CN110427875A - Infrared image object detection method based on depth migration study and extreme learning machine - Google Patents
Infrared image object detection method based on depth migration study and extreme learning machine Download PDFInfo
- Publication number
- CN110427875A CN110427875A CN201910704660.0A CN201910704660A CN110427875A CN 110427875 A CN110427875 A CN 110427875A CN 201910704660 A CN201910704660 A CN 201910704660A CN 110427875 A CN110427875 A CN 110427875A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- sample
- aiming field
- migration
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Photometry And Measurement Of Optical Pulse Characteristics (AREA)
Abstract
The present invention relates to a kind of infrared image object detection methods based on depth migration study and extreme learning machine, include the following steps: visible images target detection model training, framework is detected using maskrcnn two-stage multitask, the training on visible light sample set D, mask exposure mask is inputted in neural network, the loss function of overall network structure is redefined；Based on the method for sample migration, by expanding aiming field, i.e., the distribution of infrared sample set T obtains the data set of transfer learning；Based on the method for model migration, the pre-training model by the target detection model with high accuracy based on visible images as the data set of the transfer learning of above-mentioned generation is trained using frame identical with visible light target detection；The full articulamentum of network is replaced using extreme learning machine, overcomes the over-fitting of small sample model transfer training.
Description
Technical field
The invention belongs to field of image detection, are related to the object detection method of infrared image.
Background technique
Small-sample learning is the research hotspot of deep learning at present, there is a large amount of unmarked images, detection in real world
The realization of task relies on a large amount of labeled data, considerably increases time and money cost.Traditional machine learning method exists
One serious shortcomings: assuming that training data and test data obey identical data distribution, but in many cases and it is unsatisfactory for this
Kind is assumed, it usually needs spends a large amount of manpower and resource to mark mass data again to meet training requirement, causes data
Waste；And transfer learning can extract from available data and migrate knowledge, for completing new learning tasks.As engineering
The branch of habit, transfer learning original intention is to save the time of artificial mark sample, in recent years due to the rapid hair of deep neural network
Exhibition, transfer learning are combined with neural network more and more, and high resource utilization attracts with the characteristics of lower trained cost
Academia and industry have carried out many correlative studys.
Transfer learning can be divided into according to concrete methods of realizing: sample migration, feature migration and model migration.When source domain and mesh
When the data in mark domain are very close, sample migration can effectively solve the problems, such as that aiming field sample is insufficient；Feature migrates across weight
Structure feature finds the shared potential feature space of source domain and aiming field to minimize the difference between field；Source is worked as in model migration
Department pattern distribution or Study first can be shared when domain sample is similar to aiming field sample distribution, between learning tasks.
More popular remote domain target migration at present, by intermediate field have been carried out face to aircraft transfer learning target
Detection, provides a solution route for deep learning labeled data difficulty.
Summary of the invention
The object of the present invention is to provide a kind of methods of infrared target detection based on depth migration study, and being primarily based on can
Light-exposed image and two-stage detection framework, training obtain the high-precision target detection model based on visible images；So
It is migrated afterwards by the sample of transfer learning, model migration various ways use parallel, it will be seen that the detection model of light migrates application
Onto small sample infrared image, the method for being eventually adding extreme learning machine improves infrared image in the case where avoiding over-fitting
Detection accuracy.Its technical solution is as follows:
1. a kind of infrared image object detection method based on depth migration study and extreme learning machine, including following step
It is rapid:
Step 1: visible images target detection model training, detects framework, In using maskrcnn two-stage multitask
Training on visible light sample set D, mask exposure mask is inputted in neural network, the loss function of overall network structure is redefined:
L=Lcls+Lbox+Lmask (1)
LclsAnd LboxRespectively represent error in classification and position error, and LmaskFor pixel error.
Using ResneXt101 network, by the backpropagation mode of stochastic gradient descent, obtain based on visible images
Target detection model D with high accuracymodel。
Second step, based on the method for sample migration, by expanding aiming field, i.e., the distribution of infrared sample set T is used
AdaBoosT assigns sample difference weight to filter out the samples different from aiming field in source domain, the source domain, that is, visible light
Sample set D weights sample in source domain again to constitute the distribution for being similar to aiming field, finally, by using source domain is come from
Weight again example and the original instance from aiming field carry out training pattern, finally obtain the data set T of transfer learningnew:
Tnew=T+Dpart (2)
Third step, based on the method for model migration, by the target detection model D with high accuracy based on visible imagesmodel
As the data set T of the transfer learning of above-mentioned generationnewPre-training model, using frame identical with visible light target detection,
It is trained, using based on L2The insertion loss function of norm, training obtain occurring the T of over-fitting situationmodel, but be not optimal
Model.
4th step needs to reconstruct the last one multi-categorizer layer, using the limit when adjusting model parameter in aiming field
Habit machine replaces the full articulamentum of network, overcomes the over-fitting of small sample model transfer training, with truncation gradient method (dropout)
Study of the deep neural network model parameter in aiming field in transfer learning is assisted, to implicitly help the mind being activated
Through meta learning, while limiting the learning of neuron not being activated.
The image detecting method of deep learning migration based on extreme learning machine of the invention, passes through a large amount of visible images
Sample carries out pre-training to convolutional neural networks, uses AdaBoost algorithm picks visible light sample by the method that sample migrates
The image for concentrating distribution similar to aiming field sample, is put into visible light model as the infrared image small sample set of pre-training model
In, pre-training network structure is merged with ELM (extreme learning machine) structure then, reduces network training process to sample number
The demand of amount, while directly may lead hairnet network using the rare training data progress retraining of target domain to overcome
Overfitting problem, the simple method using a kind of novel, limit the parameter in target domain can studying space, make its
As close possible to the parameter learnt in source domain in adaptive process.Feature of the invention: 1) expand infrared small sample distribution；
2) infrared migration models strong robustness, detection accuracy is high, is not susceptible to over-fitting.
Detailed description of the invention
Fig. 1 two-stage detects framework overall flow figure
Fig. 2 transfer learning schematic diagram
Fig. 3 extreme learning machine (ELM) structure chart
The infrared sample transfer learning overall flow figure of Fig. 4
Specific embodiment
By retrieval, it is found that the patent of infrared small sample target detection is less, this patent learns from transfer learning and the limit
Infrared small sample problem is solved in terms of machine.In real world visible images in occupation of image distribution absolute ratio, at present
The detection of deep learning image object, segmentation, the tasks such as tracking are built upon mostly on the visible images largely marked, but can
Light-exposed image detects accuracy of identification under night, the situations such as complicated weather and substantially reduces, the spy that infrared image passes through energy imaging
Point has filled up the defect of visible light significantly, but infrared sample marks higher cost, lack in real world largely mark it is infrared
Data set.By visible light to infrared migration, the target detection precision to infrared image is improved.
This patent be it is a kind of based on depth migration study infrared target detection method, be primarily based on visible images and
Two-stage detection framework, training obtain the high-precision target detection model based on visible images；Then pass through migration
The sample of study migrates, and model migration various ways use parallel, it will be seen that it is red that the detection model migration of light is applied to small sample
On outer image, the method for being eventually adding extreme learning machine improves the detection essence of infrared image in the case where avoiding over-fitting
Degree.Scheme is as follows:
The first step, it is seen that light model training
This step is the purpose is to use the visible light sample image training objective detection model largely marked, using the two-stage
Detection framework maskrcnn (deep learning image detection algorithm two-stage mainstream framework), is enhanced, feature pyramid using data,
Model Fusion etc., in visible images domain, training detection accuracy is higher, the visible detection model of model strong robustness.Based on two
The detection framework in stage, in addition the Detection task of mask (mask), finer positioning target position, use the side ROIAlign
Method eliminates integer operation, remains decimal, is added after maskhead, algorithm whole loss function becomes following formula:
L=Lcls+Lbox+Lmask
LclsAnd LboxRespectively represent error in classification and position error.And LmaskAssuming that a shared k classification, then mask points
The output dimension for cutting branch is k*m*m, and for each point in m*m, all exporting k two-value mask, (each classification is used
Sigmoid exports successive value).When calculating loss, which class which belongs to, and the sigmoid output of which class will just be counted
Calculate loss.
Second step, sample migration
This step is moved the purpose is to obtain visible images similar with infrared sample image by the depth that sample migrates
Shifting method, chooses the visible images sample set for matching infrared small sample image, and this method is based on AdaBoost (Integrated Algorithm)
Sample difference weight is assigned to filter out the samples different from aiming field (infrared image) in source domain (visible images), in source
Sample is weighted again in domain to constitute the distribution for being similar to aiming field.Finally, by using the weighting sample again from source domain
This and the original sample from aiming field carry out training pattern.
It specifically can formal definitions are as follows: source domain Ds, originating task Ts, aiming field Dt, domain goal task Tt.One kind is selected first
Base learner (base learner), then according to the performance of some base learners before, to train current base learner, and
The adjustment sample weights of iteration.If Y is classification space, aiming field has identical data distribution space, i.e. target numeric field data, is denoted as X_
S, source domain are different distribution spaces, i.e. source domain data, are denoted as X_d.Entire training data space is training data T
S={ (xi)}xi∈Xs(i=1,2 ..., k)
Wherein training set T can be divided into the data T_s of the data T_d from different distributions and same distribution,
Firstly, normalizing the weight of each data, a distribution is become
Then Weak Classifier is called.Integrally it regard the data of T_d and T_s as training data, this step is also source domain data
The place worked to model.
Error rate ε is calculated againt.Only calculate the data extracted in T_s, that is, target numeric field data.Source domain data do not enter
It calculates.And it needs again to normalize the extraction data weighting in T_s when calculating error rate.
Calculate separately the rate of T_s and T_d weight adjustment.The weight adjustment rate of iteration each time, T_s is all different,
It and in T_d is the same.Final execute most has strategy, selection and the most like source domain data sample of aiming field, with aiming field
Sample is combined into a new sample set.
Third step, model migration
The purpose is to the model obtained through the above steps and sample sets for this step, migrate the preceding two layers of mesh of visible light model
Network framework, the new sample set of re -training.It will be seen that the training pattern of light, moves on infrared sample set, sentences to learn to have
The feature of other property, we introduce insertion loss function.Insertion loss function can not only make representative characteristic criterion
Change, and the variance in each individual class can be effectively reduced, while expanding the variance between Different Individual class.Based on L2
The insertion loss function form of norm
F is the output of full articulamentum, they can be regarded as respectively from picture XiAnd XjThe representative spy of middle extraction
Sign.yij=1 and yij=0 respectively represents whether two pictures belong to the same individual.Wherein hyper parameter m refers to spacing parameter.
Total target loss function is as follows:
Loss front portion is the loss function of cross entropy, and rear portion is L2 loss function, the model damage redefined
Lose function, can better specification features so that the infrared sample set that visible light model adaptation is new.
4th step, extreme learning machine replace the full articulamentum of network
After being trained in source domain to deep neural network model, intended in next step it will be seen that the migration of optical mode shape parameter
To aiming field.But the sample distribution in target domain may and source domain in data distribution have significantly deviation and
Training data sample is rare in aiming field, and directly carrying out retraining using the rare training data of target domain may cause
The over-fitting of network, and collapse dept neural network model is in the argument structure of source domain.
This step retains the parameter information W learnt in source domains, the parameter W in source domain middle school is constrained on aiming fieldsIn mesh
Mark is finely tuned in field.Over-fitting in order to prevent, and reduce difference of the source domain in aiming field in samples pictures distribution, this step
Full articulamentum is replaced using extreme learning machine, passes through the parameter W learnt in constraint source domainsStudying space to target lead
Fine tuning obtains actual parameter in domain, wherein | | Wt-Ws| | L can be used1Or L2Normal form.
J(Wt,bt)=Loss+ ∈ | | Wt-Ws||2
This method limits the parameter W in target domaintCan studying space, make its during adaptive as far as possible
The parameter W learnt close in source domains.Rare sample is recycled to be trained, WtIt can be carried out in reasonable variation range
Fine tuning.
When adjusting model in aiming field, since the classification number of picture in the classification number of picture in source domain and aiming field may
Difference, so needing to reconstruct the last one multi-categorizer layer.The decline of model gradient updates as follows: L1:J(Wt,bt)=Loss+ | |
Wt-Ws||
In above-mentioned formula, whenWhen, the element W in W matrixijValue be 1, remaining situation WijValue be 0.
In addition, the selective activation performance of neuron, this step assists moving using truncation gradient method (dropout)
Study of the deep neural network model parameter in aiming field in study is moved, to implicitly help the neuron being activated
It practises, while the learning of neuron that limitation is not activated.Gradient decline updates:
Pass through limited model parameter WtStudying space in target domain, and the neuron for only allowing to be activated passes through
Gradient calligraphy learning is truncated, the above method is conducive to collaboration and carries out adaptive learning of the parameter in aiming field.
In conjunction with the embodiments, if infrared sample set is T, it is seen that light sample set is D, and step of the invention is as follows:
Step 1: visible images target detection model training, detailed process is as shown in Figure 1, use maskrcnn second order
Section multitask detects framework, and mask exposure mask is inputted in neural network, redefines entirety by the training on visible light sample set D
The loss function of network structure:
L=Lcls+Lbox+Lmask (1)
LclsAnd LboxRespectively represent error in classification and position error.And LmaskFor pixel error.
Using ResneXt101 network, by the backpropagation mode of stochastic gradient descent, obtain based on visible images
Target detection model D with high accuracymodel。
Second step, based on the method for sample migration, as shown in Fig. 2, passing through point for expanding aiming field (infrared sample set T)
It is different from aiming field in source domain (visible light sample set D) to filter out to assign sample difference weight using AdaBoost for cloth
Sample weights sample in source domain again to constitute the distribution for being similar to aiming field.Finally, by using the weight from source domain
New weighting example and the original instance from aiming field carry out training pattern.Finally obtain new sample set:
Tnew=T+Dpart (2)
Third step, the method based on model migration, it will be seen that the model D of lightmodelAs the transfer learning of above-mentioned generation
Data set TnewPre-training model as shown in Figure 1, be trained, used using frame identical with visible light target detection
Insertion loss function can not only make representative characteristic criterion, but also can be effectively reduced in each individual class
Variance, while expanding the variance between Different Individual class.Based on L2The insertion loss function form of norm:
Total target loss function:
The training of this step obtains occurring the T of over-fitting situationmodel, but be not optimal models.
4th step, when adjusting model parameter in aiming field, due to picture in the classification number of picture in source domain and aiming field
Classification number may be different, so needing to reconstruct the last one multi-categorizer layer.It is replaced using extreme learning machine is (as shown in Figure 3)
The full articulamentum of network, overcomes the over-fitting of small sample model transfer training, and overall flow figure is as shown in figure 4, its model
It is updated as its gradient declines:
L1:J(Wt,bt)=Loss+ | | Wt-Ws|| (5)
Extreme learning machine replaces the full articulamentum of migration models network, effectively avoids infrared sample set because of negligible amounts,
With overfitting problem caused by the reasons such as visible images variance is larger, it is strong to obtain model generalization ability, detection accuracy and model
The high T of robustnessmodel。
Claims (1)
1. a kind of infrared image object detection method based on depth migration study and extreme learning machine, including the following steps:
Step 1: visible images target detection model training, detects framework using maskrcnn two-stage multitask, visible
Training on light sample set D, mask exposure mask is inputted in neural network, the loss function of overall network structure is redefined:
L=Lcls+Lbox+Lmask (1)
LclsAnd LboxRespectively represent error in classification and position error, and LmaskFor pixel error；
The essence based on visible images is obtained by the backpropagation mode of stochastic gradient descent using ResneXt101 network
Spend high target detection model Dmodel；
Second step, based on the method for sample migration, by expanding aiming field, i.e., the distribution of infrared sample set T uses AdaBoost
Sample difference weight is assigned to filter out the samples different from aiming field in source domain, the source domain, that is, visible light sample set D,
Sample is weighted again in source domain to constitute the distribution for being similar to aiming field, finally, by using adding again from source domain
Power example and the original instance from aiming field carry out training pattern, finally obtain the data set T of transfer learningnew:
Tnew=T+Dpart (2)
Third step, based on the method for model migration, by the target detection model D with high accuracy based on visible imagesmodelAs
The data set T of the transfer learning of above-mentioned generationnewPre-training model carried out using frame identical with visible light target detection
Training, using based on L2The insertion loss function of norm, training obtain occurring the T of over-fitting situationmodel, but be not optimal mould
Type；
4th step needs to reconstruct the last one multi-categorizer layer, using extreme learning machine when adjusting model parameter in aiming field
Instead of the full articulamentum of network, the over-fitting of small sample model transfer training is overcome, with truncation gradient method (dropout) come auxiliary
Study of the deep neural network model parameter in aiming field in transfer learning is helped, to implicitly help the neuron being activated
Study, while the learning of neuron that limitation is not activated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910704660.0A CN110427875B (en) | 2019-07-31 | 2019-07-31 | Infrared image target detection method based on deep migration learning and extreme learning machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910704660.0A CN110427875B (en) | 2019-07-31 | 2019-07-31 | Infrared image target detection method based on deep migration learning and extreme learning machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110427875A true CN110427875A (en) | 2019-11-08 |
| CN110427875B CN110427875B (en) | 2022-11-11 |
Family
ID=68413609
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910704660.0A Active CN110427875B (en) | 2019-07-31 | 2019-07-31 | Infrared image target detection method based on deep migration learning and extreme learning machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110427875B (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111339836A (en) * | 2020-02-07 | 2020-06-26 | 天津大学 | SAR image ship target detection method based on transfer learning |
| CN111401454A (en) * | 2020-03-19 | 2020-07-10 | 创新奇智(重庆)科技有限公司 | Few-sample target identification method based on transfer learning |
| CN111582477A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111723823A (en) * | 2020-06-24 | 2020-09-29 | 河南科技学院 | Underwater target detection method based on third-party transfer learning |
| CN111721536A (en) * | 2020-07-20 | 2020-09-29 | 哈尔滨理工大学 | Rolling bearing fault diagnosis method for improving model migration strategy |
| CN111767868A (en) * | 2020-06-30 | 2020-10-13 | 创新奇智(北京)科技有限公司 | Face detection method and device, electronic equipment and storage medium |
| CN111882055A (en) * | 2020-06-15 | 2020-11-03 | 电子科技大学 | Method for constructing target detection self-adaptive model based on cycleGAN and pseudo label |
| CN111915566A (en) * | 2020-07-03 | 2020-11-10 | 天津大学 | Infrared sample target detection method based on cyclic consistency countermeasure network |
| CN112287839A (en) * | 2020-10-29 | 2021-01-29 | 广西科技大学 | SSD infrared image pedestrian detection method based on transfer learning |
| CN112506667A (en) * | 2020-12-22 | 2021-03-16 | 北京航空航天大学杭州创新研究院 | Deep neural network training method based on multi-task optimization |
| CN112949387A (en) * | 2021-01-27 | 2021-06-11 | 西安电子科技大学 | Intelligent anti-interference target detection method based on transfer learning |
| CN113076895A (en) * | 2021-04-09 | 2021-07-06 | 太原理工大学 | Conveyor belt longitudinal damage vibration sensing method based on infrared computer vision |
| CN113159300A (en) * | 2021-05-15 | 2021-07-23 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
| CN113221993A (en) * | 2021-05-06 | 2021-08-06 | 西安电子科技大学 | Large-view-field small-sample target detection method based on meta-learning and cross-stage hourglass |
| CN113344868A (en) * | 2021-05-28 | 2021-09-03 | 山东大学 | Label-free cell classification screening system based on mixed transfer learning |
| CN113537403A (en) * | 2021-08-14 | 2021-10-22 | 北京达佳互联信息技术有限公司 | Training method and device and prediction method and device of image processing model |
| CN113627541A (en) * | 2021-08-13 | 2021-11-09 | 北京邮电大学 | Light path transmission quality prediction method based on sample migration screening |
| CN113762466A (en) * | 2021-08-02 | 2021-12-07 | 国网河南省电力公司信息通信公司 | Electric power internet of things flow classification method and device |
| CN114783072A (en) * | 2022-03-17 | 2022-07-22 | 哈尔滨工业大学（威海） | Image identification method based on remote domain transfer learning |
| CN115037641A (en) * | 2022-06-01 | 2022-09-09 | 网络通信与安全紫金山实验室 | Network traffic detection method and device based on small samples, electronic equipment and medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104657596A (en) * | 2015-01-27 | 2015-05-27 | 中国矿业大学 | Model-transfer-based large-sized new compressor performance prediction rapid-modeling method |
| CN106096004A (en) * | 2016-06-23 | 2016-11-09 | 北京工业大学 | A kind of method setting up extensive cross-domain texts emotional orientation analysis framework |
| CN107463954A (en) * | 2017-07-21 | 2017-12-12 | 华中科技大学 | A kind of template matches recognition methods for obscuring different spectrogram picture |
| US20180341832A1 (en) * | 2017-05-24 | 2018-11-29 | Hrl Laboratories, Llc | Transfer learning of convolutional neural networks from visible color (rbg) to infrared (ir) domain |
| CN109308483A (en) * | 2018-07-11 | 2019-02-05 | 南京航空航天大学 | Double source image characteristics extraction and fusion identification method based on convolutional neural networks |
| CN109583482A (en) * | 2018-11-13 | 2019-04-05 | 河海大学 | A kind of infrared human body target image identification method based on multiple features fusion Yu multicore transfer learning |
-
2019
- 2019-07-31 CN CN201910704660.0A patent/CN110427875B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104657596A (en) * | 2015-01-27 | 2015-05-27 | 中国矿业大学 | Model-transfer-based large-sized new compressor performance prediction rapid-modeling method |
| CN106096004A (en) * | 2016-06-23 | 2016-11-09 | 北京工业大学 | A kind of method setting up extensive cross-domain texts emotional orientation analysis framework |
| US20180341832A1 (en) * | 2017-05-24 | 2018-11-29 | Hrl Laboratories, Llc | Transfer learning of convolutional neural networks from visible color (rbg) to infrared (ir) domain |
| CN107463954A (en) * | 2017-07-21 | 2017-12-12 | 华中科技大学 | A kind of template matches recognition methods for obscuring different spectrogram picture |
| CN109308483A (en) * | 2018-07-11 | 2019-02-05 | 南京航空航天大学 | Double source image characteristics extraction and fusion identification method based on convolutional neural networks |
| CN109583482A (en) * | 2018-11-13 | 2019-04-05 | 河海大学 | A kind of infrared human body target image identification method based on multiple features fusion Yu multicore transfer learning |
Non-Patent Citations (7)
| Title |
|---|
| QIAN WENG等: ""Land-Use Classification via Extreme Learning Classifier Based on Deep Convolutional Features"", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
| WENYUAN DAI等: ""Boosting for Transfer Learning"", 《PROCEEDINGS OF THE 24TH INTERNATIONAL CONFERENCE ON MACHINE LEARNING》 * |
| 曹越: ""基于迁移学习的小样本人脸识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 朱大炜: ""基于深度学习的红外图像飞机目标检测方法"", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
| 林鸿生等: ""基于迁移学习的红外图像多目标检测技术"", 《红外》 * |
| 闫河等: ""基于深度 CNN 和极限学习机相结合的实时文档分类"", 《计算机应用与软件》 * |
| 陆鑫伟: ""基于迁移学习的雷达辐射源识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111339836A (en) * | 2020-02-07 | 2020-06-26 | 天津大学 | SAR image ship target detection method based on transfer learning |
| CN111401454A (en) * | 2020-03-19 | 2020-07-10 | 创新奇智(重庆)科技有限公司 | Few-sample target identification method based on transfer learning |
| CN111582477A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111582477B (en) * | 2020-05-09 | 2023-08-29 | 北京百度网讯科技有限公司 | Training method and device for neural network model |
| CN111882055B (en) * | 2020-06-15 | 2022-08-05 | 电子科技大学 | Method for constructing target detection self-adaptive model based on cycleGAN and pseudo label |
| CN111882055A (en) * | 2020-06-15 | 2020-11-03 | 电子科技大学 | Method for constructing target detection self-adaptive model based on cycleGAN and pseudo label |
| CN111723823A (en) * | 2020-06-24 | 2020-09-29 | 河南科技学院 | Underwater target detection method based on third-party transfer learning |
| CN111767868B (en) * | 2020-06-30 | 2024-06-11 | 创新奇智(北京)科技有限公司 | Face detection method and device, electronic equipment and storage medium |
| CN111767868A (en) * | 2020-06-30 | 2020-10-13 | 创新奇智(北京)科技有限公司 | Face detection method and device, electronic equipment and storage medium |
| CN111915566B (en) * | 2020-07-03 | 2022-03-15 | 天津大学 | Infrared sample target detection method based on cyclic consistency countermeasure network |
| CN111915566A (en) * | 2020-07-03 | 2020-11-10 | 天津大学 | Infrared sample target detection method based on cyclic consistency countermeasure network |
| CN111721536A (en) * | 2020-07-20 | 2020-09-29 | 哈尔滨理工大学 | Rolling bearing fault diagnosis method for improving model migration strategy |
| CN112287839A (en) * | 2020-10-29 | 2021-01-29 | 广西科技大学 | SSD infrared image pedestrian detection method based on transfer learning |
| CN112506667A (en) * | 2020-12-22 | 2021-03-16 | 北京航空航天大学杭州创新研究院 | Deep neural network training method based on multi-task optimization |
| CN112949387A (en) * | 2021-01-27 | 2021-06-11 | 西安电子科技大学 | Intelligent anti-interference target detection method based on transfer learning |
| CN112949387B (en) * | 2021-01-27 | 2024-02-09 | 西安电子科技大学 | Intelligent anti-interference target detection method based on transfer learning |
| CN113076895A (en) * | 2021-04-09 | 2021-07-06 | 太原理工大学 | Conveyor belt longitudinal damage vibration sensing method based on infrared computer vision |
| CN113076895B (en) * | 2021-04-09 | 2022-08-02 | 太原理工大学 | Conveyor belt longitudinal damage vibration sensing method based on infrared computer vision |
| CN113221993A (en) * | 2021-05-06 | 2021-08-06 | 西安电子科技大学 | Large-view-field small-sample target detection method based on meta-learning and cross-stage hourglass |
| CN113221993B (en) * | 2021-05-06 | 2023-08-01 | 西安电子科技大学 | Large-view-field small-sample target detection method based on meta-learning and cross-stage hourglass |
| CN113159300B (en) * | 2021-05-15 | 2024-02-27 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
| CN113159300A (en) * | 2021-05-15 | 2021-07-23 | 南京逸智网络空间技术创新研究院有限公司 | Image detection neural network model, training method thereof and image detection method |
| CN113344868A (en) * | 2021-05-28 | 2021-09-03 | 山东大学 | Label-free cell classification screening system based on mixed transfer learning |
| CN113344868B (en) * | 2021-05-28 | 2023-08-25 | 山东大学 | Label-free cell classification screening system based on mixed transfer learning |
| CN113762466B (en) * | 2021-08-02 | 2023-06-20 | 国网河南省电力公司信息通信公司 | Electric power internet of things flow classification method and device |
| CN113762466A (en) * | 2021-08-02 | 2021-12-07 | 国网河南省电力公司信息通信公司 | Electric power internet of things flow classification method and device |
| CN113627541B (en) * | 2021-08-13 | 2023-07-21 | 北京邮电大学 | Optical path transmission quality prediction method based on sample migration screening |
| CN113627541A (en) * | 2021-08-13 | 2021-11-09 | 北京邮电大学 | Light path transmission quality prediction method based on sample migration screening |
| CN113537403A (en) * | 2021-08-14 | 2021-10-22 | 北京达佳互联信息技术有限公司 | Training method and device and prediction method and device of image processing model |
| CN114783072B (en) * | 2022-03-17 | 2022-12-30 | 哈尔滨工业大学（威海） | Image identification method based on remote domain transfer learning |
| CN114783072A (en) * | 2022-03-17 | 2022-07-22 | 哈尔滨工业大学（威海） | Image identification method based on remote domain transfer learning |
| CN115037641A (en) * | 2022-06-01 | 2022-09-09 | 网络通信与安全紫金山实验室 | Network traffic detection method and device based on small samples, electronic equipment and medium |
| CN115037641B (en) * | 2022-06-01 | 2024-05-03 | 网络通信与安全紫金山实验室 | Network traffic detection method and device based on small sample, electronic equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110427875B (en) | 2022-11-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110427875A (en) | Infrared image object detection method based on depth migration study and extreme learning machine | |
| CN110457982B (en) | Crop disease image identification method based on feature migration learning | |
| CN106650813B (en) | A kind of image understanding method based on depth residual error network and LSTM | |
| CN109558942B (en) | Neural network migration method based on shallow learning | |
| CN103345656B (en) | A kind of data identification method based on multitask deep neural network and device | |
| CN110580496A (en) | Deep migration learning system and method based on entropy minimization | |
| CN108984745A (en) | A kind of neural network file classification method merging more knowledge mappings | |
| CN110377686A (en) | A kind of address information Feature Extraction Method based on deep neural network model | |
| CN109829541A (en) | Deep neural network incremental training method and system based on learning automaton | |
| CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
| CN110135295A (en) | A kind of unsupervised pedestrian recognition methods again based on transfer learning | |
| CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
| CN110348624B (en) | Sand storm grade prediction method based on Stacking integration strategy | |
| CN107644235A (en) | Image automatic annotation method based on semi-supervised learning | |
| CN110516095A (en) | Weakly supervised depth Hash social activity image search method and system based on semanteme migration | |
| CN107194433A (en) | A kind of Radar range profile's target identification method based on depth autoencoder network | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN109697469A (en) | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint | |
| CN106203318A (en) | The camera network pedestrian recognition method merged based on multi-level depth characteristic | |
| CN105701507A (en) | Image classification method based on dynamic random pooling convolution neural network | |
| CN109919205A (en) | Based on bull from the convolution echo state network timing classification method of attention mechanism | |
| CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
| CN106446927A (en) | Self-paced reinforcement image classification method and system | |
| CN107133955A (en) | A kind of collaboration conspicuousness detection method combined at many levels | |
| CN102314614A (en) | Image semantics classification method based on class-shared multiple kernel learning (MKL) |
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 |