CN109033321A - It is a kind of that image is with natural language feature extraction and the language based on keyword indicates image partition method - Google Patents
It is a kind of that image is with natural language feature extraction and the language based on keyword indicates image partition method Download PDFInfo
- Publication number
- CN109033321A CN109033321A CN201810790480.4A CN201810790480A CN109033321A CN 109033321 A CN109033321 A CN 109033321A CN 201810790480 A CN201810790480 A CN 201810790480A CN 109033321 A CN109033321 A CN 109033321A
- Authority
- CN
- China
- Prior art keywords
- feature
- image
- keyword
- word
- language
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The present invention provides a kind of image and natural language feature extraction and the language based on keyword indicates image partition method, on the basis of image characteristics extraction and natural language feature extraction, to the natural language of input picture and input, the keyword according to included in natural language, by the feature f of keyword correspondence image region ii, keyword weighting after sentence feature qiAnd the corresponding vision contextual feature c based on keywordiThree features are cascaded altogether；Feature input multi-layer perception (MLP) after cascade is classified, and segmentation result is obtained.Compared with prior art, the feature extraction of image and natural language can be easy to implement the language instruction image partition method based on keyword；Language of the invention indicates image partition method, reduces the processing difficulty to long sentence, improves the accuracy of object positioning and identification, and then improves language instruction image segmentation precision.
Description
Technical field
The present invention relates to a kind of image and natural language feature extraction and the language based on keyword indicates image segmentation side
Method is related to image procossing, computer vision, image segmentation, language and image Combined Treatment and leads.
Background technique
With the arrival of big data era, different types of mass data circulates in network, by different types of data phase
In conjunction with being the new demand of big data era.Wherein, image procossing has received widespread attention in conjunction with natural language.Language instruction
Image segmentation refers to, is partitioned into the object of natural language description in image, is the committed step in language and image Combined Treatment.
The technology for solving language instruction image segmentation problem at present mainly utilizes deep neural network to extract nature respectively
Language and characteristics of image, then by natural language and feature combinations, image is split as new feature.It specifically can be with
It is divided into two classes: the language instruction image partition method based on sentence and the language instruction image partition method based on word.It is based on
The language instruction image partition method of sentence extracts the feature of entire sentence, with feature combinations；Language based on word refers to
Show that image partition method extracts the feature of each word, by its respectively with feature combinations.These methods are primarily present two
Defect:
1, the difference of importance between word is had ignored, each word of equality processing causes to be difficult to handle to long sentence；
2, the context relations such as appearance, the position inside image between different zones are not accounted for, and these vision contexts close
It is often most important for finding the object of natural language description in the picture.
Summary of the invention
The present invention provides a kind of images and natural language feature extracting method, and having can be easy to implement based on keyword
Language indicate image partition method the characteristics of.
The present invention also provides a kind of, and the language based on keyword indicates image partition method, and having reduces to long sentence
Processing difficulty improves the characteristics of object is positioned at the accuracy of identification.
A kind of image for there is provided according to the present invention and natural language feature extracting method, including image characteristic extracting method and
Natural language feature extracting method；Wherein,
Image characteristic extracting method includes, and for input picture, extracts characteristics of image F using depth convolutional neural networks；Institute
Stating characteristics of image is a two dimensional character figure, the inside each feature vector fiEncode the feature of corresponding region i in image；Root
The location information of object is needed according to natural language feature instruction image segmentation task；
Natural language feature extracting method includes, for the natural language of input, by each word be encoded to one-hot feature to
After amount, dimensionality reduction is carried out with word insertion；Word after dimensionality reduction presses the sequence in former sentence, sequentially inputs Recognition with Recurrent Neural Network；For
T-th of word in sentence, Recognition with Recurrent Neural Network learn the feature q to wordt；The feature q of the wordtEncode word t itself
Contextual information with entire sentence of semantic information and word t itself；The feature vector of various words constitutes a matrix
Q indicates the feature of entire sentence.
The method of the location information that object is needed according to natural language feature instruction image segmentation task includes extracting
The relative position coordinates of each image-region cascade with feature F, obtain the final visual signature V of each image-region.
A kind of language instruction image partition method based on keyword provided according to the present invention, based on above-mentioned image and certainly
Right language feature extracting method realizes, specific method includes,
To the natural language of input picture and input, the keyword according to included in natural language, by keyword correspondence image
The feature f of region ii, keyword weighting after sentence feature qiAnd the corresponding vision contextual feature c based on keywordiAltogether
Three features are cascaded；Feature input multi-layer perception (MLP) after cascade is classified, and segmentation result is obtained；
The multi-layer perception (MLP) is made of two layers of neural network, and first layer includes ReLU activation primitive, and the second layer includes sigmoid
Activation primitive；
Wherein, the acquisition methods of image-region i corresponding to keyword include, for key corresponding to each image-region i
Word is trained extraction, and the process that training is extracted includes,
For the feature of obtained each word, keyword is extracted using language attention model；The language attention model
It is made of two layers of neural network, first layer includes tanh activation primitive, and the second layer does not have activation primitive；For each image-region
I first cascades the feature of each word t and the image-region, and then input language attention model, the power that gains attention are beaten
Point；To the attention marking be normalized, the marking value after normalization between 0 to 1, closer to 1 illustrate word t for
Image-region i is more crucial；Conversely, illustrating that word t is more inessential for image-region i closer to 0；
It is given a mark with attention and corrects the feature of sentence, improved influence power of the keyword in sentence, reduce the influence of non-key word
Power；Attention after normalization is given a mark and corresponding word feature qtIt is multiplied, word feature is weighted；Then all words are added
Feature after power is added, and generates entire sentence for the sentence feature q of image-region ii；
It sets a keyword screening threshold value and illustrates image district if the attention marking after normalization is greater than the threshold value
Domain i thinks that word t is keyword；
For each word t, all image-regions for thinking that it is keyword are found out, the context relation in these regions is learnt；
The feature in these regions is averaged first, for integrating area information；Then with a full articulamentum based on the spy after being averaged
Sign study contextual feature gt ；
Learn vision contextual feature g corresponding to each keywordtAfterwards, be integrated into entire sentence it is corresponding visually under
Literary feature；For image-region i, the vision contextual feature g of corresponding each keywordtIt is added, generates entire sentence pair
The vision contextual feature c answeredi。
The method also includes being normalized attention marking using softmax.
The method also includes it is 0.05 that setting keyword, which screens threshold value,.
Compared with prior art, the feature extraction of image and natural language can be easy to implement the language based on keyword
Indicate image partition method；Language of the invention indicates image partition method, reduces the processing difficulty to long sentence, improves
The accuracy of object positioning and identification, and then improve language instruction image segmentation precision.
Detailed description of the invention
Fig. 1 is the schematic illustration of a wherein embodiment of the invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention.
Any feature disclosed in this specification (including abstract and attached drawing) unless specifically stated can be equivalent by other
Or the alternative features with similar purpose are replaced.That is, unless specifically stated, each feature is a series of equivalent or class
Like an example in feature.
A kind of image and natural language feature extracting method, including image characteristic extracting method and natural language feature extraction
Method；Wherein,
Image characteristic extracting method includes, and for input picture, extracts characteristics of image using depth convolutional neural networks (CNN)
F；Described image is characterized in a two dimensional character figure, the inside each feature vector fiEncode the spy of corresponding region i in image
Sign；Indicate that image segmentation task needs the location information of object according to natural language feature；
Natural language feature extracting method includes, for the natural language of input, by each word be encoded to one-hot feature to
After amount, dimensionality reduction is carried out with word insertion (word embedding)；Word after dimensionality reduction presses the sequence in former sentence, sequentially inputs and follows
Ring neural network (RNN)；For t-th of word in sentence, Recognition with Recurrent Neural Network learns the feature q to wordt；The word
Feature qtEncode contextual information of semantic information and the word t itself of word t with entire sentence itself；Various words
Feature vector constitute a matrix Q, indicate the feature of entire sentence.
Based on above-mentioned image characteristic extracting method and natural language feature extracting method, it is easy to implement the language based on keyword
Speech instruction image partition method.
It is described to indicate that image segmentation task needs object according to natural language feature as one embodiment of the present invention
The method of location information include extracting the relative position coordinates of each image-region, cascaded with feature F, obtain each image
The final visual signature V in region.
As shown in Figure 1, a kind of language based on keyword indicates image partition method, it is based on above-mentioned image and natural language
Feature extracting method realizes that specific method includes,
To the natural language of input picture and input, the keyword according to included in natural language, by keyword correspondence image
The feature f of region ii, keyword weighting after sentence feature qiAnd the corresponding vision contextual feature c based on keywordiAltogether
Three features are cascaded；Feature input multi-layer perception (MLP) (MLP) after cascade is classified, and segmentation result is obtained；
The multi-layer perception (MLP) is made of two layers of neural network, and first layer includes ReLU activation primitive, and the second layer includes sigmoid
Activation primitive；
Wherein, the acquisition methods of image-region i corresponding to keyword include, for key corresponding to each image-region i
Word is trained extraction, and the process that training is extracted includes,
For the feature of obtained each word, keyword is extracted using language attention model；The language attention model
It is made of two layers of neural network, first layer includes tanh activation primitive, and the second layer does not have activation primitive；For each image-region
I first cascades the feature of each word t and the image-region, and then input language attention model, the power that gains attention are beaten
Point；To the attention marking be normalized, the marking value after normalization between 0 to 1, closer to 1 illustrate word t for
Image-region i is more crucial；Conversely, illustrating that word t is more inessential for image-region i closer to 0；
It is given a mark with attention and corrects the feature of sentence, improved influence power of the keyword in sentence, reduce the influence of non-key word
Power；Attention after normalization is given a mark and corresponding word feature qtIt is multiplied, word feature is weighted；Then all words are added
Feature after power is added, and generates entire sentence for the sentence feature q of image-region ii；
It sets a keyword screening threshold value and illustrates image district if the attention marking after normalization is greater than the threshold value
Domain i thinks that word t is keyword；
For each word t, all image-regions for thinking that it is keyword are found out, the context relation in these regions is learnt；
The feature in these regions is averaged first, for integrating area information；Then with a full articulamentum based on the spy after being averaged
Sign study contextual feature gt ；
Learn vision contextual feature g corresponding to each keywordtAfterwards, be integrated into entire sentence it is corresponding visually under
Literary feature；For image-region i, the vision contextual feature g of corresponding each keywordtIt is added, generates entire sentence pair
The vision contextual feature c answeredi。
Prior art one side equality handles each of sentence word, causes to be difficult to handle to long sentence；Another party
Face does not account for the context relations such as appearance, the position inside image between different zones, these context relations are for scheming
It is positioned as in and the object of identification natural language description is most important.The present invention proposes the language instruction image point based on keyword
Algorithm is cut, the keyword in natural language is extracted, to reduce the processing difficulty to long sentence.And learn based on keyword
Vision context relation improves the accuracy of object positioning and identification, and then improves language and indicate image segmentation precision.
As one embodiment of the present invention, the method also includes returning attention marking using softmax
One change processing.
As one embodiment of the present invention, the method also includes it is 0.05 that setting keyword, which screens threshold value,.
It is described in more detail below with a specific embodiment.
It determines database, determines that language indicates image segmentation database, such as Google Referit database.
Data prediction pre-processes database, extracts original image, natural language, segmentation result.It is wherein original
Image need to extract the relative position coordinates of each point.Natural language need to convert one-hot vector for word each in sentence.
Build depth network model.Wherein convolutional neural networks (CNN) select DeepLab101, export 60 × 60 figures
As region, each provincial characteristics fi is set as 1000 dimensions.Recognition with Recurrent Neural Network (RNN) selects long memory unit (LSTM) in short-term,
Every sentence maximum word number is set as 20, and each word feature qt is set as 1000 dimensions.
Determine keyword threshold value.Keyword threshold value Thr is set as 0.05.
Model initialization, the model initialization parameter of convolutional neural networks (CNN) pre-training on ImageNet.Model its
Remaining part divides random initializtion.
Learning rate is set and gradient decline strategy, convolutional neural networks (CNN), are based on keyword at language attention model
Vision context relational learning model, multi-layer perception (MLP) (MLP) learning rate be set as 0.0001, Recognition with Recurrent Neural Network
(RNN) learning rate is set as 0.001.Optimal way is using ADAM gradient decline strategy.
Training pattern, model buildings and initialization finish, and after determining that study and gradient decline are tactful, are trained.It will count
Model, 5 epoch of training are sequentially input according to the data of training set in library.
Test model, after model training, the image and sentence of test set in input database obtain language instruction figure
As segmentation result.
Claims (5)
1. a kind of image and natural language feature extracting method, including image characteristic extracting method and natural language feature extraction side
Method；Wherein,
Image characteristic extracting method includes, and for input picture, extracts characteristics of image F using depth convolutional neural networks；Institute
Stating characteristics of image is a two dimensional character figure, the inside each feature vector fiEncode the feature of corresponding region i in image；Root
The location information of object is needed according to natural language feature instruction image segmentation task；
Natural language feature extracting method includes, for the natural language of input, by each word be encoded to one-hot feature to
After amount, dimensionality reduction is carried out with word insertion；Word after dimensionality reduction presses the sequence in former sentence, sequentially inputs Recognition with Recurrent Neural Network；For
T-th of word in sentence, Recognition with Recurrent Neural Network learn the feature q to wordt；The feature q of the wordtEncode word t itself
Contextual information with entire sentence of semantic information and word t itself；The feature vector of various words constitutes a matrix
Q indicates the feature of entire sentence.
2. image according to claim 1 and natural language feature extracting method, described to be indicated according to natural language feature
It includes extracting the relative position coordinates of each image-region that image segmentation task, which needs the method for the location information of object, with spy
F cascade is levied, the final visual signature V of each image-region is obtained.
3. a kind of language based on keyword indicates image partition method, based on image of any of claims 1 or 2 and natural language
Say that feature extracting method realizes that specific method includes,
To the natural language of input picture and input, the keyword according to included in natural language, by keyword correspondence image
The feature f of region ii, keyword weighting after sentence feature qiAnd the corresponding vision contextual feature c based on keywordiAltogether
Three features are cascaded；Feature input multi-layer perception (MLP) after cascade is classified, and segmentation result is obtained；
The multi-layer perception (MLP) is made of two layers of neural network, and first layer includes ReLU activation primitive, and the second layer includes sigmoid
Activation primitive；
Wherein, the acquisition methods of image-region i corresponding to keyword include, for key corresponding to each image-region i
Word is trained extraction, and the process that training is extracted includes,
For the feature of obtained each word, keyword is extracted using language attention model；The language attention model
It is made of two layers of neural network, first layer includes tanh activation primitive, and the second layer does not have activation primitive；For each image-region
I first cascades the feature of each word t and the image-region, and then input language attention model, the power that gains attention are beaten
Point；To the attention marking be normalized, the marking value after normalization between 0 to 1, closer to 1 illustrate word t for
Image-region i is more crucial；Conversely, illustrating that word t is more inessential for image-region i closer to 0；
It is given a mark with attention and corrects the feature of sentence, improved influence power of the keyword in sentence, reduce the influence of non-key word
Power；Attention after normalization is given a mark and corresponding word feature qtIt is multiplied, word feature is weighted；Then all words are added
Feature after power is added, and generates entire sentence for the sentence feature q of image-region ii；
It sets a keyword screening threshold value and illustrates image district if the attention marking after normalization is greater than the threshold value
Domain i thinks that word t is keyword；
For each word t, all image-regions for thinking that it is keyword are found out, the context relation in these regions is learnt；
The feature in these regions is averaged first, for integrating area information；Then with a full articulamentum based on the spy after being averaged
Sign study contextual feature gt；
Learn vision contextual feature g corresponding to each keywordtAfterwards, be integrated into entire sentence it is corresponding visually under
Literary feature；For image-region i, the vision contextual feature g of corresponding each keywordtIt is added, generates entire sentence pair
The vision contextual feature c answeredi。
4. language according to claim 3 indicates image partition method, the method also includes using softmax to note
Meaning power marking is normalized.
5. language according to claim 3 or 4 indicates image partition method, the method also includes setting keyword is sieved
Selecting threshold value is 0.05.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810790480.4A CN109033321B (en) | 2018-07-18 | 2018-07-18 | Image and natural language feature extraction and keyword-based language indication image segmentation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810790480.4A CN109033321B (en) | 2018-07-18 | 2018-07-18 | Image and natural language feature extraction and keyword-based language indication image segmentation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109033321A true CN109033321A (en) | 2018-12-18 |
| CN109033321B CN109033321B (en) | 2021-12-17 |
Family
ID=64643921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810790480.4A Active CN109033321B (en) | 2018-07-18 | 2018-07-18 | Image and natural language feature extraction and keyword-based language indication image segmentation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109033321B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711463A (en) * | 2018-12-25 | 2019-05-03 | 广东顺德西安交通大学研究院 | Important object detection method based on attention |
| CN111598155A (en) * | 2020-05-13 | 2020-08-28 | 北京工业大学 | Fine-grained image weak supervision target positioning method based on deep learning |
| CN112037239A (en) * | 2020-08-28 | 2020-12-04 | 大连理工大学 | Text guidance image segmentation method based on multi-level explicit relation selection |
| CN114299348A (en) * | 2022-02-21 | 2022-04-08 | 山东力聚机器人科技股份有限公司 | Image classification method and device based on restoration self-supervision task |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103559193A (en) * | 2013-09-10 | 2014-02-05 | 浙江大学 | Topic modeling method based on selected cell |
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
| CN106599198A (en) * | 2016-12-14 | 2017-04-26 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Image description method for multi-stage connection recurrent neural network |
| CN106778835A (en) * | 2016-11-29 | 2017-05-31 | 武汉大学 | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic |
| CN107391709A (en) * | 2017-07-28 | 2017-11-24 | 深圳市唯特视科技有限公司 | A kind of method that image captions generation is carried out based on new attention model |
| CN107608943A (en) * | 2017-09-08 | 2018-01-19 | 中国石油大学(华东) | Merge visual attention and the image method for generating captions and system of semantic notice |
| CN107688821A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | View-based access control model conspicuousness and across the modality images natural language description methods of semantic attribute |
| US9939272B1 (en) * | 2017-01-06 | 2018-04-10 | TCL Research America Inc. | Method and system for building personalized knowledge base of semantic image segmentation via a selective random field approach |
| CN107909115A (en) * | 2017-12-04 | 2018-04-13 | 上海师范大学 | A kind of image Chinese subtitle generation method |
| CN108009154A (en) * | 2017-12-20 | 2018-05-08 | 哈尔滨理工大学 | A kind of image Chinese description method based on deep learning model |
| CN108228686A (en) * | 2017-06-15 | 2018-06-29 | 北京市商汤科技开发有限公司 | It is used to implement the matched method, apparatus of picture and text and electronic equipment |
-
2018
- 2018-07-18 CN CN201810790480.4A patent/CN109033321B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103559193A (en) * | 2013-09-10 | 2014-02-05 | 浙江大学 | Topic modeling method based on selected cell |
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
| CN106778835A (en) * | 2016-11-29 | 2017-05-31 | 武汉大学 | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic |
| CN106599198A (en) * | 2016-12-14 | 2017-04-26 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Image description method for multi-stage connection recurrent neural network |
| US9939272B1 (en) * | 2017-01-06 | 2018-04-10 | TCL Research America Inc. | Method and system for building personalized knowledge base of semantic image segmentation via a selective random field approach |
| CN108228686A (en) * | 2017-06-15 | 2018-06-29 | 北京市商汤科技开发有限公司 | It is used to implement the matched method, apparatus of picture and text and electronic equipment |
| CN107688821A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | View-based access control model conspicuousness and across the modality images natural language description methods of semantic attribute |
| CN107391709A (en) * | 2017-07-28 | 2017-11-24 | 深圳市唯特视科技有限公司 | A kind of method that image captions generation is carried out based on new attention model |
| CN107608943A (en) * | 2017-09-08 | 2018-01-19 | 中国石油大学(华东) | Merge visual attention and the image method for generating captions and system of semantic notice |
| CN107909115A (en) * | 2017-12-04 | 2018-04-13 | 上海师范大学 | A kind of image Chinese subtitle generation method |
| CN108009154A (en) * | 2017-12-20 | 2018-05-08 | 哈尔滨理工大学 | A kind of image Chinese description method based on deep learning model |
Non-Patent Citations (3)
| Title |
|---|
| LIANG-CHIEH CHEN ET AL.: "Attention to Scale: Scale-aware Semantic Image Segmentation", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| RONGHANG HU ET AL.: "Segmentation from Natural Language", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
| 李志欣 等: "图像检索中语义映射方法综述", 《计算机辅助设计与图形学学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711463A (en) * | 2018-12-25 | 2019-05-03 | 广东顺德西安交通大学研究院 | Important object detection method based on attention |
| CN109711463B (en) * | 2018-12-25 | 2023-04-07 | 广东顺德西安交通大学研究院 | Attention-based important object detection method |
| CN111598155A (en) * | 2020-05-13 | 2020-08-28 | 北京工业大学 | Fine-grained image weak supervision target positioning method based on deep learning |
| CN112037239A (en) * | 2020-08-28 | 2020-12-04 | 大连理工大学 | Text guidance image segmentation method based on multi-level explicit relation selection |
| CN114299348A (en) * | 2022-02-21 | 2022-04-08 | 山东力聚机器人科技股份有限公司 | Image classification method and device based on restoration self-supervision task |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109033321B (en) | 2021-12-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
| CN109033321A (en) | It is a kind of that image is with natural language feature extraction and the language based on keyword indicates image partition method | |
| CN111177446B (en) | Method for searching footprint image | |
| CN111126069B (en) | Social media short text named entity identification method based on visual object guidance | |
| CN107273864B (en) | Face detection method based on deep learning | |
| CN106960206A (en) | Character identifying method and character recognition system | |
| RU2707147C1 (en) | Neural network training by means of specialized loss functions | |
| CN112733866A (en) | Network construction method for improving text description correctness of controllable image | |
| CN107491729B (en) | Handwritten digit recognition method based on cosine similarity activated convolutional neural network | |
| CN110781897A (en) | Semantic edge detection method based on deep learning | |
| CN112560710B (en) | Method for constructing finger vein recognition system and finger vein recognition system | |
| CN109885796A (en) | A kind of Internet news figure matching detection method based on deep learning | |
| Yu et al. | Exemplar-based recursive instance segmentation with application to plant image analysis | |
| CN111126155B (en) | Pedestrian re-identification method for generating countermeasure network based on semantic constraint | |
| CN109508640A (en) | A kind of crowd's sentiment analysis method, apparatus and storage medium | |
| CN117149944A (en) | Multi-mode situation emotion recognition method and system based on wide time range | |
| CN115170403A (en) | Font repairing method and system based on deep meta learning and generation countermeasure network | |
| CN110503090A (en) | Character machining network training method, character detection method and character machining device based on limited attention model | |
| CN113449776A (en) | Chinese herbal medicine identification method and device based on deep learning and storage medium | |
| CN116385832A (en) | Bimodal biological feature recognition network model training method | |
| EP4288910A1 (en) | Continual learning neural network system training for classification type tasks | |
| WO2020224244A1 (en) | Method and apparatus for obtaining depth-of-field image | |
| CN113569867A (en) | Image processing method and device, computer equipment and storage medium | |
| CN113129399A (en) | Pattern generation | |
| CN113792703B (en) | Image question-answering method and device based on Co-Attention depth modular network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |