CN105678340B - A kind of automatic image marking method based on enhanced stack autocoder - Google Patents
A kind of automatic image marking method based on enhanced stack autocoder Download PDFInfo
- Publication number
- CN105678340B CN105678340B CN201610035975.7A CN201610035975A CN105678340B CN 105678340 B CN105678340 B CN 105678340B CN 201610035975 A CN201610035975 A CN 201610035975A CN 105678340 B CN105678340 B CN 105678340B
- Authority
- CN
- China
- Prior art keywords
- model
- training
- stack autocoder
- stack
- image
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
Abstract
The present invention relates to a kind of automatic image marking methods based on enhanced stack autocoder: aiming at the problem that tradition SAE model in deep learning, which is difficult to effectively train, inclined data set, it proposes a kind of balance stack autocoder for promoting low frequency tags accuracy rate, preferably improves the mark effect of low frequency tags.Then aiming at the problem that single B-SAE model is unstable to cause mark effect easily to vary widely with parameter change, it is proposed that a kind of enhancing for image labeling task balances stack autocoder, by being grouped sequentially training, the optimal B-SAE submodel of weighted accumulation each group, stable annotation results are obtained.This method is by layer-by-layer pre-training weight and with Back Propagation Algorithm entirety tuning, improve that traditional shallow Model generalization ability is weak, is difficult to the problems such as converging to best extreme point, and reinforce the training of weak exemplar in the training process, improve the mark effect of entire model, this method simple and flexible has stronger practicability.
Description
Technical field
The present invention relates to pattern-recognitions and computer vision field, especially a kind of to be based on enhanced stack autocoder
Automatic image marking method.
Background technique
As multimedia image technology is fast-developing, image information is in explosive increase on internet.These digital pictures
It is very widely used, such as business, news media, medicine, education etc..Therefore, how user to be helped quickly and accurately to look for
Become one of the hot subject of multimedia research in recent years to the image of needs.And it solves the most important technology of this project and is exactly
Image retrieval and automatic image annotation technology.
Automatic image annotation, which refers to, adds several keywords to image to indicate the semantic content of image automatically.Automated graphics
Mark can use the image set marked, and the automatic relational model for learning semantic concept space and visual signature space is used in combination
The image of this model mark unknown semantics.On the one hand, automatic image annotation attempts to high-level semantics features and bottom visual signature
Between establish a bridge block, therefore, it can solve semanteme existing for most of Content-Based Image Retrieval methods to a certain degree
Divide problem, and there is preferable objectivity.On the other hand, text relevant to picture material can be generated in automatic image annotation
Word information has better accuracy.If being able to achieve automatic image annotation, existing image retrieval problem actually may be used
To be converted to more mature text retrieval problem.Therefore, automatic image annotation technology can easily be realized based on keyword
Image retrieval, meet the retrieval habit of people.Generally speaking, automatic image annotation is related to computer vision, machine learning, letter
Breath retrieval etc. it is multi-party and content, there is very strong researching value and potential business application, as image classification, image retrieval,
Image understanding and intelligent image analysis etc..
According to the main implementation feature of existing automatic image marking method, two classes can be divided into: the mark based on probability statistics
Injecting method and mask method based on machine learning.Although the method based on probability statistics can very easily expand to big data
Collection, but overall mark effect is not ideal enough.Method based on machine learning, once model training finishes, so that it may it carries out quick
Mark, and Most current classification, return etc. learning methods be shallow structure algorithm, for its generalization ability of complicated classification problem by
It is restricted to certain.In recent years, innovation algorithm of the deep learning as machine learning, is widely used in target identification, image classification, language
Sound identification and other fields, but rarely have and be applied in image labeling problem.Since deep learning can train deep layer, complexity
Model, processing big data problem on have great advantage.The two models of DBN and CNN are label is less, feature is simple, special
Better effects can be obtained by levying in complete identification mission, and image labeling problem label is numerous, characteristics of image is various and complicated,
And it greatly affected in real world images there is also noise problems such as a large amount of each class texts, network address, two dimensional code and image watermarks
The application effect of DBN and CNN.And SAE network, the more approximate expression between emphasis feature, model is easily adjusted by the defeated of complexity
Enter to be expressed as preferably to export and be applied to particular condition, therefore, this patent selects SAE model to solve the problems, such as image labeling.
Summary of the invention
The purpose of the present invention is to provide a kind of automatic image marking method based on enhanced stack autocoder, with
Overcome defect existing in the prior art, solves the problems, such as the automatic image annotation for multipair as multi-tag.
To achieve the above object, the technical scheme is that it is a kind of based on the automatic of enhanced stack autocoder
Image labeling method is realized in accordance with the following steps:
Step S1: building stack autocoder model differentiates weak label sample on the stack autocoder model
This, and noise is added to increase the frequency of training of the weak exemplar, and then constructs balance stack autocoder model；
Step S2: automatic to training image station work quantum balancing stack by the balance stack autocoder model
Encoder model, the optimal submodel of weighted accumulation each group obtain enhancing balance stack autocoder model；
Step S3: unknown images are input to the enhancing balance stack autocoder model and export annotation results.
In an embodiment of the present invention, in the step S1, further include following steps:
Step S11: encoder f is definedθWith decoder gθ'；The encoder fθInput picture x is converted into hidden layer expression
H, decoder gθ'By the hidden layer expression h be reconstructed into the consistent vector x of input picture x dimension '；Wherein, fθ(x)=σ
(Wx+b), θ={ W, b }, W are network weight, meet W'=WT, b is bias vector,To activate letter
Number；θ '={ W', b'}；
Step S12: one function of study makes to export x'=gθ'(fθ(x)) and the input picture x is approximate, and
Definition loss function is L (x, x')=(x-x')2, and learnt by minimizing loss function:
Step S13: note has L layers for the SAE model of image labeling, and is indicated with serial number l ∈ { 1 ..., L }；Use hlTable
Show l layers of output vector, WlAnd blNetwork weight and the biasing for indicating l layers, by autocoder to { Wl,bl, l ∈
{ 1 ..., L } layer-by-layer pre-training；
Step S14: feed forward process is executed and with Back Propagation Algorithm tuning；The feedforward of the stack autocoder model
Operation statement are as follows: hl+1=σ (Wl+1hl+bl+1), l ∈ { 0 ..., L-1 }；The back-propagating of the stack autocoder model
The statement of algorithm tuning are as follows:Wherein,It is multiple autocodings
The composite function of device model, and θlFor parameter { Wl,bl, l ∈ { 1 ..., L }, loss function is L (x, y)=(x-y)2；
Step S15: bound variable is defined, vector C=(c is enabled1,c2,...,cM), indicate keyword yiIn training
The number occurred in collection P, indicates the average frequency of occurrence of keyword；Vector C=(c1,c2,...,cM) indicate
I-th width image xiEach keyword Yi j, number Y that j ∈ { 1,2 ..., M } occurs in training setC,i=C*Yi；To
To in image xiThe minimum keyword of middle frequency of occurrence is
Step S16: define Φ (x) function, the stack autocoder model in the training process to training sample into
Row judgement adds noise appropriate to input picture x if input picture x includes the number of low frequency tags more than k；It is fixed
Adopted Γ (x) function increases training strength to input picture x, if the frequency of occurrence of the included label of input picture x is lower than default
Threshold value (generally takes), then increase frequency of training, wherein function gamma (x) are as follows:
Wherein, α and β is constant coefficient, and for β for determining the sample for needing to aggravate training, α, which is used to control, needs to aggravate training
The training strength of sample；
Function phi (x) are as follows:
Wherein, χ is constant coefficient, and for controlling the intensity of noise addition, d is image xiThe dimension of feature,Indicate image xi
The value of j-th of dimension, Ran () are random number functions；
Step S17: adjusting and optimizing equation is balanced stack autocoder model；It willIt is adjusted toIt willIt is adjusted toAfter model training is good, institute
State the prediction distribution D of the output as keyword of forecast image of the last layer of balance stack autocoder model.
In an embodiment of the present invention, further include following steps in the step S2:
Step S21: station work quantum balancing stack autocoder model will balance stack autocoder model
For model by the different different groups of model split of making an uproar that adds, each group interior according to different hidden neuron number division submodelsT indicates that balance stack autocoder model adds the mode of making an uproar using t kind, and k indicates k-th of sub- B-SAE mould
The hidden neuron number of type setting；
Step S22: setting initial weight simultaneously calculates quantum balancing stack autocoder model category of model error rate,
It is as follows to training data setting weight:
W=(w11,...,w1i,...,w1N),
The error in classification rate of calculating:Wherein,It indicates: assuming that image xiTrue tag collection YiComprising c keyword,
And tally set Y is obtained by model predictioni *Number be also c, if Yi=Yi *, then for
Otherwise false is true；
Step S23: calculated equilibrium stack autocoder model weight, and update training data weight；According to group
Interior all sonsThe error in classification rate of model, the minimum Model B of available group classification error rate-SAEtAnd it is corresponding
Error in classification rate et, calculate B-SAEtWeight:After the model training of t group is complete, training number is updated
According to weight, to obtain the weight of next group model, the mode for updating training data weight is as follows:
Wt+1={ wt+1,1,...,wt+1,i,...,wt+1,N},
Step S24: weighted accumulation quantum balancing stack autocoder model obtains enhancing balance stack autocoding
Device model arrives keyword prediction distribution after all having trained for all groups:
Compared to the prior art, the invention has the following advantages: one kind proposed by the invention is based on enhanced stack
The automatic image marking method of formula autocoder, using the powerful feature representation ability of SAE deep neural network, based on to certainly
Motion video mark, multi-tag are classified and the understanding of stack autocoder, proposes for image data set label imbalance, is difficult to
The automatic image marking method of the enhanced stack autocoder for the problems such as effectively training big image data, finally obtains one kind
Deep layer, complicated automatic image annotation model, especially a kind of automated graphics mark based on enhanced stack autocoder
Injecting method.This method is simple, realizes that flexibly practicability is stronger.
Detailed description of the invention
Fig. 1 is the flow chart of the automatic image marking method based on enhanced stack autocoder in the present invention.
Specific embodiment
With reference to the accompanying drawing, technical solution of the present invention is specifically described.
The invention proposes a kind of automatic image marking methods based on enhanced stack autocoder, first against depth
Tradition SAE (Stacked Auto-Encoder, SAE) model, which is difficult to effectively training, in degree study the problem of inclined data set, mentions
A kind of balance stack autocoder (Balance Stacked Auto-Encoder, B- promoting low frequency tags accuracy rate out
SAE), preferably improve the mark effect of low frequency tags.Then for single B-SAE model it is unstable (model is complicated, parameter compared with
It is more) cause to mark the problem of effect is easily varied widely with parameter change, propose a kind of increasing for image labeling task
Strong balance stack autocoder (Enhanced Balance StackedAuto-Encoder, EB-SAE), is pressed by grouping
Sequence training, the optimal B-SAE submodel of weighted accumulation each group, obtain stable annotation results.Specific step is as follows:
S1: first constructing SAE model, and weak exemplar is then differentiated on SAE model and adds the weak label sample of noise increase
This frequency of training, constructs B-SAE model with this；
S2: B-SAE model B-SAE model to training image station work, weighted accumulation each group are obtained using step S1
Optimal submodel obtains EB-SAE model, as shown in Figure 1；
S3: unknown images are input to EB-SAE model that step S2 is obtained and export annotation results.
Further, in the present embodiment, building B-SAE model is realized according to the following steps in step sl:
Step S11: encoder f is definedθWith decoder gθ', encoder fθInput picture x is converted into hidden layer table
Up to h, decoder gθ'By h be reconstructed into the consistent vector x of x dimension '.fθ(x)=σ (Wx+b), wherein θ=
{ W, b }, W are network weight, meet W'=WT, b is bias vector, For activation primitive.Wherein, θ '={ W', b'}.
Step S12: one function of study makes to export x'=gθ'(fθ(x)) and x is approximate, defines loss function
For L (x, x')=(x-x')2, then the model can be learnt by minimizing loss function:
Step S13: feed forward process is executed and with Back Propagation Algorithm tuning, it is assumed that the SAE model for image labeling has L
Layer is indicated with serial number l ∈ { 1 ..., L }.Use hlIndicate l layers of output vector (h0=x indicates input, hLIndicate output).Wl
And blIndicate l layers of network weight and biasing.According to noted earlier, { Wl,bl, l ∈ { 1 ..., L } is successively instructed in advance using AE
Practice.The feedforward operation of SAE can state are as follows: hl+1=σ (Wl+1hl+bl+1), l ∈ { 0 ..., L-1 }, entire model use after to biography
Broadcast algorithm tuning:Wherein,It is the synthesis of multiple AE models
Function, and θlFor parameter { Wl,bl, l ∈ { 1 ..., L }, loss function is defined as L (x, y)=(x-y)2。
Step S14: bound variable is defined, vector C=(c is enabled1,c2,...,cM),Indicate keyword yiIn training set
The number occurred in P,Indicate the average frequency of occurrence of keyword.In this way, our available vector,
Indicate the i-th width image xiEach keyword Yi j, number Y that j ∈ { 1,2 ..., M } occurs in training setC,i=C*Yi(* table
Show that two vector corresponding points are multiplied to obtain a new vector).To obtain in image xiThe minimum keyword of middle frequency of occurrence is
Step S15: defining Φ (x) function, model allowed to judge in the training process training sample, if sample x,
That is input picture x, the number comprising low frequency tags then add noise appropriate to the sample more than k.Γ (x) function is defined,
Training strength is increased to sample x, if the frequency of occurrence of the included label of the sample is lower than certain threshold value, increases its training time
Number, in the present embodiment, which generally takes
Wherein, α and β is constant coefficient, and for determining which sample needs to aggravate training, α needs to aggravate training β for controlling
Sample training strength.
Wherein, χ is constant coefficient, and for controlling the intensity of noise addition, d is image xiThe dimension of feature,Indicate image
xiThe value of j-th of dimension, Ran () are random number functions, for example, the desirable random letter for obeying (0,1) Gaussian Profile of Ran ()
The equally distributed random function that several or value is 0 to 1.
Step S16: adjusting and optimizing equation obtains B-SAE model, is adjusted to It is adjusted toAfter model training is good, the output of the last layer of B-SAE is prognostic chart
The prediction distribution D of the keyword of picture.
Further, in the present embodiment, training EB-SAE model is realized in step S2 according to the following steps:
Step S21: the sub- B-SAE model of station work, B-SAE model add the different group of model split of making an uproar by different, often
Submodel is divided according to different hidden neuron numbers in one groupT indicates that Model B-SAE adds the side of making an uproar using t kind
Formula, k indicate the hidden neuron number of k-th of sub- B-SAE model setting.
Step S22: setting initial weight simultaneously calculates sub- B-SAE category of model error rate, such as to training data setting weight
Under:
W=(w11,...,w1i,...,w1N),
In this way, error in classification rate can calculate in this way:Its
In,The meaning of expression is, it is assumed that image xiTrue tag collection YiInclude
C keyword, and tally set Y is obtained by model predictioni *Number be also c, if Yi=Yi *, then it is false, is otherwise true.
Step S23: calculating B-SAE Model Weight and updates training data weight, according to all sons in organizingModel
Error in classification rate, the minimum Model B of available group classification error rate-SAEtWith corresponding error in classification rate et, B-SAEt
Weight can calculate in this way:After the model training of t group is complete, need to update the weight of training data,
Preferably to obtain the weight of next group model, the mode for updating training data weight is as follows:
Wt+1={ wt+1,1,...,wt+1,i,...,wt+1,N},
Step S24: the sub- B-SAE model of weighted accumulation obtains EB-SAE model, after all having trained for all groups, so that it may
To keyword prediction distribution:
The above are preferred embodiments of the present invention, all any changes made according to the technical solution of the present invention, and generated function is made
When with range without departing from technical solution of the present invention, all belong to the scope of protection of the present invention.
Claims (1)
1. a kind of automatic image marking method based on enhanced stack autocoder, which is characterized in that in accordance with the following steps
It realizes:
Step S1: building stack autocoder model differentiates weak exemplar on the stack autocoder model, and
Addition noise constructs balance stack autocoder model to increase the frequency of training of the weak exemplar；
Step S2: by the balance stack autocoder model to training image station work quantum balancing stack autocoding
Device model, the optimal submodel of weighted accumulation each group obtain enhancing balance stack autocoder model；
Step S3: unknown images are input to the enhancing balance stack autocoder model and export annotation results；
In the step S1, further include following steps:
Step S11: encoder f is definedθWith decoder gθ'；The encoder fθInput picture x is converted into hidden layer expression h, decoding
Device gθ'By the hidden layer expression h be reconstructed into the consistent vector x of input picture x dimension '；Wherein, fθ(x)=σ (Wx+
B), θ={ W, b }, W are network weight, meet W'=WT, b is bias vector,For activation primitive；θ '={ W', b'}；
Step S12: one function of study makes to export x'=gθ'(fθ(x)) and the input picture x is approximate, and defines loss function
For L (x, x')=(x-x')2, and learnt by minimizing loss function:
Step S13: note has L layers for the SAE model of image labeling, and is indicated with serial number l ∈ { 1 ..., L }；Use hlIndicate l
The output vector of layer, WlAnd blNetwork weight and the biasing for indicating l layers, by autocoder to { Wl,bl, l ∈
{ 1 ..., L } layer-by-layer pre-training；
Step S14: feed forward process is executed and with Back Propagation Algorithm tuning；The feedforward operation of the stack autocoder model
Statement are as follows: hl+1=σ (Wl+1hl+bl+1), l ∈ { 0 ..., L-1 }；The Back Propagation Algorithm of the stack autocoder model
Tuning statement are as follows:Wherein,It is multiple autocoder moulds
The composite function of type, and θlFor parameter { Wl,bl, l ∈ { 1 ..., L }, loss function is L (x, y)=(x-y)2；
Step S15: bound variable is defined, vector C=(c is enabled1,c2,...,cM),Indicate keyword yiIn training set P
The number of appearance,Indicate the average frequency of occurrence of keyword；Vector C=(c1,c2,...,cM) indicate the i-th width
Image xiEach keyword Yi j, number Y that j ∈ { 1,2 ..., M } occurs in training setC,i=C*Yi；To obtain scheming
As xiThe minimum keyword of middle frequency of occurrence is
Step S16: defining Φ (x) function, and the stack autocoder model in the training process sentences training sample
It is disconnected, if input picture x includes the number of low frequency tags more than k, noise appropriate is added to input picture x；Define Γ
(x) function increases training strength to input picture x, if the frequency of occurrence of the included label of input picture x is lower than default threshold
Value, then increase frequency of training, wherein function gamma (x) are as follows:
Wherein, α and β is constant coefficient, and β is for determining that the sample for needing to aggravate training, α are used to control the sample for needing to aggravate training
Training strength；
Function phi (x) are as follows:
Wherein, χ is constant coefficient, and for controlling the intensity of noise addition, d is image xiThe dimension of feature,Indicate image xiJth
The value of a dimension, Ran () are random number functions；
Step S17: adjusting and optimizing equation is balanced stack autocoder model；It willIt is adjusted toIt willIt is adjusted toIt is described after model training is good
Balance the prediction distribution D of the output as keyword of forecast image of the last layer of stack autocoder model；
Further include following steps in the step S2:
Step S21: station work quantum balancing stack autocoder model will balance stack autocoder model
By the different different groups of model split of making an uproar that adds, submodel is divided according to different hidden neuron numbers in each groupT indicates that balance stack autocoder model adds the mode of making an uproar using t kind, and k indicates k-th of sub- B-SAE mould
The hidden neuron number of type setting；
Step S22: setting initial weight simultaneously calculates quantum balancing stack autocoder model category of model error rate, to instruction
It is as follows to practice data setting weight:
It calculatesError in classification rate:Wherein, It indicates: assuming that image xiTrue tag collection YiComprising c keyword, and pass through modelPrediction
Obtain tally set Yi *Number be also c, if Yi=Yi *, thenIt is otherwise true for false；
Step S23: calculated equilibrium stack autocoder model weight, and update training data weight；According to institute in group
There is sonThe error in classification rate of model, the minimum Model B-SAE of available group classification error ratetAnd it is corresponding
Error in classification rate et, calculate B-SAEtWeight:After the model training of t group is complete, training data is updated
Weight, to obtain the weight of next group model, the mode for updating training data weight is as follows:
Step S24: weighted accumulation quantum balancing stack autocoder model obtains enhancing balance stack autocoder mould
Type arrives keyword prediction distribution after all having trained for all groups:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610035975.7A CN105678340B (en) | 2016-01-20 | 2016-01-20 | A kind of automatic image marking method based on enhanced stack autocoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610035975.7A CN105678340B (en) | 2016-01-20 | 2016-01-20 | A kind of automatic image marking method based on enhanced stack autocoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105678340A CN105678340A (en) | 2016-06-15 |
| CN105678340B true CN105678340B (en) | 2018-12-25 |
Family
ID=56301673
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610035975.7A Active CN105678340B (en) | 2016-01-20 | 2016-01-20 | A kind of automatic image marking method based on enhanced stack autocoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105678340B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250915B (en) * | 2016-07-22 | 2019-08-09 | 福州大学 | A kind of automatic image marking method of fusion depth characteristic and semantic neighborhood |
| CN109271539B (en) * | 2018-08-31 | 2020-11-24 | 华中科技大学 | Image automatic labeling method and device based on deep learning |
| CN114035098A (en) * | 2021-12-14 | 2022-02-11 | 北京航空航天大学 | Lithium battery health state prediction method integrating future working condition information and historical state information |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
| CN104166859A (en) * | 2014-08-13 | 2014-11-26 | 西安电子科技大学 | Polarization SAR image classification based on SSAE and FSALS-SVM |
| CN104679863A (en) * | 2015-02-28 | 2015-06-03 | 武汉烽火众智数字技术有限责任公司 | Method and system for searching images by images based on deep learning |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AUPQ684400A0 (en) * | 2000-04-11 | 2000-05-11 | Telstra R & D Management Pty Ltd | A gradient based training method for a support vector machine |
-
2016
- 2016-01-20 CN CN201610035975.7A patent/CN105678340B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104166859A (en) * | 2014-08-13 | 2014-11-26 | 西安电子科技大学 | Polarization SAR image classification based on SSAE and FSALS-SVM |
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
| CN104679863A (en) * | 2015-02-28 | 2015-06-03 | 武汉烽火众智数字技术有限责任公司 | Method and system for searching images by images based on deep learning |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105678340A (en) | 2016-06-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109447140B (en) | Image identification and cognition recommendation method based on neural network deep learning | |
| CN109508375A (en) | A kind of social affective classification method based on multi-modal fusion | |
| CN112749608B (en) | Video auditing method, device, computer equipment and storage medium | |
| CN107657008B (en) | Cross-media training and retrieval method based on deep discrimination ranking learning | |
| CN109614842A (en) | The machine learning of candidate video insertion object type for identification | |
| CN110110323B (en) | Text emotion classification method and device and computer readable storage medium | |
| CN106960206A (en) | Character identifying method and character recognition system | |
| CN109684912A (en) | A kind of video presentation method and system based on information loss function | |
| CN109933664A (en) | A kind of fine granularity mood analysis improved method based on emotion word insertion | |
| CN110704601A (en) | Method for solving video question-answering task requiring common knowledge by using problem-knowledge guided progressive space-time attention network | |
| CN104572631B (en) | The training method and system of a kind of language model | |
| CN111460157B (en) | Cyclic convolution multitask learning method for multi-field text classification | |
| CN108228758A (en) | A kind of file classification method and device | |
| CN111488524B (en) | Attention-oriented semantic-sensitive label recommendation method | |
| CN105701516B (en) | A kind of automatic image marking method differentiated based on attribute | |
| CN110825850B (en) | Natural language theme classification method and device | |
| CN109598586A (en) | A kind of recommended method based on attention model | |
| CN109858505A (en) | Classifying identification method, device and equipment | |
| CN105678340B (en) | A kind of automatic image marking method based on enhanced stack autocoder | |
| CN109740151A (en) | Public security notes name entity recognition method based on iteration expansion convolutional neural networks | |
| CN105701225A (en) | Cross-media search method based on unification association supergraph protocol | |
| CN109271513A (en) | A kind of file classification method, computer-readable storage media and system | |
| CN104077408B (en) | Extensive across media data distributed semi content of supervision method for identifying and classifying and device | |
| CN111985520A (en) | Multi-mode classification method based on graph convolution neural network | |
| CN113297387B (en) | News detection method for image-text mismatching based on NKD-GNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |