CN114943889A - SAR image target identification method based on small sample incremental learning - Google Patents

SAR image target identification method based on small sample incremental learning Download PDF

Info

Publication number
CN114943889A
CN114943889A CN202210296918.XA CN202210296918A CN114943889A CN 114943889 A CN114943889 A CN 114943889A CN 202210296918 A CN202210296918 A CN 202210296918A CN 114943889 A CN114943889 A CN 114943889A
Authority
CN
China
Prior art keywords
prototype
base class
training
residual error
learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210296918.XA
Other languages
Chinese (zh)
Inventor
周峰
王力
杨鑫瑶
谭浩月
白雪茹
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202210296918.XA priority Critical patent/CN114943889A/en
Publication of CN114943889A publication Critical patent/CN114943889A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/243Classification techniques relating to the number of classes
    • G06F18/2431Multiple classes
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent

Landscapes

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

Abstract

The invention relates to an SAR image target identification method based on small sample increment learning, which comprises the following steps: step 1: constructing a depth residual error network based on a prototype idea; step 2: respectively carrying out base class training and incremental learning on the depth residual error network based on the prototype idea by utilizing a base class data set and an incremental data set; and step 3: and inputting the SAR image to be detected into a depth residual error network which is completed by training and learning and based on the prototype idea to obtain a prediction classification result of the SAR image to be detected. The SAR image target recognition method based on small sample incremental learning designs a depth residual error network based on a prototype idea, the network can have the capability of automatically extracting the SAR image features even under the condition of less image samples, compared with the prior art, the method for mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of the number of samples to a certain extent, and the recognition capability of the whole network to the small sample image is improved.

Description

SAR image target identification method based on small sample increment learning
Technical Field
The invention belongs to the field of radar image processing, and particularly relates to an SAR image target identification method based on small sample incremental learning.
Background
Synthetic Aperture Radar (SAR) is an active microwave imaging radar, has all-time and all-weather multi-dimensional earth surface information acquisition capability and certain earth surface and vegetation penetration capability. Synthetic aperture radar imaging technology is currently an indispensable technical means in the field of earth observation. Automatic Target Recognition (ATR) is a technique for determining image class by capturing image feature information, and it is still quite difficult to perform real-time automatic target recognition on radar images in a complex electromagnetic environment.
Currently, deep learning based methods provide new approaches for automatic target recognition. The method based on deep learning does not require a good model to be established for the radar image, but trains the model by using the image obtained by the radar imaging technology on the basis of initializing the model, so that the model learns the characteristics of the radar image, and the automatic identification of the radar image is completed.
Due to the excellent feature extraction and recognition capability of deep learning, the application of the deep learning method to the SAR ATR draws more attention, and many methods have been proposed for the automatic target recognition of the SAR image. The deep convolution neural network has good performance in the aspect of feature extraction, and the feedforward neural network has good performance on target identification.
Because deep learning often requires a lot of time to train a network, and meanwhile, SAR images are difficult to acquire, and acquisition cost is low, when the network encounters a new task or a new class, the idea of adding a small sample new class sample into a training sample and then retraining the network is unrealistic, and if the network is not retrained any more, the network cannot identify the new class, and the new class can be wrongly classified into a known class.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides an SAR image target identification method based on small sample increment learning. The technical problem to be solved by the invention is realized by the following technical scheme:
the invention provides an SAR image target identification method based on small sample increment learning, which comprises the following steps:
step 1: constructing a depth residual error network based on a prototype idea;
step 2: carrying out base class training and incremental learning on the depth residual error network based on the prototype idea by utilizing a base class data set and an incremental data set;
and step 3: and inputting the SAR image to be detected into a depth residual error network which is completed by training and learning and based on a prototype idea to obtain a prediction classification result of the SAR image to be detected.
In one embodiment of the present invention, the depth residual error network based on prototype idea comprises a convolution module, a depth residual error network, an adaptive average pooling layer, a full connection layer and a classifier connected in sequence, wherein,
the convolution module comprises a first convolution layer and a first ReLU active layer which are connected in sequence;
the depth residual error network comprises a plurality of residual error blocks which are connected in sequence and used for extracting the image characteristics of the input image;
the full-connection layer is used for storing class prototypes, a graph attention network is integrated in the full-connection layer, and the graph attention network is used for adjusting the positions of the class prototypes stored on the full-connection layer in an increment learning stage;
the classifier is a cosine classifier and is used for outputting a prediction classification result.
In one embodiment of the present invention, the residual block includes a second convolution layer, a first batch of normalization layers, a second ReLU active layer, a third convolution layer, a second batch of normalization layers, and a third ReLU active layer, which are connected in sequence;
and the result of the addition of the input of the residual block and the output of the residual block sequentially passes through the second convolution layer, the first batch of normalization layers, the second ReLU activation layer, the third convolution layer and the second batch of normalization layers and is output after passing through the third ReLU activation layer.
In an embodiment of the present invention, the residual block further includes a channel number conversion unit, the channel number conversion unit is connected between the input end of the second convolution layer and the output end of the second batch of normalization layers, and the channel number conversion unit includes a convolution layer and a batch of normalization layers which are connected in sequence;
and when the input channel number of the residual block is not equal to the output channel number of the residual block, converting the input of the residual block into the same channel number as the output of the residual block through the channel number conversion unit, and then outputting a result of adding the result with the output of the second batch of normalization layers after passing through the third ReLU activation layer.
In an embodiment of the present invention, each of the base class data set and the incremental data set includes a plurality of classes of SAR images with attached classification tags, and the incremental data set is a small sample data set of 1-way-5-shot or 1-way-1-shot.
In one embodiment of the present invention, the step 2 comprises:
step 2.1: in the base class training stage, the base class data set is used for training and learning the depth residual error network based on the prototype thought, and the network weight of the depth residual error network based on the prototype thought is updated through back propagation to obtain a network model in the base class training stage;
step 2.2: in the incremental learning stage, the incremental data set is used for training and learning the network model in the base class training stage, and the weight of the full connection layer is updated through back propagation to obtain the depth residual error network based on the prototype idea after training and learning.
In one embodiment of the invention, said step 2.1 comprises:
step 1 a: initializing network parameters of the depth residual error network based on the prototype idea;
step 1 b: obtaining a base class training sample in the base class data set and a corresponding classification label according to an index, inputting the base class training sample into the depth residual error network based on the prototype idea, and obtaining the mapping of the base class training sample in a feature space after the base class training sample passes through the convolution module and the depth residual error network;
step 1 c: the mapping passes through the self-adaptive average pooling layer and the full connection layer and then is input into the classifier, and the cosine similarity between the base class training sample and the class prototype is output through the classifier;
step 1 d: taking the class corresponding to the index with the maximum cosine similarity as a prediction classification result of the base class training sample, and calculating a cross entropy loss function and accuracy according to a classification label corresponding to the base class training sample;
step 1 e: and updating the network weight of the depth residual error network based on the prototype idea through back propagation according to the cross entropy loss function, taking the mean value of the mapping expression of the base class training sample in the feature space as a class prototype, and storing the mean value as the weight of the full connection layer to complete the base class training stage of the depth residual error network based on the prototype idea, so as to obtain a network model of the base class training stage.
In one embodiment of the invention, said step 2.2 comprises:
step 2 a: acquiring an incremental training sample in the incremental data set and a corresponding classification label thereof, and inputting the incremental training sample into the network model of the base class training stage to obtain the mapping of the incremental training sample in a feature space;
and step 2 b: taking the mean value of the mapping expression of the same type of samples to obtain a new type prototype of the new type in the feature space, and taking the new type prototype as the newly added weight of the full connection layer;
and step 2 c: inputting the base class training samples and the increment training samples into the network model obtained in the step 2b, and calculating a cross entropy loss function according to the base class training samples, the increment training samples and corresponding classification labels;
step 2 d: and updating the weight of the full connection layer through back propagation according to the cross entropy loss function to obtain the depth residual error network based on the prototype thought, which is completed by training and learning.
Compared with the prior art, the invention has the beneficial effects that:
1. the SAR image target recognition method based on small sample incremental learning designs a depth residual error network based on prototype thought, the network can also have the capability of automatically extracting SAR image features even under the condition of less image samples, compared with the prior art, the method for mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of the number of samples to a certain extent, and the recognition capability of the whole network to the small sample image is improved;
2. according to the SAR image target recognition method based on small sample incremental learning, the continuous evolution classifier is designed to be capable of continuously learning a new class in a trained network, and the setting is more in line with the requirement of continuously learning the new SAR image information of the small sample in reality;
3. the SAR image target recognition method based on small sample increment learning adopts an image attention network to store various prototypes, fine adjusts the distance between the prototypes in the image attention network to obtain a better classification boundary, the storage mode is directly integrated in a full connection layer of the network, and is trained along with the network according to a training sample, so that the SAR image target recognition method has a better decision boundary compared with the network of a fixed prototype, and the recognition accuracy is improved.
The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly understood, the present invention may be implemented in accordance with the content of the description, and in order to make the above and other objects, features, and advantages of the present invention more clearly understood, the following preferred embodiments are described in detail with reference to the accompanying drawings.
Drawings
Fig. 1 is a schematic diagram of an SAR image target identification method based on small sample incremental learning according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a depth residual error network based on a prototype idea according to an embodiment of the present invention;
FIG. 3 is a schematic flowchart of training and testing a deep residual error network based on prototype idea according to an embodiment of the present invention;
fig. 4 is a schematic structural diagram illustrating a training and learning phase of an attention network according to an embodiment of the present invention.
Detailed Description
In order to further explain the technical means and effects of the present invention adopted to achieve the predetermined invention purpose, a method for identifying an SAR image target based on small sample incremental learning according to the present invention is described in detail below with reference to the accompanying drawings and the detailed embodiments.
The foregoing and other technical contents, features and effects of the present invention will be more clearly understood from the following detailed description of the embodiments taken in conjunction with the accompanying drawings. The technical means and effects of the present invention adopted to achieve the predetermined purpose can be more deeply and specifically understood through the description of the specific embodiments, however, the attached drawings are provided for reference and description only and are not used for limiting the technical scheme of the present invention.
Example one
Referring to fig. 1, fig. 1 is a schematic diagram of a method for identifying a target in an SAR image based on small sample incremental learning according to an embodiment of the present invention, where as shown in the figure, the method for identifying a target in an SAR image based on small sample incremental learning according to the embodiment includes:
step 1: constructing a depth residual error network based on a prototype idea;
step 2: respectively carrying out base class training and incremental learning on the depth residual error network based on the prototype idea by utilizing a base class data set and an incremental data set;
and step 3: and inputting the SAR image to be detected into a depth residual error network which is completed by training and learning and based on a prototype idea to obtain a prediction classification result of the SAR image to be detected.
Referring to fig. 2 in combination, fig. 2 is a schematic structural diagram of a depth residual error network based on a prototype idea provided in an embodiment of the present invention, and as shown in the drawing, the depth residual error network based on the prototype idea in the embodiment includes a convolution module, a depth residual error network, an adaptive average pooling layer, a full connection layer, and a classifier, which are connected in sequence. The convolution module comprises a first convolution layer and a first ReLU active layer which are sequentially connected; the depth residual error network comprises a plurality of residual error blocks which are connected in sequence and used for extracting the image characteristics of the input image; the full connection layer is used for storing class prototypes, an image attention network is integrated in the full connection layer, and the image attention network is used for adjusting the positions of the class prototypes stored on the full connection layer in an incremental learning stage; the classifier is a cosine classifier and is used for outputting a prediction classification result.
Specifically, the residual block comprises a second convolution layer, a first batch of normalization layers, a second ReLU active layer, a third convolution layer, a second batch of normalization layers and a third ReLU active layer which are connected in sequence. And the result of the addition of the input of the residual block and the output of the input of the residual block sequentially passing through the second convolution layer, the first batch of normalization layers, the second ReLU activation layer, the third convolution layer and the second batch of normalization layers is output after passing through the third ReLU activation layer.
Furthermore, the residual block further comprises a channel number conversion unit, the channel number conversion unit is connected between the input end of the second convolution layer and the output end of the second batch of normalization layers, and the channel number conversion unit comprises a convolution layer and a batch of normalization layers which are sequentially connected. In this embodiment, since the input and the output of the residual block are additive, when the number of input channels of the residual block is not equal to the number of output channels thereof, the result of the addition of the input of the residual block converted into the same number of channels as the output thereof by the channel number conversion unit and the output of the second normalization layer is output after passing through the third ReLU activation layer.
Further, in this embodiment, the base class dataset and the incremental dataset each include a plurality of categories of SAR images with classification tags attached thereto, and the incremental dataset is a small sample dataset of 1-way-5-shot or 1-way-1-shot.
Optionally, the training samples of the base class data set and the incremental data set are both SAR images obtained from the MSTAR data set, and the incremental data set is an SAR image observed at different pitch angles.
It should be noted that, because the image data sent to the depth residual error network based on the prototype idea is different at each stage of training and learning, the image data indexes needed at each stage need to be obtained from the MSTAR radar image data set in advance.
Further, in this embodiment, step 2 includes:
step 2.1: in the base class training stage, a base class data set is used for training and learning the deep residual error network based on the prototype idea, and a network weight of the deep residual error network based on the prototype idea is updated through back propagation to obtain a network model of the base class training stage;
specifically, step 2.1 comprises:
step 1 a: initializing network parameters of a depth residual error network based on a prototype idea;
step 1 b: acquiring a base class training sample and a corresponding classification label in a base class data set according to the index, inputting the base class training sample into a depth residual error network based on a prototype idea, and mapping the base class training sample in a feature space after the base class training sample passes through a convolution module and the depth residual error network;
step 1 c: mapping, passing through a self-adaptive average pooling layer and a full-connection layer, inputting the mapping into a classifier, and outputting cosine similarity between a base class training sample and a class prototype through the classifier;
step 1 d: taking the class corresponding to the index with the largest cosine similarity as a prediction classification result of the base class training sample, and calculating a cross entropy loss function and accuracy according to a classification label corresponding to the base class training sample;
step 1 e: and updating the network weight of the depth residual error network based on the prototype idea through back propagation according to the cross entropy loss function, taking the mean value of the mapping expression of the base class training sample in the feature space as a class prototype, and storing the class prototype as the weight of the full connection layer to complete the base class training stage of the depth residual error network based on the prototype idea and obtain the network model of the base class training stage.
Step 2.2: in the incremental learning stage, the incremental data set is used for training and learning the network model in the base class training stage, and the weight of the full connection layer is updated through back propagation to obtain the depth residual error network based on the prototype thought, which is completed by training and learning.
Specifically, step 2.2 comprises:
step 2 a: acquiring an incremental training sample in an incremental dataset and a corresponding classification label thereof, inputting the incremental training sample into a network model of a base class training stage, and obtaining the mapping of the incremental training sample in a feature space;
and step 2 b: taking the mean value of the mapping expression of the same type of samples to obtain a new type prototype of the new type in the feature space, and taking the new type prototype as a newly added weight of the full connection layer;
and step 2 c: inputting the base class training samples and the increment training samples into the network model obtained in the step 2b, and calculating a cross entropy loss function according to the base class training samples, the increment training samples and the corresponding classification labels;
and step 2 d: and updating the weight of the full connection layer through back propagation according to the cross entropy loss function to obtain the depth residual error network based on the prototype thought, which is completed by training and learning.
It should be noted that, in step 3, if the to-be-detected SAR image relates to a new class, incremental learning needs to be performed on the new class of SAR image.
Further, with reference to fig. 3 and fig. 4, a process of base class training and incremental learning of the deep residual error network based on the prototype idea in this embodiment is specifically described, fig. 3 is a schematic flowchart of training and testing of the deep residual error network based on the prototype idea provided in the embodiment of the present invention, and fig. 4 is a schematic structural diagram of a training and learning stage of the attention network provided in the embodiment of the present invention.
In this embodiment, the image data indexes needed by each stage are acquired from the MSTAR radar image dataset in advance. Setting a base class data set as 7 classes, setting an incremental data set as 3 classes, and dividing incremental training samples in the 3 classes of incremental data sets into three incremental stages for learning by a classifier, wherein base class pictures are enough, all training samples of the first 7 classes are used, and the training is consistent with the training of a common closed set; 3 types of increment training samples in the increment learning stage all adopt a 1-way-5-shot small sample training mode. Meanwhile, according to different stages, the test set samples are used as a verification set, so that the overfitting problem of sample learning is reduced as much as possible while the accuracy is ensured.
Specifically, a training set with a pitch angle of 17 ° in the MSTAR radar image dataset is used as the training set of this embodiment, where the top 7 classes are used as the base class training samples D b (ii) a The last three classes are used as incremental training samples D i The incremental learning stage belongs to small sample incremental learning, specifically to small sample learning of 1-way-5-shot, and training samples for the small sample learning are taken from incremental training samples according to random indexes.
Meanwhile, a test set with a pitch angle of 15 ° in the MSTAR radar image data set is taken as a verification set of this embodiment. As the continuous evolution classifier is used for learning the SAR image recognition capacity in stages, the verification set used in each stage is different, and the verification set used in the base class training stage is V b The verification set used in the kth incremental learning stage is
Figure BDA0003563815510000101
In this embodiment, the depth residual error network includes 4 residual error blocks, and the specific structure of the depth residual error network based on the prototype idea is as follows: first convolution layer → first ReLU activation layer → first residual block → second residual block → third residual block → fourth residual block → adaptive averaging pooling layer → fully connected layer → graph attention network → cosine distance classifier.
Wherein, the specific structure of each residual block is as follows: second convolution layer → first batch of normalization layers → second ReLU active layer → third convolution layer → second batch of normalization layers → third ReLU active layer. In this embodiment, it is necessary to determine whether the number of input channels of the residual block is equal to the number of output channels, and if so, the input of the residual block and the input pass through the second convolution layer → the first normalization layer → the second ReLU active layer → the third convolution layer → the second normalization layer, and the addition result passes through the third ReLU active layer and is then output. If not, the input of the residual block is passed through an additional convolution layer and an additional batch of normalization layers (channel number conversion unit), and then the result of the addition is output through the third ReLU activation layer with the output of the input passed through the second convolution layer → the first batch of normalization layers → the second ReLU activation layer → the third convolution layer → the second batch of normalization layers.
This is because the residual block requires that its input and output are additive during design, so when the number of channels is different, the input will go through a convolutional layer and a batch normalization layer, and the number of channels will be equal to the number of channels output at that time.
The training sample input into the depth residual error network is an SAR image with the channel number of 3 and the size of 128 x 128. Each residual block comprises two convolutional layers, the number of output channels of the residual block is determined by the number of output channels of the second convolutional layer, namely: the number of convolution kernels of the second convolution layer is determined; the number of convolution kernels of the second convolution layer in the first residual block, the second residual block, the third residual block and the fourth residual block is respectively as follows: 32. 64, 128, 256; the convolution kernels inside the residual block are all 3 × 3, and the convolution kernels of the additional convolution layers are 1 × 1; the number of convolution kernels of the first convolution layer is 16, and the size of the convolution kernels is 3 multiplied by 3; the SAR image is changed into 2 multiplied by 2 characteristic information after passing through all residual blocks, and then passes through a self-adaptive average pooling layer with the set output of 1 multiplied by 1; the full connection layer comprises 256 nodes and is linearly connected to 10 nodes.
In this embodiment, first, all base class training samples x are extracted according to a pre-obtained index b And its corresponding classification label y b Sending the data into a depth residual error network based on a prototype idea to obtain a class prototype prto of the base class b And taking another test sample as a verification set to prevent overfitting, so as to obtain a classifier capable of identifying the base class (namely a network model in the training stage of the base class). Secondly, a small number of incremental training samples x are further added according to the new class index i And its corresponding classification label y i Feeding the baseIn the network model in the class training stage, the network model continues to learn new class characteristics to obtain a new class prototype prto i Thus, the continuous evolution classifier learns the ability to identify new classes during the class increment learning phase.
Specifically, the specific steps of base class training are as follows:
1a) initializing network parameters of a depth residual error network based on a prototype idea, and setting the iteration number T of each stage to be more than or equal to 100.
1b) Extracting a base class training sample x according to a pre-obtained index b And its corresponding classification label y b Training the base class sample x b Sending the data into a depth residual error network based on the prototype idea, and obtaining the mapping f (x) of the sample in the feature space through a convolution module and the depth residual error network b ) Wherein f (x) b ) Is a mapping expression of one dimension N-256;
wherein, f () is a mapping function of the network, and the expression of the mapped sample is:
encoder=f(x b ):x b ∈R 128×128 →f(x b )∈R 256 (1)。
1c) the full connection layer is
Figure BDA0003563815510000121
In the base class training stage, 7 initialized class prototypes are stored in the weight of the full connection layer as the weight of the full connection layer, and the calculation formula of the class prototypes is as follows:
Figure BDA0003563815510000122
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003563815510000123
is a k-th class prototype of the base class,
Figure BDA0003563815510000124
is a k-th class training set of the base class,
Figure BDA0003563815510000125
is composed of
Figure BDA0003563815510000126
Total number of samples.
1d) The cosine similarity d (f (x) of the sample and the 7-class prototype can be obtained by multiplying the 256-dimensional mapping expression of the training sample by 7 weights b ),prto b ) The formula for calculating the cosine similarity is as follows:
Figure BDA0003563815510000127
if the cosine similarity is larger, the cosine distance is smaller, and the cosine similarity and the cosine distance are more relevant. So based on the cosine similarity d (f (x) between the training sample and the class prototype b ),prto b ) The index of the maximum value can obtain the prediction classification result of the sample, and then the classification label y of the sample is combined b Then, the cross entropy loss function loss and the accuracy acc can be calculated;
wherein, the prediction classification result of the sample is as follows:
Figure BDA0003563815510000131
the training error calculation formula is as follows:
Figure BDA0003563815510000132
1e) according to the cross entropy loss function loss, the network parameters of the depth residual error network based on the prototype idea are updated by back propagation, and the network parameters comprise class prototype prto stored in a full connection layer b . Therefore, the feature mapping of the base class training samples is closer to the corresponding class prototypes, so that clearer decision boundaries among the base class training samples are obtained, the base class training stage is completed, and the network model of the base class training stage is obtained.
Specifically, the incremental learning comprises the following specific steps:
2a) training a small number of incremental samples x according to the new class index i And its corresponding classification label y i Sending the network model into a network model in a base class training stage, mapping the network model into a feature space to obtain a mapping expression f (x) i )。
2b) Averaging the mapping expression of the same class of samples to obtain a new class prototype prto of the new class in a feature space i And prto the new prototype i And the 8 th weight of the full connection layer is used, wherein the first 7 weights are class prototypes of the base class training stage.
2c) A new parameter is introduced into the network, 8 types of samples (including 7 types of training samples in a base class training stage and 1 type of incremental training samples in an incremental learning stage) in the front of a training set need to be input into the network after the new parameter is introduced, a loss function is calculated by combining classification labels corresponding to the training samples, and then 8 weights of the full connection layer are updated through back propagation.
2d) When entering a second incremental learning stage, repeating the steps 2a) -2c) to obtain a new class prototype by averaging the mapping of the new class sample, wherein the new class prototype is used as a 9 th weight of the full connection layer, and the 9 weights are updated by using the previous 9 classes of training set samples.
2e) And in the third incremental learning stage, repeating the steps 2a) -2c) to obtain a new prototype by averaging the mapping of the new samples, using the new prototype as the 10 th weight of the fully-connected layer, and updating the 10 weights by using the previous 10 training set samples to obtain the prototype-thought-based deep residual error network which is trained and learned.
In this embodiment, a depth residual network based on prototype ideas is used to perform training using MSTAR image samples in stages and to perform validation using a test set. The classifier with excellent recognition performance can be obtained in the base class training stage, and the new class data of the small sample is continuously sent into the network, namely: the performance of the classifier is degraded in the incremental learning stage, and the forgetting of the knowledge is necessary. Therefore, in this embodiment, the prototype of the top 7 classes is also used as the weight of the network full-connection layer during base class training; a new class prototype is obtained in each class increment learning stage, fine adjustment is carried out on the obtained class prototype, the process is equivalent to that in a graph attention network, a forward propagation error is calculated by using a training sample, then the error is propagated reversely to update the weight of a network full-connection layer, namely the relative position of the fine adjustment class prototype, so that a better classification boundary is obtained, a continuous evolution classifier is enabled to obtain the capability of accurately identifying the new class sample, and the network identification performance is improved.
In the SAR image target recognition method based on small sample incremental learning, a depth residual error network based on a prototype idea is designed, the network can have the capability of automatically extracting the SAR image features even under the condition of less image samples, compared with the prior art, the method for mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of the number of samples to a certain extent, and the recognition capability of the whole network on the small sample image is improved;
in the SAR image target recognition method based on small sample incremental learning in the embodiment, a continuous evolution classifier is designed to be capable of continuously learning a new class in a trained network, and the setting is more in line with the requirement of continuously learning the new SAR image information of the small sample in reality. And moreover, an attention network is adopted to store various prototypes, the distance between the prototypes is finely adjusted in the attention network to obtain a better classification boundary, the storage mode is directly integrated on a full-connection layer of the network, and the prototypes are trained along with the network according to training samples, so that the network has a better decision boundary than the network of fixed prototypes, and the recognition accuracy is improved.
Example two
The present embodiment explains, by using a simulation experiment, an effect of the SAR image target identification method based on small sample incremental learning in the first embodiment.
(1) Conditions of the experiment
The hardware platform of the simulation experiment of this embodiment is: the GPU is NVIDIA GeForce RTX 2080Ti and 20 cores, the main frequency is 2.2GHz, and the memory size is 128 GB; the video memory size is 11 GB.
The software platform of the simulation experiment of the embodiment is as follows: the operating system is windows 10.
The base class training samples of the simulation experiment of the embodiment are selected from 7 types of SAR images observed by radar in MSTAR data set under a pitch angle of 17 degrees, namely BMP-2(TG1), BTR-70(TG2), T-72(TG3), 2S1(TG4), BRDM-2(TG5), ZSU-234(TG6) and ZIL-131(TG 10); the incremental training samples used in the incremental learning stage are the remaining three types, namely BTR-60(TG7), D-7(TG8) and T-62(TG 9).
The test set samples are the test sets in the MSTAR dataset with a pitch angle of 15 °.
(2) Emulated content
In order to verify the small sample identification capability of this embodiment, two small sample incremental learning modes, namely, 1-way-1-shot and 1-way-5-shot, are respectively set in a simulation experiment, and the simulation result is shown in table 1:
TABLE 1 simulation results
Figure BDA0003563815510000151
Figure BDA0003563815510000161
It can be seen from table 1 that the continuous evolution classifier also reaches a higher level when the number of samples in the incremental learning phase is small. As can be seen from the average accuracy of the test, even if only 1 sample is used in each iteration, the average accuracy of the identification is only a little lower than that of 5 samples used in each iteration, which shows that the depth residual error network based on the prototype idea can overcome the overfitting phenomenon caused by small samples and can generate a better decision boundary for SAR image classification.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Furthermore, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that an article or device that comprises a list of elements does not include only those elements but may include other elements not expressly listed. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of additional like elements in the article or device in which the element is included. The terms "connected" or "coupled" and the like are not restricted to physical or mechanical connections, but may include electrical connections, whether direct or indirect.
The foregoing is a further detailed description of the invention in connection with specific preferred embodiments and it is not intended to limit the invention to the specific embodiments described. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.

Claims (8)

1. A SAR image target identification method based on small sample increment learning is characterized by comprising the following steps:
step 1: constructing a depth residual error network based on a prototype idea;
step 2: carrying out base class training and incremental learning on the depth residual error network based on the prototype idea by utilizing a base class data set and an incremental data set;
and step 3: and inputting the SAR image to be detected into a depth residual error network which is completed by training and learning and based on a prototype idea to obtain a prediction classification result of the SAR image to be detected.
2. The SAR image target recognition method based on small sample incremental learning of claim 1, wherein the depth residual error network based on prototype idea comprises a convolution module, a depth residual error network, an adaptive average pooling layer, a full connection layer and a classifier which are connected in sequence,
the convolution module comprises a first convolution layer and a first ReLU activation layer which are sequentially connected;
the depth residual error network comprises a plurality of residual error blocks which are connected in sequence and used for extracting the image characteristics of the input image;
the full-connection layer is used for storing class prototypes, a graph attention network is integrated in the full-connection layer, and the graph attention network is used for adjusting the positions of the class prototypes stored on the full-connection layer in an increment learning stage;
the classifier is a cosine classifier and is used for outputting a prediction classification result.
3. The SAR image target recognition method based on small sample incremental learning of claim 2, wherein the residual block comprises a second convolution layer, a first batch of normalization layers, a second ReLU activation layer, a third convolution layer, a second batch of normalization layers and a third ReLU activation layer which are connected in sequence;
and the result of the addition of the input of the residual block and the output of the residual block sequentially passes through the second convolution layer, the first batch of normalization layers, the second ReLU activation layer, the third convolution layer and the second batch of normalization layers and is output after passing through the third ReLU activation layer.
4. The SAR image target recognition method based on small sample incremental learning of claim 3, wherein the residual block further comprises a channel number conversion unit, the channel number conversion unit is connected between the input end of the second convolutional layer and the output end of the second normalization layer, and the channel number conversion unit comprises a convolutional layer and a batch normalization layer which are connected in sequence;
and when the input channel number of the residual block is not equal to the output channel number of the residual block, converting the input of the residual block into the same channel number as the output of the residual block through the channel number conversion unit, and then outputting a result of adding the result with the output of the second batch of normalization layers after passing through the third ReLU activation layer.
5. The SAR image target recognition method based on small sample incremental learning of claim 1, wherein the base class dataset and the incremental dataset each comprise a plurality of classes of SAR images with classification labels attached, and the incremental dataset is a small sample dataset of 1-way-5-shot or 1-way-1-shot.
6. The SAR image target recognition method based on small sample incremental learning of claim 2, wherein the step 2 comprises:
step 2.1: in the base class training stage, the base class data set is used for training and learning the depth residual error network based on the prototype thought, and the network weight of the depth residual error network based on the prototype thought is updated through back propagation to obtain a network model in the base class training stage;
step 2.2: in the incremental learning stage, the incremental data set is used for training and learning the network model in the base class training stage, and the weight of the full connection layer is updated through back propagation to obtain the depth residual error network based on the prototype idea after training and learning.
7. The SAR image target recognition method based on small sample incremental learning of claim 6, wherein the step 2.1 comprises:
step 1 a: initializing network parameters of the depth residual error network based on the prototype idea;
step 1 b: obtaining a base class training sample in the base class data set and a corresponding classification label according to the index, inputting the base class training sample into the depth residual error network based on the prototype idea, and obtaining the mapping of the base class training sample in a feature space after the base class training sample passes through the convolution module and the depth residual error network;
step 1 c: the mapping passes through the self-adaptive average pooling layer and the full connection layer and then is input into the classifier, and the cosine similarity between the base class training sample and the class prototype is output through the classifier;
step 1 d: taking the class corresponding to the index with the maximum cosine similarity as a prediction classification result of the base class training sample, and calculating a cross entropy loss function and accuracy according to a classification label corresponding to the base class training sample;
step 1 e: and updating the network weight of the depth residual error network based on the prototype idea through back propagation according to the cross entropy loss function, taking the mean value of the mapping expression of the base class training sample in the feature space as a class prototype, and storing the mean value as the weight of the full connection layer to complete the base class training stage of the depth residual error network based on the prototype idea, so as to obtain a network model of the base class training stage.
8. The SAR image target recognition method based on small sample incremental learning of claim 7, wherein the step 2.2 comprises:
step 2 a: acquiring an incremental training sample in the incremental data set and a corresponding classification label thereof, and inputting the incremental training sample into the network model of the base class training stage to obtain the mapping of the incremental training sample in a feature space;
and step 2 b: taking the mean value of the mapping expression of the same type of samples to obtain a new type prototype of the new type in the feature space, and taking the new type prototype as the newly added weight of the full connection layer;
and step 2 c: inputting the base class training samples and the increment training samples into the network model obtained in the step 2b, and calculating a cross entropy loss function according to the base class training samples, the increment training samples and corresponding classification labels;
step 2 d: and updating the weight of the full connection layer through back propagation according to the cross entropy loss function to obtain the depth residual error network based on the prototype thought, which is completed by training and learning.
CN202210296918.XA 2022-03-24 2022-03-24 SAR image target identification method based on small sample incremental learning Pending CN114943889A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210296918.XA CN114943889A (en) 2022-03-24 2022-03-24 SAR image target identification method based on small sample incremental learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210296918.XA CN114943889A (en) 2022-03-24 2022-03-24 SAR image target identification method based on small sample incremental learning

Publications (1)

Publication Number Publication Date
CN114943889A true CN114943889A (en) 2022-08-26

Family

ID=82906089

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210296918.XA Pending CN114943889A (en) 2022-03-24 2022-03-24 SAR image target identification method based on small sample incremental learning

Country Status (1)

Country Link
CN (1) CN114943889A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116524322A (en) * 2023-04-10 2023-08-01 北京盛安同力科技开发有限公司 SAR image recognition method based on deep neural network
CN117975203A (en) * 2024-04-02 2024-05-03 山东大学 Small sample image type increment learning method and system based on data enhancement

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116524322A (en) * 2023-04-10 2023-08-01 北京盛安同力科技开发有限公司 SAR image recognition method based on deep neural network
CN116524322B (en) * 2023-04-10 2024-07-12 北京盛安同力科技开发有限公司 SAR image recognition method based on deep neural network
CN117975203A (en) * 2024-04-02 2024-05-03 山东大学 Small sample image type increment learning method and system based on data enhancement

Similar Documents

Publication Publication Date Title
CN109086700B (en) Radar one-dimensional range profile target identification method based on deep convolutional neural network
CN109766835B (en) SAR target recognition method for generating countermeasure network based on multi-parameter optimization
CN113378632B (en) Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method
CN110232341B (en) Semi-supervised learning image identification method based on convolution-stacking noise reduction coding network
CN110045015B (en) Concrete structure internal defect detection method based on deep learning
CN113326731B (en) Cross-domain pedestrian re-identification method based on momentum network guidance
CN114943889A (en) SAR image target identification method based on small sample incremental learning
CN111160176B (en) Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network
CN106709486A (en) Automatic license plate identification method based on deep convolutional neural network
CN111079847B (en) Remote sensing image automatic labeling method based on deep learning
CN111582236B (en) LPI radar signal classification method based on dense convolutional neural network
CN106951915B (en) One-dimensional range profile multi-classifier fusion recognition method based on category confidence
CN112764024B (en) Radar target identification method based on convolutional neural network and Bert
CN112990334A (en) Small sample SAR image target identification method based on improved prototype network
CN111598167B (en) Small sample image identification method and system based on graph learning
CN112464717A (en) Remote sensing image target detection method, system, electronic equipment and storage medium
CN112684427A (en) Radar target identification method based on serial quadratic reinforcement training
CN115034257B (en) Cross-modal information target identification method and device based on feature fusion
CN113406623A (en) Target identification method, device and medium based on radar high-resolution range profile
CN115311531A (en) Ground penetrating radar underground cavity target automatic detection method of RefineDet network model
CN112835008B (en) High-resolution range profile target identification method based on attitude self-adaptive convolutional network
CN113239959A (en) Radar HRRP target identification method based on decoupling representation variational self-coding machine
CN116563636A (en) Synthetic aperture radar image generation method and system
CN113240081B (en) High-resolution range profile target robust identification method aiming at radar carrier frequency transformation
CN116047418A (en) Multi-mode radar active deception jamming identification method based on small sample

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