CN114529767A - Small sample SAR target identification method based on two-stage comparison learning framework - Google Patents

Small sample SAR target identification method based on two-stage comparison learning framework Download PDF

Info

Publication number
CN114529767A
CN114529767A CN202210152954.9A CN202210152954A CN114529767A CN 114529767 A CN114529767 A CN 114529767A CN 202210152954 A CN202210152954 A CN 202210152954A CN 114529767 A CN114529767 A CN 114529767A
Authority
CN
China
Prior art keywords
strategy
sar
learning
class
module
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
Application number
CN202210152954.9A
Other languages
Chinese (zh)
Inventor
丁兴号
黄悦
黄乐兴
蔡森林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xiamen University
Original Assignee
Xiamen University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xiamen University filed Critical Xiamen University
Priority to CN202210152954.9A priority Critical patent/CN114529767A/en
Publication of CN114529767A publication Critical patent/CN114529767A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

A small sample SAR target identification method based on a two-stage comparison learning framework relates to image identification. Aiming at the practical problems that an SAR data set is difficult to obtain and labeled data consumes time and labor, a comparative learning framework based on a two-stage training strategy is provided for solving SAR target recognition under the condition of small sample data. Decoupling the training process into a characterization learning phase and a classification learning phase, comprising: 1) in the characterization learning stage, a representation module of a network is trained on an original SAR data set by adopting supervised contrast learning; 2) and in the classification learning stage, the coding module is fixed and a classification module of the network is trained by adopting a rebalance strategy. The method is established on the basis of contrast learning in machine learning and combined with the thought of decoupling learning, has strong practicability and good stability, and can meet the requirements of high standards of SAR target identification precision under the conditions of unbalanced categories, small sample data and the like.

Description

Small sample SAR target identification method based on two-stage comparison learning framework
Technical Field
The invention belongs to the field of image recognition, relates to an SAR target recognition technology under the condition of small sample data, and particularly relates to a decoupling learning and contrast learning method.
Background
Synthetic Aperture Radar (SAR) is a high resolution radar device based on the aperture principle. Compared with optical remote sensing, the method can overcome the influence of severe weather factors and day and night factors, and is carried on flight platforms such as satellites and unmanned aerial vehicles. The uniqueness of SAR makes it highly useful in military, surveying and mapping, agriculture, etc. (a. moreira, p.prats-Iraola, m.younis, g.krieger, i.ha-jnsek, and k.p.papathionassu, "a national on synthetic aperture radar," IEEE Geoscience and remote sensing, vol.1, No.1, pp.6-43,2013). In recent years, with the development of convolutional neural networks, many deep learning algorithms applied to SAR target recognition appear, however, most of the algorithms are proposed based on equalized data sets. In real scenarios, SAR data is typically unbalanced because some targets are less likely to occur or are more difficult to collect, resulting in tail-like small sample data. Under the condition of small sample data, the model is easy to be over-fitted to the head class, and a serious challenge is brought to the traditional deep learning algorithm. In addition, a large amount of manpower and material resources are required to collect balanced data sets, and therefore the method is impractical; therefore, the method has important practical significance in training a deep learning model with good generalization performance by fully utilizing the small sample SAR data.
Rebalancing strategies (e.g., re-weighting and re-sampling) are outstanding and effective methods to solve the unbalanced training data problem, but most conventional schemes are classification net learning, and it is expected to increase the weight of the tail class, thereby reducing the bias of the model to the head class. While the rebalancing strategy can significantly facilitate classification learning of deep networks, it also unexpectedly impairs the ability to characterize deep features of interest to some extent (Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-MinChen, "Bbn: binary-bridge network with collective visual Recognition," advanced-entries of the IEEE/CVF Conference on Computer vision and Pattern Recognition,2020, pp.9719-9728).
Disclosure of Invention
The invention aims to solve the problems that in the prior art, the SAR data set is difficult to obtain, the labeled data is time-consuming and labor-consuming, the data imbalance causes poor generalization capability of a model to a small sample SAR target, and the like, and provides a small sample SAR target recognition method based on a two-stage contrast learning framework, which can improve the generalization performance when a deep learning model is used for carrying out small sample SAR target recognition, improve the target recognition accuracy of a small sample, meet the actual application requirements, and improve the overall performance of the model.
The invention decouples the training process into a representation learning stage and a classification learning stage, and specifically comprises the following steps:
1) in the characterization learning stage, a representation module of a network is trained on an original SAR data set by adopting supervised contrast learning;
2) and in the classification learning stage, the coding module is fixed and a classification module of the network is trained by adopting a rebalance strategy.
In the step 1), the representation module for training the network on the original SAR data set by adopting Supervised contrast Learning firstly adopts a contrast Learning algorithm Supervised contrast Learning as a training strategy of the representation module, designs a data enhancement strategy aiming at the characteristics of the SAR target, and further optimizes the training effect of the contrast Learning, so that the representation module can extract the characteristics with higher generalization for subsequent classification tasks; the specific steps of the characterization learning stage are as follows:
(1) the SAR image is obtained by imaging the synthetic aperture radar on a flight platform at a bird's-eye view angle, the angle of a target in the image is uncertain, and in order to identify the SAR target by missing partial image information, the following data enhancement strategy is designed: random cropping and scaling, random flipping and random rotation; and respectively applying a data enhancement strategy twice to each SAR image to obtain 2N SAR images.
(2) Respectively sending the 2N SAR images into a coding module in the network to obtain a characteristic vector f corresponding to each SAR imageiThen the feature vector f is divided intoiSending the data to a projection module, mapping the data to another representation space to obtain a non-linear representation zi＝Proj(xi)＝W(2)σ(W(1)fi) Where σ is the ReLu activation function and W is a single linear layer; the adopted projection module is realized by two layers of perceptrons with a hidden layer;
(3) use of
Figure BDA0003511324730000021
Norm normalization ziThen using normalized ziAnd calculating the loss of the supervised contrast learning, so as to train the network until convergence, wherein the loss of the supervised contrast learning is calculated as follows:
Figure BDA0003511324730000022
wherein P (i) represents a sum ofiAll sample subscript sets of the same class, | P (i) | is the size of this set, τ is a temperature over-parameter.
In the step 2), the rebalance strategy is used for improving the identification precision of the small sample SAR target, and the rebalance strategy comprises a class balance sampling strategy and a cost sensitive cross entropy loss strategy, and adopts one of the classification modules of the training network;
the class balance sampling strategy comprises the following specific steps:
(1) carrying out class balance sampling on the training set, randomly selecting N SAR images, sending the N SAR images into a coding module to generate a characteristic vector fiMeanwhile, a gradient interruption strategy is adopted for the coding module, so that adverse effects generated by a subsequent training classification module are avoided. For class-balanced sampling, the probability that the ith sample is chosen is as follows:
Figure BDA0003511324730000031
wherein, yiIs a category of the ith sample,
Figure BDA0003511324730000032
is yiThe number of samples of a class;
(2) based on the generated feature vector fiThe invention adopts a layer of linear layer as a classification module, and obtains a characteristic vector f through the classification moduleiAnd (4) predicting the probability of each corresponding class, and then training the network by taking the cross entropy as the loss until convergence.
The specific steps of the cost-sensitive cross entropy loss strategy are as follows:
(1) similar to step (1) of the class-balanced sampling strategy, except that the class-balanced sampling is replaced with random sampling, i.e., without modification to the original training data;
(2) similar to step (2) of the class-balanced sampling strategy, the difference is that cost-sensitive cross-entropy is used as the loss of training, i.e. normal cross-entropy loss is weighted, and the cost-sensitive cross-entropy loss is defined as follows:
Figure BDA0003511324730000033
wherein n isminIs the minimum of all the categories, LCEIs a cross entropy loss function.
Compared with the prior art, the invention has the advantages and technical effects that:
(1) aiming at the background knowledge of the SAR image, the invention provides a data amplification strategy facing the SAR image, which comprises random cutting and scaling, random turning and random rotation.
(2) Different from the traditional small sample training strategy, the method is based on contrast learning in machine learning, combines the thought of decoupling learning, adopts a two-stage training strategy, combines two advantages of contrast learning and a rebalance strategy, and has excellent performance on the SAR image data set. The method improves generalization performance of small sample SAR target identification, and improves the accuracy of the invention in identifying the target with small data volume.
(3) The method solves the problem that the generalization capability of the model to the small sample SAR target is poor due to data imbalance in practical problems, the network can learn the characterization capability with strong generalization from unbalanced data, the classification capability of the network is trained through the traditional rebalance strategy while the characterization capability of the network is maintained, and the decision boundary of the network is finely adjusted, so that the identification performance of the network to the small sample SAR target is further improved, the practicability is strong, the stability is good, the generalization performance of the deep learning model during small sample SAR target identification can be improved, the target identification precision of the small sample is improved, and the standard requirement of high SAR target identification precision under the conditions of unbalanced category, small sample data and the like is met.
Drawings
Fig. 1 is a schematic diagram of a two-stage contrast-based learning framework.
Fig. 2 is a schematic diagram of a small sample SAR data distribution.
Fig. 3 is a TSNE schematic.
Fig. 4 is a comparison graph of the effect of the classification module ablation experiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the following embodiments will further describe the present invention with reference to the accompanying drawings.
The invention provides a small sample SAR target recognition method based on a two-stage contrast learning framework. Through the two-stage training strategy, the accuracy of the head class of the model can be kept, and the recognition precision of the tail class small sample is improved, so that the overall performance of the model is improved.
The embodiment of the invention comprises the following steps:
A. a characterization learning stage;
B. a classification learning stage;
in order to deal with the difficulty that the generalization capability of the model to the small sample SAR target is poor due to data imbalance in the actual problem, the invention hopes that the network can learn the characterization capability with strong generalization from unbalanced data, and the classification capability of the network is trained through the traditional rebalancing strategy while the characterization capability of the network is maintained, so as to finely adjust the decision boundary of the network, thereby further improving the recognition performance of the network to the small sample SAR target. The training process is as shown in figure 1, and in the characterization learning stage, the invention adopts supervised contrast learning to train a representation module of a network on an original SAR data set; in the classification learning stage, the invention fixes the coding module and trains the classification module of the network by adopting a class balance sampling strategy or a cost sensitive cross entropy loss strategy. In order to more clearly describe the subsequent content, the total number of the classes is defined as C classes, and the number of SAR images trained in each round is N.
The specific steps of the step A are as follows:
although the actual SAR data set is unbalanced, the actual SAR data set comprises real associated information in the data, and a better characterization capability is beneficial to network learning. The invention hopes that the representation module trains and learns in the original SAR data, obtains a characterization capability with stronger generalization through real and accurate data distribution information, and avoids the representation module from learning wrong distribution information. In the step, the invention firstly adopts the most popular contrast Learning algorithm Supervised contrast Learning (Prannay Khosla, Piotr Terwak, Chen Wang, AaronSarna, Yonglong Tian, Phillip Isola, Aaron Maschinet, Ce Liu, and Dilip Krishan, "Supervised contrast Learning," arXiv preprintiv arXiv:2004.11362,2020) as the training strategy of the representation module, and designs a data enhancement strategy according to the characteristics of the SAR target, further optimizes the training effect of the contrast Learning, so that the representation module can extract the characteristics with higher generalization for the subsequent classification task. The method comprises the following specific steps:
the first step is as follows: the SAR image is obtained by imaging the synthetic aperture radar on a flight platform at a bird's eye view angle, and the angle of a target in the image is uncertain. In addition, missing part of the image information should also identify SAR targets. Therefore, the present invention is directed to the above situation, and designs the following data enhancement strategy: random cropping and scaling, random flipping and random rotation. And respectively applying a data enhancement strategy twice to each SAR image to obtain 2N SAR images.
The second step is that: respectively sending the 2N SAR images into a coding module (such as AlexNet, VGG, ResNet, EfficientNet) in the network to obtain a feature vector f corresponding to each SAR imagei. Then the feature vector fiSending the data to a projection module, mapping the data to another representation space to obtain a non-linear representation zi＝Proj(xi)＝W(2)σ(W(1)fi) Where σ is the ReLu activation function and W is a single linear layer. Therefore, the projection module adopted by the invention is realized by two layers of perceptrons with a hidden layer.
The third step: for the subsequent representation of the similarity between two vectors by dot multiplication, the invention uses
Figure BDA0003511324730000053
Norm to normalize zi. Then using normalized ziThe loss of supervised contrast learning is calculated, and the network is trained until convergence. The loss of supervised contrast learning is calculated as follows:
Figure BDA0003511324730000051
wherein P (i) represents a sum ofiAll sample indices of the same class are set, | P (i) | is the size of this set, τ is a temperature over-parameter.
The concrete steps of the step B are as follows:
the rebalance strategy can effectively improve the identification accuracy of the small sample SAR target, but the rebalance strategy can distort the original distribution information of data and destroy the learning of the model characterization capability. In order to better utilize the advantages of the rebalancing strategy and avoid the influence of the rebalancing strategy on the characterization capability, the invention fixes the coding module and trains the classification module through one of the rebalancing strategies. The rebalancing strategy adopted by the invention comprises the following steps: class-balanced sampling strategies and cost code-sensitive cross-entropy loss strategies.
The class balance sampling strategy comprises the following specific steps:
the first step is as follows: carrying out class balance sampling on the training set, then randomly selecting N SAR images, sending the N SAR images into a coding module to generate a characteristic vector fiMeanwhile, a gradient interruption strategy is adopted for the coding module, so that adverse effects generated by a subsequent training classification module are avoided. For class-balanced sampling, the probability that the ith sample is chosen is as follows:
Figure BDA0003511324730000052
wherein, yiIs a category of the ith sample,
Figure BDA0003511324730000061
is yiNumber of samples of a class.
The second step is that: based on the generated feature vector fiThe invention adopts a linear layer as a classification module, and obtains a characteristic vector f through the classification moduleiAnd (4) predicting the probability of each corresponding class, and then training the network by taking the cross entropy as the loss until convergence.
The specific steps of the cost-sensitive cross entropy loss strategy are as follows:
the first step is as follows: similar to the first step of class-balanced sampling, except that the class-balanced sampling is replaced with random sampling, i.e., without modification to the original training data.
The second step is that: similar to the second step of class-balanced sampling, except that cost-sensitive cross-entropy is used as the training penalty, i.e., the normal cross-entropy penalty is weighted. The cost sensitive cross-entropy loss is defined as follows:
Figure BDA0003511324730000062
wherein n isminIs the minimum of all the categories, LCEIs a cross entropy loss function.
Specific examples are given below.
1. In the characterization learning phase, the specific experimental setup is as follows: by usingSGD as optimizer, where momentum is 0.9 and weight decade is 1 × 10-4The batch size is 64 passes to train the feature network 1000. In addition, the initial learning rate is 0.05, and the learning rate is adjusted by adopting a cosine attenuation strategy. In addition, the temperature coefficient tau is set to be 0.1 and is used for supervising and contrasting learning loss function optimization.
2. In the classifier learning phase, the specific experimental details are as follows: the weight parameters of the coding module are fixed, the batch size is 128, and the learning rate is 0.01 to train the 100 rounds of the classifier. The learning rate in the training process is adjusted to be a cosine attenuation strategy, the optimizer also uses SGD, the parameter momentum is 0.9, and the weight decay is 5 multiplied by 10-5。
In order to verify that the method has better generalization performance on the small sample SAR data, the Exp distribution and the Step distribution are adopted to simulate the small sample data distribution, the specific distribution is shown in figure 2, and other traditional methods for solving the class imbalance problem are compared. Table 1 gives experimental comparisons of some prior works under small sample SAR data. As can be seen from the table, the present invention performs optimally as a whole compared to other methods. Particularly, the two-stage comparison learning framework can keep the accuracy of the head class and effectively improve the identification precision and the overall performance of the tail class small sample.
TABLE 1
Figure BDA0003511324730000063
In order to explore the influence of the rebalancing strategy on the coding module in the invention, different strategies are designed for training the coding module, and a fixed class balancing strategy is used for training the classification module. Table 2 shows the model identification accuracy under different strategy training under Exp distribution. As shown in the table, the baseline approach (without any strategy to directly train the coding module) is much better at tail class small samples and overall performance than the approach of adding the rebalancing strategy. This indicates that good features produce good classification, and the specific feature representation is shown in fig. 3. Specifically, the coding module directly trained by the raw data can obtain good feature representation, thereby improving the recognition performance of the model.
TABLE 2
Figure BDA0003511324730000071
In order to discuss the influence of the rebalancing strategy on the classification module in the invention, various strategies are designed to carry out experiments on the classification module. FIG. 4 shows the classification effect of the classification module under Exp distribution by adopting different strategies. As shown, the model can achieve better performance when the classification module employs a re-weighting or resampling method. This shows that the rebalancing strategy helps to change the decision boundary of the classification module, thereby improving the recognition performance of the model.
The invention improves the potential of contrast learning and obtains better SAR image representation capability. The method adopts a two-stage training strategy, combines two advantages of contrast learning and a rebalance strategy, and has excellent performance on the SAR image data set. The generalization performance of the small-sample SAR target recognition is improved, and the target recognition accuracy of the invention with small data volume is improved.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims (6)

1. The small sample SAR target identification method based on the two-stage comparison learning framework is characterized by comprising the following steps of:
1) in the characterization learning stage, a representation module of a network is trained on an original SAR data set by adopting supervised contrast learning;
2) and in the classification learning stage, the coding module is fixed and a classification module of the network is trained by adopting a rebalance strategy.
2. The small-sample SAR target recognition method based on the dual-stage contrast Learning framework as claimed in claim 1, characterized in that in step 1), the representation module adopting the Supervised contrast Learning to train the network on the original SAR data set first adopts a contrast Learning algorithm Supervised contrast Learning as a training strategy of the representation module, and designs a data enhancement strategy according to the characteristics of the SAR target, further optimizes the training effect of the contrast Learning, so that the representation module can extract the features with higher generalization for the subsequent classification task.
3. The method for identifying a small sample SAR target based on a two-stage contrast learning framework as claimed in claim 1, wherein in step 1), the specific steps of the characterization learning stage are as follows:
(1) the SAR image is obtained by imaging the synthetic aperture radar on a flight platform at a bird's-eye view angle, the angle of a target in the image is uncertain, and in order to identify the SAR target by missing partial image information, the following data enhancement strategy is designed: random cropping and scaling, random flipping and random rotation; respectively applying a data enhancement strategy twice to each SAR image to obtain 2N SAR images;
(2) respectively sending the 2N SAR images into a coding module in the network to obtain a characteristic vector f corresponding to each SAR imageiThen the feature vector f is divided intoiSending the data to a projection module, mapping the data to another representation space to obtain a non-linear representation zi＝Proj(xi)＝W(2)σ(W(1)fi) Where σ is the ReLu activation function and W is a single linear layer; the adopted projection module is realized by two layers of perceptrons with a hidden layer;
(3) use of2Norm normalization ziThen using normalized ziAnd calculating the loss of the supervised contrast learning, so as to train the network until convergence, wherein the loss of the supervised contrast learning is calculated as follows:
Figure FDA0003511324720000011
wherein P (i) represents a sum ofiAll sample subscript sets of the same class, | P (i) | is the size of this set, τ is a temperature over-parameter.
4. The method for identifying a small sample SAR target based on a two-stage contrast learning framework as claimed in claim 1, wherein in step 2), the rebalancing strategy is used for improving the identification precision of the small sample SAR target, the rebalancing strategy comprises a class-balanced sampling strategy and a cost-sensitive cross-entropy loss strategy, and one of the class-balanced sampling strategy and the cost-sensitive cross-entropy loss strategy is adopted as a classification module of a training network.
5. The small-sample SAR target recognition method based on the two-stage contrast learning framework as claimed in claim 4, characterized in that the class-balanced sampling strategy comprises the following specific steps:
(1) carrying out class balance sampling on the training set, randomly selecting N SAR images, sending the N SAR images into a coding module to generate a characteristic vector fiMeanwhile, a gradient interruption strategy is adopted for the coding module, so that adverse effects generated by a subsequent training classification module are avoided; for class-balanced sampling, the probability that the ith sample is chosen is as follows:
Figure FDA0003511324720000021
wherein, yiIs a category of the i-th sample,
Figure FDA0003511324720000022
is yiThe number of samples of a class;
(2) based on the generated feature vector fiThe invention adopts a layer of linear layer as a classification module, and obtains a characteristic vector f through the classification moduleiAnd (4) predicting the probability of each corresponding class, and then training the network by taking the cross entropy as the loss until convergence.
6. The small-sample SAR target recognition method based on the two-stage contrast learning framework as claimed in claim 4 or 5, characterized in that the cost-sensitive cross-entropy loss strategy comprises the following specific steps:
(1) similar to step (1) of the class-balanced sampling strategy, except that the class-balanced sampling is replaced with random sampling, i.e., without modification to the original training data;
(2) similar to step (2) of the class-balanced sampling strategy, the difference is that cost-sensitive cross-entropy is used as the loss of training, i.e. normal cross-entropy loss is weighted, and the cost-sensitive cross-entropy loss is defined as follows:
Figure FDA0003511324720000023
wherein n isminIs the minimum of all the categories, LCEIs a cross entropy loss function.
CN202210152954.9A 2022-02-18 2022-02-18 Small sample SAR target identification method based on two-stage comparison learning framework Pending CN114529767A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210152954.9A CN114529767A (en) 2022-02-18 2022-02-18 Small sample SAR target identification method based on two-stage comparison learning framework

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210152954.9A CN114529767A (en) 2022-02-18 2022-02-18 Small sample SAR target identification method based on two-stage comparison learning framework

Publications (1)

Publication Number Publication Date
CN114529767A true CN114529767A (en) 2022-05-24

Family

ID=81623114

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210152954.9A Pending CN114529767A (en) 2022-02-18 2022-02-18 Small sample SAR target identification method based on two-stage comparison learning framework

Country Status (1)

Country Link
CN (1) CN114529767A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114937288A (en) * 2022-06-21 2022-08-23 四川大学 Atypical class data set balancing method, device and medium
CN115409124A (en) * 2022-09-19 2022-11-29 小语智能信息科技(云南)有限公司 Small sample sensitive information identification method based on fine-tuning prototype network

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114937288A (en) * 2022-06-21 2022-08-23 四川大学 Atypical class data set balancing method, device and medium
CN115409124A (en) * 2022-09-19 2022-11-29 小语智能信息科技(云南)有限公司 Small sample sensitive information identification method based on fine-tuning prototype network
CN115409124B (en) * 2022-09-19 2023-05-23 小语智能信息科技(云南)有限公司 Small sample sensitive information identification method based on fine tuning prototype network

Similar Documents

Publication Publication Date Title
CN107862705B (en) Unmanned aerial vehicle small target detection method based on motion characteristics and deep learning characteristics
CN112308019B (en) SAR ship target detection method based on network pruning and knowledge distillation
Wang et al. Tropical cyclone intensity estimation from geostationary satellite imagery using deep convolutional neural networks
CN114529767A (en) Small sample SAR target identification method based on two-stage comparison learning framework
CN110132263B (en) Star map identification method based on representation learning
Shi et al. Objects detection of UAV for anti-UAV based on YOLOv4
CN110018453A (en) Intelligent type recognition methods based on aircraft track feature
Ye et al. CT-Net: An efficient network for low-altitude object detection based on convolution and transformer
CN111046756A (en) Convolutional neural network detection method for high-resolution remote sensing image target scale features
Liu et al. R2YOLOX: a lightweight refined anchor-free rotated detector for object detection in aerial images
CN111062310B (en) Few-sample unmanned aerial vehicle image identification method based on virtual sample generation
Li et al. Improved lightweight YOLOv5 using attention mechanism for satellite components recognition
Chen et al. Ship Detection with Optical Image Based on Attention and Loss Improved YOLO
Ozaki et al. DNN-based self-attitude estimation by learning landscape information
CN113627356A (en) Intelligent sea surface ship identification method
CN113111706A (en) SAR target feature unwrapping and identifying method for continuous missing of azimuth angle
Chen et al. Class-aware domain adaptation for coastal land cover mapping using optical remote sensing imagery
Li et al. I-gans for infrared image generation
Li et al. Remote sensing image scene classification model based on dual knowledge distillation
Zhai et al. MF‐SarNet: Effective CNN with data augmentation for SAR automatic target recognition
Yin et al. M2F2-RCNN: Multi-functional faster RCNN based on multi-scale feature fusion for region search in remote sensing images
CN109829385A (en) A kind of spectral patterns recognition methods of fine identification extraterrestrial target
CN113724325B (en) Multi-scene monocular camera pose regression method based on graph convolution network
Zhang et al. Flight target recognition via neural networks and information fusion
Thoreau et al. SaRNet: A dataset for deep learning assisted search and rescue with satellite imagery

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
CB03 Change of inventor or designer information

Inventor after: Huang Yue

Inventor after: Cai Senlin

Inventor after: Huang Lexing

Inventor after: Ding Xinghao

Inventor before: Ding Xinghao

Inventor before: Huang Yue

Inventor before: Huang Lexing

Inventor before: Cai Senlin

CB03 Change of inventor or designer information