CN112686305A - Semi-supervised learning method and system under assistance of self-supervised learning - Google Patents
Semi-supervised learning method and system under assistance of self-supervised learning Download PDFInfo
- Publication number
- CN112686305A CN112686305A CN202011597645.XA CN202011597645A CN112686305A CN 112686305 A CN112686305 A CN 112686305A CN 202011597645 A CN202011597645 A CN 202011597645A CN 112686305 A CN112686305 A CN 112686305A
- Authority
- CN
- China
- Prior art keywords
- data
- image
- supervised learning
- network
- semi
- 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
Images
Landscapes
- Image Analysis (AREA)
Abstract
A semi-supervised learning method under the assistance of self-supervised learning comprises the following steps: dividing the training samples into a plurality of batches, wherein each batch comprises marked data or unmarked data; randomly overturning the data of the current batch, and then combining the data with the original data to be used as current input data; extracting the characteristics of the current input data by using a characteristic extraction network to obtain abstract characteristics; and sending the extracted abstract features into an image category classification network or an image angle overturning classification network to judge the image category and the image angle overturning, if the input data is marked data, sending the abstract features into the image category classification network and the image angle overturning classification network simultaneously to judge the image category and the image angle overturning, and if the input data is unmarked data, sending the abstract features into the image angle overturning classification network only to judge the image angle overturning. The method improves the image category classification performance of semi-supervised learning.
Description
Technical Field
The invention relates to the field of machine learning algorithms, in particular to a semi-supervised learning method and a semi-supervised learning system under the assistance of self-supervised learning.
Background
Current methods based on deep learning have enjoyed great success in computer vision-related tasks, such as: the accuracy of face recognition has exceeded the human level. However, deep learning models often require a huge amount of labeled data, and in practical situations the collection of samples of certain categories (e.g., abnormal behavior data) is difficult, and labeling a large number of samples is often time-consuming and laborious. In contrast, our human typically only needs a very small number of samples to learn a new class. Inspired by this learning mode of human beings, the concept of semi-supervised learning has been proposed in the field of machine learning for a long time, i.e. only a small part of the data (usually around 10%) is tagged in a large amount of sample data, and the tagged data and the untagged data have the same data distribution. Semi-supervised learning is expected to learn a coarse pattern distribution of training data from a small amount of labeled data, and finally achieve a training result that can distinguish all samples (labeled/unlabeled) by continuously performing alternating learning with labeled and unlabeled data. In recent two years, the deep learning field also provides a paradigm of self-supervision learning. The self-supervision learning does not need to label the sample, and the high-level visual semantic information can be learned only through some self-defined preposition tasks (pretext tasks). At present, the self-supervision learning is usually used for training an initial model, the model can rapidly transfer knowledge in subsequent visual tasks (image classification, object recognition and the like), and the training purpose can be achieved only by a small amount of samples. Currently, self-supervised learning and semi-supervised learning are usually trained separately for different learning tasks. The self-supervised learning and the semi-supervised learning are two different learning paradigms, but the two are combined from the learning purpose (the depiction is from the bottom layer (the middle layer (the high layer mode distribution)).
Disclosure of Invention
The invention provides a semi-supervised learning method under the assistance of self-supervised learning, which can be used for training by simultaneously utilizing data with labels and data without labels, and for the data with labels, self-supervised loss and classification loss are simultaneously used for training; for unlabeled data, training was performed using only the loss of self-supervision. Through the combined training mode, the model can obtain better generalization performance on data without labels, and the classification performance of the self-supervision learning model is improved.
The technical scheme of the invention is as follows:
according to an aspect of the present invention, there is provided a semi-supervised learning method with the assistance of self-supervised learning, comprising the steps of: s1: obtaining sample data: dividing the training samples into a plurality of batches, wherein each batch comprises marked data or unmarked data; s2: and (3) turning over the random data and combining the random data with the original data: randomly overturning the data of the current batch, and then combining the data with the original data to be used as current input data; s3: extracting the characteristics of the current input data by using a characteristic extraction network to obtain abstract characteristics; and S4: and if the input data is label-free data, only the abstract features are sent to the image angle turnover classification network to judge the image angle turnover.
Preferably, in the semi-supervised learning method described above, in step S2, the angle at which the data is flipped is randomly selected from the set {90 °, 180 °, 270 °, 0 ° }.
Preferably, in the semi-supervised learning method, an auto-supervised learning loss function is added under an existing semi-supervised image classification model, so that labeled and unlabeled images can be simultaneously utilized in the training process of the model, wherein the auto-supervised learning loss function is expressed as:
Loss＝Ll(Dl，θl)+μLu(Du，θu) (1)
where μ is the harmonic coefficient of the two term loss functions, typically taken to be 0.5,
and wherein Ll(Dl，θl) Network for classifying image classes, thetalFor corresponding network parameters, only the marking data DlThe cross entropy loss function is adopted, and is specifically shown as the following formula (2):
Ll(Dl，θ)＝-∑ici log(pi) (2)
wherein, ciIs a class of sample i, piIs that the sample i is classified into class ciThe probability of (a) of (b) being,
Lu(Du，θu) The classification network is flipped over in correspondence with the angle of the image, thetauFor corresponding network parameters, the data D are adapted to be labeledlAnd label-free data DuSpecifically, the following formula (3) is shown below:
wherein, R is a selected angle set {90 degrees, 180 degrees, 270 degrees, 0 degrees }, R is a specific rotation angle, drThe representation adopts a rotation form represented by r for data d, and L adopts a cross entropy loss function.
Preferably, the semi-supervised learning method includes: the image angle overturning and classifying method comprises a feature extraction network, an image category classifying network and an image angle overturning and classifying network, wherein: the system comprises a feature extraction network, a data processing network and a data processing network, wherein the feature extraction network is used for extracting features of current input data to obtain abstract features, training samples are divided into a plurality of batches, each batch comprises marked data or unmarked data, the data of the current batch are randomly overturned and then are combined with original data to serve as the current input data; and the image category classification network and the image angle overturning classification network are used for carrying out image category and image angle overturning judgment on the extracted abstract features, if the input data is the labeled data, the abstract features are simultaneously sent into the image category classification network and the image angle overturning classification network to carry out image category and image angle overturning judgment, and if the input data is the unlabeled data, only the abstract features are sent into the image angle overturning classification network to carry out image angle overturning judgment.
Preferably, in the semi-supervised learning method, the image classification network includes: 2 fully-connected layers, each fully-connected layer containing 2048 neurons; and 1 image classification network softmax output layer, wherein the number of the neurons contained in the image classification network softmax output layer is the number of the classes to be classified plus 1.
Preferably, in the semi-supervised learning method, the image angle reversal classification network includes: 2 fully-connected layers, each fully-connected layer containing 2048 neurons; and 1 angle flip classification network softmax output layer, which contains 4 neurons, the output corresponding to one of {90 °, 180 °, 270 °, 0 ° } respectively.
According to the technical scheme of the invention, the beneficial effects are as follows:
compared with the prior semi-supervised learning method, the semi-supervised learning method under the self-supervision assistance provided by the invention can be used for learning the distribution modes of labeled data and unlabelled data at the same time in the training stage, so that the feature extraction network depended by the semi-unsupervised learning has stronger generalization capability on the unlabelled data, and the image category classification performance of the semi-supervised learning is improved.
The invention is further illustrated by way of example in the following with reference to the accompanying drawings:
drawings
In order to more clearly illustrate the detailed description of the invention or the technical solutions in the prior art, the drawings that are needed in the detailed description of the invention or the prior art will be briefly described below.
FIG. 1 is a flow chart of a semi-supervised learning method with the assistance of self-supervised learning proposed by the present invention;
FIG. 2 is a block diagram of a semi-supervised learning system with the assistance of the self-supervised learning proposed by the present invention;
FIG. 3 is a diagram of an image angle flip classification network architecture for self-supervised learning;
fig. 4 is a diagram of an image category classification network structure for semi-supervised learning.
Detailed Description
In the semi-supervised learning method under the assistance of the self-supervised learning, the self-supervised learning and the semi-supervised learning are put into the same learning task frame for training, and the generalization capability of a model learned by a small amount of samples in a semi-supervised learning mode is improved by extracting visual information of a bottom layer/middle layer of an image through the self-supervised learning.
The method is realized by adding an additional loss function on the currently popular semi-supervised learning model by utilizing a model learning framework of self-supervised learning and semi-supervised learning combined training, namely, the self-supervised learning loss function is added under the existing semi-supervised image classification model, so that the labeled and unlabeled images can be simultaneously utilized in the training process of the model, and the generalization capability of the semi-supervised image classification model to the unlabeled images is improved.
The principle of the semi-supervised learning method under the assistance of the self-supervised learning of the invention is as follows: because the self-supervised learning does not need additional sample labels, the self-supervised learning can train a feature extraction network by using a large number of unlabelled samples in a data set; since the same feature extraction network is used for the self-supervised learning and the semi-supervised learning in the model frame, the semi-supervised learning also has the pattern extraction capability of the self-supervised learning from a large number of unlabelled samples, so that the classification performance of the semi-supervised learning can be well popularized to the unlabelled samples.
In the prior art, only labeled sample data is used for training a classifier in the initial stage of semi-supervised learning, and the method can simultaneously use data with labels and data without labels to train a feature extraction network; the prior method only has one class classification loss supervision item in the training stage, and the method comprises the class classification loss supervision item and an angle turnover classification loss self-supervision item.
For the semi-supervised image classification task, the training samples comprise labeled data and unlabeled data. The semi-supervised learning method under the self-supervision assistance can process the two types of data simultaneously. Firstly, dividing a training sample into a plurality of batches, wherein each batch comprises marked data or unmarked data; randomly overturning the data of the current batch, and then combining the data with the original data to be used as current input data; secondly, extracting the characteristics of the current input data by using a characteristic extraction network to obtain abstract characteristics; and then the extracted abstract features are sent to an image category classification network or an image angle overturning classification network to judge the image category and the image angle overturning, if the input data is marked data, the abstract features are simultaneously sent to the two classification networks, and if the input data is unmarked data, only the abstract features are sent to the image angle overturning classification network.
The invention brings the self-supervision learning and the semi-supervision learning into the same model training frame, and trains the feature extraction network by using the labeled data and the unlabeled data, thereby improving the classification performance of the semi-supervision image classification network on the unlabeled data. Fig. 1 is a flowchart of a semi-supervised learning method under the assistance of the self-supervised learning of the present invention, and fig. 2 is a frame diagram of a semi-supervised learning system under the assistance of the self-supervised learning proposed by the present invention, as shown in fig. 1 and fig. 2, the method of the present invention includes the following steps:
s1: sample data (annotated data or annotated-free data) is obtained. Specifically, the training samples are divided into a plurality of batches, and each batch comprises labeled data 2 or unlabeled data 1;
s2: and the random data is overturned and combined with the original data. Specifically, the data of the current batch is randomly inverted and then merged with the original data as the current input data, i.e., the training data Dt. Wherein the turning angle is randomly selected from the set {90 °, 180 °, 270 °, 0 ° };
s3: and extracting the characteristics of the current input data by using a characteristic extraction network to obtain abstract characteristics. In particular, the network f is extracted by featuresθ(shown as 3 in FIG. 2) to DtExtracting the features to obtain abstract features Ft。fθMay be taken as VGG-16 network ("Very Deep conditional Networks for Large-Scale Image registration", Karen Simnyan, Andrew Zisserman, arxiv. org/abs/1409.1556 /);
s4: and if the input data is label-free data, only the abstract features are sent to the image angle overturning classification network to judge the image angle overturning. As shown in the figure, if DtTo label data 1, corresponding to the labeled data in FIG. 1, F will betSimultaneously sending the image classification network C (5 in figure 2) and the image angle overturning classification network
(shown in 4 in fig. 2), performing image type and image angle flip judgment; if D istIf there is no label number 2, only F will be addedtSending the image angle to a turnover classification network
And carrying out image angle overturning judgment.
The invention relates to a semi-supervised learning system under the assistance of self-supervised learning, which comprises: the image angle overturning and classifying method comprises a feature extraction network, an image category classifying network and an image angle overturning and classifying network, wherein:
and the characteristic extraction network is used for extracting the characteristics of the current input data to obtain abstract characteristics. Dividing training samples into a plurality of batches, wherein each batch comprises labeled data 2 or unlabeled data 1; randomly overturning the data of the current batch, and then combining the data with the original data to be used as current input data;
and the image category classification network and the image angle overturning classification network are used for carrying out image category and image angle overturning judgment on the extracted abstract features, if the input data is marked data, the abstract features are simultaneously sent into the image category classification network and the image angle overturning classification network to carry out image category and image angle overturning judgment, and if the input data is unmarked data, only the abstract features are sent into the image angle overturning classification network to carry out image angle overturning judgment.
The structure of the image classification network C is shown in fig. 4, and specifically includes: 2 fully-connected layers 8, each fully-connected layer 8 containing 2048 neurons; 1 image classification network softmax output layer 9, the number of the neurons included in the output layer is the number of the classes to be classified plus one. Image angle flip classification network
As shown in fig. 3, the structure of (a) specifically includes: 2 fully-connected layers 6, each fully-connected layer 6 containing 2048 neurons; the 1-degree turnover classification network softmax output layer 7 comprises 4 neurons, and the output corresponds to one of {90 degrees, 180 degrees, 270 degrees, 0 degrees } respectively.
The loss function of the semi-supervised learning method under the assistance of the self-supervised learning is shown as the formula 1:
Loss＝Ll(Dl，θl)+μLu(Du，θu) (1)
where μ in equation 1 is the harmonic coefficient of the two-term loss function, typically taken to be 0.5, Ll(Dl，θl) Classifying the network C (theta) corresponding to the image classlAs corresponding network parameters) only for the annotation data DlA cross entropy loss function is adopted, specifically the form shown in formula 2:
Ll(Dl，θ)＝-∑ici log(pi) (2)
wherein c isiIs a class of sample i, piIs that the sample i is classified into class ciThe probability of (c).
Lu(Du，θu) Sorting network for image angle flipping
(θuAs corresponding network parameters) to the annotation data DiAnd label-free data DuIn the form shown in formula 3:
wherein R is a selected angle set {90, 180, 270, 0 }, R is a specific rotation angle, drThe representation adopts a rotation form represented by r for data d, and L adopts a cross entropy loss function.
After training is finished, input data do not need to be turned over when reasoning is carried out, and meanwhile, the features output by the feature extraction network only need to be input into the image category classification network for reasoning. Specifically, the inference process includes the following 2 steps:
1.) inputting an image to be inferred into a feature extraction network to extract image features;
2.) inputting the extracted image features into an image category classification network to obtain a classification result.
The method of the invention carries out algorithm verification on ILSVRC-2012 ("image target Scale Visual Recognition Challenge", Olga Russakovsky, Jia Deng, Hao Su et al, International Journal of Computer Vision) semi-supervised learning data set, and under the condition that the annotation data is only 10% in training, the mAP (mean Average precision) of Top-5 is 83.64%. The method of the present invention was compared with the existing semi-supervised learning method for performance on the ILSVRC-2012 dataset, and the results are shown in table 1. The comparison result shows that the method of the invention is superior to some semi-supervised learning methods in the semi-supervised image classification task.
TABLE 1 comparison of image Classification Performance on ILSVRC-2012 datasets
| Participatory comparison method | mAP under 10% data annotation for ILSVRC-2012 dataset |
| Pseudolabels[1] | 82.41 |
| VAT[2] | 82.78 |
| The method of the invention | 83.64 |
Documents participating in the comparison:
[1]Dong-Hyun Lee.Pseudo-label:The simple and efficient semi-supervised learning method for deep neural networks.ICML 2013Workshop:Challenges in Representation Learning(WREPL),07,2013.
[2]Takeru Miyato,Shin-ichih Maeda,Masanori Koyama,and Shin Ishii.Virtual adversarial training:A regularization method for supervised and semi-supervised learning.arXiv preprint arXiv:1704.03976,2017.
the foregoing description is of the preferred embodiment of the concept and principles of operation in accordance with the present invention. The above-described embodiments should not be construed as limiting the scope of the claims, and other embodiments and combinations of implementations according to the inventive concept are within the scope of the invention.
Claims (6)
1. A semi-supervised learning method under the assistance of self-supervised learning is characterized by comprising the following steps:
s1: obtaining sample data: dividing the training samples into a plurality of batches, wherein each batch comprises marked data or unmarked data;
s2: and (3) turning over the random data and combining the random data with the original data: randomly overturning the data of the current batch, and then combining the data with the original data to be used as current input data;
s3: extracting the characteristics of the current input data by using a characteristic extraction network to obtain abstract characteristics; and
s4: and sending the extracted abstract features into an image category classification network or an image angle overturning classification network to judge the image category and the image angle overturning, if the input data is marked data, sending the abstract features into the image category classification network and the image angle overturning classification network simultaneously to judge the image category and the image angle overturning, and if the input data is unmarked data, sending the abstract features into the image angle overturning classification network only to judge the image angle overturning.
2. The semi-supervised learning method of claim 1, wherein in step S2, the angle at which the data is flipped is randomly selected from the set {90 °, 180 °, 270 °, 0 °.
3. The semi-supervised learning method of claim 1, wherein an auto-supervised learning loss function is added under the existing semi-supervised image classification model, so that labeled and unlabeled images can be utilized simultaneously in the training process of the model, wherein the auto-supervised learning loss function is expressed as:
Loss＝Ll(Dl，θl)+μLu(Du，θu) (1)
where μ is the harmonic coefficient of the two term loss functions, typically taken to be 0.5,
and wherein Ll(Dl，θl) A network of classification corresponding to said image classes, θlFor corresponding network parameters, only the marking data DlThe cross entropy loss function is adopted, and is specifically shown as the following formula (2):
Ll(Dl，θ)＝-∑icilog(pi) (2)
wherein, ciIs a class of sample i, piIs that the sample i is classified into class ciThe probability of (a) of (b) being,
Lu(Du，θu) The classification network is flipped over in correspondence with the angle of the image, thetauFor corresponding network parameters, the data D are adapted to be labeledlAnd label-free data DuSpecifically, the following formula (3) is shown below:
wherein, R is a selected angle set {90, 180, 270, 0 }, R is a specific rotation angle, drThe representation adopts a rotation form represented by r for data d, and L adopts a cross entropy loss function.
4. A semi-supervised learning system with the assistance of self-supervised learning, comprising: the image angle overturning and classifying method comprises a feature extraction network, an image category classifying network and an image angle overturning and classifying network, wherein:
the feature extraction network is used for extracting features of current input data to obtain abstract features, wherein training samples are divided into a plurality of batches, each batch comprises marked data or unmarked data, the data of the current batch are randomly overturned and then are combined with original data to serve as the current input data; and
the image type classification network and the image angle overturning classification network are used for carrying out image type and image angle overturning judgment on the extracted abstract features, if input data is marked data, the abstract features are simultaneously sent into the image type classification network and the image angle overturning classification network for carrying out image type and image angle overturning judgment, and if the input data is unmarked data, only the abstract features are sent into the image angle overturning classification network for carrying out image angle overturning judgment.
5. The semi-supervised learning system of claim 4, wherein the image category classification network comprises:
2 fully-connected layers, each of said fully-connected layers comprising 2048 neurons; and
1 image classification network softmax output layer, wherein the number of the neurons contained in the image classification network softmax output layer is the number of the categories to be classified plus 1.
6. The semi-supervised learning system of claim 4, wherein the image angle flip classification network comprises:
2 fully-connected layers, each of said fully-connected layers comprising 2048 neurons; and
1 angle flip sort network softmax output layer, which contains 4 neurons, the outputs corresponding to one of {90 °, 180 °, 270 °, 0 ° } respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011597645.XA CN112686305A (en) | 2020-12-29 | 2020-12-29 | Semi-supervised learning method and system under assistance of self-supervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011597645.XA CN112686305A (en) | 2020-12-29 | 2020-12-29 | Semi-supervised learning method and system under assistance of self-supervised learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112686305A true CN112686305A (en) | 2021-04-20 |
Family
ID=75454150
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011597645.XA Pending CN112686305A (en) | 2020-12-29 | 2020-12-29 | Semi-supervised learning method and system under assistance of self-supervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112686305A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113128601A (en) * | 2021-04-22 | 2021-07-16 | 北京百度网讯科技有限公司 | Training method of classification model and method for classifying images |
| CN113469296A (en) * | 2021-09-03 | 2021-10-01 | 广东众聚人工智能科技有限公司 | Image classification method and system based on small amount of label data self-supervision joint learning |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108520535A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | Object classification method based on depth recovery information |
| CN109002845A (en) * | 2018-06-29 | 2018-12-14 | 西安交通大学 | Fine granularity image classification method based on depth convolutional neural networks |
| CN110263845A (en) * | 2019-06-18 | 2019-09-20 | 西安电子科技大学 | SAR image change detection based on semi-supervised confrontation depth network |
| CN110569882A (en) * | 2019-08-15 | 2019-12-13 | 杨春立 | Image information classification method and device |
| CN110852295A (en) * | 2019-10-15 | 2020-02-28 | 深圳龙岗智能视听研究院 | Video behavior identification method based on multitask supervised learning |
| CN111126361A (en) * | 2019-11-19 | 2020-05-08 | 西安电子科技大学 | SAR target identification method based on semi-supervised learning and feature constraint |
| CN111179227A (en) * | 2019-12-16 | 2020-05-19 | 西北工业大学 | Mammary gland ultrasonic image quality evaluation method based on auxiliary diagnosis and subjective aesthetics |
-
2020
- 2020-12-29 CN CN202011597645.XA patent/CN112686305A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108520535A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | Object classification method based on depth recovery information |
| CN109002845A (en) * | 2018-06-29 | 2018-12-14 | 西安交通大学 | Fine granularity image classification method based on depth convolutional neural networks |
| CN110263845A (en) * | 2019-06-18 | 2019-09-20 | 西安电子科技大学 | SAR image change detection based on semi-supervised confrontation depth network |
| CN110569882A (en) * | 2019-08-15 | 2019-12-13 | 杨春立 | Image information classification method and device |
| CN110852295A (en) * | 2019-10-15 | 2020-02-28 | 深圳龙岗智能视听研究院 | Video behavior identification method based on multitask supervised learning |
| CN111126361A (en) * | 2019-11-19 | 2020-05-08 | 西安电子科技大学 | SAR target identification method based on semi-supervised learning and feature constraint |
| CN111179227A (en) * | 2019-12-16 | 2020-05-19 | 西北工业大学 | Mammary gland ultrasonic image quality evaluation method based on auxiliary diagnosis and subjective aesthetics |
Non-Patent Citations (2)
| Title |
|---|
| HANKOOK LEE ET AL.: "Self-supervised Label Augmentation via Input Transformations", 《ARXIV》 * |
| XIAOHUA ZHAI ET AL.: "S4L: Self-Supervised Semi-Supervised Learning", 《ARXIV》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113128601A (en) * | 2021-04-22 | 2021-07-16 | 北京百度网讯科技有限公司 | Training method of classification model and method for classifying images |
| CN113469296A (en) * | 2021-09-03 | 2021-10-01 | 广东众聚人工智能科技有限公司 | Image classification method and system based on small amount of label data self-supervision joint learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109271522B (en) | Comment emotion classification method and system based on deep hybrid model transfer learning | |
| Wang et al. | Deep learning algorithms with applications to video analytics for a smart city: A survey | |
| CN113177132B (en) | Image retrieval method based on depth cross-modal hash of joint semantic matrix | |
| CN110135461B (en) | Hierarchical attention perception depth measurement learning-based emotion image retrieval method | |
| CN112733965B (en) | Label-free image classification method based on small sample learning | |
| CN114048568A (en) | Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework | |
| CN108427740B (en) | Image emotion classification and retrieval algorithm based on depth metric learning | |
| CN111275092A (en) | Image classification method based on unsupervised domain adaptation | |
| CN110163286A (en) | Hybrid pooling-based domain adaptive image classification method | |
| CN111832573B (en) | Image emotion classification method based on class activation mapping and visual saliency | |
| Guo et al. | An ensemble broad learning scheme for semisupervised vehicle type classification | |
| CN114067385A (en) | Cross-modal face retrieval Hash method based on metric learning | |
| CN112686305A (en) | Semi-supervised learning method and system under assistance of self-supervised learning | |
| CN114092747A (en) | Small sample image classification method based on depth element metric model mutual learning | |
| Manohar et al. | Convolutional neural network with SVM for classification of animal images | |
| Zhou et al. | Sampling-attention deep learning network with transfer learning for large-scale urban point cloud semantic segmentation | |
| Long et al. | Deep neural networks in fully connected crf for image labeling with social network metadata | |
| CN112883931A (en) | Real-time true and false motion judgment method based on long and short term memory network | |
| CN113947725B (en) | Hyperspectral image classification method based on convolution width migration network | |
| CN111310820A (en) | Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration | |
| CN115331284A (en) | Self-healing mechanism-based facial expression recognition method and system in real scene | |
| Huynh et al. | An efficient model for copy-move image forgery detection | |
| CN112883930A (en) | Real-time true and false motion judgment method based on full-connection network | |
| CN115640401B (en) | Text content extraction method and device | |
| Deng et al. | Behavior recognition based on category subspace in crowded videos |
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 |