CN109375186B - Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network - Google Patents
Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network Download PDFInfo
- Publication number
- CN109375186B CN109375186B CN201811405815.2A CN201811405815A CN109375186B CN 109375186 B CN109375186 B CN 109375186B CN 201811405815 A CN201811405815 A CN 201811405815A CN 109375186 B CN109375186 B CN 109375186B
- Authority
- CN
- China
- Prior art keywords
- scale
- layer
- neural network
- radar target
- center
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a radar target identification method based on a depth residual error multi-scale one-dimensional convolution neural network. The radar target has translation sensitivity and attitude angle sensitivity, so that the identification difficulty is high, and the method can automatically extract the invariant features of the radar target HRRP. The method designs two one-dimensional residual multi-scale blocks and two one-dimensional multi-scale down-sampling layers by utilizing the characteristics of convolution kernels and down-sampling operation, and constructs a neural network model by taking the two blocks as cores, and on the basis, a new loss function is provided to improve the separability of the characteristics. The method can effectively extract the depth invariant feature of the target, has high recognition accuracy, and has good robustness and generalization performance.
Description
Technical Field
The invention belongs to the radar target automatic identification technology, and provides a radar target identification method based on a deep learning framework, which solves the problems of target HRRP (high resolution range) feature extraction and classification identification under the condition of designing a full-angle domain.
Background
The existing deep learning-based radar target HRRP feature extraction method is basically based on a self-coding model and a variant thereof, and self-coding consists of an encoder and a decoder. The encoding process is the mapping of input to a feature layer, the output of an encoder is the feature extracted from encoding, and when the number of neurons in the feature layer (hidden layer) is smaller than the dimension of input data, the encoding can be regarded as a dimension reduction operation, similar to a Principal Component Analysis (PCA).
At present, the problems of radar target HRRP identification are mainly divided into two aspects, namely attitude angle sensitivity of HRRP, angular domain division is needed in the identification process, and generalization performance of a second identification model is poor. Therefore, the number of hidden layers and neurons thereof needs to be increased by using self-coding for identification, but the overfitting phenomenon is easily generated in the model due to the limited number of training data. The self-coding based radar target HRRP identification is therefore mainly directed to the smaller angular domain.
The method utilizes a one-dimensional depth residual error multi-scale coiling machine neural network model to extract full-angle domain complex features of HRRP, the model has good generalization performance due to the local connection characteristic of a convolution kernel, and meanwhile, the cosine central loss function provided by the method improves the feature separability of different types of targets, so that the method can improve the target identification accuracy.
Disclosure of Invention
The invention aims to provide a radar target identification method based on a depth residual error multi-scale one-dimensional convolution neural network aiming at the identification difficulty of full-angle-domain HRRP, and the identification accuracy of the HRRP target under the full-angle-domain condition is improved.
The technical solution of the invention is as follows: and constructing a one-dimensional depth residual multi-scale convolutional neural network, designing a loss function which is more beneficial to the fine division of the target, and training the model by using the training sample to obtain the model which can be used for end-to-end identification.
In order to achieve the purpose, the invention comprises the following steps:
step 1: and constructing a depth one-dimensional residual convolution neural network model, and initializing model parameters.
Step 2: and (5) forward propagation, and calculating a loss function L in the iterative process.
And step 3: and (4) backward propagation, wherein parameters in the model are updated by adopting a chain rule.
And 4, step 4: and (5) repeating the steps 2 and 3 until the loss function is converged to obtain a model for radar target identification.
Compared with the prior art, the invention has the following advantages:
(1) the depth residual error multi-scale convolution neural network takes a one-dimensional convolution kernel as a core, the convolution kernel has the characteristic of sharing a local visual field and a weight, the convolution kernels with different scales can extract features with different fineness, the two features are utilized to enable a model to extract the full-angle-domain complex features of the HRRP by utilizing fewer parameters, and the generalization capability of the model is improved.
(2) A new loss function-cosine center loss is designed, and the loss function is utilized to train a model, so that the inter-class distance of the features can be increased, the intra-class distance of the features can be reduced, and the identification accuracy of the target can be improved.
Drawings
FIG. 1: multi-scale convolutional layer I.
FIG. 2: multi-scale convolutional layer II.
FIG. 3: residual multi-scale block schematic.
FIG. 4: layer I is multi-scale down sampled.
FIG. 5: layer II is multi-scale down sampled.
FIG. 6: the model provided by the method is a schematic diagram.
Detailed Description
The invention is described in further detail below with reference to the drawings. The model constructed by the invention is explained as follows by referring to the attached drawings of the specification:
1. construction of residual multi-scale volume blocks
Two multi-scale convolutional layers are designed, and the structures of the layers are shown in figures 1 and 2. Except for the modules that label the downsampled layers, the other modules represent convolutional layers.
Each of the multi-scale convolutional layers I, II includes four branches, where (from left to right), branch one and branch two of multi-scale convolutional layer I, II are the same, branch one includes one convolutional layer, the convolutional kernel scale is 1 × 1, and branch two includes two convolutional layers, and the convolutional kernel scales (from top to bottom) are 1 × 1 and 3 × 3, respectively. The difference in the multi-scale convolutional layer I, II is mainly reflected in branch three and branch four. The three branches contain three convolutional layers, the sizes of the convolutional cores from top to bottom of the three branches of the multi-scale convolutional layer I, II are 1 x 1, 3 x 1, 5 x 1 and 5 x 1, 3 x 1 and 1 x 1 respectively, the four branches contain one down-sampling layer and one convolutional layer, and the convolutional cores of the convolutional layers in the four branches of the multi-scale convolutional layers I and II are 3 x 1 and 1 x 1 respectively. And the down-sampling layers of the branch four are the same, and the function of the layer is to remove redundant information and better keep effective information of the characteristics. And finally, directly splicing the features of each branch together, taking the multi-scale convolutional layer I as an example, and obtaining an output feature with the number of channels of the input feature passing through the multi-scale convolutional layer as 16+16+16+ 64. The output of each convolution layer and down-sampling layer uses zero padding operation, namely the dimension of the input and output of the multi-scale convolution layer is not changed, and only the number of channels is changed.
The residual block is composed of convolutional layers, and may include one or more convolutional layers, and in order to ensure that the network does not degrade, the number of convolutional layers in the residual block is not too large, and is usually two. The output of the residual block is the sum of the convolution layer output and the residual block input, and the use of the residual block can avoid the situation that the model has gradient explosion or disappears. Through convolution operation, the length of the characteristic vector is reduced, so that the zero filling method is utilized to ensure that the number and the dimensionality of the output characteristic vector are the same as those of the input characteristic vector. The plus sign in the figure indicates that the elements of the corresponding feature vector are added one by one. The expression of the residual block is
xl+1=F(xl,θr)+Wrxl
Wherein x isl∈Rk×mRepresenting input feature vectors, xl+1∈Rh×mRepresenting the output eigenvectors, k and h represent the number of input and output eigenvectors, respectively. Thetar={k1,k2,b1,b2Denotes the set of parameters, W, of the residual blockr∈Rh×kFor adjusting xlDimension of (2) is such that it is in contact with xl+1Are the same, here we setWrThe constant matrix is not needed to be updated in the process of back propagation.
The residual multi-scale convolution block can be obtained by replacing the convolution layer in the residual block with a multi-scale convolution layer, as shown in fig. 3.
2. Construction of multiscale downsampling layers
Two multi-scale down-sampling layers are designed as shown in fig. 4 and 5. Except for the modules that label the downsampled layers, the other modules represent convolutional layers.
The multi-scale downsampling layers I, II each include three branches, where (from left to right) branch one includes one convolution layer with convolution kernel sizes of 3 x 1 and 1 x 1, respectively. The branch two contains two convolution layers, the convolution kernel sizes are 5 x 1, 3 x 1 and 3 x 1, 1 x 1 respectively, the branch three consists of one convolution layer and one down-sampling layer, the convolution kernel sizes are 3 x 1 and 1 x 1 respectively, and the window length of the down-sampling layer is 3 x 1. The step length of the last layer of each branch is 2, so that the output dimensionality is one half of the input characteristic dimensionality, and the function of the multi-scale down-sampling layer not only reduces the data dimensionality, but also extracts deeper characteristics.
3. Loss function design
The Loss function design uses Large boundary Cosine Loss (Large Margin Cosine Loss) in the face recognition method for reference, and on the basis, the constraint of the intra-class distance of the features is considered, and a Loss function-Cosine center Loss function based on the Cosine distance and the feature center is provided. The expression for the loss function is as follows:
where m is the number of training data, xiRepresenting the output characteristics, y, of the fully connected layer preceding the output layer for the ith sample dataiThe true label of the ith sample data represents the true category to which the ith sample data belongs. Wj∈RdIs the weight matrix W of the full connection layer belongs to Rd×nThe jth column of (1) is the weight vector corresponding to the jth class target.Represents Wj,xiCosine values between, s and alpha controlling L respectivelyLMCLAnd LcenterThe amplitude of the normalized feature is adjusted to avoid the problem of unconvergence of the loss function after feature normalization, and a is a constant and is used for enhancing the feature xiAnd weight WjIs constrained.Is the label y of the ith sampleiIs correspondingly provided withThe feature centers of the classes, λ is the weight of the central loss term, the larger λ, the more clustered the features within the class.
The loss function is divided into two parts LLMCLAnd Lcenter,LLMCLCalled Large Margin Cosine Loss, abbreviated as (LMCL), and the expression of the traditional softmax-Loss isWherein the content of the first and second substances,cosθjthe similarity between the weight W and the feature x can be considered as a similarity measurement, and in order to train the model by making full use of the angle information, we want the norm of the weight matrix W and the norm of the feature vector x to be 1, so that the normalization operation is performed on W and x, and W is equal to W*/||W*||,x=x*/||x*If | then getBecause of the fact that Is monotonically decreasing, the characteristic xiAnd weight WjThe smaller the included angle of (a) is,the larger, the feature xiThe greater the probability that the corresponding sample i belongs to the class j, when satisfiedWhen sample i belongs to the yiAnd (4) class. In the proposed method, when satisfiedWhen the sample i belongs to the yiAnd (4) class.
The effect of the hyperparameter a > 0 is to increaseStrong characteristic xiAnd weight WjThe included angle of (a) is constrained, and under the ideal condition, the value range of a isWhere n is the number of categories. In the feature space, a is 1-cos (2 pi/n) if and only if the included angle between each adjacent feature center is the same and the included angle between the feature of each type of sample and the corresponding feature center is 0.
The parameter s solves the problem of non-convergence of the loss function, and the size of the parameter s satisfiesm is the number of training samples in min-batch, and ε is a very small constant, the closer to 0 the better.
LcenterC is normalized characteristic center, and the initialized expression isWherein, δ (y)iJ) represents when yiClass j, δ (-) equals 1, otherwise δ (-) equals 0, mjIndicating the number of samples contained in the j-th class target in min-batch,feature x representing sample iiFeature center corresponding theretoThe function penalizes samples with larger included angle between the feature vector and the feature center.
Claims (2)
1. A radar target identification method based on a depth residual error multi-scale one-dimensional convolution neural network is characterized by comprising the following steps:
step 1, constructing a depth one-dimensional residual convolution neural network model, and initializing model parameters;
step 2, forward propagation is carried out, and a loss function L in the iterative process is calculated;
the calculation method of the loss function specifically comprises the following steps:
wherein L isLMCLIs LargeMargin Cosine Loss, LcenterFor the angle center loss function, m is the number of training data in one min-batch, xiRepresenting the output characteristics, y, of the fully connected layer preceding the output layer for the ith sample dataiThe real label of the ith sample data represents the category of the ith sample data; w is a group ofj∈RdIs the weight matrix W of the full connection layer belongs to Rd×nThe jth column of (1) represents a weight vector corresponding to the jth class target;represents WjAnd xiThe cosine value between s and alpha controls LLMCLAnd LcenterDetermining cosine boundaries between the categories by taking the amplitude of the normalized features as a constant;is the label y of the ith sample dataiThe feature center of the corresponding class, λ is the weight of the center loss term;feature x representing sample iiFeature center corresponding theretoCosine value of the included angle of (a);
step 3, backward propagation, and updating parameters in the model by adopting a chain rule;
and 4, repeating the steps 2 and 3 until the loss function is converged to obtain a model for radar target identification.
2. The radar target recognition method of claim 1, wherein the neural network model in step 1 is composed of a residual multi-scale convolutional layer, a multi-scale down-sampling layer, a convolutional layer, a down-sampling layer, and a full-connection layer.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811405815.2A CN109375186B (en) | 2018-11-22 | 2018-11-22 | Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811405815.2A CN109375186B (en) | 2018-11-22 | 2018-11-22 | Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109375186A CN109375186A (en) | 2019-02-22 |
CN109375186B true CN109375186B (en) | 2022-05-31 |
Family
ID=65383391
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811405815.2A Active CN109375186B (en) | 2018-11-22 | 2018-11-22 | Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109375186B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109901129A (en) * | 2019-03-06 | 2019-06-18 | 中国人民解放军海军航空大学 | Object detection method and system in a kind of sea clutter |
CN109828251B (en) * | 2019-03-07 | 2022-07-12 | 中国人民解放军海军航空大学 | Radar target identification method based on characteristic pyramid light-weight convolution neural network |
CN110069985B (en) * | 2019-03-12 | 2020-08-28 | 北京三快在线科技有限公司 | Image-based target point position detection method and device and electronic equipment |
CN109977871B (en) * | 2019-03-27 | 2021-01-29 | 中国人民解放军战略支援部队航天工程大学 | Satellite target identification method based on broadband radar data and GRU neural network |
CN110837777A (en) * | 2019-10-10 | 2020-02-25 | 天津大学 | Partial occlusion facial expression recognition method based on improved VGG-Net |
CN110942777B (en) * | 2019-12-05 | 2022-03-08 | 出门问问信息科技有限公司 | Training method and device for voiceprint neural network model and storage medium |
CN110929697B (en) * | 2019-12-17 | 2021-04-13 | 中国人民解放军海军航空大学 | Neural network target identification method and system based on residual error structure |
CN112712126B (en) * | 2021-01-05 | 2024-03-19 | 南京大学 | Picture identification method |
CN112346056B (en) * | 2021-01-11 | 2021-03-26 | 长沙理工大学 | Resolution characteristic fusion extraction method and identification method of multi-pulse radar signals |
CN113050077B (en) * | 2021-03-18 | 2022-07-01 | 电子科技大学长三角研究院(衢州) | MIMO radar waveform optimization method based on iterative optimization network |
CN113486917B (en) * | 2021-05-17 | 2023-06-02 | 西安电子科技大学 | Radar HRRP small sample target recognition method based on metric learning |
CN114636995A (en) * | 2022-03-16 | 2022-06-17 | 中国水产科学研究院珠江水产研究所 | Underwater sound signal detection method and system based on deep learning |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015176305A1 (en) * | 2014-05-23 | 2015-11-26 | 中国科学院自动化研究所 | Human-shaped image segmentation method |
CN108776835A (en) * | 2018-05-28 | 2018-11-09 | 嘉兴善索智能科技有限公司 | A kind of deep neural network training method |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103544296B (en) * | 2013-10-22 | 2017-02-15 | 中国人民解放军海军航空工程学院 | Adaptive intelligent integration detection method of radar range extension target |
CN106097355A (en) * | 2016-06-14 | 2016-11-09 | 山东大学 | The micro-Hyperspectral imagery processing method of gastroenteric tumor based on convolutional neural networks |
US10007865B1 (en) * | 2017-10-16 | 2018-06-26 | StradVision, Inc. | Learning method and learning device for adjusting parameters of CNN by using multi-scale feature maps and testing method and testing device using the same |
CN108169745A (en) * | 2017-12-18 | 2018-06-15 | 电子科技大学 | A kind of borehole radar target identification method based on convolutional neural networks |
CN108062754B (en) * | 2018-01-19 | 2020-08-25 | 深圳大学 | Segmentation and identification method and device based on dense network image |
CN108416378B (en) * | 2018-02-28 | 2020-04-14 | 电子科技大学 | Large-scene SAR target recognition method based on deep neural network |
CN108734660A (en) * | 2018-05-25 | 2018-11-02 | 上海通途半导体科技有限公司 | A kind of image super-resolution rebuilding method and device based on deep learning |
-
2018
- 2018-11-22 CN CN201811405815.2A patent/CN109375186B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015176305A1 (en) * | 2014-05-23 | 2015-11-26 | 中国科学院自动化研究所 | Human-shaped image segmentation method |
CN108776835A (en) * | 2018-05-28 | 2018-11-09 | 嘉兴善索智能科技有限公司 | A kind of deep neural network training method |
Non-Patent Citations (1)
Title |
---|
《CosFace: Large Margin Cosine Loss for Deep Face Recognition》;H. Wang et al;《2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition》;20180723;5265-5274 * |
Also Published As
Publication number | Publication date |
---|---|
CN109375186A (en) | 2019-02-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109375186B (en) | Radar target identification method based on depth residual error multi-scale one-dimensional convolution neural network | |
Leng et al. | Local feature descriptor for image matching: A survey | |
Song et al. | A survey of remote sensing image classification based on CNNs | |
Yuan et al. | Remote sensing image scene classification using rearranged local features | |
Song et al. | Hyperspectral image classification with deep feature fusion network | |
CN108009559B (en) | Hyperspectral data classification method based on space-spectrum combined information | |
CN107316013B (en) | Hyperspectral image classification method based on NSCT (non-subsampled Contourlet transform) and DCNN (data-to-neural network) | |
CN111191583B (en) | Space target recognition system and method based on convolutional neural network | |
Wang et al. | Remote sensing landslide recognition based on convolutional neural network | |
CN103971123B (en) | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) | |
Qayyum et al. | Scene classification for aerial images based on CNN using sparse coding technique | |
CN108229551B (en) | Hyperspectral remote sensing image classification method based on compact dictionary sparse representation | |
CN108460400B (en) | Hyperspectral image classification method combining various characteristic information | |
CN113139512B (en) | Depth network hyperspectral image classification method based on residual error and attention | |
CN112115806B (en) | Remote sensing image scene accurate classification method based on Dual-ResNet small sample learning | |
CN111652273A (en) | Deep learning-based RGB-D image classification method | |
Zhao et al. | PCA dimensionality reduction method for image classification | |
CN112836671A (en) | Data dimension reduction method based on maximization ratio and linear discriminant analysis | |
CN115019145A (en) | Method for monitoring settlement deformation of road surface of Qinghai-Tibet plateau highway | |
CN109034213B (en) | Hyperspectral image classification method and system based on correlation entropy principle | |
CN115131580A (en) | Space target small sample identification method based on attention mechanism | |
CN104809471A (en) | Hyperspectral image residual error fusion classification method based on space spectrum information | |
Ni et al. | Scene classification from remote sensing images using mid-level deep feature learning | |
Ding et al. | Efficient vanishing point detection method in unstructured road environments based on dark channel prior | |
Ergun | Segmentation of wood cell in cross-section using deep convolutional neural networks |
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 |