CN109190684B - SAR image sample generation method based on sketch and structure generation countermeasure network - Google Patents
SAR image sample generation method based on sketch and structure generation countermeasure network Download PDFInfo
- Publication number
- CN109190684B CN109190684B CN201810928531.5A CN201810928531A CN109190684B CN 109190684 B CN109190684 B CN 109190684B CN 201810928531 A CN201810928531 A CN 201810928531A CN 109190684 B CN109190684 B CN 109190684B
- Authority
- CN
- China
- Prior art keywords
- sketch
- sar image
- block
- network
- representing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an SAR image sample generation method based on a sketch and structure generation countermeasure network, which mainly solves the problem of sample imbalance in SAR image semantic segmentation, and comprises the following implementation steps: (1) the SAR image is subjected to sketch to obtain a sketch map; (2) extracting a small extremely inhomogeneous region according to a regional map of the SAR image; (3) constructing a paired data set in the shape of a sketch block-SAR image block; (4) selecting samples in the data set to form a training set and a testing set; (5) constructing a generation countermeasure network based on sketch information and structural constraints; (6) alternately training the sketch fitting network, the judging network and the generating network in sequence through sketch line loss, confrontation loss and generator loss; (7) inputting a test sketch block to a trained generation network to obtain a generated SAR image block; the SAR image sample inosculated with the ground object structure of the original SAR image can be generated according to the sketch map, and the problem of sample imbalance of classification of extremely uneven regions of the SAR image can be solved.
Description
Technical Field
The invention belongs to the technical field of image generation, and further relates to an image generation method for deep learning and generating confrontation network GANs (genetic adaptive networks) and synthetic Aperture radar (synthetic Aperture radar) images.
Background
Sample imbalance is a common and difficult to avoid problem in synthetic aperture radar SAR image classification and segmentation. Common machine learning models usually assume that samples are unbiased samples of true distributions, while most models have very poor recall of rare classes on the test set due to the extreme difference in samples between rare and rich classes in image classification. Similarly, the existence of small regions puts higher requirements on the segmentation technology of the SAR image, if the image segmentation is regarded as a classification problem at the image pixel level, the number of samples of the small regions is far less than that of the regions with larger areas, so that the small regions of the test set tend to be divided into large regions, especially for extremely heterogeneous regions in the SAR image, such as forests, urban areas and the like, the number of samples of the extremely heterogeneous regions is extremely different, and the difficulty of the SAR image segmentation is greatly increased.
In the face of the sample imbalance problem of SAR image classification and segmentation, the existing solutions are generally divided into the following two categories: under-sampling and over-sampling at the data level, and class weight changes at the model level.
Under-sampling and over-sampling at the data level mean that the class distribution of samples is changed through a certain strategy so as to achieve the purpose of converting the samples with unbalanced distribution into the samples with relatively balanced distribution.
And the sample undersampling is to randomly select a small number of samples from a plurality of classes and combine the samples with the original small number of classes to form a new training set. Sample oversampling, in contrast, performs sample enhancement by copying samples or performing various affine transformations on a few classes of samples. The method has the disadvantages that the undersampling of the samples can reduce the scale of the training set, thereby causing information loss, and the samples which are not sampled have important information; the samples generated by common affine transformation adopted in the sample oversampling are almost the same as original images, or are seriously distorted, even do not accord with an SAR image imaging mechanism, and the overfitting of the model is easily caused.
The class weight change at the model level refers to that the cost function of the learning model is modified so that the learning model is more sensitive to the cost of the rare class. The method has the disadvantages that the complexity of the model is increased, the parameter adjustment of the model is more difficult, and the classification problem of extremely different sample numbers is invalid.
Disclosure of Invention
The invention aims to provide an SAR image sample generation method for generating a countermeasure network based on sketch and structure, which solves the defect that classification and segmentation cannot be carried out on an SAR image due to the existing sample imbalance.
In order to achieve the purpose, the invention adopts the technical scheme that:
the invention provides an SAR image sample generation method based on a sketch and structure generation countermeasure network, which comprises the following steps:
Step 2, sketch I in the step 1sCarrying out regionalization to obtain a regional image, and extracting an extremely heterogeneous region set of an original SAR image I from the regional image;
step 3, constructing a paired data set in the form of a sketch block-SAR image block according to the extremely heterogeneous region set of the original SAR image I obtained in the step 2;
step 4, in the data set constructed in the step 3, randomly selecting 70% of data samples as a training data set, and using the rest data samples as a testing data set;
step 5, constructing a generation countermeasure network of sketch information and structural constraint, wherein the generation countermeasure network comprises a generator network, a discriminator network and a sketch fitter network;
step 6, training the generated countermeasure network constructed in the step 5 by using the training data set in the step 4;
and 7, inputting the test data set in the step 4 into the generated countermeasure network obtained by training in the step 6 to obtain a generated SAR image block.
Preferably, in step 1, the specific method for performing the sketch processing on the original SAR image I is as follows:
obtaining a sketch model of the original SAR image according to the distribution characteristics of the original SAR image I, and performing sketch processing on the original SAR image I by using the sketch model to obtain a sketch I of the original SAR image Is。
Preferably, in step 2, the specific method for extracting the extremely inhomogeneous region set of the original SAR image I is as follows:
firstly, the sketch I obtained in the step 1 is regionalized by sketch linesCarrying out regionalization treatment to obtain a regional image of the original SAR image, wherein the regional image comprises an aggregation region, a sketch-free region and a structural region;
secondly, mapping the regional image into an original SAR image I to obtain a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace of the SAR image I;
and finally, extracting a mixed pixel subspace of the SAR image I as an extremely heterogeneous region set of the original SAR image I, wherein the mixed pixel subspace is generally composed of a plurality of extremely heterogeneous regions which are not communicated with each other and have different sizes.
Preferably, in step 3, the specific steps of constructing the paired data sets in the form of "sketch block-SAR image block" are as follows:
1, selecting a very inhomogeneous region from the very inhomogeneous regions obtained in the step 2, wherein the very inhomogeneous region refers to any one of 1/3 extremely inhomogeneous regions with the area smaller than the area of the maximum very inhomogeneous region in the very inhomogeneous region set; then, carrying out interval dicing on the extremely inhomogeneous region by using a rectangular window with the size of 128 multiplied by 128, wherein the cutting interval size is 32, and then obtaining a plurality of cut SAR image blocks y;
step 2, according to the position of the SAR image block y in the original SAR image I, obtaining a sketch I from the step 1sAnd drawing a sketch block x corresponding to the position of the SAR image block y, and constructing a paired data set in the form of a sketch block-SAR image block.
Preferably, in step 5, the generator network includes an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a first maximum pooling layer, a fifth convolution layer, a sixth convolution layer, a second maximum pooling layer, a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a seventh convolution layer, and an output layer, which are connected in sequence; wherein, the input of the generator network is 8-dimensional Gaussian random noise z and a sketch block x in a training set, and the output is the SAR image block
The discriminator network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a fifth convolution layer, a sixth convolution layer, a seventh convolution layer and a classifier which are connected in sequence; wherein, the input of the discriminator network is SAR image block y in the training set and the generated SAR image blockThe output is a single node which represents the probability that the SAR image block conforms to the distribution of the real SAR sample;
the sketch fitter network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a first deconvolution layer, a second deconvolution layer, a fourth convolution layer and an output layer which are connected in sequence; wherein, the input of the sketch fitting device network is SAR image block y in the training set and SAR image block generationThe output is a fitted sketch block of the SAR image.
Preferably, in step 6, the specific method for training the constructed generation countermeasure network is as follows:
first, a sketch line loss function L is constructedSAntagonistic loss function LGANAnd a generator loss function L;
secondly, by the line loss function LSAntagonistic loss function LGANAnd a generator loss function L and the training data set obtained in the step 4 and the 8-dimensional Gaussian noise vector generated randomly are combined with a small batch random gradient descent method to alternately train the sketch fitter network, the discriminator network and the generator network in sequence, and finally the weight values of the trained sketch fitter network, the discriminator network and the generator network are obtained.
Preferably, the generator loss function L is a geometric window loss function LWSketch line loss function LSAnd a loss-fighting function LGAGThe weighted sum yields:
wherein, thetaGRepresenting the parameters of the generator G, thetaDRepresenting parameters of a discriminator D, wherein alpha, beta and gamma are all parameters; specifically, the method comprises the following steps:
geometric window loss function LWExpressed as:
wherein, thetaGParameters representing generator G, pz、pXAnd pYRespectively representing the distribution of noise z, sketch blocks x in the training set and SAR image blocks y in the training set, G (z | x) representing a SAR image generated conditioned on the original SAR image sketch blocks, | indicating an element-by-element multiplication, | | > |1L representing a vector1Norm, E [. C]Expressing an average value; n is a radical ofxThe number of line segments.
Preferably, geometric window lossesFunction LWThe design method specifically comprises the following steps:
step 2, obtaining N of SAR image block y in the training set from the geometric structure window in step 1xBinary split mask matrixWherein the content of the first and second substances,the matrix size of the SAR image block y is the same as that of the SAR image block y, the value of the structural area of the SAR image block is 1, and the value of the non-structural area of the SAR image block is 0;
step 3, binary division mask matrix M obtained from step 2xTo obtain a geometric structure window loss function LWThe mathematical expression of (a):
wherein, thetaGParameters representing generator G, pz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, the sketch block x in the training set and the SAR image block y in the training set, G (z | x) representing the SAR image block generated with the condition of the sketch block x in the training set, indicating element-by-element multiplication, < | >, respectively1L representing a vector1Norm, E [. C]Indicating averaging.
Preferably, the sketch line loss function L used in training the generatorSThe design method specifically comprises the following steps:
sketch line loss function LSIs expressed as a sketch block x of the original SAR image and a fitting sketch block for generating the SAR imageAfter weightingL of2Loss:
wherein, thetaGParameters representing generator G, pzAnd pXRespectively representing the distribution, E [ ·, of 8-dimensional Gaussian noise z and sketch blocks x in the training set]Indicating mean, y+And y-Respectively representing the inner part of the geometric structure window and the outer part of the geometric structure window, x in the SAR image block yiAndrespectively representing original SAR image sketch block x and generating SAR image fitting sketch blockA value at a specified location i;
setting a sketch block x in a training set, and expressing the ratio lambda of the number of pixels of a sketch part of an SAR image block y in the training set to the total number of pixels of the image as follows:
setting a sample x of a sketch block of the original SAR image, 8-dimensional Gaussian random noise z, a generator G and a sketch fitting device S, and generating a fitting sketch block of the SAR imageCan be expressed as:
sketch loss function L 'used in training sketch fitter'SThe design method is as follows:
wherein, thetaSParameters, p, representing a sketch fitter SXAnd pYRespectively representing the distribution of the sketch blocks x and the SAR image blocks y in the training set, E [ ·]Indicating mean, y+And y-Respectively representing the inside part and the outside part of a geometric structure window in the SAR image block y, xiAndrespectively representing original SAR image sketch block x and generating SAR image fitting sketch blockThe definition of the value, λ, at the given position i is the sketch line loss function L used in training the generatorSThe design method of (1).
Preferably, the penalty function L is resistedGANThe design method specifically comprises the following steps:
the input of the discriminator D is SAR image block y in the training set and the generated SAR image blockOutputting the probability that the sample accords with the original SAR image distribution; wherein, according to the definition of the condition GAN, the loss resisting function LGANThe mathematical expression of (a) is:
wherein, thetaGRepresenting the parameters of the generator G, thetaDParameter, p, representing discriminator Dz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, a sketch block x in a training set and an SAR image block y in the training set; g (z | x) represents a SAR image block generated conditioned on a sketch block x in the training setD (y | x) represents the output of the discriminator when inputting SAR image block y in the training set, and D (G (z | x)) represents the input of the discriminator to generate SAR image blockAn output of time; e [. C]Indicating averaging.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a SAR image sample generation method based on sketch and structure generation confrontation network, which obtains a generation confrontation network of sketch information and structure constraint by the steps 5 and 6, the network samples from a small extremely heterogeneous region in the SAR image mixed pixel subspace for unsupervised training, and the trained generator network is used for generating a sample which is matched with the original image structure and is distributed uniformly, thereby overcoming the defects that the sample generated by the existing oversampling technology is the same as the original image or seriously distorted and can not improve the training effect of a classification model, and leading the invention to better improve the problem of sample imbalance in SAR image segmentation;
meanwhile, step 5 also obtains a sketch fitter network, which is used for sampling and training from the original SAR image and the sketch obtained by the sketch model thereof, and the fitted sketch can be directly obtained from the SAR image by using the trained sketch fitter network, so that the defect that the sketch model adopted in the prior art is not tiny is overcome, the SAR image sketch line constraint can be applied to a system optimized by using a gradient descent method by adopting the method, and the sensitivity of the system to SAR image structure information is further improved;
furthermore, the training process of generating the confrontation network is guided by simultaneously using the geometric structure window loss function, the sketch line loss function and the confrontation loss function, so that the defect that the confrontation network is generated only by using the confrontation loss function constraint in the prior art without paying attention to the structural characteristics of the SAR image is overcome, the structural characteristics of the SAR image generated by the confrontation network can be effectively constrained and generated by adopting the method, and the fidelity of the generated sample is improved.
Drawings
FIG. 1 is a flow chart of generation of a countermeasure network for sketch information and structural constraints in the present invention, wherein a solid arrow represents forward propagation and a dashed arrow represents a backward propagation update generator G;
FIG. 2 is a schematic diagram of a network architecture of the present invention, wherein 2a is a network architecture of a generative model; 2b is the network architecture of the discriminant model; 2c is the network architecture of the sketch fitting model;
FIG. 3 is a schematic view of a geometric window of the present invention;
fig. 4 is a partial simulation result diagram of the present invention, wherein, viewed in columns, the first column from left to right is an original SAR image block, the second column is a sketch block of the original SAR image block, the third column is a generated SAR image block, and the fourth column is a fitting sketch block of the generated SAR image block.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
As shown in fig. 1, the method for generating an SAR image sample based on a sketch and structure generation countermeasure network provided by the present invention includes the following steps:
obtaining a sketch model of the input original SAR image according to the distribution characteristics of the input original SAR image I, and performing sketch processing on the input original SAR image by using the sketch model to obtain a sketch I of the input SAR images(ii) a Sketch IsThe method is a binary image, wherein the value of the sketch-able part is 1, which indicates that the brightness of the original SAR image I is suddenly changed at the position, and the value of the non-sketch part is 0, which indicates that the brightness of the original SAR image I is unchanged or slowly changed at the position.
The synthetic aperture radar SAR image sketch model used by the invention is a model proposed by Jie-Wu et al in 2014 in an article Local maximum geographic region search for SAR specific reduction with sketch-based geographic region search function in IEEE Transactions on Geoscience and Remote Sensing journal.
Step 2, sketch I in the step 1sRegionalizing to obtain a regional image, and extracting an extremely heterogeneous region set of the original SAR image I from the regional image, wherein the specific method comprises the following steps:
firstly, a sketch I is formed by adopting a sketch line regionalization methodsCarrying out regionalization treatment to obtain a regional image of the SAR image comprising an aggregation region, a sketch-free region and a structural region;
secondly, mapping the obtained region map to an original SAR image I to obtain a mixed aggregation structure ground object pixel subspace (hereinafter referred to as mixed pixel subspace), a homogeneous region pixel subspace (hereinafter referred to as homogeneous pixel subspace) and a structural pixel subspace of the original SAR image I;
and finally, extracting a mixed pixel subspace of the original SAR image I, wherein the subspace generally consists of a plurality of extremely heterogeneous regions which are different in size and are not communicated with one another.
The SAR image regionalization processing method used by the invention is a method proposed by an article 'historical semantic model and statistical mechanism based PolSAR image classification' published in Pattern Recognition journal in 2016 by Liu F and Shi J F and the like.
Step 3, constructing a paired data set in the form of a sketch block-SAR image block;
the specific steps of constructing the paired data sets in the form of "sketch blocks-SAR image blocks" are as follows:
step 2, according to the position of the SAR image block y in the original SAR image I, obtaining a sketch I from the step 1sThe method comprises the steps of extracting a sketch block x corresponding to the position y of an SAR image block, and constructing the sketch block-SAR image blockFor the data set.
And 4, selecting a sample.
Randomly selected 70% of the samples in the data set constitute the training data set, and the remaining 30% of the samples constitute the testing data set.
And 5, constructing a generation countermeasure network of sketch information and structural constraint.
Constructing a deep convolution neural network, inputting 8-dimensional Gaussian random noise z and a sketch block x in a training set, and outputting the deep convolution neural network as a generated SAR image blockAnd randomly initializing parameters of each convolution kernel in the network to obtain an initialized generator network.
FIG. 2 is a schematic diagram of a network architecture for a generative model, a discriminative model and a sketch-fitting model.
As shown in fig. 2a, the generator network includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first max pooling layer, a fifth convolutional layer, a sixth convolutional layer, a second max pooling layer, a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a seventh convolutional layer, and an output layer, which are connected in sequence. The input of the input layer is 8-dimensional Gaussian noise and a sketch block with the size of 128 multiplied by 1, the filter sizes of 11 layers between the input layer and the output layer are respectively 3,3,3,3,1,3,3,3,3,3,3, and 32,32,64,64,80,192, 512,256,128,1, and the output layer outputs a gray scale map with the size of 128 multiplied by 1.
Constructing a deep convolutional neural network, and respectively converting SAR image blocks y and SAR image blocks in a training setAnd as an input, outputting a single node which represents the probability that an input image block accords with the distribution of a real SAR sample, and randomly initializing the parameters of each convolution kernel in the network to obtain an initialized discriminator network.
As shown in fig. 2b, the discriminator network includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer and a classifier, which are connected in sequence. The input of the input layer is SAR image blocks of size 128 x 1, the filter sizes of 7 layers between the input layer and the output layer are respectively 4,4,4,4,4,4,1, the number of feature maps is respectively 32,64,128,256,512,1024,1, and the binary classifier outputs a scalar.
Constructing a deep convolutional neural network, and respectively converting SAR image blocks y and SAR image blocks in a training setThe input is the probability that each pixel in the SAR image block can be sketched (hereinafter referred to as the fitting sketch block of the SAR image), and the parameters of each convolution kernel in the network are initialized randomly to obtain the initialized sketch fitter network.
As shown in fig. 2c, the sketch fitter network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a first deconvolution layer, a second deconvolution layer, a fourth convolution layer and an output layer, which are connected in sequence. The input of the input layer is SAR image blocks with the size of 128 × 128 × 1, the filter sizes of 6 layers between the input layer and the output layer are respectively 7,3,3,3,3 and 7, the number of feature maps is respectively 32,64,128,64,32 and 1, and the output layer outputs gray-scale maps with the size of 128 × 128 × 1.
And 6, training and learning the generator network, the discriminator network and the sketch fitter network.
FIG. 1 is a flow diagram of generation of a countermeasure network for sketch information and structural constraints.
As shown in fig. 1, the line loss function L is plotted against the lineSAntagonistic loss function LGANAnd a generator loss function L, alternately training the sketch fitter network, the discriminator network and the generator network in sequence by using the training data set in the paired data sets of the sketch block-SAR image block constructed in the step 4 and the randomly generated 8-dimensional Gaussian noise vector by using a small batch random gradient descent method to obtain the weights of the trained sketch fitter network, the discriminator network and the generator network.
Said generator lossesLoss function L is a function of loss from a geometric windowWSketch line loss function LSAnd a penalty function LGANThe weighted sum yields:
wherein, thetaGRepresenting the parameters of the generator G, thetaDParameters representing the discriminator D, α, β and γ are parameters for balancing the 3 loss functions.
Fig. 3 is a schematic view of a geometry window.
As shown in FIG. 3, the geometry window loss function L used in training the generatorwThe design method is as follows:
step 2, defining a region surrounded by the geometric structure window constructed in the step 1 as a local geometric structure region with abrupt brightness change (hereinafter referred to as a structure region), and defining a region outside the window as a region with unchanged brightness or slowly changed brightness (hereinafter referred to as a non-structure region);
step 3, setting a sketch block x, 8-dimensional Gaussian noise z, an SAR image block y and a generator G as known, and assuming that the number of sketch line segments in the sketch block x is NxAccording to NxLine segment structure NxA geometric window. And taking a single geometric structure window, setting the value of the area in the window to be 1, and setting the values of other areas outside the window to be 0, and defining the binary matrix to be a binary segmentation mask matrix of the original SAR image block y. Thereby according to the above-mentioned NxN of original SAR image block y can be obtained through each geometric structure windowxBinary split mask matrixWherein the content of the first and second substances,is the same as the SAR image block y, and has a value of 1 for the area inside the window and 0 for the other areas outside the window.
By dividing the mask matrix M in binaryxGeometric structure window loss function LWCan be expressed as an original SAR image block y and a generated SAR image blockIn NxL of pixels in a structural area surrounded by a geometric structural window1The sum of the losses.
Geometric window loss function LWCan be expressed as:
wherein, thetaGParameters representing generator G, pz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, the original SAR image sketch block samples x and the original SAR image block samples y, G (z |) representing the SAR image block generated conditioned on the original SAR image sketch block, indicating an element-by-element multiplication, | > | | > |1L representing a vector1Norm, E [. C]Indicating averaging.
Sketch line loss function L used in training generatorsSThe design method is as follows:
step 2, giving original SAR image pixel description block samplesx, 8-dimensional Gaussian random noise z, a generator G, a sketch fitter S, and a fitting sketch block for generating SAR imagesCan be expressed as:
step 3, final sketch line loss function LsIs expressed as a sketch block x of the original SAR image and a fitting sketch block for generating the SAR imageWeighted L2Loss:
wherein, thetaGParameters representing generator G, pzAnd pXRespectively representing the distribution, E [ ·, of 8-dimensional Gaussian noise z and sketch blocks x in the training set]Indicating mean, y+And y-Respectively representing the inner part of the geometric structure window and the outer part of the geometric structure window, x in the SAR image block yiAndrespectively representing original SAR image sketch block x and generating SAR image fitting sketch blockA value at a specified location i;
penalty function L used in training generatorsGANThe design method is as follows:
the input of the discriminator D is an original SAR image block sample y and a generated SAR image block sampleAnd outputting the probability that the sample accords with the original SAR image distribution. According toDefinition of the conditional GAN, will oppose the loss function LGANThe form of the zero sum game is expressed as follows:
wherein, thetaGRepresenting the parameters of the generator G, thetaDParameter, p, representing discriminator Dz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, original SAR image sketch block samples x and original SAR image block samples y; g (z | x) represents an SAR image block generated by taking the original SAR image pixel description block x as a condition, D (y | x) represents the output of the discriminator when the original SAR image block sample is input by taking the original SAR image pixel description block x as a condition, and D (G (z | x)) represents the output of the discriminator when the SAR image block is generated by input; e [. C]Indicating averaging.
The sketch loss function used in training the sketch fitter is different from the sketch loss function used in training the generator. The sketch line loss function knows the original SAR image sketch block sample x, the original SAR image block sample y and the sketch fitter S, at this time, the fitting sketch block of the original SAR imageIt should be expressed as:
the sketch line loss function L'SShould be expressed as the original SAR image sketch block x and the original SAR image fitting sketch blockWeighted L2Loss:
wherein, thetaSParameters representing a sketch fitter S,pXAnd pYRespectively representing the distribution of the sketch blocks x and the SAR image blocks y in the training set, E [ ·]Indicating mean, y+And y-Respectively representing the inside part and the outside part of a geometric structure window in the SAR image block y, xiAndrespectively representing a sketch block x in a training set and generating a SAR image fitting sketch blockThe definition of the value, λ, at the given position i is the sketch line loss function L used in training the generatorSThe design method of (1).
The method for alternately training the sketch fitter network, the discriminator network and the generator network in sequence by means of the small-batch random gradient descent method comprises the following specific steps:
step 2, from the training data set constructed in step 4: randomly sampling a batch of n samples { (X) in a { SAR image sketch block sample set X, SAR image block sample set Y }(1),y(1)),…,(x(n),y(n))};
Step 3, from the noise distribution pzIn a batch of n samples { z }(1),…,z(n)};
And 4, updating the sketch fitter network S by a small batch random gradient descent method:
and 5, updating the discriminator network D by a small batch random gradient descent method:
and 6, updating a generator network G by a small batch random gradient descent method:
step 7, repeating the steps 2 to 6 until the iteration number k is reached;
And 7, generating the SAR image block by using the generator network.
Inputting the sketch block in the test data set in the paired data set of the sketch block-SAR image block constructed in the step 4 and 8-dimensional Gaussian noise into the generator network obtained by training in the step 6 to obtain a generated SAR image block.
The effect of the present invention will be further described with reference to the simulation diagram.
1. Simulation conditions are as follows:
the hardware platform for simulation of the invention is as follows: HP Z840; the software platform is as follows: PyTorch; synthetic aperture radar SAR images used in the simulation of the invention are as follows: and (3) a pyramides graph with the resolution of the Terras SAR X band being 1 meter.
2. Simulation content and results:
the method is used for carrying out experiments under the simulation condition, 70% of paired samples (accounting for 70% of total number of all samples of 5200 and about 3600 paired samples) are selected from a paired data set of 'sketch block-SAR image block' to carry out alternate training on a sketch fitter network, a discriminator network and a generator network in sequence, and the rest samples are used as test samples to obtain the SAR image block generation result shown in figure 4.
Looking at fig. 4 by columns, the first column from left to right is the original SAR image block, the second column is the sketch block of the original SAR image block (input to the generator network), the third column is the generated SAR image block (output of the generator network), and the fourth column is the fitted sketch block that generates the SAR image block (output of the sketch fitter network).
As can be seen from fig. 4: the generated SAR image block has clear edges and rich details, has more difference with the original SAR image block, and meanwhile, the comparison of the sketch fitting graph of the generated SAR image block and the sketch graph of the original SAR image block can find that the two are similar in structure, thereby ensuring the structural consistency of the generated SAR image block and the original SAR image block.
In conclusion, the method uses the generation countermeasure network of the sketch information and the structural constraint, can learn the generation of the extremely inhomogeneous region of the SAR image without supervision, and simultaneously, the two structural constraints effectively solve the problems of fuzzy and serious deformation of the image structure generated by the generation countermeasure network GAN, namely the method has good modeling capacity on the SAR image in structure and surface, can generate the SAR image sample which is consistent with the ground object structure of the original SAR image according to the sketch, and solves the problem of sample imbalance in the SAR image semantic segmentation.
Claims (8)
1. A SAR image sample generation method based on sketch and structure generation countermeasure network is characterized by comprising the following steps:
step 1, performing sketch processing on an original SAR image I to obtain a sketch I of the SAR image Is;
Step 2, sketch I in the step 1sCarrying out regionalization to obtain a regional image, and extracting an extremely heterogeneous region set of an original SAR image I from the regional image;
step 3, constructing a paired data set in the form of a sketch block-SAR image block according to the extremely heterogeneous region set of the original SAR image I obtained in the step 2;
step 4, randomly selecting 70% of data samples as a training data set and the rest data samples as a testing data set in the data set constructed in the step 3;
step 5, constructing a generation countermeasure network of sketch information and structural constraint, wherein the generation countermeasure network comprises a generator network, a discriminator network and a sketch fitter network;
step 6, training the generated countermeasure network constructed in the step 5 by using the training data set in the step 4;
step 7, inputting the test data set in the step 4 into the generated countermeasure network obtained by training in the step 6 to obtain a generated SAR image block;
in the step 2, the specific method for extracting the extremely heterogeneous region set of the original SAR image I is as follows:
firstly, the sketch I obtained in the step 1 is regionalized by sketch linesCarrying out regionalization treatment to obtain a regional image of the original SAR image, wherein the regional image comprises an aggregation region, a sketch-free region and a structural region;
secondly, mapping the regional image into an original SAR image I to obtain a mixed pixel subspace, a homogeneous pixel subspace and a structural pixel subspace of the SAR image I;
finally, extracting a mixed pixel subspace of the SAR image I as an extremely heterogeneous region set of the original SAR image I, wherein the mixed pixel subspace is composed of a plurality of extremely heterogeneous regions which are not communicated with each other and have different sizes;
in step 5, the generator network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a first maximum pooling layer, a fifth convolution layer, a sixth convolution layer, a second maximum pooling layer, a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a seventh convolution layer and an output layer which are connected in sequence; wherein, the input of the generator network is 8-dimensional Gaussian random noise z and a sketch block x in a training set, and the output is the SAR image block
The discriminator network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a fifth convolution layer, a sixth convolution layer, a seventh convolution layer and a classifier which are connected in sequence; wherein, the input of the discriminator network is SAR image block y in the training set and the generated SAR image blockThe output is a single node which represents the probability that the SAR image block conforms to the distribution of the real SAR samples;
the sketch fitter network comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer, a first deconvolution layer, a second deconvolution layer, a fourth convolution layer and an output layer which are connected in sequence; wherein, the input of the sketch fitting device network is SAR image block y in the training set and SAR image block generationThe output is a fitted sketch block of the SAR image.
2. The method for generating the SAR image sample based on the sketch and the structure generation countermeasure network according to claim 1, wherein in step 1, the specific method for performing the sketch processing on the original SAR image I is as follows:
obtaining a sketch model of the original SAR image according to the distribution characteristics of the original SAR image I, and performing sketch processing on the original SAR image I by using the sketch model to obtain a sketch I of the original SAR image Is。
3. The method for generating SAR image samples based on sketch and structure generation countermeasure network as claimed in claim 1, wherein in step 3, the specific steps of constructing the paired data sets in the form of "sketch block-SAR image block" are as follows:
1, selecting a very inhomogeneous region from the very inhomogeneous regions obtained in the step 2, wherein the very inhomogeneous region refers to any one of 1/3 extremely inhomogeneous regions with the area smaller than the area of the maximum very inhomogeneous region in the very inhomogeneous region set; then, carrying out interval dicing on the extremely inhomogeneous region by using a rectangular window with the size of 128 multiplied by 128, wherein the cutting interval size is 32, and then obtaining a plurality of cut SAR image blocks y;
step 2, according to the position of the SAR image block y in the original SAR image I, obtaining a sketch I from the step 1sAnd extracting a sketch block x corresponding to the position of the SAR image block y, and constructing a paired data set in the form of a sketch block-SAR image block.
4. The SAR image sample generation method based on sketch and structure generation countermeasure network as claimed in claim 1, characterized in that in step 6, the specific method for training the constructed generation countermeasure network is:
first, a sketch line loss function L is constructedSAntagonistic loss function LGANAnd a generator loss function L;
secondly, by the line loss function LSAntagonistic loss function LGANAnd a generator loss function L and the training data set obtained in the step 4 and the 8-dimensional Gaussian noise vector generated randomly are combined with a small batch random gradient descent method to alternately train the sketch fitter network, the discriminator network and the generator network in sequence, and finally the weight of the trained sketch fitter network, the discriminator network and the generator network is obtained.
5. The SAR image sample generation method based on sketch and structure generation countermeasure network as claimed in claim 4, wherein the generator loss function L is a geometric structure window loss function LWSketch line loss function LSAnd a penalty function LGANThe weighted sum yields:
wherein, thetaGRepresenting the parameters of the generator G, thetaDRepresenting parameters of a discriminator D, wherein alpha, beta and gamma are all parameters; specifically, the method comprises the following steps:
geometric window loss function LWExpressed as:
wherein, thetaGParameters representing generator G, pz、pXAnd pYRespectively representing the distribution of noise z, sketch blocks x in the training set and SAR image blocks y in the training set, G (z |) representing a SAR image generated conditioned on the original SAR image sketch blocks, | representing element-by-element multiplication, | · |1L representing a vector1Norm, E [. C]Expressing an average value; n is a radical ofxThe number of line segments to be sketched;and representing a binary segmentation mask matrix corresponding to the ith sketch block.
6. The SAR image sample generation method based on sketch and structure generation countermeasure network as claimed in claim 5, characterized in that, the geometric structure window loss function LWThe design method specifically comprises the following steps:
step 1, constructing a geometric structure window by utilizing any sketch line segment in a sketch block x in a training set, wherein the geometric structure window surrounds the sketch line segment, an area surrounded by the geometric structure window is defined as a structure area, and an area outside the window is defined as a non-structure area;
step 2, obtaining N of SAR image block y in the training set from the geometric structure window in step 1xBinary split mask matrixWherein the content of the first and second substances,the matrix size of (a) is the same as that of the SAR image block x, the value of the structural area of the SAR image block x is 1, and the value of the non-structural area of the SAR image block x is 0;
step 3, binary division mask matrix M obtained from step 2xTo obtain a geometric structure window loss function LWThe mathematical expression of (a):
wherein, thetaGParameters representing generator G, pz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, the sketch block x in the training set and the SAR image block y in the training set, G (z | x) representing the SAR image block generated conditioned on the sketch block x in the training set, | indicating element-by-element multiplication, | | > |1L representing a vector1Norm, E [. C]Indicating averaging.
7. The SAR image sample generation method based on sketch and structure generation confrontation network as claimed in claim 5, wherein the sketch line loss function L used in training the generatorSThe design method specifically comprises the following steps:
sketch line loss function LSIs expressed as a sketch block x of the original SAR image and a fitting sketch block for generating the SAR imageWeighted L2Loss:
wherein, thetaGParameters representing generator G, pzAnd pXRespectively representing the distribution of 8-dimensional Gaussian noise z and the sketch block x in the training set, E [ ·]Indicating mean, y+And y-Respectively representing the inside part and the outside part of a geometric structure window in the SAR image block y, xiAndrespectively representing original SAR image sketch block x and generating SAR image fitting sketch blockA value at a specified location i;
setting a sketch block x in a training set, and expressing the ratio lambda of the number of pixels of a sketch part of an SAR image block y in the training set to the total number of pixels of the image as follows:
setting a sample x of a sketch block of the original SAR image, 8-dimensional Gaussian random noise z, a generator G and a sketch fitting device S, and generating a fitting sketch block of the SAR imageCan be expressed as:
sketch loss function L 'used in training sketch fitter'SThe design method is as follows:
wherein, thetaSParameters, p, representing a sketch fitter SXAnd pYRespectively representing the distribution of the sketch blocks x and the SAR image blocks y in the training set, E [ ·]Indicating mean, y+And y-Respectively representing the inside part and the outside part of a geometric structure window in the SAR image block y, xiAndrespectively representing original SAR image sketch block x and generating SAR image fitting sketch blockThe definition of the value, λ, at the given position i is the sketch line loss function L used in training the generatorSThe design method of (1).
8. The SAR image sample generation method based on sketch and structure generation countermeasure network as claimed in claim 5, characterized in that the countermeasure loss function LGANThe design method specifically comprises the following steps:
the input of the discriminator D is SAR image block y in the training set and the generated SAR image blockOutputting the probability that the sample accords with the original SAR image distribution; wherein, according to the definition of the condition GAN, the loss resisting function LGANThe mathematical expression of (a) is:
wherein, thetaGRepresenting the parameters of the generator G, thetaDParameter, p, representing discriminator Dz、pXAnd pYRespectively representing the distribution of 8-dimensional Gaussian noise z, a sketch block x in a training set and an SAR image block y in the training set; g (z | x) represents a SAR image block generated conditioned on a sketch block x in the training setD (y | x) represents the output of the discriminator when inputting SAR image block y in the training set, and D (G (z | x)) represents the input of the discriminator to generate SAR image blockAn output of time; e [. C]Indicating averaging.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810928531.5A CN109190684B (en) | 2018-08-15 | 2018-08-15 | SAR image sample generation method based on sketch and structure generation countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810928531.5A CN109190684B (en) | 2018-08-15 | 2018-08-15 | SAR image sample generation method based on sketch and structure generation countermeasure network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109190684A CN109190684A (en) | 2019-01-11 |
CN109190684B true CN109190684B (en) | 2022-03-04 |
Family
ID=64935926
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810928531.5A Active CN109190684B (en) | 2018-08-15 | 2018-08-15 | SAR image sample generation method based on sketch and structure generation countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109190684B (en) |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109903242A (en) * | 2019-02-01 | 2019-06-18 | 深兰科技(上海)有限公司 | A kind of image generating method and device |
CN109871898B (en) * | 2019-02-27 | 2020-04-07 | 南京中设航空科技发展有限公司 | Method for generating deposit training sample by using generated confrontation network |
CN109816614A (en) * | 2019-02-28 | 2019-05-28 | 乐山知行智能科技有限公司 | Synthetic method, device and the storage medium of image |
CN110097059B (en) * | 2019-03-22 | 2021-04-02 | 中国科学院自动化研究所 | Document image binarization method, system and device based on generation countermeasure network |
CN109978893B (en) * | 2019-03-26 | 2023-06-20 | 腾讯科技(深圳)有限公司 | Training method, device, equipment and storage medium of image semantic segmentation network |
CN110096994B (en) * | 2019-04-28 | 2021-07-23 | 西安电子科技大学 | Small sample PolSAR image classification method based on fuzzy label semantic prior |
CN110210486B (en) * | 2019-05-15 | 2021-01-01 | 西安电子科技大学 | Sketch annotation information-based generation countermeasure transfer learning method |
CN110298384B (en) * | 2019-06-03 | 2021-03-12 | 西华大学 | Countermeasure sample image generation method and apparatus |
CN110210418B (en) * | 2019-06-05 | 2021-07-23 | 西安电子科技大学 | SAR image airplane target detection method based on information interaction and transfer learning |
CN110197517B (en) * | 2019-06-11 | 2023-01-31 | 常熟理工学院 | SAR image coloring method based on multi-domain cycle consistency countermeasure generation network |
CN110555811A (en) * | 2019-07-02 | 2019-12-10 | 五邑大学 | SAR image data enhancement method and device and storage medium |
CN110610124B (en) * | 2019-07-30 | 2021-11-30 | 珠海亿智电子科技有限公司 | Image generation method based on generation countermeasure network |
CN111027292B (en) * | 2019-11-29 | 2021-05-28 | 北京邮电大学 | Method and system for generating limited sampling text sequence |
CN111951187B (en) * | 2020-07-21 | 2023-04-18 | 电子科技大学 | SAR image enhancement method based on transformation gradient domain |
CN112560795B (en) * | 2020-12-30 | 2022-07-26 | 南昌航空大学 | SAR image target recognition algorithm based on CN-GAN and CNN |
CN112766381B (en) * | 2021-01-22 | 2023-01-24 | 西安电子科技大学 | Attribute-guided SAR image generation method under limited sample |
CN115374859A (en) * | 2022-08-24 | 2022-11-22 | 东北大学 | Method for classifying unbalanced and multi-class complex industrial data |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106611420A (en) * | 2016-12-30 | 2017-05-03 | 西安电子科技大学 | SAR image segmentation method based on deconvolution network and sketch direction constraint |
CN106611422A (en) * | 2016-12-30 | 2017-05-03 | 西安电子科技大学 | Stochastic gradient Bayesian SAR image segmentation method based on sketch structure |
CN107341813A (en) * | 2017-06-15 | 2017-11-10 | 西安电子科技大学 | SAR image segmentation method based on structure learning and sketch characteristic inference network |
CN107563428A (en) * | 2017-08-25 | 2018-01-09 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on generation confrontation network |
CN107886491A (en) * | 2017-11-27 | 2018-04-06 | 深圳市唯特视科技有限公司 | A kind of image combining method based on pixel arest neighbors |
CN108399625A (en) * | 2018-02-28 | 2018-08-14 | 电子科技大学 | A kind of SAR image orientation generation method generating confrontation network based on depth convolution |
-
2018
- 2018-08-15 CN CN201810928531.5A patent/CN109190684B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106611420A (en) * | 2016-12-30 | 2017-05-03 | 西安电子科技大学 | SAR image segmentation method based on deconvolution network and sketch direction constraint |
CN106611422A (en) * | 2016-12-30 | 2017-05-03 | 西安电子科技大学 | Stochastic gradient Bayesian SAR image segmentation method based on sketch structure |
CN107341813A (en) * | 2017-06-15 | 2017-11-10 | 西安电子科技大学 | SAR image segmentation method based on structure learning and sketch characteristic inference network |
CN107563428A (en) * | 2017-08-25 | 2018-01-09 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on generation confrontation network |
CN107886491A (en) * | 2017-11-27 | 2018-04-06 | 深圳市唯特视科技有限公司 | A kind of image combining method based on pixel arest neighbors |
CN108399625A (en) * | 2018-02-28 | 2018-08-14 | 电子科技大学 | A kind of SAR image orientation generation method generating confrontation network based on depth convolution |
Non-Patent Citations (5)
Title |
---|
Generative adversarial network-based restoration of speckled SAR images;Puyang Wang等;《2017 IEEE 7th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)》;20180312;第1-5页 * |
Local Maximal Homogeneous Region Search for SAR Speckle Reduction With Sketch-Based Geometrical Kernel Function;Jie Wu等;《IEEE Transactions on Geoscience and Remote Sensing》;20140106;第52卷(第9期);第5751-5764页 * |
基于深度学习和层次语义模型的极化SAR分类;石俊飞等;《自动化学报》;20170215;第43卷(第2期);第215-226页 * |
基于深度学习网络的SAR图像目标识别研究;赵菲妮;《中国优秀硕士学位论文全文数据库 (信息科技辑)》;20180415;第I136-2394页 * |
基于素描稀疏表示和低秩分解的SAR图像目标检测;闫晓莉;《中国优秀硕士学位论文全文数据库 (信息科技辑)》;20170315;第I136-2215页 * |
Also Published As
Publication number | Publication date |
---|---|
CN109190684A (en) | 2019-01-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109190684B (en) | SAR image sample generation method based on sketch and structure generation countermeasure network | |
CN110135267B (en) | Large-scene SAR image fine target detection method | |
CN113159051B (en) | Remote sensing image lightweight semantic segmentation method based on edge decoupling | |
CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
CN108154192B (en) | High-resolution SAR terrain classification method based on multi-scale convolution and feature fusion | |
CN107798381B (en) | Image identification method based on convolutional neural network | |
CN110136154B (en) | Remote sensing image semantic segmentation method based on full convolution network and morphological processing | |
CN111612807B (en) | Small target image segmentation method based on scale and edge information | |
CN106228185B (en) | A kind of general image classifying and identifying system neural network based and method | |
CN111563902A (en) | Lung lobe segmentation method and system based on three-dimensional convolutional neural network | |
CN111310666B (en) | High-resolution image ground feature identification and segmentation method based on texture features | |
CN111612017B (en) | Target detection method based on information enhancement | |
CN111832501A (en) | Remote sensing image text intelligent description method for satellite on-orbit application | |
CN110929602A (en) | Foundation cloud picture cloud shape identification method based on convolutional neural network | |
CN111369442B (en) | Remote sensing image super-resolution reconstruction method based on fuzzy kernel classification and attention mechanism | |
CN107808138B (en) | Communication signal identification method based on FasterR-CNN | |
CN110598806A (en) | Handwritten digit generation method for generating countermeasure network based on parameter optimization | |
CN104866868A (en) | Metal coin identification method based on deep neural network and apparatus thereof | |
CN111368935B (en) | SAR time-sensitive target sample amplification method based on generation countermeasure network | |
CN111275640B (en) | Image enhancement method for fusing two-dimensional discrete wavelet transform and generation of countermeasure network | |
CN112233129B (en) | Deep learning-based parallel multi-scale attention mechanism semantic segmentation method and device | |
CN107967474A (en) | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks | |
CN113095333B (en) | Unsupervised feature point detection method and unsupervised feature point detection device | |
CN112950780B (en) | Intelligent network map generation method and system based on remote sensing image | |
CN113743417B (en) | Semantic segmentation method and semantic segmentation device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |