CN107291845A - A kind of film based on trailer recommends method and system - Google Patents
A kind of film based on trailer recommends method and system Download PDFInfo
- Publication number
- CN107291845A CN107291845A CN201710408481.3A CN201710408481A CN107291845A CN 107291845 A CN107291845 A CN 107291845A CN 201710408481 A CN201710408481 A CN 201710408481A CN 107291845 A CN107291845 A CN 107291845A
- Authority
- CN
- China
- Prior art keywords
- film
- user
- trailer
- movie
- picture
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/73—Querying
- G06F16/738—Presentation of query results
- G06F16/739—Presentation of query results in form of a video summary, e.g. the video summary being a video sequence, a composite still image or having synthesized frames
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/7867—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using information manually generated, e.g. tags, keywords, comments, title and artist information, manually generated time, location and usage information, user ratings
Abstract
The present invention provides a kind of film based on trailer and recommends method and system, visual information to movie trailer is extracted, and be combined the visual information of movie trailer with the existing scoring of film, come the film progress score in predicting not scored each user, movie trailer usually can reflect the content of whole film, but it is shorter than the video playback time of whole film, therefore extracted compared to the existing information to whole film, reduce workload;The visual information to movie trailer is extracted simultaneously, the contextual information used compared to existing film proposed algorithm, the characteristics of visual information more can characterize film, it is possible to increase user is to the accuracy for the film score value prediction do not scored, and the recommendation to user's film is more accurate.
Description
Technical field
The present invention relates to information recommendation technology, recommend method more particularly, to a kind of film based on trailer and be
System.
Background technology
Sparse sex chromosome mosaicism is always a problem in film commending system, but is not still obtained so far very
Good solution.In actual film commending system, data often than sparse, traditional proposed algorithm performance by very big
Restriction.Therefore, the sparse sex chromosome mosaicism for solving data is extremely important to the performance for improving film commending system.For example, relatively more normal
In data set MovieLens-1m, the film number that each user's average score is crossed is about 164, with film number altogether
3544, mesh differs greatly.It is therefore seen that, score it is sparse in the case of, the scoring of simple dependence to film carries out recommendation analysis
It is that inaccurate, sparse scoring can not meet the requirement of commending system.In this case, aided in if extra information
Scoring carries out film recommendation, as a result will differ widely.Under this background, context-aware (context-aware) is pushed away
Algorithm is recommended to arise at the historic moment.
The most common algorithm that traditional film recommends related algorithm to use is the film proposed algorithm based on context, institute
The contextual information of use has:User related information, user social contact relation, film review etc..Although these information are believed as auxiliary
The performance of commending system can be improved by ceasing, but the information included relative to film itself, and it is useful that these information are included
Information is sub-fraction, in the correlative study that traditional film is recommended, and few researchers pay close attention to the spy of film itself
Property..
The content of the invention
The present invention provides a kind of electricity based on trailer for overcoming above mentioned problem or solving the above problems at least in part
Shadow recommends method and system.
Recommend method there is provided a kind of film based on trailer according to an aspect of the present invention, including:
S1, obtains scoring of each user in user's set to each film in movie collection, obtains multiple
User is to the rating matrixs of multiple films, and obtains the trailer of each film in movie collection;
S2, extracts the visual information of each movie trailer;
S3, according to the visual information of each movie trailer and the rating matrix, calculates each user to not commenting
The score in predicting value of the film divided;
S4, according to each user to the score in predicting value for the film not scored, film is recommended to user.
Beneficial effects of the present invention are:Visual information to movie trailer is extracted, and regarding movie trailer
Feel that information is combined with the existing scoring of film, come the film progress score in predicting not scored each user, preview
Piece usually can reflect the content of whole film, but shorter than the video playback time of whole film, therefore compared to existing
Information to whole film is extracted, and reduces workload;The visual information to movie trailer is extracted simultaneously, is compared
The contextual information used in existing film proposed algorithm, the characteristics of visual information more can characterize film, it is possible to increase
User is to the accuracy for the film score value prediction do not scored, and the recommendation to user's film is more accurate.
On the basis of above-mentioned technical proposal, the present invention can also make following improvement.
Further, the step S2 is specifically included:
Using the part picture in deep neural network extraction movie trailer as picture set, and extract picture set
In each pictures visual information.
Further, the deep neural network describes network and text analyzing network, the use depth including picture
Neutral net extracts part picture in movie trailer as picture set, and extracts each pictures in picture set
Visual information is specifically included:
S21, obtains the part picture in movie trailer as picture set, and it is pre- to film to use picture to describe network
The picture accused in piece is labeled, so as to obtain the corresponding sentence description of each pictures in picture set, composition film is pre-
Accuse the sentence description collections of piece;
The sentence description collections of movie trailer are analyzed and processed, obtain film pre- by S22 using text analyzing network
Accuse the visual information of piece.
Further, the picture describes combination of the network for convolutional neural networks and Recognition with Recurrent Neural Network, the text
Analysis network is made up of multilayer convolutional neural networks.
Further, also include before the step S21:
The frame picture of the part of representative of selected movie trailer, is used as the picture set for extracting visual information.
Further, the step S3 is specifically included:
S31, according to the visual information of each movie trailer, and multiple users are to the scoring square of multiple films
Battle array, is calculated using probability matrix decomposition algorithm and obtains each user characteristics vector sum each movie features vector, and then
To the eigenmatrix and the eigenmatrix of film of user;
S32, according to the eigenmatrix of user and the eigenmatrix of film, calculates each user to the film that does not score
Score in predicting value.
Further, each user is calculated to the scoring for the film not scored by equation below in the step S32
Predicted value:
Wherein,Predicted values of the user i to film j scoring is represented, μ is that all users are commented all films
The average value divided, biFor the deviation between the user i average values scored all films and μ, bjIt is all users to film j
Deviation between the average value and μ of scoring, UiFor user i characteristic vector, VjRepresent film j characteristic vector.
According to another aspect of the present invention there is provided a kind of film commending system based on trailer, including:
Acquisition module, for obtaining each user commenting to each film in movie collection in user's set
Point, rating matrix of multiple users to multiple films is obtained, and obtain the trailer of each film in movie collection;
Optimize computing module, for the visual information according to each movie trailer and the rating matrix, calculate every
One user is to the score in predicting value for the film not scored;
Recommending module, for, to the score in predicting value for the film not scored, film being recommended to user according to each user.
Further, the extraction module includes:
First extraction unit, obtains the part picture in movie trailer and describes net as picture set, and using picture
Network is labeled to the picture in movie trailer, so that the corresponding sentence description of each pictures in picture set is obtained,
Constitute the sentence description collections of movie trailer;
Second extraction unit, for being carried out using text analyzing network to the sentence description collections of movie trailer at analysis
Reason, extracts the visual information of movie trailer.
Further, the optimization computing module includes:
First computing unit, for the visual information according to each movie trailer, and the rating matrix, is used
Probability matrix decomposition algorithm, which is calculated, obtains each user characteristics vector sum each movie features vector, and then obtains user's
The eigenmatrix of eigenmatrix and film;
Second computing unit, for the eigenmatrix and the eigenmatrix of film according to user, calculates each user couple
The score in predicting value for the film not scored.
Brief description of the drawings
Fig. 1 recommends method flow diagram for the film based on trailer of one embodiment of the invention;
Fig. 2 connects block diagram for the film commending system based on trailer of another embodiment of the present invention;
Fig. 3 connects frame for the inside of extraction module in the film commending system based on trailer of one embodiment of the invention
Figure;
Fig. 4 for another embodiment of the present invention the film commending system based on trailer in optimize computing module inside
Connect block diagram.
Embodiment
With reference to the accompanying drawings and examples, the embodiment to the present invention is described in further detail.Implement below
Example is used to illustrate the present invention, but is not limited to the scope of the present invention.
Referring to Fig. 1, the film based on trailer that Fig. 1 provides one embodiment of the invention recommends method, including:S1,
Scoring of each user in user's set to each film in movie collection is obtained, multiple users are obtained to multiple electricity
The rating matrix of shadow, and obtain the trailer of each film in movie collection;S2, extracts each movie trailer
Visual information;S3, according to the visual information of each movie trailer and the rating matrix, calculates each user to not commenting
The score in predicting value of the film divided;S4, the score in predicting value for the film not scored according to each user recommends electricity to user
Shadow.
Existing film recommends method mainly to study the contextual informations such as user related information, user social contact relation, film review,
And the characteristic of film itself is paid close attention to as a kind of few researchs of critically important information, so the research self-contained letter of film
Breath is very important.But, film is typically long, and complexity accordingly can be very high in terms of processing, so the present invention chooses electricity
Shadow trailer is used as research object, on the one hand, movie trailer can be with the essential information of underlying cover film, while what is be related to is interior
Hold is usually that relatively excellent, great film is representational;On the other hand, movie trailer is more brief, relative to whole film
For the complexity studied can substantially reduce.Furthermore, it is contemplated that the Consumer's Experience of film is directly influenceed by visual information, so
Visual information in main research trailer of the invention.
The present embodiment is using the visual information of movie trailer is combined with the scoring of film, for different users,
Score value of the different users to unknown film is predicted, and by score in predicting value of each user to unknown film, next pair
Different users are targetedly recommended.
Concentrated for film, there are multiple films, and each film scored it by multiple users, that is,
Say, for user's collection and film collection, a user scored multiple films therein, but not to therein all
Film scored.Because the visual signature information progress extraction comparison to whole movie is difficult, time-consuming, and it may extract
To many useless interference informations, increase the difficulty recommended.Accordingly, it is considered to arrive this problem, the present embodiment is mainly pre- from film
Accuse to extract in visual signature information, movie trailer in piece and contain part most excellent in whole film, can cover whole
The essential characteristic of film.
All films concentrated for film, obtain the corresponding movie trailer of each film from video website, and
Extract the visual information of each movie trailer.Wherein, the scoring according to user to a part of film, by these scorings with using
Family and film are mapped, and can obtain rating matrix of the user to film.
Then, according to the visual information of each movie trailer and the rating matrix of multiple films of multiple users couple,
The eigenmatrix of user and the eigenmatrix of film, and eigenmatrix and the feature square of film according to user are obtained by optimization
Battle array, calculates each user to the score in predicting value for the film not scored.Scoring according to each user to different films
Predicted value, the recommendation of film is carried out to each user, so the film of user is recommended according to specific aim, can be reached more preferably
Effect.
The present embodiment is extracted to the visual information of movie trailer, and by the visual information and film of movie trailer
Existing scoring is combined, and come the film progress score in predicting not scored each user, movie trailer usually can be anti-
The content of whole film is reflected, but shorter than the video playback time of whole film, therefore compared to existing to whole film
Information is extracted, and reduces workload;The visual information to movie trailer is extracted simultaneously, compared to existing film
The contextual information that proposed algorithm is used, the characteristics of visual information more can characterize film, it is possible to increase user is not to scoring
The prediction of film score value accuracy, the recommendation to user's film is more accurate.
In one embodiment of the invention, the step S2 is specifically included:Film is extracted using deep neural network pre-
The part picture in piece is accused as picture set, and extracts the visual information of each pictures in picture set.
After movie trailer has been captured from video website, each movie trailer is extracted using deep neural network
In part picture as picture set, and extract the visual information of each pictures in picture set, be follow-up analysis
Data are provided to support.Simply extract the part picture in each movie trailer to extract visual information in the present embodiment, this
Sample extracts visual information compared to all pictures in movie trailer are extracted, and extracts picture and extracts the work of visual information
Work amount can be less, improves the efficiency of whole film recommendation process.
In another embodiment of the present invention, the deep neural network describes network and text analyzing net including picture
Network.Part picture in the use deep neural network extraction movie trailer extracts picture set as picture set
In the visual information of each pictures specifically include:S21, from movie trailer extraction unit component piece as picture set,
And use picture describes network and each pictures in picture set are stated, and obtain each figure in picture set
The corresponding sentence description of piece, constitutes the sentence description collections of movie trailer;S22, using text analyzing network to preview
The sentence description collections of piece are analyzed and processed, and obtain the visual information of movie trailer.
The deep neural network being previously mentioned in above-described embodiment mainly describes network and text analyzing network including picture.This
The content information that the visual information in trailer refers to obtain based on trailer picture visual signature is extracted in embodiment.Obtain pre-
The method for accusing piece picture visual signature is the deep neural network commonly used using computer vision field, such as convolutional neural networks
(Convolutional Neural Network,CNN).Convolutional neural networks CNN as a kind of ripe picture classification network,
It has been widely used in picture processing and analysis at present, network is modified on the basis of convolutional neural networks CNN, obtained
The deep neural network of different target must be applied to.The function that the deep neural network that the present embodiment is used has is to pre-
The picture accused in piece extracts visual signature, and obtains its content information.At present, in computer vision field, picture description
(image captioning) is namely based on description of the picture visual signature generation to picture.Therefore, the hair based on this technology
Exhibition, the present embodiment describes network using picture and carries out visual information abstracting to trailer.First, the portion in movie trailer is obtained
Component piece describes network as picture set, by picture and the picture in trailer is labeled, and obtains in picture set
The description of each pictures corresponding sentence, constitute the sentence description collections of movie trailer, wherein, each pictures are to that should have
One sentence description.Finally, the sentence in sentence description collections regarding in trailer has just been obtained into by text analyzing network
Feel information.
Picture describe network be based on deep neural network, due to picture describe not only be related to picture feature extraction but also
It is related to the processing of natural language, therefore it is usually convolutional neural networks CNN and Recognition with Recurrent Neural Network that picture, which describes network,
The combination of (Recurrent Neural Network, RNN).For example, the picture that Google is proposed describes network Show and
Tell is the network that GoogleNet and Long Short Term Memory (LSTM) network integration gets up.GoogleNet is
One improved convolutional neural networks, LSTM networks are an improved Recognition with Recurrent Neural Network.Show and Tell with
Based on GoogleNet, by its it is last it is several layers of block, then get up with LSTM network connections.Such combination is to realize
Network structure end to end, so that two networks can mutually improve performance.Wherein, figure is mainly extracted in GoogleNet parts
Piece feature rather than output category result.
Sentence description collections, refer to the processing that network is described by picture, obtain the sentence description of picture in trailer, often
Pictures produce a sentence, therefore for a movie trailer with regard to a sentence description collections can be obtained.
The function that text analyzing network is realized is that, by this network, the sentence that picture can be described to network output is retouched
Set is stated, processing is further analyzed, the final visual information for obtaining trailer.Text analyzing network is one and is based on convolution
Neutral net CNN network, will can be extracted by this network in the sub- description collections of the characterising clause of trailer.This reality
It is a network being made up of multilayer convolutional neural networks to apply a model for Chinese version analysis network use, such as, by five layers
Convolutional neural networks CNN is constituted, and it is output as hiding characteristic vector.Hiding feature refers to, is extracted from text or picture
That comes is different from the feature of film protagonist or subject matter.
In one embodiment of the invention, also include before the step S21:The part of selected movie trailer has
Representational frame picture, is used as the picture set for extracting visual information.
Because the video in movie trailer is made up of many frame sequence of pictures, therefore, to movie trailer
When carrying out visual information abstracting, representative partial frame picture is chosen from frame sequence of pictures, subsequent extracted vision is used as
The picture set of information.Grabbed from video website after movie trailer, it is necessary to pre-processed to movie trailer, by
It is that vision extraction is carried out to each pictures in movie trailer, therefore, in advance when subsequently being extracted to visual signature
Need to extract the frame in movie trailer, for each film, generate one group of sequence of pictures, it is special for subsequent extracted vision
Levy.
In another embodiment of the present invention, the step S3 is specifically included:S31, according to each movie trailer
Visual information, and multiple users are to the rating matrix of multiple films, calculated and obtained using probability matrix decomposition algorithm
Each user characteristics vector sum each movie features vector, and then obtain the eigenmatrix of user and the feature square of film
Battle array;S32, according to the eigenmatrix of user and the eigenmatrix of film, calculates each user to the scoring for the film not scored
Predicted value.
Due to each user be characterized in it is different, the feature of each film be also it is different, therefore, the present embodiment
It is extracted after the visual signature of each movie trailer, by the visual information of each movie trailer and data with existing collection
In rating matrix combine, and each user characteristics vector sum each film is calculated using probability matrix decomposition algorithm
Characteristic vector.I.e. according to the characteristic vector that analyzes each user of each user to many films, and according to
The characteristic vector of the film is analyzed in scoring of many users to same film, and the eigenmatrix and electricity of user have been obtained accordingly
The eigenmatrix of shadow.
Then eigenmatrix and the eigenmatrix of film further according to user predicts each user for each not
Know the score value of film, and according to score in predicting value of the user to unknown film, film recommendation is carried out to user.It is this to be based in advance
The film for accusing piece recommends method due to considering trailer visual information, can more accurately describe the feature of film, accordingly
Hobby of the user to different films can be more accurately obtained, compared to other contextual informations, the effect of recommendation is more preferable.
In another embodiment, each user is calculated to not commenting by equation below in the step S32
The score in predicting value of the film divided:
Wherein,Predicted values of the user i to film j scoring is represented, μ is that all users are commented all films
The average value divided, biFor the deviation between the user i average values scored all films and μ, bjIt is all users to film j
Deviation between the average value and μ of scoring, UiFor user i characteristic vector, VjRepresent film j characteristic vector.
Each user according to calculating the eigenmatrix of user and the eigenmatrix of film, is being calculated to not scoring
During the score in predicting value of film, using the scoring history of each user once to multiple films, to analyze each
The characteristic vector of user itself.Likewise, for each film, the scoring according to different users to the film, to analyze
The film characteristic vector of itself.Then, according to analyze come each user characteristic vector and each film spy
Vector is levied, the eigenmatrix of user and the eigenmatrix of film have been obtained accordingly.
According to the eigenmatrix of user and the eigenmatrix of film, come commenting for the film that does not score also each user couple
Divide and be predicted.During score in predicting value is calculated, when calculating score in predicting value by above-mentioned formula, wherein that uses is every
The characteristic vector of one user and the characteristic vector of each film can be obtained from user characteristics matrix and movie features matrix
Obtain.The visual information of movie trailer is combined with the existing scoring of film, come the electricity not scored each user
Shadow carries out score in predicting, and movie trailer usually can reflect the content of whole film, but than the video playback of whole film
Time is short, therefore is extracted compared to the existing information to whole film, reduces workload.
Wherein, obtain after the visual information of trailer, it is necessary to it is combined with scoring, be that user does and recommended.Often
Method is matrix decomposition, and an advantage of matrix decomposition is to be easy to extension.For basic matrix decomposition algorithm, it is not to
The predicted value for knowing scoring is:
Wherein,Predicted values of the user i to film j scoring is represented, μ is that all users are commented all films
The average value divided, biFor the deviation between the user i average values scored all films and μ, bjIt is all users to film j
Deviation between the average value and μ of scoring, UiFor user i characteristic vector, VjRepresent film j characteristic vector.If for
Family or film have other features to add, then score in predicting value formula expression has corresponding change.It is related to add user
Exemplified by feature, its score in predicting value formula is expanded to:
Due to its good scalability, the method that the present embodiment is used is probability matrix decomposition algorithm.Wherein, based on pre-
The probability matrix decomposition algorithm for accusing piece visual information refers to, adds visual information on the basis of probability matrix decomposition algorithm
Algorithm.Assuming that have N number of user and M film in commending system, R ∈ RN×MRepresent user's rating matrix, wherein matrix element Rij
Represent scorings of the user i to film j.U∈RD×NWith V ∈ RD×MThe hiding eigenmatrix of user and film, U are represented respectivelyiAnd Vj
Represent that user and film hide characteristic vector respectively.In probability matrix decomposition model, it is known that the conditional probability distribution of scoring is determined
Justice is:
Wherein N (x | μ, σ2) it is that average is μ, variance is σ2Gaussian Profile probability density function, IijIt is an instruction
Function, its value represents that user i carried out scoring i.e. scorings of the user i to film j to film j when being 1 be known, and its value is 0
When represent that scorings of the user i to film j is unknown.
For user characteristics vector, its conditional probability distribution is:
For movie features vector, it is necessary to visual information be added, so movie features vector becomes:
Vj=εj+ηj; (5)
Wherein, VjRepresent film j global feature, εjRepresent film j essential characteristic, ηjRepresent that film j trailer is regarded
Feel feature.For feature εj, probability distribution is:
So, movie features VjProbability distribution be:
Obtain after above-mentioned probability distribution, variable optimized by maximum prior probably estimation optimization method,
Negative log-likelihood is taken to formula (9) and in the case where hyper parameter is fixed, optimization problem is with minimizing mean square error
And problem equivalent, so it is as follows to obtain object function:
Wherein,IijAn indicator variable, be worth for 1 when represent user i to film j
Carried out scoring, be worth for 0 when represent that user i did not carried out scoring to film j.The optimization method that we use is that coordinate decline is excellent
Change, U is obtained by iterationiAnd Vj, U is updated in the case where its dependent variable is fixed in iterative process each timeiOr Vj.Wherein Ui
And VjMore new formula be:
Ui←(VIiVT+λUIK)-1VRi (11)
Vj←(UIjUT+λVIK)-1(URi+λVηj) (12)
Wherein, IiIt is a diagonal matrix, its diagonal element is Iij, j=1 ..., M, RiIt is user i scoring vector, its
Element is Rij, j=1 ..., M.IjAnd RjDefinition respectively with IiAnd RiIt is similar.
The film commending system based on trailer of another embodiment of the present invention is provided referring to Fig. 2, Fig. 2, including is grabbed
Modulus block 21, extraction module 22, optimization computing module 23 and recommending module 24.
Acquisition module 21, for obtaining each user commenting to each film in movie collection in user's set
Point, obtain rating matrix of multiple users to multiple films;It is additionally operable to obtain the trailer of each film in movie collection.
Extraction module 22, the visual information for extracting each movie trailer.
Optimize computing module 23, for the visual information according to each movie trailer and the rating matrix, calculate
Each user is to the score in predicting value for the film not scored.
Recommending module 24, for, to the score in predicting value for the film not scored, electricity being recommended to user according to each user
Shadow.
Referring to Fig. 3, extraction module 22 includes the first extraction unit 221 and the second extraction unit 222.Wherein, first extract
Unit 221, obtains the part picture in movie trailer and describes network to movie trailer as picture set, and using picture
In picture be labeled, so as to obtain the corresponding sentence description of each pictures in picture set, constitute movie trailer
Sentence description collections.
Second extraction unit 222, for being divided using text analyzing network the sentence description collections of movie trailer
Analysis is handled, and extracts the visual information of movie trailer.
Referring to Fig. 4, optimization computing module 23 includes the first computing unit 231 and the second computing unit 232.Wherein, first
Computing unit 231, for the visual information according to each movie trailer, and the rating matrix, using probability matrix
Decomposition algorithm, which is calculated, obtains each user characteristics vector sum each movie features vector, and then obtains the eigenmatrix of user
With the eigenmatrix of film.
Second computing unit 232, for the eigenmatrix and the eigenmatrix of film according to user, calculates each user
To the score in predicting value for the film not scored.
Specifically, the second computing unit 232, each user is calculated to the film that does not score especially by equation below
Score in predicting value:
Wherein,Predicted values of the user i to film j scoring is represented, μ is that all users are commented all films
The average value divided, biFor the deviation between the user i average values scored all films and μ, bjIt is all users to film j
Deviation between the average value and μ of scoring, UiFor user i characteristic vector, VjRepresent film j characteristic vector.
A kind of film based on trailer that the present invention is provided recommends method and system, to the visual information of movie trailer
Extracted, and the visual information of movie trailer is combined with the existing scoring of film, not score each user
Film carry out score in predicting, movie trailer usually can reflect the content of whole film, but than the video of whole film
Reproduction time is short, therefore is extracted compared to the existing information to whole film, reduces workload;It is simultaneously pre- to film
The visual information for accusing piece is extracted, the contextual information used compared to existing film proposed algorithm, and visual information can
The characteristics of more characterizing film, it is possible to increase user is to the accuracy for the film score value prediction do not scored, to user's film
It is more accurate to recommend.
Finally, the present processes are only preferably embodiment, are not intended to limit the scope of the present invention.It is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements made etc. should be included in the protection of the present invention
Within the scope of.
Claims (10)
1. a kind of film based on trailer recommends method, it is characterised in that including:
S1, obtains scoring of each user in user's set to each film in movie collection, obtains multiple users
To the rating matrix of multiple films, and obtain the trailer of each film in movie collection;
S2, extracts the visual information of each movie trailer;
S3, according to the visual information of each movie trailer and the rating matrix, calculates each user to not scoring
The score in predicting value of film;
S4, according to each user to the score in predicting value for the film not scored, film is recommended to user.
2. the film based on trailer recommends method as claimed in claim 1, it is characterised in that the step S2 is specifically wrapped
Include:
Using the part picture in deep neural network extraction movie trailer as picture set, and extract in picture set
The visual information of each pictures.
3. the film based on trailer recommends method as claimed in claim 2, it is characterised in that the deep neural network bag
Include picture and describe network and text analyzing network, the part picture that the use deep neural network is extracted in movie trailer is made
For picture set, and extract the visual information of each pictures in picture set and specifically include:
S21, obtains the part picture in movie trailer and describes network to movie trailer as picture set, and using picture
In picture be labeled, so as to obtain the corresponding sentence description of each pictures in picture set, constitute movie trailer
Sentence description collections;
The sentence description collections of movie trailer are analyzed and processed, obtain movie trailer by S22 using text analyzing network
Visual information.
4. the film based on trailer recommends method as claimed in claim 3, it is characterised in that the picture describes network and is
The combination of convolutional neural networks and Recognition with Recurrent Neural Network, the text analyzing network is made up of multilayer convolutional neural networks.
5. the film based on trailer recommends method as claimed in claim 4, it is characterised in that before the step S21 also
Including:
The frame picture of the part of representative of selected movie trailer, is used as the picture set for extracting visual information.
6. the film based on trailer recommends method as claimed in claim 5, it is characterised in that the step S3 is specifically wrapped
Include:
S31, according to the visual information of each movie trailer, and multiple users are to the rating matrix of multiple films,
Calculated using probability matrix decomposition algorithm and obtain each user characteristics vector sum each movie features vector, and then used
The eigenmatrix at family and the eigenmatrix of film;
S32, according to the eigenmatrix of user and the eigenmatrix of film, calculates each user to the scoring for the film not scored
Predicted value.
7. the film based on trailer recommends method as claimed in claim 6, it is characterised in that pass through in the step S32
Equation below calculates each user to the score in predicting value for the film not scored:
Wherein,Predicted values of the user i to film j scoring is represented, μ is what all users were scored all films
Average value, biFor the deviation between the user i average values scored all films and μ, bjFilm j is scored for all users
Average value and μ between deviation, UiFor user i characteristic vector, VjRepresent film j characteristic vector.
8. a kind of film commending system based on trailer, it is characterised in that including:
Acquisition module, obtains scoring of each user in user's set to each film in movie collection, obtains many
Rating matrix of the individual user to multiple films;It is additionally operable to obtain the trailer of each film in movie collection;
Extraction module, the visual information for extracting each movie trailer;
Optimize computing module, for the visual information according to each movie trailer and the rating matrix, calculate each
User is to the score in predicting value for the film not scored;
Recommending module, for, to the score in predicting value for the film not scored, film being recommended to user according to each user.
9. the film commending system as claimed in claim 8 based on trailer, it is characterised in that the extraction module includes:
First extraction unit, net is described for obtaining the part picture in movie trailer as picture set, and using picture
Network is labeled to the picture in movie trailer, so that the corresponding sentence description of each pictures in picture set is obtained,
Constitute the sentence description collections of movie trailer;
Second extraction unit, for being analyzed and processed using text analyzing network to the sentence description collections of movie trailer,
Extract the visual information of movie trailer.
10. the film commending system as claimed in claim 9 based on trailer, it is characterised in that the optimization computing module
Including:
First computing unit, for the visual information according to each movie trailer, and the rating matrix, using probability
Matrix decomposition algorithm, which is calculated, obtains each user characteristics vector sum each movie features vector, and then obtains the feature of user
The eigenmatrix of matrix and film;
Second computing unit, for the eigenmatrix and the eigenmatrix of film according to user, calculates each user to not commenting
The score in predicting value of the film divided.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710408481.3A CN107291845A (en) | 2017-06-02 | 2017-06-02 | A kind of film based on trailer recommends method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710408481.3A CN107291845A (en) | 2017-06-02 | 2017-06-02 | A kind of film based on trailer recommends method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107291845A true CN107291845A (en) | 2017-10-24 |
Family
ID=60094145
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710408481.3A Pending CN107291845A (en) | 2017-06-02 | 2017-06-02 | A kind of film based on trailer recommends method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107291845A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108334638A (en) * | 2018-03-20 | 2018-07-27 | 桂林电子科技大学 | Collaborative Filtering method based on shot and long term Memory Neural Networks and interest migration |
CN108388624A (en) * | 2018-02-12 | 2018-08-10 | 科大讯飞股份有限公司 | Multimedia messages recommend method and device |
CN108490784A (en) * | 2018-04-19 | 2018-09-04 | 云南佳叶现代农业发展有限公司 | Tobacco flue-curing curve based on intensified learning recommends method |
CN108595550A (en) * | 2018-04-10 | 2018-09-28 | 南京邮电大学 | A kind of music commending system and recommendation method based on convolutional neural networks |
CN109446430A (en) * | 2018-11-29 | 2019-03-08 | 西安电子科技大学 | Method, apparatus, computer equipment and the readable storage medium storing program for executing of Products Show |
CN110457517A (en) * | 2019-08-19 | 2019-11-15 | 山东云缦智能科技有限公司 | The implementation method of the similar suggested design of program request based on picture similitude |
CN111581333A (en) * | 2020-05-07 | 2020-08-25 | 重庆大学 | Text-CNN-based audio-video play list pushing method and audio-video play list pushing system |
CN112800344A (en) * | 2021-01-29 | 2021-05-14 | 重庆邮电大学 | Deep neural network-based movie recommendation method |
CN116401458A (en) * | 2023-04-17 | 2023-07-07 | 南京工业大学 | Recommendation method based on Lorenz chaos self-adaption |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105930841A (en) * | 2016-05-13 | 2016-09-07 | 百度在线网络技术(北京)有限公司 | Method and device for automatic semantic annotation of image, and computer equipment |
CN106021485A (en) * | 2016-05-19 | 2016-10-12 | 中国传媒大学 | Multi-element attribute movie data visualization system |
CN106169083A (en) * | 2016-07-05 | 2016-11-30 | 广州市香港科大霍英东研究院 | The film of view-based access control model feature recommends method and system |
-
2017
- 2017-06-02 CN CN201710408481.3A patent/CN107291845A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105930841A (en) * | 2016-05-13 | 2016-09-07 | 百度在线网络技术(北京)有限公司 | Method and device for automatic semantic annotation of image, and computer equipment |
CN106021485A (en) * | 2016-05-19 | 2016-10-12 | 中国传媒大学 | Multi-element attribute movie data visualization system |
CN106169083A (en) * | 2016-07-05 | 2016-11-30 | 广州市香港科大霍英东研究院 | The film of view-based access control model feature recommends method and system |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108388624A (en) * | 2018-02-12 | 2018-08-10 | 科大讯飞股份有限公司 | Multimedia messages recommend method and device |
CN108334638A (en) * | 2018-03-20 | 2018-07-27 | 桂林电子科技大学 | Collaborative Filtering method based on shot and long term Memory Neural Networks and interest migration |
CN108334638B (en) * | 2018-03-20 | 2020-07-28 | 桂林电子科技大学 | Project score prediction method based on long-term and short-term memory neural network and interest migration |
CN108595550A (en) * | 2018-04-10 | 2018-09-28 | 南京邮电大学 | A kind of music commending system and recommendation method based on convolutional neural networks |
CN108490784A (en) * | 2018-04-19 | 2018-09-04 | 云南佳叶现代农业发展有限公司 | Tobacco flue-curing curve based on intensified learning recommends method |
CN108490784B (en) * | 2018-04-19 | 2024-01-12 | 云南佳叶现代农业发展有限公司 | Tobacco leaf baking curve recommendation method based on reinforcement learning |
CN109446430A (en) * | 2018-11-29 | 2019-03-08 | 西安电子科技大学 | Method, apparatus, computer equipment and the readable storage medium storing program for executing of Products Show |
CN109446430B (en) * | 2018-11-29 | 2021-10-01 | 西安电子科技大学 | Product recommendation method and device, computer equipment and readable storage medium |
CN110457517A (en) * | 2019-08-19 | 2019-11-15 | 山东云缦智能科技有限公司 | The implementation method of the similar suggested design of program request based on picture similitude |
CN111581333A (en) * | 2020-05-07 | 2020-08-25 | 重庆大学 | Text-CNN-based audio-video play list pushing method and audio-video play list pushing system |
CN111581333B (en) * | 2020-05-07 | 2023-05-26 | 重庆大学 | Text-CNN-based video and audio play list pushing method and video and audio play list pushing system |
CN112800344A (en) * | 2021-01-29 | 2021-05-14 | 重庆邮电大学 | Deep neural network-based movie recommendation method |
CN112800344B (en) * | 2021-01-29 | 2022-03-22 | 重庆邮电大学 | Deep neural network-based movie recommendation method |
CN116401458A (en) * | 2023-04-17 | 2023-07-07 | 南京工业大学 | Recommendation method based on Lorenz chaos self-adaption |
CN116401458B (en) * | 2023-04-17 | 2024-01-09 | 南京工业大学 | Recommendation method based on Lorenz chaos self-adaption |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107291845A (en) | A kind of film based on trailer recommends method and system | |
CN111050219B (en) | Method and system for processing video content using a spatio-temporal memory network | |
AU2019201787B2 (en) | Compositing aware image search | |
CN107291945B (en) | High-precision clothing image retrieval method and system based on visual attention model | |
CN111581510A (en) | Shared content processing method and device, computer equipment and storage medium | |
CN103608826B (en) | Annotated using product in the video of Web information mining | |
CN111079601A (en) | Video content description method, system and device based on multi-mode attention mechanism | |
CN111178458B (en) | Training of classification model, object classification method and device | |
CN109033140A (en) | A kind of method, apparatus, equipment and the computer storage medium of determining search result | |
WO2022199504A1 (en) | Content identification method and apparatus, computer device and storage medium | |
JP2020523699A (en) | Generate point of interest copy | |
CN113343982B (en) | Entity relation extraction method, device and equipment for multi-modal feature fusion | |
CN111626362B (en) | Image processing method, device, computer equipment and storage medium | |
JP2009282980A (en) | Method and apparatus for image learning, automatic notation, and retrieving | |
CN108419094A (en) | Method for processing video frequency, video retrieval method, device, medium and server | |
CN111259940A (en) | Target detection method based on space attention map | |
CN109829353B (en) | Face image stylizing method based on space constraint | |
Wang et al. | Aspect-ratio-preserving multi-patch image aesthetics score prediction | |
CN113343981A (en) | Visual feature enhanced character recognition method, device and equipment | |
CN114996511A (en) | Training method and device for cross-modal video retrieval model | |
CN112836515A (en) | Text analysis method, recommendation device, electronic equipment and storage medium | |
WO2024027347A9 (en) | Content recognition method and apparatus, device, storage medium, and computer program product | |
CN110990630B (en) | Video question-answering method based on graph modeling visual information and guided by using questions | |
CN111651660A (en) | Method for cross-media retrieval of difficult samples | |
CN110321565B (en) | Real-time text emotion analysis method, device and equipment based on deep learning |
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: 20171024 |
|
RJ01 | Rejection of invention patent application after publication |