CN109828251A - Radar target identification method based on feature pyramid light weight convolutional neural networks - Google Patents
Radar target identification method based on feature pyramid light weight convolutional neural networks Download PDFInfo
- Publication number
- CN109828251A CN109828251A CN201910173559.7A CN201910173559A CN109828251A CN 109828251 A CN109828251 A CN 109828251A CN 201910173559 A CN201910173559 A CN 201910173559A CN 109828251 A CN109828251 A CN 109828251A
- Authority
- CN
- China
- Prior art keywords
- convolution
- depth
- radar target
- feature
- hrrp
- 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
Abstract
The invention belongs to radar target automatic identification technologies to provide a kind of radar target identification method based on feature pyramid light weight convolutional neural networks for the target HRRP feature extraction and classification and identification under Low SNR.To take into account target HRRP local message and global information, it is inputted using the multichannel of HRRP multi-scale Representation building model, and convolutional layer is separated based on depth and designs light weight convolutional neural networks, effectively reduce number of parameters, improve Generalization Capability, the fusion of feature pyramid is added simultaneously to extract the robust features of target, improves model stability.
Description
Technical field
The invention belongs to radar target automatic identification technology, under Low SNR target HRRP feature extraction and
Classification and identification provides a kind of radar target identification method based on feature pyramid light weight convolutional neural networks.
Background technique
In recent years, deep learning related algorithm is widely used in computer vision field, such as target detection and positioning, mesh
Mark classification and identification etc..Deep learning algorithm obtains required model using data-driven, and the further feature extracted is more preferable because of energy
Characterization target essential information and have good robustness.
HRRP (High Resolution Range Profiles) includes a large amount of object constructions, target scattering point intensity etc.
Information is based primarily upon stack from encoding model applied to the deep learning algorithm of radar HRRP target identification technology at present.It is self-editing
Code is a kind of unsupervised feature extracting method, cannot utilize target category label information well, and because stack certainly adopt by coding
It is successively trained with greedy algorithm, the feature extracted is easy to produce Problem of Failure with number of plies increase.To solve the above problems, this
Invention, which proposes, a kind of has supervision target identification method based on convolutional neural networks.
Conventional depth convolutional network carries out target classification and identification merely with most further output feature, and depth network
Every layer of output is target signature, and shallow-layer feature includes the information such as profile and edge more, and further feature is mostly high-level semantics information.
For the feature for making full use of each layer to extract, the present invention uses for reference scale invariant feature and becomes on the basis of conventional depth neural network
(Scale-invariant feature transform, SIFT) feature extracting method is changed, is proposed a kind of based on feature gold word
Tower merges the radar target identification method of light weight convolutional neural networks, and this method is more using HRRP multi-scale Representation building model
Channel input to take into account target HRRP local message and global information, and separates convolutional layer design light weight volume based on depth
Product neural network, effectively reduces number of parameters, improves model generalization performance, while the fusion of feature pyramid is added to extract target
Robust features, improve model stability.
Summary of the invention
It is an object of the present invention to which the problem low for the discrimination of HRRP under Low SNR, provides one kind and is based on
The radar target identification method of feature pyramid light weight convolutional neural networks, improves the robustness and Generalization Capability of algorithm.
Technical solution of the invention are as follows: the HRRP data based on multiscale space construction multichannel, building depth can
Separate convolution characteristic extracting module, establish be based on feature pyramid light weight convolutional neural networks, using training sample to model into
Row training, obtains end-to-end HRRP radar target recognition model.
To achieve the above object, the present invention realizes that steps are as follows:
Step 1: the HRRP data based on multiscale space construction multichannel；
Step 2: building depth separates convolution characteristic extracting module, establishes feature pyramid light weight convolutional neural networks,
And model parameter is initialized；
Step 3: propagated forward calculates loss function in iterative process；
Step 4: back-propagating is updated the parameter in model using chain rule；
Step 5: repeating step 2,3 until loss function convergence, obtain the model that can be used for radar target recognition.
The present invention has the following technical effect that compared with prior art
(1) institute's climbing form type is the end-to-end model of data-driven, and the model after training can automatically extract the deep layer of target
Feature.
(2) mentioned method is inputted using the multichannel that the multi-scale Representation of HRRP constitutes model, to guarantee institute's climbing form type
It is extracting target detail feature simultaneously, is taking into account object construction feature, helping to extract robust features.
(3) compared with traditional convolutional neural networks, convolutional layer is separated based on depth and devises light weight convolutional neural networks,
Reduce number of parameters, improves computational efficiency and model generalization.
(4) institute's climbing form type takes full advantage of each layer feature, improves institute's climbing form type by the method that feature pyramid merges
Robustness and convergence rate.
Detailed description of the invention
Fig. 1: it is based on feature pyramid light weight convolutional neural networks model framework chart；
Fig. 2: the convolution process schematic diagram in single depth convolutional layer；
Fig. 3: depth separates convolution characteristic extracting module structure chart；
Fig. 4: feature pyramid fusion structure schematic diagram.
Specific embodiment
Below in conjunction with Figure of description, present invention is further described in detail.Referring to Figure of description, institute's structure of the present invention
The specification of a model built is as follows:
Then the multi-scale Representation for obtaining HRRP the invention firstly uses gaussian kernel function is designed as the input of model
Construction feature pyramid light weight convolutional neural networks model is for extracting target signature and carrying out Classification and Identification to target.Wherein institute
It mentions convolutional neural networks model to be mainly made of four separable convolution characteristic extracting modules, number is respectively module 1,2,3,4.
Wherein the output of first three module is respectively as the input of branch 1,2,3, and each branch is using depth convolution to input progress
It is down-sampled.The output of branch 1,2,3 obtains a feature vector by parallel with the output of module 4, using convolution to the spy
Each channel characteristics of sign vector are merged, and obtained feature vector is launched into the full articulamentum of one-dimensional vector Yu next layer
It is attached, is finally exported by output layer as a result, the general frame of institute's climbing form type is as shown in Figure 1.DC table in module in figure
Show that depth convolutional layer, PC indicate that point convolutional layer, P indicate pond layer, step-length is 2.In order to guarantee the output vector of branch 1,2,3
Length is identical as the output vector length of module 4, and the step-length of the depth convolutional layer of branch 1,2,3 is respectively 8,4,2.FC indicates complete
Articulamentum, neuron number 50, full articulamentum is traditional neural network, for making classifier.Institute's climbing form type is that end is arrived
End, that is, it enters data into trained model, can directly export recognition result.
Mentioned method will be described in detail and be analyzed in terms of four below: 1, the multi-scale Representation of HRRP data
Method, 2, depth separate convolution characteristic extracting module building, 3, feature pyramid fusion specific method, 4, institute's climbing form type
With the computation complexity of other depth models.
The 1 HRRP multichannel building based on multiscale space
The multi-scale Representation of signal can be obtained by carrying out convolution using the Gauss checking signal of different parameters.In lesser scale
Under, signal includes more detailed information (local feature), conversely, then including more structure features, therefore, is believed with single scale
It number compares, the multi-scale information for comprehensively utilizing signal is easier to obtain the substantive characteristics of signal.Existing research achievement usually only mentions
Take the feature of original scale HRRP data, characteristic present it is mostly be HRRP local detail information, easily ignore the overall situation of HRRP
Information reduces Generalization Capability.Gaussian kernel is to realize the unique linear core of multi-scale Representation, by carrying out Gaussian kernel volume to signal
Product, can abandon signal section detailed information with the radio-frequency component in trap signal, and the global characteristics of signal will not be because of Gauss
Convolution and change.Therefore convolution operation is carried out to HRRP data with Gaussian kernel herein to obtain the multi-scale Representation of HRRP,
The wherein Convolution Formula of Gaussian kernel and HRRP are as follows:
Wherein, G (x, σ) is one-dimensional gaussian kernel function, and a indicates the amplitude of Gaussian kernel, and σ indicates the scale parameter of Gaussian kernel, I
(x) input signal is indicated, L (x, σ) is the signal after Gaussian Blur.
The present invention has selected the Gaussian kernel of three different scales, and amplitude a=1, width 3, scale σ is respectively σ0, σ021/2, σ022/2, wherein σ0Equal to 1.After the HRRP of different scale is done normalized, by composing in parallel Three-channel data
As mode input.
2 separable convolution characteristic extracting modules
Model in the present invention has used a kind of new characteristic extracting module, which separates convolutional layer and pond by depth
Change layer to be composed, wherein depth separates convolutional layer and is mainly responsible for feature extraction, and pond layer, which is mainly responsible for, reduces the superfluous of feature
Remaining.Depth separates convolution and one complete convolution algorithm is resolved into the progress of two steps, i.e. depth convolution sum point convolution.Depth
Convolution is responsible for extracting the feature of each input channel, and point convolution is responsible for merging the feature in each channel, passes through the combination of the two
To in feature spatial information and depth information carry out uncoupling.Fig. 2 describes depth convolution in single depth convolutional layer
Specific operation process wherein input the one-dimensional vector for 3 channels, the size of the convolution kernel of depth convolutional layer is 3 × 1, depth volume
It is long-pending with traditional convolution different IPs, the port number of convolution kernel is always 1 in depth convolutional layer, and each convolution is only led to the single of input
Road carries out convolution algorithm, therefore the port number of the output feature of depth convolutional layer is identical as the port number of input vector.
Depth convolution only carries out independent convolution algorithm to each channel of input, but will be not different in same spatial location
The characteristic information in channel is merged, it is therefore desirable to put convolution and each channel characteristics are combined into new feature.Point convolution is tradition
The special case of convolution, calculating process is identical as traditional convolution, and the size of convolution kernel is fixed as 1 × 1, to multi-channel feature vector into
Row point convolution, which is equivalent to, sums each channel weighting of feature to obtain new feature vector.
The number of parameters that depth separates convolutional layer is the parameter summation of depth convolution sum point convolution, it is assumed that convolution kernel is (deep
Spend convolution) size be nk× 1, the port number of input/output feature is respectively ni, no, then depth separates convolution layer parameter
Number is nink+nino。
Depth separates the composition of convolution characteristic extracting module as shown in figure 3, comprising a depth convolutional layer, and a point is rolled up
Lamination, a pond layer.The size of depth convolution kernel is fixed as 3 × 1, and number is identical as corresponding input vector port number.Pond
The step-length for changing layer is 2, and for feature vector after pond layer (pooling) is down-sampled, port number is constant, and dimension becomes original two points
One of.
3 feature pyramid fusion methods
Traditional convolutional neural networks are made of convolutional layer, pond layer, full articulamentum etc., each volume of convolutional neural networks
The feature that lamination extracts is target signature, and what shallow-layer convolutional layer extracted is mostly the low orders information such as objective contour, edge, deep layer volume
What lamination extracted is mostly high-order semantic information.Traditional convolutional neural networks merely with the feature that the last one convolutional layer extracts come into
Row target identification, and more advanced semantic feature is obtained by the depth of increase layer, to improve correct recognition rata, but at this
During a, the feature of other each layers is not fully used.In order to make full use of each layer feature, a kind of feature gold is proposed
Word tower fusion method, feature pyramid fusion structure schematic diagram are as shown in Figure 4.
Level 1~4 respectively represents the output feature vector of module 1~4 in Fig. 4, the feature vector of level i+1 by
The feature vector of level i through module i convolution and it is down-sampled obtain, vector dimension be level i half, therefore, level
1~4 combination of eigenvectors can be referred to as to be characterized pyramid.The specific method is as follows for the fusion of feature pyramid:
(1) feature vector of feature pyramid level 1~3 is down-sampled, make its characteristic dimension and 4 feature dimensions of level
It spends identical.Due to directly using pond layer to shallow-layer feature carry out it is down-sampled, be easy to cause the loss of part effective information, because
This, it is down-sampled to the feature vector of level1~3 using depth convolution herein, as shown in Figure 1, the corresponding depth of branch 1,2,3
Convolution kernel size is respectively 9 × 1,5 × 1,3 × 1, and step-length is respectively 8,4,2.
(2) multi-channel feature of each level is in parallel, if the port number of 1~4 feature of level is respectively c1~c4, in parallel
The channel of feature vector is afterwards
(3) it is merged using each channel of the convolution to vector after parallel connection.The full connection of vector input that fusion is obtained
Last recognition result can be obtained.
The comparison of 4 model structures is analyzed with computation complexity
The number of parameters of institute's climbing form type and traditional convolutional neural networks is far smaller than from encoding model.Wherein, institute's climbing form
Type is less than traditional convolutional neural networks because using depth to separate its number of parameters of convolutional layer.
Assuming that training sample number is N, the characteristic dimension that outputs and inputs of single layer is respectively D and T, then from encoding model
Single layer computation complexity is O (NDT), and wherein DT is the number of parameters of hidden layer；For traditional convolutional neural networks, it is assumed that convolution
Core size is nk, the port number of input and output is respectively niAnd no, the computation complexity of convolutional layer is O (NDnknino), nkninoFor
The number of parameters of convolutional layer；For institute's climbing form type, it is assumed that the convolution kernel size of depth convolutional layer is nk, the port number of input and output
Respectively niAnd no, then it is O (ND (n that depth, which separates the computation complexity of convolutional layer,kni+nino)), wherein nkni+ninoFor depth
Spend the number of parameters of separable convolutional layer.Usual nknino> nkni+nino> T, and the neural network number of plies based on convolution kernel is more
In from encoding model, therefore corresponding computation complexity is greater than from encoding model.
In conclusion the number of parameters of institute's climbing form type is less, computation complexity is low, belongs to light weight convolutional neural networks, tool
There is preferable Generalization Capability.
Claims (7)
1. a kind of radar target identification method based on feature pyramid light weight convolutional neural networks, which is characterized in that including with
Lower step:
Step 1, the HRRP data based on multiscale space construction multichannel；
Step 2, building depth separates convolution characteristic extracting module, establishes feature pyramid light weight convolutional neural networks, and right
Model parameter is initialized；
Step 3, propagated forward calculates loss function in iterative process；
Step 4, back-propagating is updated the parameter in model using chain rule；
Step 5, step 3,4 are repeated until loss function convergence, obtain the model that can be used for radar target recognition.
2. radar target identification method as described in claim 1, which is characterized in that step 1 specifically: using different scale
Gaussian kernel obtains the multi-scale Representation of HRRP to HRRP data progress convolution operation, wherein the Convolution Formula of Gaussian kernel and HRRP
Are as follows:
Wherein, G (x, σ) is one-dimensional gaussian kernel function, and a indicates the amplitude of Gaussian kernel, and σ indicates the scale parameter of Gaussian kernel, I (x)
Indicate input signal,Convolution algorithm is represented, L (x, σ) is the signal after Gaussian Blur, and the HRRP of different scale is normalized
After processing, by composing in parallel multi-channel data as mode input.
3. radar target identification method as described in claim 1, which is characterized in that it is special that the depth in step 2 separates convolution
Levying extraction module includes a depth convolutional layer, convolutional layer, a pond layer.
4. radar target identification method as claimed in claim 3, which is characterized in that the convolution kernel size in depth convolutional layer is solid
It is set to 3 × 1, number is identical as corresponding input vector port number.
5. radar target identification method as claimed in claim 3, which is characterized in that the step-length of pond layer is 2.
6. radar target identification method as described in claim 1, which is characterized in that be based on feature pyramid light weight in step 2
The method for building up of convolutional neural networks specifically: the neural network separates convolution characteristic extracting module by 4 depth and forms,
Wherein the output of first three module forms new spy together with the output-parallel of the 4th module respectively after a branch
Vector is levied, after being merged each channel of feature vector with convolutional layer, utilizes the non-thread classification of full articulamentum and output layer composition
Device obtains classification recognition result.
7. radar target identification method as claimed in claim 6, which is characterized in that branch is made of a depth convolutional layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910173559.7A CN109828251B (en) | 2019-03-07 | 2019-03-07 | Radar target identification method based on characteristic pyramid light-weight convolution neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910173559.7A CN109828251B (en) | 2019-03-07 | 2019-03-07 | Radar target identification method based on characteristic pyramid light-weight convolution neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109828251A true CN109828251A (en) | 2019-05-31 |
| CN109828251B CN109828251B (en) | 2022-07-12 |
Family
ID=66865703
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910173559.7A Active CN109828251B (en) | 2019-03-07 | 2019-03-07 | Radar target identification method based on characteristic pyramid light-weight convolution neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109828251B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111091091A (en) * | 2019-12-16 | 2020-05-01 | 北京迈格威科技有限公司 | Method, device and equipment for extracting target object re-identification features and storage medium |
| CN111582095A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111914993A (en) * | 2020-06-05 | 2020-11-10 | 西安理工大学 | Multi-scale light-weight deep convolution neural network model based on non-uniform grouping |
| CN112114295A (en) * | 2020-07-29 | 2020-12-22 | 中国人民解放军海军航空大学 | Target identification method and system for full-polarization radar |
| CN112230210A (en) * | 2020-09-09 | 2021-01-15 | 南昌航空大学 | HRRP radar target identification method based on improved LSGAN and CNN |
| CN112287784A (en) * | 2020-10-20 | 2021-01-29 | 哈尔滨工程大学 | Radar signal classification method based on deep convolutional neural network and feature fusion |
| CN112580660A (en) * | 2020-11-17 | 2021-03-30 | 上海闻泰信息技术有限公司 | Image processing method, image processing device, computer equipment and readable storage medium |
| CN112882010A (en) * | 2021-01-12 | 2021-06-01 | 西安电子科技大学 | High-resolution range profile target identification method based on signal-to-noise ratio field knowledge network |
| CN112966667A (en) * | 2021-04-06 | 2021-06-15 | 中国人民解放军海军航空大学 | Method for identifying one-dimensional distance image noise reduction convolution neural network of sea surface target |
| CN113376610A (en) * | 2021-06-22 | 2021-09-10 | 西安电子科技大学 | Narrow-band radar target detection method based on signal structure information |
| CN113627401A (en) * | 2021-10-12 | 2021-11-09 | 四川大学 | Myoelectric gesture recognition method of feature pyramid network fused with double-attention machine system |
| CN113959071A (en) * | 2021-07-21 | 2022-01-21 | 北京金茂绿建科技有限公司 | Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system |
| CN116091854A (en) * | 2022-12-14 | 2023-05-09 | 中国人民解放军空军预警学院 | Method and system for classifying space targets of HRRP sequence |
| CN113640764B (en) * | 2021-08-09 | 2023-08-11 | 中国人民解放军海军航空大学航空作战勤务学院 | Radar one-dimensional range profile identification method and device based on multi-dimension one-dimensional convolution |
| CN116593980A (en) * | 2023-04-20 | 2023-08-15 | 中国人民解放军93209部队 | Radar target recognition model training method, radar target recognition method and device |
| CN117574136A (en) * | 2024-01-16 | 2024-02-20 | 浙江大学海南研究院 | Convolutional neural network calculation method based on multi-element Gaussian function space transformation |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105116397A (en) * | 2015-08-25 | 2015-12-02 | 西安电子科技大学 | Radar high-resolution range profile target recognition method based on MMFA model |
| CN106909936A (en) * | 2017-01-21 | 2017-06-30 | 江苏大学 | A kind of vehicle checking method based on double vehicle deformable part models |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| WO2018120740A1 (en) * | 2016-12-29 | 2018-07-05 | 深圳光启合众科技有限公司 | Picture classification method, device and robot |
| CN108427920A (en) * | 2018-02-26 | 2018-08-21 | 杭州电子科技大学 | A kind of land and sea border defense object detection method based on deep learning |
| US10083171B1 (en) * | 2017-08-03 | 2018-09-25 | Gyrfalcon Technology Inc. | Natural language processing using a CNN based integrated circuit |
| CN109063666A (en) * | 2018-08-14 | 2018-12-21 | 电子科技大学 | The lightweight face identification method and system of convolution are separated based on depth |
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
-
2019
- 2019-03-07 CN CN201910173559.7A patent/CN109828251B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105116397A (en) * | 2015-08-25 | 2015-12-02 | 西安电子科技大学 | Radar high-resolution range profile target recognition method based on MMFA model |
| WO2018120740A1 (en) * | 2016-12-29 | 2018-07-05 | 深圳光启合众科技有限公司 | Picture classification method, device and robot |
| CN106909936A (en) * | 2017-01-21 | 2017-06-30 | 江苏大学 | A kind of vehicle checking method based on double vehicle deformable part models |
| US10083171B1 (en) * | 2017-08-03 | 2018-09-25 | Gyrfalcon Technology Inc. | Natural language processing using a CNN based integrated circuit |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| CN108427920A (en) * | 2018-02-26 | 2018-08-21 | 杭州电子科技大学 | A kind of land and sea border defense object detection method based on deep learning |
| CN109063666A (en) * | 2018-08-14 | 2018-12-21 | 电子科技大学 | The lightweight face identification method and system of convolution are separated based on depth |
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
Non-Patent Citations (5)
| Title |
|---|
| A. P. WITKIN: "《Scale-space filtering》", 《PROC. INT. JOINT CONF. ARTIF. INTELL.》 * |
| LIANG-CHIEH CHEN, YUKUN ZHU: "《Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation》", 《PROCEEDINGS OF THE EUROPEAN CONFERENCE ON COMPUTER VISION (ECCV)》 * |
| 任之俊; 蔺素珍: "《基于改进特征金字塔的Mask R-CNN目标检测方法》", 《激光与光电子学进展》 * |
| 杨俊: "《基于卷积神经网络的目标检测研究》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 柳超; 张华: "《面向识别与定位的轻量化卷积神经网络》", 《工业控制计算机》 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111091091A (en) * | 2019-12-16 | 2020-05-01 | 北京迈格威科技有限公司 | Method, device and equipment for extracting target object re-identification features and storage medium |
| CN111582095B (en) * | 2020-04-27 | 2022-02-01 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111582095A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Light-weight rapid detection method for abnormal behaviors of pedestrians |
| CN111914993A (en) * | 2020-06-05 | 2020-11-10 | 西安理工大学 | Multi-scale light-weight deep convolution neural network model based on non-uniform grouping |
| CN111914993B (en) * | 2020-06-05 | 2024-04-16 | 西安理工大学 | Multi-scale deep convolutional neural network model construction method based on non-uniform grouping |
| CN112114295A (en) * | 2020-07-29 | 2020-12-22 | 中国人民解放军海军航空大学 | Target identification method and system for full-polarization radar |
| CN112114295B (en) * | 2020-07-29 | 2023-12-22 | 中国人民解放军海军航空大学 | Method and system for identifying all-polarized radar target |
| CN112230210A (en) * | 2020-09-09 | 2021-01-15 | 南昌航空大学 | HRRP radar target identification method based on improved LSGAN and CNN |
| CN112287784A (en) * | 2020-10-20 | 2021-01-29 | 哈尔滨工程大学 | Radar signal classification method based on deep convolutional neural network and feature fusion |
| CN112287784B (en) * | 2020-10-20 | 2022-05-31 | 哈尔滨工程大学 | Radar signal classification method based on deep convolutional neural network and feature fusion |
| CN112580660A (en) * | 2020-11-17 | 2021-03-30 | 上海闻泰信息技术有限公司 | Image processing method, image processing device, computer equipment and readable storage medium |
| CN112580660B (en) * | 2020-11-17 | 2023-03-24 | 上海闻泰信息技术有限公司 | Image processing method, image processing device, computer equipment and readable storage medium |
| CN112882010A (en) * | 2021-01-12 | 2021-06-01 | 西安电子科技大学 | High-resolution range profile target identification method based on signal-to-noise ratio field knowledge network |
| CN112882010B (en) * | 2021-01-12 | 2022-04-05 | 西安电子科技大学 | High-resolution range profile target identification method based on signal-to-noise ratio field knowledge network |
| CN112966667A (en) * | 2021-04-06 | 2021-06-15 | 中国人民解放军海军航空大学 | Method for identifying one-dimensional distance image noise reduction convolution neural network of sea surface target |
| CN113376610A (en) * | 2021-06-22 | 2021-09-10 | 西安电子科技大学 | Narrow-band radar target detection method based on signal structure information |
| CN113959071B (en) * | 2021-07-21 | 2023-05-26 | 北京金茂绿建科技有限公司 | Centralized water chilling unit air conditioning system operation control optimization method based on machine learning assistance |
| CN113959071A (en) * | 2021-07-21 | 2022-01-21 | 北京金茂绿建科技有限公司 | Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system |
| CN113640764B (en) * | 2021-08-09 | 2023-08-11 | 中国人民解放军海军航空大学航空作战勤务学院 | Radar one-dimensional range profile identification method and device based on multi-dimension one-dimensional convolution |
| CN113627401A (en) * | 2021-10-12 | 2021-11-09 | 四川大学 | Myoelectric gesture recognition method of feature pyramid network fused with double-attention machine system |
| CN116091854B (en) * | 2022-12-14 | 2023-09-22 | 中国人民解放军空军预警学院 | Method and system for classifying space targets of HRRP sequence |
| CN116091854A (en) * | 2022-12-14 | 2023-05-09 | 中国人民解放军空军预警学院 | Method and system for classifying space targets of HRRP sequence |
| CN116593980A (en) * | 2023-04-20 | 2023-08-15 | 中国人民解放军93209部队 | Radar target recognition model training method, radar target recognition method and device |
| CN116593980B (en) * | 2023-04-20 | 2023-12-12 | 中国人民解放军93209部队 | Radar target recognition model training method, radar target recognition method and device |
| CN117574136A (en) * | 2024-01-16 | 2024-02-20 | 浙江大学海南研究院 | Convolutional neural network calculation method based on multi-element Gaussian function space transformation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109828251B (en) | 2022-07-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109828251A (en) | Radar target identification method based on feature pyramid light weight convolutional neural networks | |
| CN110135295A (en) | A kind of unsupervised pedestrian recognition methods again based on transfer learning | |
| CN110070073A (en) | Pedestrian's recognition methods again of global characteristics and local feature based on attention mechanism | |
| CN108345860A (en) | Personnel based on deep learning and learning distance metric recognition methods again | |
| CN108510012A (en) | A kind of target rapid detection method based on Analysis On Multi-scale Features figure | |
| CN111832546B (en) | Lightweight natural scene text recognition method | |
| CN109615008B (en) | Hyperspectral image classification method and system based on stack width learning | |
| CN109919174A (en) | A kind of character recognition method based on gate cascade attention mechanism | |
| CN104268593A (en) | Multiple-sparse-representation face recognition method for solving small sample size problem | |
| CN110390952A (en) | City sound event classification method based on bicharacteristic 2-DenseNet parallel connection | |
| CN103617609B (en) | Based on k-means non-linearity manifold cluster and the representative point choosing method of graph theory | |
| CN106203483A (en) | A kind of zero sample image sorting technique of multi-modal mapping method of being correlated with based on semanteme | |
| CN109993269A (en) | Single image people counting method based on attention mechanism | |
| Xiao et al. | Enhancing multiscale representations with transformer for remote sensing image semantic segmentation | |
| CN104881682A (en) | Image classification method based on locality preserving mapping and principal component analysis | |
| CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
| CN110414587A (en) | Depth convolutional neural networks training method and system based on progressive learning | |
| CN110969121A (en) | High-resolution radar target recognition algorithm based on deep learning | |
| CN105956610B (en) | A kind of remote sensing images classification of landform method based on multi-layer coding structure | |
| CN115965864A (en) | Lightweight attention mechanism network for crop disease identification | |
| Xiong et al. | RGB-D scene recognition via spatial-related multi-modal feature learning | |
| Yuan et al. | Few-shot scene classification with multi-attention deepemd network in remote sensing | |
| CN115035508A (en) | Topic-guided remote sensing image subtitle generation method based on Transformer | |
| CN109635726A (en) | A kind of landslide identification method based on the symmetrical multiple dimensioned pond of depth network integration | |
| Li et al. | An object-oriented CNN model based on improved superpixel segmentation for high-resolution remote sensing image classification |
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 |