CN109409442A - Convolutional neural networks model selection method in transfer learning - Google Patents
Convolutional neural networks model selection method in transfer learning Download PDFInfo
- Publication number
- CN109409442A CN109409442A CN201811396539.8A CN201811396539A CN109409442A CN 109409442 A CN109409442 A CN 109409442A CN 201811396539 A CN201811396539 A CN 201811396539A CN 109409442 A CN109409442 A CN 109409442A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- accuracy rate
- pretest
- task
- 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
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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
Transfer learning (Transfer Learning, TL) is the method handled using the existing training pattern of other field this field task, but existing network model is excessively abundant, causes confusion choosing Shi Zhongyi, is unfavorable for the completion of task.The invention proposes the methods selected in transfer learning convolutional neural networks (Convolutional Neural Networks, CNN) model.Key step are as follows: step 1 sets task object and Primary Reference index；Step 2 selects pretest model according to every million parameter accuracy rate of the CNN model under former training set；Step 3, pretest model are tested under task target detection collection, are obtained every million parameters accuracy rate, are selected pre-training model；Step 4, the fine tuning of pre-training model, the later training under task object training set；Step 5, pre-training model measurement see whether to meet target.The present invention can be widely used in image classification process field, such as the processing of low probability of intercept radar (Low Probability Intercept, LPI) image classification, medical conditions classification etc..
Description
Technical field
The present invention relates to the transfer learning in machine learning, the method for specifically a kind of convolutional neural networks model selection.
Background technique
Transfer learning (Transfer Learning, TL) is using the existing training pattern of other field to this field task
It is handled.Currently, various industries all begin to use transfer learning to solve the problems, such as, such as field of biomedicine, Haijun Lei
Et al identifies Hep-2 cell by transfer learning；In transport field, Javad Abbasi Aghamaleki et
Al identifies noise-containing ground traffic tools picture by transfer learning；In police field, Christian Galea
Et al et al. matches suspect by transfer learning and its relevant personage draws a portrait.But current existing deep learning net
Network model is abundant, is easy to appear confusion during choosing training pattern, is unfavorable for fast and efficiently completing task object.Cause
This, chooses corresponding network model for different task objects, for fast and efficiently completing task object with important
Realistic meaning and application value.
Convolutional neural networks (Convolution neural network, CNN) have been widely used in image classification,
With higher accuracy compared with conventional method.CNN is a kind of multilevel structure being made of Multilevel method unit, mainly includes
There are many convolution kernels, these convolution kernels respectively carry out input special in convolutional layer, pond layer and nonlinear transformation convolutional layer
Sign is extracted, and a variety of different features can be extracted.Pond layer screens feature, filters out more representational feature,
Simultaneously to input dimensionality reduction, reduce complexity.Nonlinear transformation mainly carries out nonlinear transformation to input, and it is empty to change feature representation
Between.Currently, CNN network structure develops to LeNet structure from two convolutional layers of beginning, to the system developed in recent years
State-of-art network is arranged, each network all has the characteristics that very outstanding.
The models such as existing CNN, such as AlexNet, VGG, Inception, ResNet are the training on ImageNet
Collection is trained, therefore very different with the target sample in many Practical Projects, cannot directly be used.And it is proposed by the present invention
A kind of convolutional neural networks model selection method based on every million parameters accuracy rate, by comparing every under former training dataset
Million parameter accuracys rate select the pre-training model for meeting task feature, cut to it, then under task target data set into
Row training, the convolutional neural networks model of most suitable task object is selected by comparing every million parameters accuracy rate.Energy of the present invention
The model for completing task object is fast and efficiently selected, is widely used in image classification process field, such as low probability of intercept thunder
Up to the processing of (Low Probability Intercept, LPI) image classification, medical conditions classification etc..
Summary of the invention
The problem to be solved in the present invention is: existing convolutional neural networks model is abundant, and is all specifically to train
It is trained under data set, when needing to apply in a certain particular task target, network model chooses difficulty, and needs to spend big
The time of amount, calculation power, manpower are one by one compared model, are unfavorable for fast and efficiently completing task object.
To solve the above problems, the invention provides the following technical scheme:
Present application example first aspect provides a kind of classification method that convolutional neural networks model is chosen, and specific steps are such as
Under:
Set task object and Primary Reference index: accuracy rate, number of parameters etc.；
Pretest model is tentatively chosen according to every million parameter accuracy rate of the CNN model under former training set；
Pretest model is tested under task target detection collection, is obtained every million parameters accuracy rate, is selected pre-training network
Model；
The fine tuning of pre-training model, the later training under task object training set；
Pre-training model measurement sees whether to meet target.
The first aspect of present application example, the accuracy rate in Primary Reference index refers in particular task target data set
The correct probability of lower picture classification.
The first aspect of present application example, every million parameters accuracy rate refers to the ratio between accuracy rate and number of parameters, wherein joining
Number unit of quantity is million, and formula is as follows:
The first aspect of present application example, in a kind of embodiment tentatively chosen to pretest model,
Using a variety of CNN models being trained on other data sets as the alternative collection of pretest model；Simultaneously
Set accuracy rate to the accuracy of the picture classification under former training dataset；Finally by every million parameter for comparing these models
Accuracy rate selects pretest model.
The first aspect of present application example, a kind of embodiment that pretest model is tested under task target detection collection
In,
For the test carried out under task target detection collection, cutting appropriate is done to pretest network architecture, it is main
It include: to remove the last layer of network, while freezing all parameters in network；One layer of new full articulamentum of addition, and
Species number to be identified is set by neuron number, and the weight of newly added full articulamentum is adjusted；Model is in office later
It is tested under business target detection collection, obtains every million parameters accuracy rate, be compared and select pre-training model.
The first aspect of present application example, it is selected according to every million parameter accuracy rate of the model under task target detection collection
Out in a kind of embodiment of pre-training model, the accuracy rate in every million parameters accuracy rate is revised accuracy rate.
The first aspect of present application example, in a kind of embodiment of pre-training network model fine tuning, only by network the
One layer of parameter is freezed, and is trained under task object training set；
The first aspect of present application example, in another embodiment of pre-training network model fine tuning, network is all
Parameter is not freezed, and is trained under task object training set；
Present application example second aspect provides a kind of accuracy rate calculation method, and this method is used in evaluation model performance
Be modified to accuracy rate: aiming at the problem that Known Species quantity is classified, when calculating accuracy rate, needing to remove does not make
With any method, pass through the influence for the accuracy rate that random guess obtains.
In a kind of embodiment of accuracy rate calculation method, following formula is can be used in the second aspect of present application example:
Accuracy rate=network class accuracy-random guess accuracy
Wherein, network class accuracy refers to the accuracy obtained using CNN network model when carrying out classification task.From
Above technical scheme and experimental result discovery, the embodiment of the present invention have the advantage that
The network model for being more suitable for task object can be found in current existing CNN model, and accuracy rate is higher, it can
Fast and efficiently complete task object.
Detailed description of the invention
Technological invention scheme in order to illustrate the embodiments of the present invention more clearly, below will be to embodiment or the prior art
Attached drawing needed in description briefly describes, it is therefore apparent that the accompanying drawings in the following description is only of the invention some
Embodiment for those of ordinary skill in the art without creative efforts, can also be attached according to these
Figure obtains other attached drawings.
Fig. 1 is that network model of the invention selects flow chart；
Fig. 2 is that the CNN model that the LPI radar waveform of the embodiment of the present invention identifies selects flow chart；
Fig. 3 is that the pretest model measurement of present example and pre-training model choose flow chart；
Fig. 4 time-frequency figure of each radar signal by PWVD processing when being the noiseless of present example；
Fig. 5 is classification accuracy and wave pattern of the 10 kinds of LPI signals of present example under different models；
Fig. 6 is the effect contrast figure of the MobileNetV2 and kongnet of present example；
Specific embodiment
It elaborates below to the embodiment of the present invention, the present embodiment carries out under the premise of the technical scheme of the present invention
Implement, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to following implementation
Example.
As shown in Fig. 2, the present embodiment key step includes: the first step, task object is set: to 10 kinds of low probability of intercept
(Low Probability Intercept, LPI) radar signal is classified, and classification accuracy is not less than 95%；Second
Step selects 5 from the model that MxNet is provided by comparing every million parameter accuracy rate of the model under ImageNet training set
Kind pretest model；Third step, 5 kinds of pretest models are tested under LPI radar test collection, obtain every million parameters accuracy rate, choosing
Select out pre-training model；4th step, the fine tuning of pre-training model row, the later training under LPI radar training collection；5th step, it is pre- to instruct
Practice model to test under LPI radar test collection, sees whether to meet expected setting target.Specific implementation step is as follows:
Step 1: setting task object: to 10 kinds of low probability of intercept (Low Probability Intercept, LPI) thunder
Classify up to signal, and classification accuracy is not less than 95%:
Step 1.1: at -10dB, -8dB, -6dB, -4dB, -2dB totally 5 kinds of LPI signal-to-noise ratio, 10 kinds of generation (BPSK,
FMCW, P1, P2, P3, P4, T1, T2, T3, T4) LPI radar signal initial data, and it is each under each signal-to-noise ratio environment
Kind signal has 1700 samples, and each signal randomly selects 70% for training, and 30% for testing.
Step 2: by comparing every million parameter accuracy rate of the model under ImageNet training set, from the mould of MxNet offer
5 kinds of pretest models are selected in type:
Step 2.1: by comparing the CNN model in MxNet applied to signal waveform classification under ImageNet data set
Every million parameters accuracy rate, select for the first time AlexNet, VGGNet-16, Inception v3, ResNet-50V2,
This 5 kinds of network structures of MobileNetV2-1.0 as pretest model, while using the kongnet model finely tuned as reference
Model；
Step 3:5 kind pretest model is tested under LPI radar test collection, is obtained every million parameters accuracy rate, is selected pre-
Training pattern:
Step 3.1: to 10 kinds of radar signal initial data processing sides PWVD under 5 kinds of signal-to-noise ratio in step 1.1
Method processing, obtains time-frequency image, as shown in figure 4, each radar signal is by PWVD treated time-frequency figure when Fig. 4 is noiseless.
Wherein PWVD processing method derives as follows:
Winger-Vile distribution (Winger-Vile Distribution, WVD) is one and is become with time and frequency for oneself
The three-dimensional function of the description signal amplitude of amount.One continuous one-dimensional WVD function are as follows:
In formula: x (t) is original signal, and t is time variable, and ω is angular frequency, and * indicates conjugation.Formula (1) shows the meter of WVD
It is non-causal at last.Therefore, which can not be used for actual WVD calculating.This limitation can be by by WVD analytic process
Middle adding window and improved, referred to as puppet Winger-Vile be distributed (Pseudo Winger-Vile Distribution, PWVD).
PWVD analysis to discrete signal are as follows:
In formula: ω (n) is the real window function and ω (0)=1 that a length is 2N-1.Use fl(n) kernel function, i.e. f are indicatedl
(n)=x (l+n) x*(l-n) then PWVD becomes ω (n) ω (- n)
The wherein selection (usually 2 of Nk, k is positive integer) it is very big on the operand of PWVD and the influence of time-frequency resolving power.By
Formula (3) is it is found that big N value can obtain high time-frequency resolving power, to generate more smooth as a result, wherein this example uses
N be 1024.Step 3.2: signal does adapting operation to data after PWVD is converted and is mapped: the figure for first changing time-frequency
As replicating one time in each channel of RGB, the image of a triple channel is formed.Then it is wanted according to the use of pre-training model
It asks, by each of the image in tri- channels RGB channel normalization to [0,1] section, and according to the requirement in each channel,
Image is carried out regular.It wherein, is 3*299*299 except Inception-v3 requires the image dimension of input, other model needs
The image dimension of input is 3*224*224；
Step 3.3: 5 kinds of pretest models in step 2 being cut: the last layer of network being removed, is added
Add one layer of new full articulamentum, and set species number to be identified for neuron number, while freezing the parameter value in primitive network,
Only the weight of newly added full articulamentum is adjusted；
Step 3.4: the LPI radar signal of the pretest model and reference model that cut in different signal-to-noise ratio is tested
Collection is tested, and is repeated test 20 times, is obtained accuracy rate and fluctuation, concrete condition is as shown in Figure 5, wherein subject to ordinate
True rate, abscissa are signal-to-noise ratio；
Step 3.5: observation and analysis chart 5 are it is found that in all pretest models, and MobileNetV2 is for extremely low noise
Discrimination than signal waveform is only second to Inception-v3, and the signal waveform discrimination compared with high s/n ratio is only second to
AlexNet, while the stability bandwidth of MobileNetV2 is not high, and every million parameters accuracy rate is highest in all pre-training models.
Comprehensively consider, selects MobileNetV2 as pre-training model.
Overall process is as shown in Figure 3.
Step 4: the fine tuning of pre-training model, the training under LPI radar training collection later:
Step 4.1: MobileNetV2 being finely adjusted: removing the last layer pond layer of MobileNetV2, and freezes
The convolution kernel of first layer convolutional layer；
Step 4.2: the MobileNetV2 finely tuned training under LPI radar signal training set.
Step 5: pre-training model is tested under LPI radar test collection, sees whether to meet expected setting target:
Step 5.1: the MobileNetV2 and reference model kongnet trained is surveyed under LPI radar signal test set
Examination, test results are shown in figure 6；
Step 5.2: observation and analysis chart 6, for the LPI signal in the case of -10dB, MobileNetV2 classification is quasi- for discovery
True rate ratio kongnet is higher by about 30%, and especially to bpsk signal, classifying quality improves nearly 40%, simultaneously for signal-to-noise ratio
Signal higher than -8dB, for MobileNetV2 discrimination close to 100%, overall effect reaches expected setting target.
Claims (5)
1. convolutional neural networks (Convolutional Neural Networks, CNN) model selects in a kind of transfer learning
Method characterized by comprising
Set task object and Primary Reference index: accuracy rate, number of parameters etc.；
Pretest model is tentatively chosen according to every million parameter accuracy rate of the CNN model under former training set；
Pretest model is tested under task target detection collection, is obtained every million parameters accuracy rate, is selected pre-training model；
The fine tuning of pre-training model, the later training under task object training set；
Pre-training model measurement sees whether to meet target.
2. the method according to claim 1, wherein
Accuracy rate in Primary Reference index refers to the correct probability of the picture classification under particular task target data set；
Every million parameters accuracy rate under former training set refers to the ratio between accuracy rate and number of parameters, and every million refer to number of parameters unit,
Accuracy rate refers to the correct probability of the picture classification under former training dataset；
Pretest model selects to be a variety of CNN models that will be trained on other data sets as pretest model
Alternative collection, while setting accuracy rate to the accuracy of the picture classification under former training dataset, by comparing these models
Every million parameters accuracy rate, selects pretest model；
Test of the pretest model under task target detection collection needs to cut model structure, specifically includes that network
The last layer remove, add one layer of new full articulamentum, and set species number to be identified for neuron number, while freezing original
Parameter value in beginning network is only adjusted the weight of newly added full articulamentum；Model is in task target detection collection later
Lower test obtains every million parameters accuracy rate, is compared and selects pre-training model；
There are two types of the fine tunings of pre-training model: only freezing the parameter of network first tier, the trained or net under task object training set
All parameters of network are not freezed, the training under task object training set.
3. according to the method described in claim 2, it is characterized in that,
It is revised accuracy rate that every million parameters accuracy rate, which selects the accuracy rate in pre-training model,.
4. according to the method described in claim 3, it is characterized in that,
When calculating accuracy rate, need to remove the influence of the accuracy rate obtained when without using any method by random guess,
Revised accuracy rate is the difference of network class accuracy and random guess accuracy.
5. according to the method described in claim 4, it is characterized in that,
Network class accuracy refers to the correct probability obtained using CNN network model when carrying out classification task.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811396539.8A CN109409442A (en) | 2018-11-21 | 2018-11-21 | Convolutional neural networks model selection method in transfer learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811396539.8A CN109409442A (en) | 2018-11-21 | 2018-11-21 | Convolutional neural networks model selection method in transfer learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109409442A true CN109409442A (en) | 2019-03-01 |
Family
ID=65474339
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811396539.8A Pending CN109409442A (en) | 2018-11-21 | 2018-11-21 | Convolutional neural networks model selection method in transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109409442A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110220709A (en) * | 2019-06-06 | 2019-09-10 | 北京科技大学 | Fault Diagnosis of Roller Bearings based on CNN model and transfer learning |
| CN110457274A (en) * | 2019-08-14 | 2019-11-15 | 北京思图场景数据科技服务有限公司 | A kind of data file processing method based on transfer learning, device, equipment and computer storage medium |
| CN111582236A (en) * | 2020-05-27 | 2020-08-25 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN112070535A (en) * | 2020-09-03 | 2020-12-11 | 常州微亿智造科技有限公司 | Electric vehicle price prediction method and device |
| CN112434462A (en) * | 2020-10-21 | 2021-03-02 | 华为技术有限公司 | Model obtaining method and device |
| CN114298278A (en) * | 2021-12-28 | 2022-04-08 | 河北工业大学 | Electric equipment performance prediction method based on pre-training model |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107451661A (en) * | 2017-06-29 | 2017-12-08 | 西安电子科技大学 | A kind of neutral net transfer learning method based on virtual image data collection |
| CN107742061A (en) * | 2017-09-19 | 2018-02-27 | 中山大学 | A kind of prediction of protein-protein interaction mthods, systems and devices |
| US20180114114A1 (en) * | 2016-10-21 | 2018-04-26 | Nvidia Corporation | Systems and methods for pruning neural networks for resource efficient inference |
| CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
| CN108596138A (en) * | 2018-05-03 | 2018-09-28 | 南京大学 | A kind of face identification method based on migration hierarchical network |
| CN108647702A (en) * | 2018-04-13 | 2018-10-12 | 湖南大学 | A kind of extensive food materials image classification method based on transfer learning |
| CN108818537A (en) * | 2018-07-13 | 2018-11-16 | 南京工程学院 | A kind of robot industry method for sorting based on cloud deep learning |
-
2018
- 2018-11-21 CN CN201811396539.8A patent/CN109409442A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180114114A1 (en) * | 2016-10-21 | 2018-04-26 | Nvidia Corporation | Systems and methods for pruning neural networks for resource efficient inference |
| CN107451661A (en) * | 2017-06-29 | 2017-12-08 | 西安电子科技大学 | A kind of neutral net transfer learning method based on virtual image data collection |
| CN107742061A (en) * | 2017-09-19 | 2018-02-27 | 中山大学 | A kind of prediction of protein-protein interaction mthods, systems and devices |
| CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
| CN108647702A (en) * | 2018-04-13 | 2018-10-12 | 湖南大学 | A kind of extensive food materials image classification method based on transfer learning |
| CN108596138A (en) * | 2018-05-03 | 2018-09-28 | 南京大学 | A kind of face identification method based on migration hierarchical network |
| CN108818537A (en) * | 2018-07-13 | 2018-11-16 | 南京工程学院 | A kind of robot industry method for sorting based on cloud deep learning |
Non-Patent Citations (6)
| Title |
|---|
| BONELEE: "一文看懂迁移学习：怎样用预训练模型搞定深度学习？——重用神经网络的结构", 《网络在线公开：HTTPS://WWW.CNBLOGS.COM/BONELEE/P/8921311.HTML》 * |
| DISHASHREE GUPTA: "Transfer learning and the art of using Pre-trained Models in Deep Learning", 《网络在线公开: HTTPS://WWW.ANALYTICSVIDHYA.COM/BLOG/2017/06/TRANSFER-LEARNING-THE-ART-OF-FINE-TUNING-A-PRE-TRAINED-MODEL/#》 * |
| HAIJUNLEI 等: "A deeply supervised residual network for HEp-2 cell classification via cross-modal transfer learning", 《PATTERN RECOGNITION》 * |
| S. H. HASANPOUR 等: "Let’s keep it simple, Using simple architectures to outperform deeper and more complex architectures", 《ARXIV平台公开：HTTPS://ARXIV.ORG/ABS/1608.06037》 * |
| 王秋然: "低信噪比低截获雷达信号的调制方式识别方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 龙满生 等: "基于卷积神经网络与迁移学习的油茶病害图像识别", 《农业工程学报》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110220709A (en) * | 2019-06-06 | 2019-09-10 | 北京科技大学 | Fault Diagnosis of Roller Bearings based on CNN model and transfer learning |
| CN110220709B (en) * | 2019-06-06 | 2020-04-21 | 北京科技大学 | Rolling bearing fault diagnosis method based on CNN model and transfer learning |
| CN110457274A (en) * | 2019-08-14 | 2019-11-15 | 北京思图场景数据科技服务有限公司 | A kind of data file processing method based on transfer learning, device, equipment and computer storage medium |
| CN111582236A (en) * | 2020-05-27 | 2020-08-25 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN111582236B (en) * | 2020-05-27 | 2022-08-02 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN112070535A (en) * | 2020-09-03 | 2020-12-11 | 常州微亿智造科技有限公司 | Electric vehicle price prediction method and device |
| CN112434462A (en) * | 2020-10-21 | 2021-03-02 | 华为技术有限公司 | Model obtaining method and device |
| WO2022083624A1 (en) * | 2020-10-21 | 2022-04-28 | 华为技术有限公司 | Model acquisition method, and device |
| CN112434462B (en) * | 2020-10-21 | 2024-07-09 | 华为技术有限公司 | Method and equipment for obtaining model |
| CN114298278A (en) * | 2021-12-28 | 2022-04-08 | 河北工业大学 | Electric equipment performance prediction method based on pre-training model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109409442A (en) | Convolutional neural networks model selection method in transfer learning | |
| CN112052755B (en) | Semantic convolution hyperspectral image classification method based on multipath attention mechanism | |
| CN110533631B (en) | SAR image change detection method based on pyramid pooling twin network | |
| Dreissigacker et al. | Deep-learning continuous gravitational waves | |
| CN114937151B (en) | Lightweight target detection method based on multiple receptive fields and attention feature pyramid | |
| CN110321963A (en) | Based on the hyperspectral image classification method for merging multiple dimensioned multidimensional sky spectrum signature | |
| CN110361778B (en) | Seismic data reconstruction method based on generation countermeasure network | |
| CN111723701B (en) | Underwater target identification method | |
| CN108446312B (en) | Optical remote sensing image retrieval method based on deep convolution semantic net | |
| CN110728656A (en) | Meta-learning-based no-reference image quality data processing method and intelligent terminal | |
| CN104732240A (en) | Hyperspectral image waveband selecting method applying neural network to carry out sensitivity analysis | |
| CN103984746B (en) | Based on the SAR image recognition methodss that semisupervised classification and region distance are estimated | |
| CN112818777B (en) | Remote sensing image target detection method based on dense connection and feature enhancement | |
| CN116047427B (en) | Small sample radar active interference identification method | |
| CN111562597A (en) | Beidou satellite navigation interference source identification method based on BP neural network | |
| CN112288700A (en) | Rail defect detection method | |
| CN106097290A (en) | SAR image change detection based on NMF image co-registration | |
| CN118332443B (en) | Pulsar data radio frequency interference detection method | |
| CN115565019A (en) | Single-channel high-resolution SAR image ground object classification method based on deep self-supervision generation countermeasure | |
| CN116299684B (en) | Novel microseismic classification method based on bimodal neurons in artificial neural network | |
| CN117523394A (en) | SAR vessel detection method based on aggregation characteristic enhancement network | |
| CN109460788B (en) | Hyperspectral image classification method based on low-rank-sparse information combination network | |
| CN116973677A (en) | Distribution network single-phase earth fault line selection method based on cavity convolution and attention mechanism | |
| CN116884435A (en) | Voice event detection method and device based on audio prompt learning | |
| CN116797928A (en) | SAR target increment classification method based on stability and plasticity of balance model |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190301 |