CN110705508A - Satellite identification method of ISAR image - Google Patents

Satellite identification method of ISAR image Download PDF

Info

Publication number
CN110705508A
CN110705508A CN201910980263.6A CN201910980263A CN110705508A CN 110705508 A CN110705508 A CN 110705508A CN 201910980263 A CN201910980263 A CN 201910980263A CN 110705508 A CN110705508 A CN 110705508A
Authority
CN
China
Prior art keywords
isar
image
satellite
data
dcnn
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
CN201910980263.6A
Other languages
Chinese (zh)
Inventor
李智
张刚
李雪薇
徐灿
尹灿斌
方宇强
林财永
程文华
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.)
Peoples Liberation Army Strategic Support Force Aerospace Engineering University
Original Assignee
Peoples Liberation Army Strategic Support Force Aerospace Engineering 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 Peoples Liberation Army Strategic Support Force Aerospace Engineering University filed Critical Peoples Liberation Army Strategic Support Force Aerospace Engineering University
Priority to CN201910980263.6A priority Critical patent/CN110705508A/en
Publication of CN110705508A publication Critical patent/CN110705508A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention provides a satellite identification method of an ISAR image, which comprises the following steps: making two-dimensional real image sample data of the ISAR; amplifying the manufactured two-dimensional real image data of the ISAR, and randomly dividing the amplified image into a training set, a verification set and a test set; manufacturing a satellite target label for the training set, the verification set and the test set; constructing a satellite recognition Deep Convolutional Neural Network (DCNN) structure for the ISAR image; setting DCNN network training parameters suitable for ISAR images; training the DCNN by using a training set and a verification set to obtain a satellite target identification network model of the ISAR image; and testing the network model by using the test set, and verifying and identifying the performance of the network model. The invention can greatly save manpower and time, reduce uncertain influence and artificial subjective judgment error caused by artificial selection of characteristics on satellite target identification results, and greatly improve the satellite target identification precision and identification efficiency of ISAR images.

Description

Satellite identification method of ISAR image
Technical Field
The invention belongs to the technical field of Radar target identification, and relates to a satellite identification method of an ISAR (Inverse synthetic Aperture Radar) image, in particular to a satellite target identification method of an ISAR two-dimensional image.
Background
With the high importance of aerospace in various countries of the world and the rapid development of civil aerospace, more and more satellites are launched and lifted off, and the space becomes increasingly crowded and more competitive and antagonistic. How to effectively sense the space situation and further control the space becomes a big problem currently faced by each aerospace large country. As an important means for effectively sensing spatial situation, the broadband radar has the characteristics of all-time, all-weather and high resolution, plays an important role in a spatial target identification system, and is widely applied.
There are two main identification methods for identifying a satellite target by using a broadband radar: ISAR/SAR (Synthetic Aperture Radar, SAR) -based two-dimensional images and HRRP (High Resolution Range Profile, HRRP) -based one-dimensional distance images. The HRRP-based satellite target identification method is the projection of a two-dimensional image of a satellite target in the vertical direction (transverse direction) and the emission direction (longitudinal direction) of a radar respectively, and the projection mode can cause the loss of partial information (texture and shape) of the satellite target. The original data of the two-dimensional image of the ISAR is complex data, and is difficult to process by using the conventional method. However, with the great success of Deep Convolutional Neural Network (DCNN) in computer vision, big data analysis, and the like, and the gpu (graphical processing unit) with strong computing power, it is possible to process ISAR two-dimensional images using DCNN. The DCNN is a nonlinear network structure composed of a convolutional layer, a pooling layer, a nonlinear activation layer and a full connection layer, and has very strong feature representation capability. And the DCNN is used for carrying out nonlinear mapping on the satellite target of the ISAR image, so that the detection, identification and classification precision and efficiency of the satellite target of the ISAR image are improved. Therefore, the research on the satellite target identification classification of the ISAR image based on the DCNN has important research significance.
The key point of the DCNN-based ISAR image satellite target identification method is the selection and extraction of satellite target characteristics. The traditional feature learning method mainly adopts artificial design features, and can be roughly divided into the following two types: one method is to identify the satellite target by extracting target airspace statistical characteristics or transform domain statistical characteristics, such as extracting the spectrum intensity, bispectrum characteristics and the number of scattering points of the target echo; another method is to extract the HRRP features of the radar through rational mathematical modeling, for example, using hidden markov models, matching pursuit models, principal component analysis models, and the like. However, the conventional method for learning the characteristics of the ISAR image requires a lot of manpower for model design, and also requires a lot of target prior knowledge, and the requirement on the expertise is very high. In addition, these feature extraction methods are based on human in-circuit design choices, which require a lot of time and effort. Meanwhile, the extracted features have great uncertainty, and the generalization capability of different target types to different radars is weak. The radar target identification method based on the shallow neural network utilizes a BP (Back propagation) neural network and a multilayer perceptron to extract radar target characteristics. The satellite target characteristics extracted by the method have the accuracy rate which is not in an ideal state in the aspect of satellite target identification of the ISAR images.
Disclosure of Invention
The invention aims to overcome the defects of the technology and provides a method for identifying the satellite target of the ISAR image. DCNN is a class of strongly supervised learning algorithms for deep learning, consisting of a series of Convolutional layers (Convolutional layer), Pooling layers (posing layer), nonlinear active layers, and Fully Connected layers (FC). The two-dimensional image of the ISAR and the satellite target label are used as the input of the network, and the deep features of the satellite target are automatically extracted through continuous convolution and pooling operations. After each convolution layer, a nonlinear activation layer is connected, so that the nonlinear feature representation capability of the network can be greatly improved. And finally, reducing the dimension of the extracted satellite target characteristics of the ISAR image through a full connection layer to finish the identification of the satellite target. The invention can eliminate the uncertain influence on the satellite target identification and classification result caused by manually selecting the features, and can also greatly improve the identification precision and the identification efficiency of the satellite target.
In order to achieve the purpose, the invention is concretely realized by the following technical scheme:
the invention provides a satellite identification method of an ISAR image, which comprises the following steps:
step one, obtaining a module value of original complex data of an ISAR image to obtain an ISAR amplitude image, and performing normalization processing on the ISAR amplitude image to obtain a normalized ISAR amplitude image;
processing the normalized ISAR amplitude image into a 0-255 gray image, and reconstructing the gray image into an RGB three-channel color image by using an openCV to obtain two-dimensional real image sample data of the ISAR;
amplifying two-dimensional real image sample data of the ISAR, and randomly dividing the amplified image data into a training set, a verification set and a test set; the training set and the verification set are used for training the Deep Convolutional Neural Network (DCNN), and the test set is used for testing the performance of a satellite target recognition network model of the DCNN; manufacturing satellite target class labels for the training set, the test set and the verification set, wherein the format of the satellite target class labels is 'ISAR two-dimensional real image sample name + satellite target class name';
step four, constructing a DCNN network structure for satellite target identification of the ISAR image;
the constructed DCNN structure for satellite target identification of the ISAR image consists of an input layer, a hidden layer and an output layer; the input of the input layer is an ISAR image and a satellite target class label; the ISAR image data format is H W C, H and W are the length and width of the input ISAR image, and C is the channel number of the input image; the number of layers of the hidden layer needs to be selected according to the size of the training set, and the output of the previous layer of the hidden layer is used as the input of the next layer; the output layer is a classification layer, and the output number of the classification layer corresponds to the number of satellite target classes;
step five, setting training parameters of the DCNN structure suitable for satellite target recognition of the ISAR images constructed in the step four, wherein the training parameters comprise a training deep learning platform of the DCNN structure suitable for satellite target recognition of the ISAR images, maximum iteration times, a learning rate, batch processing parameters (batch size), momentum, an optimization algorithm of network parameter weight and the like;
step six, training the DCNN structure of the satellite target recognition of the ISAR image constructed in the step four according to the training parameters set in the step five by using a training set and a verification set to obtain a satellite target recognition network model of the ISAR image;
and step seven, testing the satellite target identification network model of the ISAR image obtained in the step six by using the test set in the step three, and verifying the satellite target identification performance of the satellite target identification network model on the satellite target in the ISAR image.
In the first step, obtaining a magnitude image of the ISAR by taking a module value from the original complex data of the ISAR image, includes:
carrying out modulus value operation on original complex data of the ISAR image, converting the complex data into real image data of the ISAR:
Figure BDA0002234960030000031
where s is the amplitude value of ISAR, R is the real part of the ISAR raw data, and I is the imaginary part of the ISAR raw data.
In the first step, the normalization processing is performed on the ISAR amplitude image to obtain a normalized ISAR amplitude image, and the method includes:
Figure BDA0002234960030000041
wherein p isnThe amplitude range is 0-1 for the normalized ISAR amplitude image; snFor the norm of the original complex data of the ISAR image, max (-) and min (-) are maximum and minimum functions, respectively.
In the second step, the normalized ISAR amplitude image is processed into a 0-255 gray image, and the gray image is reconstructed into an RGB three-channel color image by using openCV, so as to obtain two-dimensional real image sample data of the ISAR, which includes:
image_r＝pn×255
image_g＝pn×255,image＝merge[image_r,image_g,image_b],n＝1,2,3,...,N；
image_b＝pn×255
wherein, image _ r, image _ g and image _ b are RGB three primary color pixel values of ISAR color image, merge [ ·]As a function of data fusion, pnThe image is the two-dimensional real image sample data of the ISAR, which is a normalized ISAR amplitude image.
In the third step, the amplification of the two-dimensional real image sample data of the ISAR comprises the following steps:
and turning, rotating, mirroring, translating, adding noise and the like on the image of the two-dimensional real image sample data of the ISAR.
In the third step, randomly dividing the amplified image data into a training set, a verification set and a test set, including:
randomly dividing the image data after data amplification into three image subsets of a training data set train, a verification data set val and a test data set test used by DCNN, wherein,
the train dataset accounts for 70% of the entire dataset, the val dataset accounts for 10% of the entire image dataset and the test dataset accounts for 20% of the entire dataset; the training data set train and the verification data set val are used for training the DCNN, and the test data set test is used for testing the final satellite target recognition network model.
In the third step, a satellite target class label is made for the training set, the test set and the verification set, and the format of the satellite target class label is 'ISAR two-dimensional real image sample name + satellite target class name', and the method comprises the following steps:
and writing a satellite target class label making program by using a python language, only making satellite target class labels on the train data set, the test data set and the val data set, and obtaining three txt documents which are named as train. txt is the ISAR two-dimensional real image sample name and satellite target class in the train dataset; txt is the ISAR two-dimensional real image sample name and the satellite target class in the test data set; txt is the ISAR two-dimensional real image sample name and satellite object class name in the val dataset.
In the fourth step, constructing a DCNN network structure for satellite target identification of the ISAR image includes:
the DCNN structure for satellite target identification of the ISAR image has ten layers in total, namely an input layer, five convolution layers, three full-connection layers and an output layer; wherein, the five convolution layers and the three full-connection layers form a hidden layer of the network; the number of layers of the DCNN network structure can be increased or decreased according to the number of images.
In the fourth step, the output layer is a classification layer, and the loss function of the output layer is SoftMaxWithLoss. The SoftMaxWithLoss function is defined as:
Figure BDA0002234960030000051
wherein the content of the first and second substances,
Figure BDA0002234960030000052
true class of satellite targets, f (z)i) Class of predicted satellite objects for the DCNN model, ziIs the input ISAR image.
In the fifth step, the setting of the training parameters of the DCNN network structure for satellite target recognition of the ISAR image includes:
the deep learning platform of the DCNN structure for training satellite target recognition of the ISAR image is Caffe, the maximum iteration number is 5000 times, the learning rate is 0.001, the batch size is 32, the momentum is 0.9, and the optimization algorithm of the network parameter weight is a random gradient optimization algorithm.
The invention has the beneficial effects that:
(1) compared with the traditional recognition algorithm, the technical scheme of the invention can utilize the superiority of DCNN in feature extraction, and realize automatic extraction of deep features of the satellite target of the ISAR image through layer-by-layer nonlinear transformation.
(2) The method takes data as drive, automatically extracts the satellite target characteristics through reasonable reconstruction ISAR image data and DCNN, gets rid of uncertainty of manual design and screening of the data characteristics, greatly saves the design cost of the characteristics, and improves the accuracy of target identification and classification.
(3) The method can realize the integration of the feature extraction and the recognition classification of the ISAR image data, simplifies the target recognition process, and is suitable for the requirement of mass and real-time target recognition.
(4) The method has good generalization capability on multi-category satellite target data, only needs to reconstruct the original target data obtained under different radar parameters into image data, does not need to modify other layers of the network, has good applicability to different parameter radars and different targets, and has strong generalization capability.
Drawings
Fig. 1 is a flowchart illustrating a satellite identification method for an ISAR image according to the present invention;
FIG. 2 shows ISAR image partial raw data and normalized data of a modulus value;
FIG. 3 shows sample data of ten kinds of ISAR two-dimensional real images;
FIG. 4 shows an example of an image augmented with Number-5 satellite target data;
FIG. 5 is a diagram illustrating a DCNN network architecture for satellite target recognition of ISAR images provided by the present invention;
FIG. 6 is a graph showing the variation of the loss function values with the number of iterations in the network structure training of the present invention;
FIG. 7 is a graph of the variation of the validation set loss function values with iteration times in the network structure training of the present invention;
FIG. 8 is a graph of validation set accuracy versus iteration number in the network structure training of the present invention;
FIG. 9 shows the test results of the satellite identification network structure model of the ISAR image according to the test set of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
The embodiment of the invention provides a satellite target identification method of ISAR images of ten satellite targets based on DCNN, which is respectively named as Number-1, Number-2, Number-3, Number-4, Number-5, Number-6, Number-7, Number-8, Number-9 and Number-10. The image sample data of each satellite target is more than 1000 pieces after being amplified, and fig. 1 shows a flow chart of a satellite target identification method of an ISAR image. As shown in fig. 1, an embodiment of the present invention discloses a satellite target identification method for an ISAR image, including:
step one, obtaining a module value of original complex data of an ISAR image to obtain an ISAR amplitude image, and performing normalization processing on the ISAR amplitude image to obtain a normalized ISAR amplitude image.
As shown in fig. 2, Raw Data (Raw Data) of an ISAR image is a Data matrix with a pattern of x ═ R + Ij (R is a real part of the Data, and I is an imaginary part of the Data), so that a modulus operation needs to be performed on the Raw Data
Figure BDA0002234960030000072
Where s is the amplitude value of ISAR, R is the real part of the ISAR raw data, and I is the imaginary part of the ISAR raw data. Converting the complex image data into real image data to obtainThe data of (2) is normalized. For the original complex data of the ISAR image, the magnitude information of the object has much greater effect than the phase information. When ISAR original complex data are processed, phase information of an ISAR image is removed, and image calculation complexity is reduced by taking a modulus value of the ISAR original complex data.
Carrying out normalization processing on the ISAR amplitude image to obtain a normalized ISAR amplitude image, wherein the steps comprise:
Figure BDA0002234960030000071
wherein p isnThe range of the amplitude is 0-1 for the normalized ISAR amplitude image. snFor the norm of the original complex data of the ISAR image, max (-) and min (-) are maximum and minimum functions, respectively.
And step two, processing the normalized ISAR amplitude image obtained in the step one into a gray image of 0-255, reconstructing the gray image into an RGB three-channel color image by using openCV, and finally obtaining two-dimensional real image sample data of the ISAR. Fig. 3 is two-dimensional real image sample data of ten kinds of ISARs used in the embodiment.
Specifically, processing the normalized ISAR amplitude image into a 0-255 gray image, and reconstructing the gray image into an RGB three-channel color image by using openCV to obtain ISAR two-dimensional real image sample data, which comprises:
image_r＝pn×255
image_g＝pn×255,image＝merge[image_r,image_g,image_b],n＝1,2,3,...,N；
image_b＝pn×255
wherein, image _ r, image _ g and image _ b are RGB three primary color pixel values of ISAR color image, merge [ ·]As a function of data fusion, pnThe image is the two-dimensional real image sample data of the ISAR, which is a normalized ISAR amplitude image.
The reason for processing the normalized ISAR amplitude image into an RGB three-channel color image is to increase the number of satellite target features. In the process of extracting the features of the satellite target of the ISAR image by using the DCNN, compared with a normalized ISAR gray image, the color image using the RGB tee joint has the advantage that the number of the features of the satellite is increased by three times. The method is beneficial to improving the satellite target identification accuracy of the ISAR image.
And step three, amplifying the two-dimensional real image sample data of the ISAR generated in the step two, and randomly dividing the amplified image data into a training set, a verification set and a test set. The training set and the verification set are used for training a Deep Convolutional Neural Network (DCNN), and the test set is used for testing the performance of a satellite target recognition model of the DCNN. And finally, manufacturing satellite target class labels for the training set, the test set and the verification set, wherein the format of the satellite target class labels is 'ISAR two-dimensional real image sample name + satellite target class name'.
The method comprises the steps of amplifying generated two-dimensional real image sample data of the ISAR, wherein a plurality of data amplification methods are adopted, and image rotation, image overturning, image mirroring, image translation and image denoising are mainly adopted. Wherein the degree of rotation of the image rotation is 45 °, 90 °, 135 °, 180 °, 225 °, and 270 °; image denoising is adding additive white gaussian noise with variance of 25 to two-dimensional real image sample data of ISAR. The software language for implementing the augmentation mode used in this example is python 3. FIG. 4 is an example of an image augmented with Number-5 satellite target data.
Randomly dividing the image data after data amplification into three image subsets, namely a training data set train, a verification data set val and a test data set test, which are used by DCNN, and including:
the train dataset accounts for 70% of the entire dataset, the val dataset accounts for 10% of the entire image dataset and the test dataset accounts for 20% of the entire dataset; the training data set train and the verification data set val are used for training the DCNN, and the test data set test is used for testing the final satellite target recognition network model.
The method for manufacturing the satellite target class label for the training set, the test set and the verification set comprises the following steps:
and making satellite object category labels for three data sets of train, val and test, wherein the labels are three text documents respectively. The satellite object class label format is image name + satellite object class name.
And writing a satellite target class label making program by using a python language, only making satellite target class labels on the train data set, the test data set and the val data set, and obtaining three txt documents which are named as train. txt is the ISAR two-dimensional real image sample name and satellite target class in the train dataset; txt is the ISAR two-dimensional real image sample name and the satellite target class in the test data set; txt is the ISAR two-dimensional real image sample name and satellite object class name in the val dataset.
Step four, constructing a DCNN network structure for satellite target identification of the ISAR image;
the constructed DCNN network structure for satellite target recognition of the ISAR image consists of an input layer, a hidden layer and an output layer. The input of the input layer is an ISAR image and a satellite target class label; the ISAR image data format is H W C, H and W are the length and width of the input ISAR image, and C is the channel number of the input image; the number of hidden layers needs to be selected according to the size of the training set, and the phenomenon of over-fitting or under-fitting during training of the constructed DCNN network structure is mainly avoided. The output of the previous layer of the hidden layer is used as the input of the next layer; the output layer is a classification layer, and the output number of the classification layer corresponds to the number of satellite target classes; comparing the test results of the DCNN network structures with different layers, the hidden layer of the DCNN network structure for satellite target recognition of the ISAR image constructed by the people finally comprises a five-layer volume base layer and three full-connection layers. Fig. 5 is a DCNN network structure of satellite target recognition of an ISAR image constructed in the present embodiment;
constructing a DCNN network structure for satellite identification of ISAR images, comprising:
the DCNN network structure for satellite identification of ISAR images has ten layers in total, namely an input layer, five convolution layers, three full-connection layers and an output layer. Wherein, five convolution layers and three full connection layers form a hidden layer of the network. The number of network structure layers of the DCNN can be increased or decreased according to the number of images.
The output layer is a classification layer whose loss function is SoftMaxWithLoss. The SoftMaxWithLoss function is defined as:
Figure BDA0002234960030000091
wherein the content of the first and second substances,
Figure BDA0002234960030000092
true class of satellite targets, f (z)i) Class of predicted satellite objects for the DCNN model, ziIs the input ISAR image.
Step five, setting training parameters of the DCNN structure suitable for satellite target recognition of the ISAR images constructed in the step four, wherein the training parameters comprise a training deep learning platform of the DCNN structure suitable for satellite target recognition of the ISAR images, maximum iteration times, a learning rate, batch processing parameters (batch size), momentum, an optimization algorithm of network parameter weight and the like; wherein, the deep learning platform is determined according to the environment of the future network model; the maximum iteration times are set according to training loss of a DCNN network structure for training satellite target recognition of the ISAR images; the learning rate is determined according to the oscillation amplitude of training loss of a DCNN network structure for training satellite target recognition of the ISAR image; the batch processing parameters are determined according to the sizes of the GPU and the CPU of the computer hardware; the optimization algorithm of the momentum and the network parameter weight is determined according to whether the training loss function is converged quickly or not.
In the DCNN network structure training parameter setting of satellite target recognition of ISAR images, the method comprises the following steps: the deep learning platform of the DCNN structure for training satellite target recognition of the ISAR image is Caffe, the maximum iteration number is 5000 times, the learning rate is 0.001, the batch size is 32, the momentum is 0.9, and the optimization algorithm of the network parameter weight is a random gradient optimization algorithm.
Step six, training the DCNN structure of the satellite target recognition of the ISAR image constructed in the step four according to the training parameters set in the step five by using a training set and a verification set to obtain a satellite target recognition network model of the ISAR image; in the training process, whether the DCNN network structure identified by the satellite target of the ISAR image is over-fit or under-fit and whether the network structure can be rapidly converged is determined according to the training loss value and the verification loss value of the network. And if any phenomenon exists, stopping network training, and modifying the training parameters of the DCNN structure of the satellite target recognition of the ISAR image.
And step seven, testing the satellite target identification network model of the ISAR image obtained in the step six by using the test set in the step three, and verifying the satellite target identification performance of the satellite target identification network model on the satellite target in the ISAR image.
In this example, the original size of the ISAR image is 300 × 800 × 3, containing ten types of satellite targets. Adjusting ISAR images in the data set obtained in step three to W1×H1(W in the present invention)1＝227,H1227) and as an input to the DCNN network structure for satellite object identification of ISAR images, the features of the satellite objects are extracted through multi-layer convolution and pooling operations. According to the actual requirement, we adopt the batch size as 32. Thus, the dimension of the input data is 4,946,784 dimensions (227 × 3 × 32 ═ 4,946,784). Fig. 5 is a diagram illustrating a DCNN network structure for satellite target recognition of an ISAR image according to the present invention. The hidden layer of the DCNN network structure for satellite target recognition of the ISAR image consists of five convolution layers and three full-connection layers. The first layer of convolutional layers input is 227 x 3 x 32, and the output characteristic is 55 x 96; the output of the first layer of convolution passes through the nonlinear active layer and the Pooling layer Powing 1, and the output characteristic is 27 × 96; the input of the second convolution layer is the output of Pooling1, i.e., 27 x 96, with an output characteristic of 27 x 256; the second convolution layer outputs the characteristic of 13 × 256 after passing through the nonlinear active layer and the Pooling layer Powing 2; the input of the third convolution layer is the output of Pooling2, namely 13 × 256, and the output characteristic is 13 × 384; the input of the fourth layer of convolution layer is 13 × 384, and the output is 13 × 384; the fifth layer of convolution input is 13 × 384, and the output is 13 × 256; the net finally passed through the Pooling layer Pooling5 with an output signature of 6 x 256; the dimensionality of the data at this point is 9, 216. Inputting the obtained 9216 dimensional data into a first layer full connection layer FC6, and performing first dimension reduction on satellite target characteristics to obtain 4096 dimensional data; however, the device is not suitable for use in a kitchenSending 4096-dimensional output of FC6 to a second layer full-connection layer for second dimension reduction to obtain 1000-dimensional satellite data; finally, the output 1000 of the FC7 is taken as data and sent to a third full-connection layer FC8 for third dimension reduction to obtain final 10-dimensional data; the obtained 10-dimensional data is sent to an output layer (classification layer) to obtain the satellite category of the final network prediction.
There are generally two approaches to the choice of DCNN network structure training for terminating satellite target recognition of ISAR images: firstly, setting fixed iteration times as a termination threshold condition, and when the Loss function value (Loss) of the network does not decrease along with the increase of the iteration times, indicating that the network has reached a convergence state; and secondly, setting a fixed Loss value as a threshold value, and stopping the training of the network when the Loss value of the network reaches the threshold value. In this example, the iteration number is set as a threshold for terminating network training, and is set to 5000 times. The results of the training set and validation set loss values and validation set accuracy as a function of the number of iterations are given in fig. 6,7, and 8. In addition, the optimizer of the DCNN network structure for satellite object recognition of ISAR images adopts a Stochastic Gradient Descent (SGD) optimizer including a momentum concept, and the activation function adopts a modified linear unit (ReLU). The RuLU has higher gradient descent and backward propagation efficiency than the traditional Sigmoid function, can effectively avoid the problems of gradient explosion and gradient disappearance, and ensures that the DCNN neural network obtains better training effect.
After the optimizer is determined, a suitable loss function needs to be selected to determine the weight space, and thus the optimization process, i.e., the process of minimizing the loss function, is performed. The DCNN network structure for satellite target identification of ISAR image takes the excellent classification cross entropy function in classification output as a loss function, and if the real class of the satellite target is
Figure BDA0002234960030000111
The satellite type of the DCNN network structure prediction of the satellite target recognition of the ISAR image is f (z)i) Then the classification cross entropy function can be defined as
Figure BDA0002234960030000121
Finally, in order to avoid overfitting in the training process, a Dropout layer is added in the network, and in the values propagated by the hidden layer, certain values are randomly discarded according to a Dropout probability value which is set to be 0.5.
For the purpose of more clearly explaining the present invention, it is convenient for those skilled in the art to understand that the present invention provides a specific example:
in this example, after the ISAR images of ten satellite targets are amplified, 10000 images are obtained in total, wherein the train data set is 7000 images, the val data set is 1000 images, and the test data set is 2000 images. The train data set and the val data set are used for training a DCNN structure of satellite target recognition of the ISAR image to obtain a DCNN structure model of satellite target recognition of the ISAR image; the test data set is used for testing a DCNN network structure model of satellite target identification of the ISAR image. FIG. 9 shows the test results of the satellite identification network structure model of the ISAR image in the test set. From the recognition results of the ten satellite targets in the graph, it can be concluded that the final average recognition rate Top-1 (former) of the satellite recognition network model of the ISAR image to the test set reaches 98.1% and Top-3 (former three) is 100%. And finally, packaging the trained satellite recognition network model of the ISAR image to be used as a target recognition network of the ISAR image.
As the invention uses DCNN to identify the satellite target of the ISAR image for the first time, the experimental result is based on the identification results of Top-1 and Top-3. Table 1 gives the identification accuracy for ten satellite targets.
TABLE 110 identification rates of satellite objects of the type
Figure BDA0002234960030000122
The result of the embodiment shows that the identification method of the invention has very high accuracy, and can improve the generalization capability and robustness of the network model by expanding the image sample. Meanwhile, the satellite target is identified without manually designed and screened ISAR image data characteristics, and the satellite target is automatically identified based on the DCNN neural network. The method comprises the steps of preprocessing original data of an ISAR image, including data modulus value taking, data normalization and image reconstruction, marking the data, and constructing a training set. And then constructing a DCNN structure of the satellite target recognition of the ISAR image, setting training parameters of the network structure, and using the training set and the verification set of the ISAR image for the training of the DCNN structure of the satellite target recognition of the ISAR image to obtain a DCNN structure model of the satellite target recognition of the ISAR image. And finally, preprocessing ISAR image data to be recognized, inputting the preprocessed ISAR image data into a DCNN structural model for satellite target recognition of the ISAR image, and judging the type of the satellite target according to the satellite target type value of the network output unit.
The invention realizes the end-to-end learning of the feature extraction and classification of the satellite target ISAR image data, simplifies the target identification process and can meet the requirement of mass and real-time target identification.
The invention has the beneficial effects that:
(1) compared with the traditional recognition algorithm, the technical scheme of the invention can realize the deepest representation of the original data by utilizing the superiority of deep learning in the aspect of feature extraction and through layer-by-layer nonlinear transformation.
(2) The method takes data as drive, automatically extracts the satellite target characteristics through reasonable reconstruction ISAR image data and DCNN, gets rid of uncertainty of manual design and screening of the data characteristics, greatly saves the design cost of the characteristics, and improves the accuracy of target identification and classification.
(3) The method can realize the integration of the feature extraction and the recognition classification of the ISAR image data, simplifies the target recognition process, and is suitable for the requirement of mass and real-time target recognition.
(4) The method has good generalization capability on multi-category satellite target data, only needs to reconstruct the original target data obtained under different radar parameters into image data, does not need to modify other layers of the network, has good applicability to different parameter radars and different targets, and has strong generalization capability.
The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.
It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the order of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required by the invention.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.

Claims (10)

1. A satellite identification method of an ISAR image is characterized by comprising the following steps:
step one, obtaining a module value of original complex data of an ISAR image to obtain an ISAR amplitude image, and performing normalization processing on the ISAR amplitude image to obtain a normalized ISAR amplitude image;
processing the normalized ISAR amplitude image into a 0-255 gray image, and reconstructing the gray image into an RGB three-channel color image by using an openCV to obtain two-dimensional real image sample data of the ISAR;
amplifying two-dimensional real image sample data of the ISAR, and randomly dividing the amplified image data into a training set, a verification set and a test set; the training set and the verification set are used for training the Deep Convolutional Neural Network (DCNN), and the test set is used for testing the performance of a satellite target recognition network model of the DCNN; manufacturing satellite target class labels for the training set, the test set and the verification set, wherein the format of the satellite target class labels is 'ISAR two-dimensional real image sample name + satellite target class name';
step four, constructing a DCNN network structure for satellite target identification of the ISAR image;
the constructed DCNN structure for satellite target identification of the ISAR image consists of an input layer, a hidden layer and an output layer; the input of the input layer is an ISAR image and a satellite target class label; the ISAR image data format is H W C, H and W are the length and width of the input ISAR image, and C is the channel number of the input image; the number of layers of the hidden layer needs to be selected according to the size of the training set, and the output of the previous layer of the hidden layer is used as the input of the next layer; the output layer is a classification layer, and the output number of the classification layer corresponds to the number of satellite target classes;
step five, setting training parameters of the DCNN structure suitable for satellite target recognition of the ISAR images constructed in the step four, wherein the training parameters comprise a training deep learning platform of the DCNN structure suitable for satellite target recognition of the ISAR images, maximum iteration times, a learning rate, batch processing parameters (batch size), momentum, an optimization algorithm of network parameter weight and the like;
step six, training the DCNN structure of the satellite target recognition of the ISAR image constructed in the step four according to the training parameters set in the step five by using a training set and a verification set to obtain a satellite target recognition network model of the ISAR image;
and step seven, testing the satellite target identification network model of the ISAR image obtained in the step six by using the test set in the step three, and verifying the satellite target identification performance of the satellite target identification network model on the satellite target in the ISAR image.
2. The method of claim 1, wherein in the first step, taking a modulus value of original complex data of an ISAR image to obtain an ISAR amplitude image, comprises:
carrying out modulus value operation on original complex data of the ISAR image, converting the complex data into real image data of the ISAR:
Figure FDA0002234960020000021
where s is the amplitude value of ISAR, R is the real part of the ISAR raw data, and I is the imaginary part of the ISAR raw data.
3. The method of claim 1, wherein in the first step, performing normalization processing on the ISAR amplitude image to obtain a normalized ISAR amplitude image comprises:
Figure FDA0002234960020000022
wherein p isnThe amplitude range is 0-1 for the normalized ISAR amplitude image; snFor the norm of the original complex data of the ISAR image, max (-) and min (-) are maximum and minimum functions, respectively.
4. The method according to claim 1, wherein in the second step, the normalized ISAR amplitude image is processed into a 0-255 gray scale image, and the gray scale image is reconstructed into an RGB three-channel color image by using openCV, so as to obtain two-dimensional real image sample data of the ISAR, including:
Figure FDA0002234960020000023
wherein, image _ r, image _ g and image _ b are RGB three primary color pixel values of ISAR color image, merge [ ·]As a function of data fusion, pnThe image is the two-dimensional real image sample data of the ISAR, which is a normalized ISAR amplitude image.
5. The method of claim 1, wherein in the third step, the amplifying two-dimensional real image sample data of ISAR comprises:
and turning, rotating, mirroring, translating, adding noise and the like on the image of the two-dimensional real image sample data of the ISAR.
6. The method of claim 1, wherein in step three, randomly dividing the augmented image data into a training set, a validation set, and a test set comprises:
randomly dividing the image data after data amplification into three image subsets of a training data set train, a verification data set val and a test data set test used by DCNN, wherein,
the train dataset accounts for 70% of the entire dataset, the val dataset accounts for 10% of the entire image dataset and the test dataset accounts for 20% of the entire dataset; the training data set train and the verification data set val are used for training the DCNN, and the test data set test is used for testing the final satellite target recognition network model.
7. The method of claim 1, wherein in the third step, the training set, the test set and the verification set are labeled with the satellite object class label in the format of "ISAR two-dimensional real image sample name + satellite object class name", and the method comprises:
writing a satellite target class label making program by using a python language, only making satellite target class labels on a train data set, a test data set and a val data set, and obtaining three txt documents which are named as train.txt, test.txt and val.txt respectively; txt is the ISAR two-dimensional real image sample name and satellite target class in the train dataset; txt is the ISAR two-dimensional real image sample name and the satellite target class in the test data set; txt is the ISAR two-dimensional real image sample name and satellite object class name in the val dataset.
8. The method as claimed in claim 1, wherein in the fourth step, constructing a DCNN network structure for satellite target recognition of ISAR images comprises:
the DCNN structure for satellite target identification of the ISAR image has ten layers in total, namely an input layer, five convolution layers, three full-connection layers and an output layer; wherein, the five convolution layers and the three full-connection layers form a hidden layer of the network; the number of layers of the DCNN network structure can be increased or decreased according to the number of images.
9. The method of claim 1, wherein in step four, the output layer is a classification layer and the loss function of the output layer is SoftMaxWithLoss. The SoftMaxWithLoss function is defined as:
Figure FDA0002234960020000031
wherein the content of the first and second substances,
Figure FDA0002234960020000032
true class of satellite targets, f (z)i) Class of predicted satellite objects for the DCNN model, ziIs the input ISAR image.
10. The method as claimed in claim 1, wherein in the fifth step, the setting of the training parameters of the DCNN network structure for satellite target recognition of ISAR images comprises:
the deep learning platform of the DCNN structure for training satellite target recognition of the ISAR image is Caffe, the maximum iteration number is 5000 times, the learning rate is 0.001, the batch size is 32, the momentum is 0.9, and the optimization algorithm of the network parameter weight is a random gradient optimization algorithm.
CN201910980263.6A 2019-10-15 2019-10-15 Satellite identification method of ISAR image Pending CN110705508A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910980263.6A CN110705508A (en) 2019-10-15 2019-10-15 Satellite identification method of ISAR image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910980263.6A CN110705508A (en) 2019-10-15 2019-10-15 Satellite identification method of ISAR image

Publications (1)

Publication Number Publication Date
CN110705508A true CN110705508A (en) 2020-01-17

Family

ID=69198881

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910980263.6A Pending CN110705508A (en) 2019-10-15 2019-10-15 Satellite identification method of ISAR image

Country Status (1)

Country Link
CN (1) CN110705508A (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112183185A (en) * 2020-08-13 2021-01-05 天津大学 Liquid leakage detection method based on optical flow method and CNN-SVM
CN112329878A (en) * 2020-11-16 2021-02-05 山西三友和智慧信息技术股份有限公司 Method for detecting sliding window of satellite image moving object based on convolutional neural network
CN112949555A (en) * 2021-03-17 2021-06-11 西安电子科技大学 Spatial target ISAR image classification method based on target prior information
CN113269135A (en) * 2021-06-17 2021-08-17 中国空间技术研究院 Satellite transponder recognition model and training method and using method thereof
CN113537020A (en) * 2021-07-07 2021-10-22 中国人民解放军国防科技大学 Complex SAR image target identification method based on improved neural network
CN115856854A (en) * 2023-02-24 2023-03-28 南京理工大学 Model and data hybrid driven radar detection method and system
CN115877344A (en) * 2023-02-21 2023-03-31 南京理工大学 Radar detection method and system for integrated processing of broadband detection, tracking and identification
CN116109944A (en) * 2023-04-12 2023-05-12 南京信息工程大学 Satellite image cloud target extraction method based on deep learning network architecture
CN116385813A (en) * 2023-06-07 2023-07-04 南京隼眼电子科技有限公司 ISAR image classification method, ISAR image classification device and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105373799A (en) * 2015-12-18 2016-03-02 北京环境特性研究所 Space target ISAR image processing method for template identification
CN108830242A (en) * 2018-06-22 2018-11-16 北京航空航天大学 SAR image targets in ocean classification and Detection method based on convolutional neural networks
CN108846409A (en) * 2018-04-28 2018-11-20 中国人民解放军国防科技大学 Radar echo extrapolation model training method based on cyclic dynamic convolution neural network
CN109753874A (en) * 2018-11-28 2019-05-14 南京航空航天大学 A kind of low slow small classification of radar targets method based on machine learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105373799A (en) * 2015-12-18 2016-03-02 北京环境特性研究所 Space target ISAR image processing method for template identification
CN108846409A (en) * 2018-04-28 2018-11-20 中国人民解放军国防科技大学 Radar echo extrapolation model training method based on cyclic dynamic convolution neural network
CN108830242A (en) * 2018-06-22 2018-11-16 北京航空航天大学 SAR image targets in ocean classification and Detection method based on convolutional neural networks
CN109753874A (en) * 2018-11-28 2019-05-14 南京航空航天大学 A kind of low slow small classification of radar targets method based on machine learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
RUI MIN 等: "A Gradually Distilled CNN for SAR Target Recognition", 《IEEE ACCESS》 *
张富生: "基于卷积神经网络的小样本ISAR图像识别", 《电脑编程技巧与维护》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112183185A (en) * 2020-08-13 2021-01-05 天津大学 Liquid leakage detection method based on optical flow method and CNN-SVM
CN112329878A (en) * 2020-11-16 2021-02-05 山西三友和智慧信息技术股份有限公司 Method for detecting sliding window of satellite image moving object based on convolutional neural network
CN112949555A (en) * 2021-03-17 2021-06-11 西安电子科技大学 Spatial target ISAR image classification method based on target prior information
CN112949555B (en) * 2021-03-17 2023-03-24 西安电子科技大学 Spatial target ISAR image classification method based on target prior information
CN113269135A (en) * 2021-06-17 2021-08-17 中国空间技术研究院 Satellite transponder recognition model and training method and using method thereof
CN113537020A (en) * 2021-07-07 2021-10-22 中国人民解放军国防科技大学 Complex SAR image target identification method based on improved neural network
CN115877344A (en) * 2023-02-21 2023-03-31 南京理工大学 Radar detection method and system for integrated processing of broadband detection, tracking and identification
CN115856854A (en) * 2023-02-24 2023-03-28 南京理工大学 Model and data hybrid driven radar detection method and system
CN116109944A (en) * 2023-04-12 2023-05-12 南京信息工程大学 Satellite image cloud target extraction method based on deep learning network architecture
CN116385813A (en) * 2023-06-07 2023-07-04 南京隼眼电子科技有限公司 ISAR image classification method, ISAR image classification device and storage medium
CN116385813B (en) * 2023-06-07 2023-08-29 南京隼眼电子科技有限公司 ISAR image space target classification method, device and storage medium based on unsupervised contrast learning

Similar Documents

Publication Publication Date Title
CN110705508A (en) Satellite identification method of ISAR image
CN108427920B (en) Edge-sea defense target detection method based on deep learning
Zhang et al. Automated pixel‐level pavement crack detection on 3D asphalt surfaces with a recurrent neural network
CN110211137B (en) Satellite image segmentation method based on residual error network and U-Net segmentation network
CN107239751B (en) High-resolution SAR image classification method based on non-subsampled contourlet full convolution network
CN111325748B (en) Infrared thermal image nondestructive testing method based on convolutional neural network
CN107563433B (en) Infrared small target detection method based on convolutional neural network
CN108921030B (en) SAR automatic target recognition method
CN110619352A (en) Typical infrared target classification method based on deep convolutional neural network
CN113240047B (en) SAR target recognition method based on component analysis multi-scale convolutional neural network
CN105117736B (en) Classification of Polarimetric SAR Image method based on sparse depth heap stack network
CN109948722B (en) Method for identifying space target
CN113536963B (en) SAR image airplane target detection method based on lightweight YOLO network
CN112699717A (en) SAR image generation method and generation device based on GAN network
CN108256557B (en) Hyperspectral image classification method combining deep learning and neighborhood integration
CN114972885A (en) Multi-modal remote sensing image classification method based on model compression
CN107680081B (en) Hyperspectral image unmixing method based on convolutional neural network
Tereikovskyi et al. The method of semantic image segmentation using neural networks
CN110647977B (en) Method for optimizing Tiny-YOLO network for detecting ship target on satellite
Li et al. Primitive fitting using deep geometric segmentation
CN113657290B (en) Snail collection and fine classification recognition system
Nanavati et al. Generative adversarial networks: A comparative analysis
CN116343016A (en) Multi-angle sonar image target classification method based on lightweight convolution network
CN115393717A (en) SAR image house extraction method and system based on evolution hybrid attention mechanism
CN113887652A (en) Remote sensing image dim target detection method based on form and multi-example learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200117

WD01 Invention patent application deemed withdrawn after publication