CN110619059A - Building marking method based on transfer learning - Google Patents
Building marking method based on transfer learning Download PDFInfo
- Publication number
- CN110619059A CN110619059A CN201910745724.1A CN201910745724A CN110619059A CN 110619059 A CN110619059 A CN 110619059A CN 201910745724 A CN201910745724 A CN 201910745724A CN 110619059 A CN110619059 A CN 110619059A
- Authority
- CN
- China
- Prior art keywords
- image
- training
- building
- network
- data set
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- 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
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Library & Information Science (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
A building calibration method based on transfer learning comprises the following steps: 1) respectively constructing a training set and a test set from a self-made data set, wherein the data set comprises a building design drawing from a related design place and a building picture shot on line, and is applied as a recommendation system test data; 2) building a building object marking technology architecture model, wherein the model comprises image feature marking, image feature extraction, database matching, Mahalanobis distance setting and final image authentication; 3) cutting the data set into small blocks, sequentially inputting the small blocks into a building recognition neural network, updating parameters by using a back propagation Adam algorithm, and training a finally changed classifier through the set number of training rounds for training; 4) and calibrating the actual building picture by the trained model.
Description
Technical Field
The invention relates to the field of image classification, in particular to a building calibration method based on transfer learning
Background
Building identification is one of research hotspots in the fields of computer vision and pattern identification, and the method can enable people to quickly acquire relevant information such as the position, name and description of a building according to an image, has important application value in the fields of building positioning and building design, building marking and the like, and effectively mark the information, which is a key problem of building identification.
The key technology for building identification is feature extraction, the most common image features include color features, texture features and shape features, most of the established image labeling and image extraction systems are based on the features, the performances of the image labeling and image extraction systems depend on the representation methods of the extracted features, and the following widely used feature descriptors are mainly used for the color features, the texture features and the shape features: HOG descriptor, LBP descriptor, HSV descriptor, SIFT descriptor. And generating an efficient feature extractor by using the descriptors and combining methods of dimension reduction, pyramid establishment, bag-of-word model construction and the like. However, the traditional features have very large limitations, the artificial feature engineering is time-consuming and labor-consuming and has high requirements on professional field knowledge, the extracted features are single, and the feature effect is worse as the complexity of the data set is increased. Therefore, the requirement of the existing computer vision field on feature extraction cannot be met by using the traditional features alone. In recent years, with the development of deep learning technology, a neural network structure based on end-to-end learning depends on big data and high-dimensional parameter space advantages, and high-level features are abstractly synthesized from the bottom layer to the top layer step by step. The data-driven self-learning mode ensures that the convolutional neural network has excellent feature extraction capability. Therefore, the deep learning is combined with the building marking, so that the method has practical application value.
Disclosure of Invention
The invention provides a building marking method based on transfer learning, aiming at solving the defects that the existing building identification technology is low in precision and needs a large amount of artificial characteristic engineering.
The method combines the convolutional neural network of the transfer learning to extract the characteristics, and performs characteristic division by a multi-characteristic calibration technology to enable the model to better capture the potential correlation information among the characteristics so as to improve the prediction precision of the model, finely adjusts the network model, obtains the input of the last layer through the inclusion network model, defines the input as Bottlenecks, trains the finally changed softmax layer by using the Bottlenecks, and performs simulation experiment on the spyder visualization platform
The technical scheme adopted by the invention for solving the technical problems is as follows:
a building marking method based on transfer learning comprises the following steps:
step 1, self-making a data set, wherein the building data set is from Google pictures and Baidu pictures, is divided into seven types of common buildings in life based on the existing data set, is respectively a church, a residential building, a hospital, a hotel, a library, a villa house and a shopping mall, and is subjected to homogenization treatment.
And 2, constructing a multi-feature calibration neural network model based on transfer learning, wherein the whole network architecture comprises image feature calibration, image feature extraction, database matching and image authentication.
Image characteristic calibration: and carrying out single characteristic calibration on buildings with obvious building styles and carrying out multi-characteristic calibration on the buildings with various styles so as to form a characteristic matrix of the picture.
Image feature extraction: the image after characteristic calibration is used as network input, the multilayer neural network learns the characteristics of the original image by the mutual matching of a series of convolution layers and down-sampling layers, the parameters are adjusted by combining the classic BP algorithm to complete weight updating, and the BP network updating weight formula is as follows:
ω(t+1)=ω(t)+ηδ(t)x(t) (1)
wherein ω (t) is a connection weight, x (t) is an output of a neuron, δ (t) represents an error term of the neuron, η represents a learning rate, and a network structure of a convolution layer in the network adopts a discrete form of convolution and is represented as:
wherein M isβRepresenting a choice of input features, k representing a convolution kernel, γ representing the number of layers of the network, b representing the added bias per input feature map, the input map features may be convolved with different convolution kernels for a particular output map. f denotes the activation function used by the convolutional neurons, here the ReLu activation function is used per block.
Database matching: the database is a feature vector of each picture obtained from a training data set, the feature information of each picture after network training is stored in a label file in a 1024-dimensional data point mode, and the feature information of a newly trained picture is matched with the database by calculating the Mahalanobis distance of the newly trained picture when the data set is tested.
Image authentication: after the image feature extraction module finishes feature extraction, two output results are obtained, namely a feature vector of the image and a label after image classification. The image authentication module firstly classifies the image to be authenticated into the corresponding category of the image library according to the existing label of the image to be authenticated, and then performs Mahalanobis distance calculation with other images of the category, and the result is marked as Di(i is 1, 2, 3 …, m), and taking out Dmin. And judging whether the image to be authenticated comes from the database or comes from the same source map as the image of the database according to the calculation result. Judging method the threshold value method, threshold value Th is usedmAccording to a large number of experimental results. If D ismin>ThmI.e. the image is at a greater distance from all images in the class and does not belong to the database. Otherwise, the image belongs to the database and is associated with DminThe corresponding images are the same image or are derived from the same image.
Definition of mahalanobis distance: known as M sample vectors X1~XmThe covariance matrix is denoted as S and the mean is denoted as the vector μ, where the sample vector X isiMahalanobis distance d (x) to μ is expressed as:
and wherein the vector XiAnd XjThe mahalanobis distance of (a) is defined as:
if the covariance matrix is an identity matrix, i.e. the sample vectors are independently and identically distributed, the formula becomes:
i.e. the euclidean distance. If the covariance matrix is a diagonal matrix, the formula becomes a normalized Euclidean distance. Therefore, the mahalanobis distance is an effective method for calculating the similarity between two unknown sample sets compared with the euclidean distance. Unlike the euclidean distance, which takes into account the link between various characteristics and is not affected by dimensional quantities, the process mainly tests the image authentication module accuracy, by carrying out operations such as small-range rotation, affine transformation, gray level transformation and the like on each picture of the self-made data set, a new data set is synthesized to simulate a tampered image, the image authentication accuracy rate mainly depends on the result of the previous image classification and the selection of the threshold, the purpose of this experiment is to find a suitable threshold value, the initial value of which is set by calculating the distance between the transformed image and the original image, statistically analyzing, through a large number of experimental analyses, the Mahalanobis distance threshold of the original image and the Mahalanobis distance threshold after the preprocessing transformation are both about 1.2, while the Mahalanobis distance of two completely different images is basically more than 3, therefore, the threshold value can be set to be relatively coarse, so that the threshold value can be set to be about 2.0 and experiments can be carried out.
Step 3, network training, comprising the following steps:
the training rounds of the network are set to 10000 times, the trained batch-size is 100, one round of prediction is carried out every 10 times, meanwhile, 80% of all data are used as main training samples T, the other 10% of all data are used as cross validation samples V in the training process, the remaining 10% of all data are used as test data sets K for predicting the performance of a classifier in the real world, the algorithm used for training is SGD (random gradient descent algorithm), the input of the network is 229x229x3, all convolutional layers in the blocks are subjected to parameter reduction by using 3x3 small convolutional kernels, GoogleNet weights from training according to imageNet are transferred and are subjected to fine tuning, the network is trained from an optimal point, the classifier is retrained, log loss function values of the training sets T are calculated, and the weights and the deviations of the multilayer neural network are updated according to a back propagation algorithm.
And calculating the log loss function value of the verification set V, judging whether the model is converged, and if the convergence model training is finished, entering data on the next training set T for training until the log loss function value on the verification set V tends to be converged. And finally, verifying the accuracy of the model by using a test set K.
And 4, calibrating the actual building picture by the trained model.
The invention has the advantages that: compared with the traditional machine learning method, the method has high universality and stability, and meanwhile, the multi-feature calibration technology provided by the invention enables the feature vector passing through the network bottleneck layer to be more representative by artificially calibrating the features of the multifunctional building, thereby greatly improving the accuracy of network classification.
Drawings
FIG. 1 is a schematic of an experimental Build-7 data set of the present invention.
Fig. 2 is a representation of a building characteristic matrix of the present invention.
FIG. 3 is a diagram of a building identification neural network of the present invention.
Fig. 4 is a building specification technical architecture diagram of the present invention.
Detailed Description
In order to better explain the technical scheme of the invention, the invention is further explained by an embodiment with the accompanying drawings.
A building calibration method based on transfer learning comprises the following steps:
step 1, constructing a training set and a test set respectively according to data shown in figure 1 from a system data set Build-7, wherein the data set comprises a building design drawing from a related design institute and a live-action building picture on line, and the data set is used as test data of a recommendation system.
And 2, performing multi-feature calibration on the picture to be trained and labeling the picture in a feature matrix mode, wherein the model feature calibration part is shown in FIG. 2, 1 represents that the building has the feature, and-1 represents that the feature does not exist. The image subjected to feature labeling is input as a network (as shown in fig. 3), the inclusion module group migrated in the network is formed by connecting a plurality of inclusion modules in series, and the structure stacks convolution (1x1, 3x3, 5x5) and pooling operation (3x3) commonly used in CNN (the sizes of the convolution and pooling are the same, and channels are added), so that the width of the network is increased on one hand, and the adaptability of the network to the scale is also increased on the other hand. The network in the network convolutional layer can extract every detail information of the input, and at the same time, the 5x5 filter can also cover most of the input of the receiving layer, for the convolutional layer, a large number of feature maps are generated by using hundreds of convolutional kernels to capture various features of the image, so that local activation values of the feature maps can be aggregated as feature vectors to construct a more distinctive expression than manually extracted descriptors and features directly extracted from CNN full-link layers.
Step 3. the labeling technique framework of the invention is as shown in fig. 4, the data set is cut into small blocks and sequentially input into the building recognition neural network, parameters are updated by using a back propagation algorithm, the input of the network is 229x229x3, all the convolution layers in the block adopt 3x3 small convolution kernels for parameter reduction, after feature calibration and feature extraction, the mahalanobis distance was set to 2.0, and further, a training round was made 10000 times, each block size (batch size) contained 100 pieces of data, 80% of the data sets are used as training sample sets, 10% of the data sets are used as cross validation sets in training, the remaining 10% of the data sets are used as test data sets for predicting the performance of a classifier in the real world, prediction is performed once every 10 rounds, the model is trained by using an Adam algorithm in the training process, the learning rate is 0.01, and when each training round is finished, the validation errors on the validation sets are calculated.
Step 4, predicting and comparing the trained model on the test set
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.
Claims (2)
1. A building marking method based on transfer learning comprises the following steps:
step 1, self-making a data set, wherein the building data set is from Google pictures and Baidu pictures, is divided into seven types of buildings which are common in life based on the existing data set, is respectively a church, a residential building, a hospital, a hotel, a library, a villa house and a shopping mall, and is subjected to homogenization treatment;
step 2, constructing a multi-feature calibration neural network model based on transfer learning, wherein the whole network architecture comprises image feature calibration, image feature extraction, database matching and image authentication;
image characteristic calibration: single characteristic calibration is carried out on buildings with obvious styles, multi-characteristic calibration is carried out on buildings with various styles, and then a characteristic matrix of the picture is formed;
image feature extraction: the image after characteristic calibration is used as network input, the multilayer neural network learns the characteristics of the original image by the mutual matching of a series of convolution layers and down-sampling layers, the parameters are adjusted by combining the classic BP algorithm to complete weight updating, and the BP network updating weight formula is as follows:
ω(t+1)=ω(t)+ηδ(t)x(t) (1)
wherein ω (t) is a connection weight, x (t) is an output of a neuron, δ (t) represents an error term of the neuron, η represents a learning rate, and a network structure of a convolution layer in the network adopts a discrete form of convolution and is represented as:
wherein M isβRepresenting one selection of input features, k representing a convolution kernel, gamma representing the number of layers of the network, b representing a bias added by each input feature map, wherein for a specific output map, the input mapping features can be obtained by applying different convolution kernels to convolve; f represents the activation function used by the convolutional neurons, here the ReLu activation function is used for each block;
database matching: the database is a feature vector of each picture obtained from a training data set, the feature information of each picture after network training is stored in a label file in a 1024-dimensional data point mode, and the feature information of a newly trained picture is matched with the database by calculating the Mahalanobis distance of the newly trained picture when the data set is tested;
image authentication: after the image feature extraction module finishes feature extraction, two output results are obtained, namely a feature vector of the image and a label after image classification; the image authentication module firstly classifies the image to be authenticated into the corresponding category of the image library according to the existing label of the image to be authenticated, and then performs Mahalanobis distance calculation with other images of the category, and the result is marked as Di(i is 1, 2, 3 …, m), and taking out Dmin(ii) a Judging whether the image to be authenticated comes from the database or comes from the same source image as the image in the database according to the calculation result; judging method the threshold value method, threshold value Th is usedmSetting according to a large number of experimental results; if D ismin>ThmThat is, the image has a large distance to all the images in the class and does not belong to the database; otherwise, the image belongs to the database and is associated with DminThe corresponding images are the same image or are derived from the same image;
definition of mahalanobis distance: known as M sample vectors X1~XmThe covariance matrix is denoted as S and the mean is denoted as the vector μ, where the sample vector X isiMahalanobis distance d (x) to μ is expressed as:
and wherein the vector XiAnd XjThe mahalanobis distance of (a) is defined as:
if the covariance matrix is an identity matrix, i.e. the sample vectors are independently and identically distributed, the formula becomes:
that is, the Euclidean distance; if the covariance matrix is a diagonal matrix, the formula becomes a standardized Euclidean distance; therefore, comparing the mahalanobis distance with the euclidean distance, it is an effective method for calculating the similarity between two unknown sample sets; different from the Euclidean distance, the method considers the relation among various characteristics and is not influenced by dimension, the accuracy of an image authentication module is tested, a new data set is synthesized to simulate a tampered image by performing small-range rotation, affine and gray level transformation on each picture of a self-made data set Build-7, the image authentication accuracy mainly depends on the result of image classification in the early stage and the selection of a threshold value, so that the experiment aims to find a proper threshold value, and the initial value of the threshold value is set after statistical analysis by calculating the distance between the transformed image and an original image;
step 3, network training, comprising the following steps:
setting the training rounds of the network as 10000 times, setting the trained batch-size as 100, performing one round of prediction every 10 times, simultaneously taking 80% of all data as a main training sample T, taking the other 10% as a cross validation sample V in the training process, taking the remaining 10% as a test data set K for predicting the expression of a classifier in the real world, wherein the algorithm used for training is a random gradient descent algorithm SGD, the input of the network is 229x229x3, all convolutional layers in a block adopt a 3x3 small convolutional kernel for parameter reduction, and the weight and the deviation of a multilayer neural network are updated according to a back propagation algorithm by transferring GoogleNet weight trained according to imageNet and performing fine tuning, so that the network is trained from an optimal point, the classifier is retrained again, the log loss function value of the training set T is calculated;
calculating a log loss function value of the verification set V, judging whether the model is converged, and if the convergence model is trained, entering data on the next training set T for training until the log loss function value on the verification set V tends to be converged; finally, verifying the accuracy of the model by using a test set K;
and 4, calibrating the actual building picture by the trained model.
2. The building marking method based on the transfer learning is characterized in that: the initial value of the threshold value in step 2 is 2.0.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910745724.1A CN110619059B (en) | 2019-08-13 | 2019-08-13 | Building marking method based on transfer learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910745724.1A CN110619059B (en) | 2019-08-13 | 2019-08-13 | Building marking method based on transfer learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110619059A true CN110619059A (en) | 2019-12-27 |
CN110619059B CN110619059B (en) | 2021-07-27 |
Family
ID=68921860
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910745724.1A Active CN110619059B (en) | 2019-08-13 | 2019-08-13 | Building marking method based on transfer learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110619059B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111582117A (en) * | 2020-04-29 | 2020-08-25 | 长江大学 | Unmanned aerial vehicle illegal building inspection method, equipment and storage medium |
CN112114231A (en) * | 2020-09-18 | 2020-12-22 | 广西大学 | CNN fault line selection method with continuous learning capability |
CN112732444A (en) * | 2021-01-12 | 2021-04-30 | 北京工业大学 | Distributed machine learning-oriented data partitioning method |
CN113128565A (en) * | 2021-03-25 | 2021-07-16 | 之江实验室 | Automatic image annotation system and device oriented to agnostic pre-training annotation data |
CN113378815A (en) * | 2021-06-16 | 2021-09-10 | 南京信息工程大学 | Model for scene text positioning recognition and training and recognition method thereof |
CN113449631A (en) * | 2021-06-25 | 2021-09-28 | 中南大学 | Image classification method and system |
CN115022049A (en) * | 2022-06-06 | 2022-09-06 | 哈尔滨工业大学 | Distributed external network traffic data detection method based on Mahalanobis distance calculation, electronic device and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106845350A (en) * | 2016-12-21 | 2017-06-13 | 浙江工业大学 | A kind of tree-shaped node recognition methods based on image procossing |
CN106991439A (en) * | 2017-03-28 | 2017-07-28 | 南京天数信息科技有限公司 | Image-recognizing method based on deep learning and transfer learning |
US20170357892A1 (en) * | 2016-06-08 | 2017-12-14 | Adobe Systems Incorporated | Convolutional Neural Network Joint Training |
CN108846444A (en) * | 2018-06-23 | 2018-11-20 | 重庆大学 | The multistage depth migration learning method excavated towards multi-source data |
CN109086745A (en) * | 2018-08-31 | 2018-12-25 | 广东工业大学 | A kind of localization method, device, equipment and computer readable storage medium |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
-
2019
- 2019-08-13 CN CN201910745724.1A patent/CN110619059B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170357892A1 (en) * | 2016-06-08 | 2017-12-14 | Adobe Systems Incorporated | Convolutional Neural Network Joint Training |
CN106845350A (en) * | 2016-12-21 | 2017-06-13 | 浙江工业大学 | A kind of tree-shaped node recognition methods based on image procossing |
CN106991439A (en) * | 2017-03-28 | 2017-07-28 | 南京天数信息科技有限公司 | Image-recognizing method based on deep learning and transfer learning |
CN108846444A (en) * | 2018-06-23 | 2018-11-20 | 重庆大学 | The multistage depth migration learning method excavated towards multi-source data |
CN109086745A (en) * | 2018-08-31 | 2018-12-25 | 广东工业大学 | A kind of localization method, device, equipment and computer readable storage medium |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
Non-Patent Citations (1)
Title |
---|
刘文涛 等: "基于全卷积神经网络的建筑物屋顶自动提取", 《地球信息科学学报》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111582117A (en) * | 2020-04-29 | 2020-08-25 | 长江大学 | Unmanned aerial vehicle illegal building inspection method, equipment and storage medium |
CN112114231A (en) * | 2020-09-18 | 2020-12-22 | 广西大学 | CNN fault line selection method with continuous learning capability |
CN112114231B (en) * | 2020-09-18 | 2023-10-10 | 广西大学 | CNN fault line selection method with continuous learning capability |
CN112732444A (en) * | 2021-01-12 | 2021-04-30 | 北京工业大学 | Distributed machine learning-oriented data partitioning method |
CN113128565A (en) * | 2021-03-25 | 2021-07-16 | 之江实验室 | Automatic image annotation system and device oriented to agnostic pre-training annotation data |
CN113378815A (en) * | 2021-06-16 | 2021-09-10 | 南京信息工程大学 | Model for scene text positioning recognition and training and recognition method thereof |
CN113378815B (en) * | 2021-06-16 | 2023-11-24 | 南京信息工程大学 | Scene text positioning and identifying system and training and identifying method thereof |
CN113449631A (en) * | 2021-06-25 | 2021-09-28 | 中南大学 | Image classification method and system |
CN115022049A (en) * | 2022-06-06 | 2022-09-06 | 哈尔滨工业大学 | Distributed external network traffic data detection method based on Mahalanobis distance calculation, electronic device and storage medium |
CN115022049B (en) * | 2022-06-06 | 2024-05-14 | 哈尔滨工业大学 | Distributed external network flow data detection method based on calculated mahalanobis distance, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110619059B (en) | 2021-07-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110619059B (en) | Building marking method based on transfer learning | |
CN111191732B (en) | Target detection method based on full-automatic learning | |
CN109919317B (en) | Machine learning model training method and device | |
CN110020682B (en) | Attention mechanism relation comparison network model method based on small sample learning | |
CN111695467B (en) | Spatial spectrum full convolution hyperspectral image classification method based on super-pixel sample expansion | |
CN108647583B (en) | Face recognition algorithm training method based on multi-target learning | |
CN111079847B (en) | Remote sensing image automatic labeling method based on deep learning | |
CN109740679B (en) | Target identification method based on convolutional neural network and naive Bayes | |
CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
CN110717526A (en) | Unsupervised transfer learning method based on graph convolution network | |
CN112347970B (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
CN110197205A (en) | A kind of image-recognizing method of multiple features source residual error network | |
CN109492750B (en) | Zero sample image classification method based on convolutional neural network and factor space | |
CN112132014B (en) | Target re-identification method and system based on non-supervised pyramid similarity learning | |
CN113239131B (en) | Low-sample knowledge graph completion method based on meta-learning | |
CN112232395B (en) | Semi-supervised image classification method for generating countermeasure network based on joint training | |
CN112364974B (en) | YOLOv3 algorithm based on activation function improvement | |
CN111340096A (en) | Weakly supervised butterfly target detection method based on confrontation complementary learning | |
CN111161244A (en) | Industrial product surface defect detection method based on FCN + FC-WXGboost | |
CN115311502A (en) | Remote sensing image small sample scene classification method based on multi-scale double-flow architecture | |
CN111310820A (en) | Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration | |
CN115984930A (en) | Micro expression recognition method and device and micro expression recognition model training method | |
CN113066528B (en) | Protein classification method based on active semi-supervised graph neural network | |
CN109784404A (en) | A kind of the multi-tag classification prototype system and method for fusion tag information | |
CN116910571B (en) | Open-domain adaptation method and system based on prototype comparison learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |