described by the way of progressive in specification, and what each embodiment stressed is and other realities
Apply the difference of example, between each embodiment identical similar portion mutually referring to.For system disclosed in embodiment
Speech, because it is corresponded to the method disclosed in Example, so description is fairly simple, related part is referring to method part illustration
.It should be pointed out that for those skilled in the art, on the premise of the application principle is not departed from, also
Some improvement and modification can be carried out to the application, these are improved and modification also falls into the application scope of the claims
It is interior.
It should also be noted that, in this manual, such as first and second or the like relational terms be used merely to by
One entity or operation make a distinction with another entity or operation, and not necessarily require or imply these entities or operation
Between any this actual relation or order be present.Moreover, term " comprising ", "comprising" or its any other variant meaning
Covering including for nonexcludability, so that process, method, article or equipment including a series of elements not only include that
A little key elements, but also the other element including being not expressly set out, or also include for this process, method, article or
The intrinsic key element of equipment.Under the situation not limited more, the key element that is limited by sentence "including a ..." is not arranged
Except other identical element in the process including the key element, method, article or equipment being also present.
Claims (10)
- A kind of 1. method of predicted pictures popularity, it is characterised in that including：Step 1：The source data of reception is pre-processed to obtain feature samples；Wherein, the feature samples include visual signature And social characteristics；Step 2：Random inactivation processing is carried out according to pre-set ratio to the feature samples, obtains the first input feature vector；Step 3：Operation is predicted to the input feature vector using default several regression models and obtains medium range forecast result, and will The medium range forecast result is combined to obtain the second input feature vector with first input feature vector；Step 4：Judge that depth stacks whether regression model restrains according to second input feature vector；If it is not, then by the centre Prediction result combines to obtain the 3rd input feature vector with second input feature vector, and using the 3rd input results as new institute State feature samples and enter step 2；If so, then enter step 5；Step 5：The corresponding picture popularity is generated according to the medium range forecast result.
- 2. method according to claim 1, it is characterised in that the source data of described pair of reception is pre-processed to obtain feature sample Originally include：The source data is divided into image data and social data；Feature extraction is carried out to the image data using artificial neural network and obtains the visual signature；The social data is converted to obtain the social characteristics using multi-time Scales scale and Z-score standardization；The visual signature and the social characteristics are spliced by preset rules to obtain the feature samples.
- 3. method according to claim 2, it is characterised in that described to be carried out using artificial neural network to the image data Feature extraction, which obtains the visual signature, to be included：Two-stage feature extraction is carried out to the image data using artificial neural network and obtains low-level features and advanced features；The low-level features and the advanced features are combined to obtain the visual signature.
- 4. method according to claim 3, it is characterised in that described to be carried out using artificial neural network to the image data Two-stage feature extraction, which obtains low-level features and advanced features, to be included：Training obtains ResNeXt models on ImageNet data sets and Place365 data sets, Xception models and DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet models are artificial neural network；The image data is input in the ResNeXt models, Xception models and DenseNet models, and by described in The value of characteristic pattern before ResNeXt models, Xception models and DenseNet models last layer is extracted and passes through feature Compression and feature selecting obtain the low-level features；The ResNeXt models, Xception models and DenseNet models are believed the scene information and classification of picture prediction Breath connection obtains the advanced features.
- A kind of 5. system of predicted pictures popularity, it is characterised in that including：Pretreatment module, for being pre-processed to obtain feature samples to the source data of reception；Wherein, the feature samples include Visual signature and social characteristics；Dropout modules, for carrying out random inactivation processing according to pre-set ratio to the feature samples, obtain the first input spy Sign；Block modules, centre is obtained for being predicted operation to first input feature vector using default several regression models Prediction result, and the medium range forecast result and the input feature vector are combined to obtain the second input feature vector；Detector modules, for judging that depth stacks whether regression model restrains according to second input feature vector；If it is not, then The medium range forecast result combines with second input feature vector to obtain the 3rd input feature vector, and by the 3rd input results Enter next layer of stacking regression model as the new feature samples；If so, phase is then generated according to the medium range forecast result The corresponding picture popularity.
- 6. system according to claim 5, it is characterised in that the pretreatment module includes：Classification submodule, for the source data to be divided into image data and social data；Visual Feature Retrieval Process submodule, it is described for being obtained using artificial neural network to image data progress feature extraction Visual signature；Social characteristics extracting sub-module, for being turned using multi-time Scales scale and Z-score standardization to the social data Change obtains the social characteristics；Splice submodule, for splicing the visual signature and the social characteristics by preset rules to obtain the feature sample This.
- 7. system according to claim 6, it is characterised in that the Visual Feature Retrieval Process submodule includes：Two-stage feature extraction unit is low for being obtained using artificial neural network to image data progress two-stage feature extraction Level feature and advanced features；Assembled unit, for being combined to obtain the visual signature by the low-level features and the advanced features.
- 8. system according to claim 7, it is characterised in that the two-stage feature extraction includes：Model training subelement, ResNeXt models are obtained for being trained on ImageNet data sets and Place365 data sets, Xception models and DenseNet models；Wherein, the ResNeXt models, Xception models and DenseNet models are equal For artificial neural network；Low level feature extraction subelement, for by the image data be input to the ResNeXt models, Xception models and In DenseNet models, and by the spy before the ResNeXt models, Xception models and DenseNet models last layer The value of sign figure is extracted and passes through Feature Compression and feature selecting obtains the low-level features；Advanced features extract subelement, for by the ResNeXt models, Xception models and DenseNet models to picture The scene information of prediction connects to obtain the advanced features with classification information.
- 9. a kind of computer-readable recording medium, is stored thereon with computer program, it is characterised in that the computer program is held The method as described in any one of Claims 1-4 is realized during row.
- A kind of 10. server, it is characterised in that including memory and processor, computer program is stored with the memory, The processor realizes the method as described in any one of Claims 1-4 when calling the computer program in the memory.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710848290.9A CN107590741A (en) | 2017-09-19 | 2017-09-19 | A kind of method and system of predicted pictures popularity |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710848290.9A CN107590741A (en) | 2017-09-19 | 2017-09-19 | A kind of method and system of predicted pictures popularity |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107590741A true CN107590741A (en) | 2018-01-16 |
Family
ID=61046930
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710848290.9A Pending CN107590741A (en) | 2017-09-19 | 2017-09-19 | A kind of method and system of predicted pictures popularity |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107590741A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108805007A (en) * | 2018-04-18 | 2018-11-13 | 浙江百世技术有限公司 | Class's line hole capital after selling all securities state judging method based on image recognition |
| CN109086932A (en) * | 2018-08-02 | 2018-12-25 | 广东工业大学 | A kind of prediction technique, system and the device of media information prevalence degree |
| CN109635990A (en) * | 2018-10-12 | 2019-04-16 | 阿里巴巴集团控股有限公司 | A kind of training method, prediction technique, device and electronic equipment |
| CN110209789A (en) * | 2019-05-29 | 2019-09-06 | 山东大学 | A kind of multi-modal dialog system and method for user's attention guidance |
| CN110503162A (en) * | 2019-08-29 | 2019-11-26 | 广东工业大学 | A kind of media information prevalence degree prediction technique, device and equipment |
| CN110956528A (en) * | 2019-10-14 | 2020-04-03 | 广东工业大学 | Recommendation method and system for e-commerce platform |
| CN112001258A (en) * | 2020-07-27 | 2020-11-27 | 上海东普信息科技有限公司 | Logistics truck punctual arrival identification method, device, equipment and storage medium |
| WO2021233389A1 (en) * | 2020-05-22 | 2021-11-25 | 支付宝(杭州)信息技术有限公司 | Method and apparatus for enhancing robustness of deep learning model |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777891A (en) * | 2016-11-21 | 2017-05-31 | 中国科学院自动化研究所 | A kind of data characteristics selection and Forecasting Methodology and device |
-
2017
- 2017-09-19 CN CN201710848290.9A patent/CN107590741A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777891A (en) * | 2016-11-21 | 2017-05-31 | 中国科学院自动化研究所 | A kind of data characteristics selection and Forecasting Methodology and device |
Non-Patent Citations (2)
| Title |
|---|
| KHALED ALMGREN 等: "Prediction of Image Popularity Over Time on Social Media Networks", 《IEEE》 * |
| 孔庆超 等: "基于动态演化的讨论帖流行度预测", 《软件学报》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108805007A (en) * | 2018-04-18 | 2018-11-13 | 浙江百世技术有限公司 | Class's line hole capital after selling all securities state judging method based on image recognition |
| CN109086932A (en) * | 2018-08-02 | 2018-12-25 | 广东工业大学 | A kind of prediction technique, system and the device of media information prevalence degree |
| CN109635990A (en) * | 2018-10-12 | 2019-04-16 | 阿里巴巴集团控股有限公司 | A kind of training method, prediction technique, device and electronic equipment |
| CN109635990B (en) * | 2018-10-12 | 2022-09-16 | 创新先进技术有限公司 | Training method, prediction method, device, electronic equipment and storage medium |
| CN110209789A (en) * | 2019-05-29 | 2019-09-06 | 山东大学 | A kind of multi-modal dialog system and method for user's attention guidance |
| CN110209789B (en) * | 2019-05-29 | 2021-08-24 | 山东大学 | Multi-modal dialog system and method for guiding user attention |
| CN110503162A (en) * | 2019-08-29 | 2019-11-26 | 广东工业大学 | A kind of media information prevalence degree prediction technique, device and equipment |
| CN110956528A (en) * | 2019-10-14 | 2020-04-03 | 广东工业大学 | Recommendation method and system for e-commerce platform |
| CN110956528B (en) * | 2019-10-14 | 2022-11-04 | 广东工业大学 | Recommendation method and system for e-commerce platform |
| WO2021233389A1 (en) * | 2020-05-22 | 2021-11-25 | 支付宝(杭州)信息技术有限公司 | Method and apparatus for enhancing robustness of deep learning model |
| CN112001258A (en) * | 2020-07-27 | 2020-11-27 | 上海东普信息科技有限公司 | Logistics truck punctual arrival identification method, device, equipment and storage medium |
| CN112001258B (en) * | 2020-07-27 | 2023-07-11 | 上海东普信息科技有限公司 | Method, device, equipment and storage medium for identifying on-time arrival of logistics truck |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107590741A (en) | A kind of method and system of predicted pictures popularity | |
| CN110837579B (en) | Video classification method, apparatus, computer and readable storage medium | |
| CN113518256B (en) | Video processing method, video processing device, electronic equipment and computer readable storage medium | |
| CN113392236A (en) | Data classification method, computer equipment and readable storage medium | |
| CN111783712A (en) | Video processing method, device, equipment and medium | |
| EP3839800A1 (en) | Recommending multimedia based on user utterances | |
| CN112995690B (en) | Live content category identification method, device, electronic equipment and readable storage medium | |
| CN113328994B (en) | Malicious domain name processing method, device, equipment and machine readable storage medium | |
| CN110609958A (en) | Data pushing method and device, electronic equipment and storage medium | |
| CN112668638A (en) | Image aesthetic quality evaluation and semantic recognition combined classification method and system | |
| CN110097472A (en) | A kind of corporations' attribute recognition approach and relevant device | |
| CN114398973A (en) | Media content label identification method, device, equipment and storage medium | |
| CN107563394A (en) | A kind of method and system of predicted pictures popularity | |
| CN106649743A (en) | Method and system for storing and sharing creative idea classified brain library resources | |
| CN111597361B (en) | Multimedia data processing method, device, storage medium and equipment | |
| CN112069230A (en) | Data analysis method, device, equipment and storage medium | |
| CN101546343A (en) | Method, device and system for matching the colors of probes | |
| CN116522131A (en) | Object representation method, device, electronic equipment and computer readable storage medium | |
| CN113312445B (en) | Data processing method, model construction method, classification method and computing equipment | |
| CN108460475A (en) | Poor student's prediction technique and device based on network playing by students behavior | |
| CN115205606A (en) | Image multi-label classification method and device and related products | |
| CN110569435B (en) | Intelligent dual-ended recommendation engine system and method | |
| CN111523034B (en) | Application processing method, device, equipment and medium | |
| CN116932873A (en) | Video account recommending method, device, equipment, storage medium and program product | |
| CN113076450A (en) | Method and device for determining target recommendation list |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180116 |
|
| RJ01 | Rejection of invention patent application after publication |