CN106096650A - Based on the SAR image sorting technique shrinking own coding device - Google Patents
Based on the SAR image sorting technique shrinking own coding device Download PDFInfo
- Publication number
- CN106096650A CN106096650A CN201610407324.6A CN201610407324A CN106096650A CN 106096650 A CN106096650 A CN 106096650A CN 201610407324 A CN201610407324 A CN 201610407324A CN 106096650 A CN106096650 A CN 106096650A
- Authority
- CN
- China
- Prior art keywords
- encoder
- self
- punctured
- representing
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000012549 training Methods 0.000 claims abstract description 85
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 35
- 230000008602 contraction Effects 0.000 claims abstract description 22
- 239000011159 matrix material Substances 0.000 claims description 26
- 230000004913 activation Effects 0.000 claims description 6
- 235000009508 confectionery Nutrition 0.000 claims description 4
- 238000005516 engineering process Methods 0.000 abstract description 3
- 230000001788 irregular Effects 0.000 abstract description 3
- 230000001427 coherent effect Effects 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 abstract 1
- 238000004088 simulation Methods 0.000 description 23
- 238000010586 diagram Methods 0.000 description 14
- 238000012360 testing method Methods 0.000 description 14
- 101001055386 Homo sapiens Melanophilin Proteins 0.000 description 9
- 102100026158 Melanophilin Human genes 0.000 description 9
- 238000000605 extraction Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 6
- 238000003384 imaging method Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000035515 penetration Effects 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
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
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Radar Systems Or Details Thereof (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of based on the SAR image sorting technique shrinking autocoder.The step that the present invention realizes is: (1) input picture;(2) stationary wavelet decomposition is carried out;(3) training sample is chosen;(4) the contraction own coding device of parallel two rank is built;(5) train the contraction of parallel two rank from coder;(6) structure sample characteristics collection;(7) training Softmax grader;(8) classification.The present invention histogrammic feature extracting method of multi-layer local pattern compared to existing technology, has nicety of grading high, and region consistency is good, the accurate advantage of marginal classification, solves coherent speckle noise impact, territorial classification confusion and the irregular problem in edge.Present invention can apply to synthetic aperture radar SAR image target detection and target recognition.
Description
Technical Field
The invention belongs to the technical field of image processing, and further relates to a Synthetic Aperture Radar (SAR) image classification method based on a shrinking self-encoder in the technical field of target identification. The method can be applied to acquisition of target identification image information and SAR image target identification, and can accurately classify different regions of the image.
Background
The synthetic aperture radar SAR is an active earth observation system, compared with other sensors such as optical sensors, infrared sensors and the like, SAR imaging is less limited by conditions such as atmosphere, illumination and the like, can perform all-weather and all-time reconnaissance on an interested target, and has various characteristics such as high resolution, large breadth and the like. The SAR can effectively identify camouflage and penetration masks, so the SAR is widely applied to the military and civil fields of remote sensing mapping, military reconnaissance, earthquake relief and the like, and becomes an indispensable tool for earth observation and military reconnaissance.
SAR image classification is a key step for realizing automatic processing of SAR images and is one of basic and key technologies in SAR image interpretation. The SAR image classification is a precondition for further interpretation of the SAR image. The purpose is to assign each pixel a class label based on the pixel's own attributes of a given image and its associated characteristics with neighboring pixels, and to cluster all pixels with the same characteristics to identify the class to which the pixel belongs. In the imaging process of the SAR image, since different ground objects have different back reflection and scattering characteristics, the SAR image often contains abundant texture information. On the other hand, because the SAR adopts a coherent imaging system, the imaged SAR image is affected by speckle noise, which brings certain difficulties to the understanding and interpretation of the SAR image. In view of the above two points, when the SAR image is used to classify the ground features, the feature extraction and selection are very difficult and challenging. At present, the feature extraction method of the SAR image is often a feature extraction method based on image sub-regions, for example:
D.X.Dai et al, in its published paper, "Multilevel local pattern histogram for SARimage Classification," (IEEE Trans on Geoscience and Remote Sensing Letters,2011,8(2), 225-. The method comprises the steps of firstly extracting image characteristics of image sub-regions by utilizing a multi-level local mode histogram, and then marking each region by utilizing an SVM classifier, thereby completing the classification of the whole image. The method has the defects that the characteristic extraction is too single, the abundant texture information in the SAR image is not considered, and the phenomena of disordered regional classification and irregular edges obviously appear in the classification result.
Schlemr-laughing et al proposed a parallel SAR image classification method based on spatial and frequency domain features in the patent of its application (patent application No.: CN201510185815.6, publication No.: CN 104751477A). The method comprises the steps of firstly calculating corresponding wavelet energy characteristics, gray level co-occurrence matrix characteristics and filtered gray level mean characteristics for each pixel in each image sub-area, then recovering the wavelet energy characteristics, the gray level co-occurrence matrix characteristics and the filtered gray level mean characteristics of each pixel in the image sub-area to obtain the wavelet energy characteristics, the gray level co-occurrence matrix characteristics and the filtered gray level mean characteristics of each pixel in the SAR image, and finally clustering the formed feature vectors to obtain a final classification result. The method has the disadvantages that the image feature extraction process excessively depends on gray features, the texture feature extraction is incomplete, and meanwhile, the unsupervised classification mode has great loss on classification precision, so that the final SAR image classification result has low precision.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an SAR image classification method based on a contraction automatic encoder and smooth wavelet transform.
The invention comprises the following steps:
(1) inputting an image:
inputting a synthetic aperture radar SAR image to be classified;
(2) performing stationary wavelet decomposition:
performing 1-layer stationary wavelet decomposition on the synthetic aperture radar SAR image to be classified, and respectively acquiring 1 low-frequency component and 3 high-frequency components of the synthetic aperture radar SAR image to be classified;
(3) selecting a training sample:
(3a) 3 high-frequency components of the synthetic aperture radar SAR image stationary wavelet decomposition to be classified are connected end to end in the column direction and combined into a total high-frequency component;
(3b) respectively and randomly selecting 10% of samples from low-frequency components and total high-frequency components of the stable wavelet decomposition of the synthetic aperture radar SAR image to be classified as low-frequency training samples and high-frequency training samples;
(4) a parallel two-level punctured self-encoder is constructed:
(4a) generating output characteristics of the first layer of the punctured self-encoder by puncturing the encoding of the first layer of the self-encoder according to:
y=f(W(1)x+b(1))
where y represents the output characteristics of the first layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation, W(1)Weight coefficients representing a matrix of all 1 values between the input unit of the punctured self-encoder and the first layer of the punctured self-encoder, x representing the original input data of the punctured self-encoder, b(1)A bias coefficient representing a matrix of all 1's values between an input unit of the punctured self-encoder and a first layer of the punctured self-encoder;
(4b) inputting output characteristics of a first layer of a punctured self-encoder to a second layer of the punctured self-encoder;
(4c) generating output characteristics of the second layer of the punctured self-encoder by decoding the second layer of the punctured self-encoder according to the following formula, and composing a two-level punctured self-encoder:
z=f(W(2)y+b(2))
where z represents the output characteristics of the second layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation operation, W(2)Weight coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, y represents an output characteristic of the first layer of the punctured self-encoder, b(2)Bias coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, b(2)The initial value of (1) is a full 1 matrix;
(4d) constructing an identical two-level shrinking self-encoder by adopting the same method of the step (4a), the step (4b) and the step (4 c);
(4e) paralleling two-level contraction self-encoders to obtain parallel two-level contraction self-encoders;
(5) training a parallel two-level punctured self-encoder:
(5a) inputting a low-frequency training sample into a two-level contraction self-encoder by adopting a contraction self-encoder network training method to obtain high-level low-frequency characteristics of the training sample;
(5b) inputting the high-frequency training sample into another two-layer level self-contraction encoder by adopting a self-contraction encoder network training method to obtain high-level high-frequency characteristics of the training sample;
(6) constructing a sample feature set:
carrying out head-to-tail joint on the high-level low-frequency features and the high-level high-frequency features of the training sample in the column direction, and combining the high-level low-frequency features and the high-level high-frequency features into a total high-level feature;
(7) training a Softmax classifier:
(7a) setting the training step length of the Softmax classifier to be 1000;
(7b) inputting the total high-grade characteristics into a Softmax classifier, and training the Softmax classifier by using a random gradient descent (SGD) method;
(7c) performing parameter fine adjustment on the whole composed of the parallel two-level contraction self-encoder and the Softmax classifier until a set step length is reached;
(8) and (4) classification:
and respectively inputting all low-frequency components and total high-frequency components acquired by the stable wavelet decomposition of the synthetic aperture radar SAR image to be classified into a trained two-level shrinking self-encoder network, and finally outputting the classification result of the synthetic aperture radar SAR image to be classified.
Compared with the prior art, the invention has the following advantages:
firstly, the method of combining the stationary wavelet decomposition technology and the shrinkage self-encoder is adopted to extract the SAR image features to be classified, so that the problem that the prior art feature extraction method is influenced by speckle noise is solved, and the SAR image to be classified extracted by the method is rich in feature information and strong in feature robustness.
Secondly, the SAR images to be classified are classified by adopting a method of combining a two-level shrinkage self-encoder with parallel training and a Softmax classifier, and the phenomena of disordered region classification and irregular edges caused by the classification method in the prior art are overcome, so that the SAR image classification method has the advantages of high classification precision, good region consistency and precise edge classification.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of the simulation of the MLPH classification method of the present invention and the prior art on a Ku-band UAVSAR SAR test image;
FIG. 3 is a simulation diagram of the MLPH classification method of the present invention and the prior art on a Terras SAR-X synthetic aperture radar SAR test image.
Detailed Description
The present invention is described in further detail below with reference to the attached drawing figures.
The steps of the present invention are described in further detail below with reference to fig. 1.
Step 1, inputting an image.
Inputting a synthetic aperture radar SAR image to be classified.
And 2, performing stationary wavelet decomposition.
And performing 1-layer stationary wavelet decomposition on the synthetic aperture radar SAR image to be classified, and respectively acquiring 1 low-frequency component and 3 high-frequency components of the synthetic aperture radar SAR image to be classified.
The specific steps of 1-layer stationary wavelet decomposition are as follows:
calculating the low-frequency component of the SAR image to be classified according to the following formula:
wherein,mth layer of representing steady wavelet decomposition of synthetic aperture radar SAR image to be classified1Line n1The low frequency component of the column, ∑ denotes the summing operation,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column.
Calculating the high-frequency component of the synthetic aperture radar SAR image to be classified in the horizontal direction according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency component of the horizontal direction of the column, ∑ denotes the summing operation,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column.
Calculating the high-frequency component of the synthetic aperture radar SAR image to be classified in the vertical direction according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency component of the vertical direction of the column, ∑ denotes the summing operation,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column.
Calculating high-frequency components in the diagonal direction of the synthetic aperture radar SAR image to be classified according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency components in the diagonal direction of the column, ∑ denotes the summing operation,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column.
And 3, selecting a training sample.
And (3) the high-frequency components of the synthetic aperture radar SAR image stationary wavelet decomposition to be classified are connected end to end in the column direction and are combined into a total high-frequency component.
Respectively and randomly selecting 10% of samples from low-frequency components and total high-frequency components of the synthetic aperture radar SAR image stationary wavelet decomposition to be classified as low-frequency training samples and high-frequency training samples.
And 4, constructing a parallel two-level shrinkage self-encoder.
First, generating output characteristics of a first layer of a punctured self-encoder by puncturing encoding of the first layer of the self-encoder according to:
y=f(W(1)x+b(1))
where y represents the output characteristics of the first layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation, W(1)Weight coefficients representing a matrix of all 1 values between the input unit of the punctured self-encoder and the first layer of the punctured self-encoder, x representing the original input data of the punctured self-encoder, b(1)And a bias coefficient representing a matrix of all 1's between an input unit of the punctured self-encoder and the first layer of the punctured self-encoder.
And a second step of inputting the output characteristics of the first layer of the punctured self-encoder to a second layer of the punctured self-encoder.
Thirdly, generating output characteristics of the second layer of the self-encoder by decoding the second layer of the self-encoder according to the following formula, and forming the two-level self-encoder:
z=f(W(2)y+b(2))
where z represents the output characteristics of the second layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation operation, W(2)Weight coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, y represents an output characteristic of the first layer of the punctured self-encoder, b(2)Bias coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, b(2)Is an all 1 matrix.
And fourthly, constructing an identical two-level contraction self-encoder by adopting the same method as the first step, the second step and the third step in the step.
And fifthly, paralleling the two-level self-contraction encoders to obtain the parallel two-level self-contraction encoder.
And 5, training a parallel two-level contraction self-encoder.
And inputting the low-frequency training sample into a two-level contraction self-encoder by adopting a contraction self-encoder network training method to obtain the high-level low-frequency characteristic of the training sample.
The method for training the self-encoder network is contracted, and the specific steps for obtaining the high-level low-frequency characteristics of the training sample are as follows: the reconstruction error of the punctured autocoder is calculated as follows:
where L (x, z) represents the reconstruction error of the punctured self-encoder, x represents the low frequency training samples, z represents the output characteristics of the second layer of the punctured self-encoder, ∑ represents the summing operation, dxRepresenting the total dimension of the feature, i representing the specific feature dimension, xiFeatures representing the ith dimension of the low frequency training sample, log representing the log taking operation, ziFeatures representing an ith dimension of the output of the second layer from the encoder are punctured.
The penalty term for the punctured autocoder is calculated as follows:
wherein,a penalty term representing a punctured self-encoder, J (x) a Jacobian matrix representing output values of the punctured self-encoder with respect to weights, x representing low frequency training samples,denotes the squaring operation of the F norm, dhRepresenting the total dimension of the output features of the first layer of the punctured self-encoder, j representing the specific feature dimension, yjRepresents the output characteristics of the jth dimension of the first layer of the shrinking self-encoder, W represents the weight coefficient of the shrinking self-encoder, | | | | | sweet wind2Representing a norm squaring operation.
The loss function of the punctured autocoder is calculated as follows:
wherein, Jcae(theta) represents a loss function of the punctured self-encoder, theta represents a parameter of the punctured self-encoder, ∑ represents a summing operation, Tn={x(1),x(2),...,x(n)Denotes a training sample set containing n samples, L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a low frequency training sample, z denotes an output characteristic of a second layer of the punctured self-encoder, λ denotes a weight adjustment parameter,a penalty term representing a punctured self-encoder, J (x) a Jacobian matrix representing output values of the punctured self-encoder with respect to weights, x representing low frequency training samples,representing the F-norm squaring operation.
Training the self-contraction encoder, optimizing a loss function of the self-contraction encoder through a random echelon descent SGD method, and adjusting a weight coefficient and a bias coefficient between layers in the self-contraction encoder to enable a reconstruction error of the self-contraction encoder to be minimum and obtain high-level low-frequency characteristics of a training sample.
And inputting the high-frequency training sample into another two-level self-contraction encoder by adopting a self-contraction encoder network training method to obtain the high-level high-frequency characteristic of the training sample.
The method for shrinking the self-encoder network training comprises the following specific steps of obtaining high-level high-frequency characteristics of a training sample: the reconstruction error of the punctured autocoder is calculated as follows:
where L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a high frequency training sample, z denotes an output characteristic of the second layer of the punctured self-encoder, ∑ denotes a summing operation, dxRepresenting the total dimension of the feature, i representing the specific feature dimension, xiFeatures representing the ith dimension of the high frequency training sample, log representing the log taking operation, ziFeatures representing an ith dimension of the output of the second layer from the encoder are punctured.
The penalty term for the punctured autocoder is calculated as follows:
wherein,represents a penalty term for the punctured autocoder, J (x) represents a Jacobian matrix of output values of the punctured autocoder for the first layer with respect to weights, x represents high frequency training samples,denotes the squaring operation of the F norm, dhRepresenting the total dimension of the output features of the first layer of the punctured self-encoder, j representing the specific feature dimension, yjRepresents the output characteristics of the jth dimension of the first layer of the shrinking self-encoder, W represents the weight coefficient of the shrinking self-encoder, | | | | | sweet wind2Representing a norm squaring operation.
The loss function of the punctured autocoder is calculated as follows:
wherein, Jcae(theta) represents a loss function of the punctured self-encoder, theta represents a parameter of the punctured self-encoder, ∑ represents a summing operation, Tn={x(1),x(2),...,x(n)Denotes a training sample set containing n samples, L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a high frequency training sample, z denotes an output characteristic of a second layer of the punctured self-encoder, λ denotes a weight adjustment parameter,represents a penalty term for the punctured autocoder, J (x) represents a Jacobian matrix of output values of the punctured autocoder for the first layer with respect to weights, x represents high frequency training samples,representing the F-norm squaring operation.
Training the self-contraction encoder, optimizing the loss function of the self-contraction encoder by a random echelon descent SGD method, and adjusting the weight coefficient and the offset coefficient between layers in the self-contraction encoder to ensure that the reconstruction error of the self-contraction encoder is minimum and obtain the high-level high-frequency characteristic of a training sample.
And 6, constructing a sample feature set.
And (4) performing head-to-tail joint on the high-level low-frequency features and the high-level high-frequency features of the training samples in the column direction, and combining the high-level low-frequency features and the high-level high-frequency features into the total high-level features.
And 7, training a Softmax classifier.
The training step size of the Softmax classifier is set to 1000.
And inputting the total high-grade features into a Softmax classifier, and training the Softmax classifier by using a random gradient descent (SGD) method.
And performing parameter fine adjustment on the whole body formed by the parallel two-level contraction self-encoder and the Softmax classifier until a set step length is reached.
And 8, classifying.
And respectively inputting all low-frequency components and total high-frequency components acquired by the stable wavelet decomposition of the synthetic aperture radar SAR image to be classified into a trained two-level shrinking self-encoder network, and finally outputting the classification result of the synthetic aperture radar SAR image to be classified.
The effects of the present invention can be further illustrated by the following simulations:
1. simulation conditions are as follows:
the simulation of the invention is carried out under the Intel (R) core (TM) i7-4720HQ CPU with main frequency of 2.6GHZ, the hardware environment of memory 8GB and the software environment of MATLAB R2014 a.
2. Simulation content:
fig. 2 is a schematic diagram of simulation on a Ku-band UAVSAR synthetic aperture radar SAR test image by the MLPH classification method of the present invention and the prior art, wherein fig. 2(a) is an original Ku-band UAVSAR polarimetric synthetic aperture radar SAR test image used in the simulation, the size is 256 × 256, and the resolution is 3 m. Fig. 2(b) is a diagram showing a simulation result of classification of the test image of fig. 2(a) by the MLPH classification method of the related art, and fig. 2(c) is a diagram showing a simulation result of classification of the test image of fig. 2(a) by the present invention.
Fig. 3 is a schematic diagram of image simulation on a terrasaar-X synthetic aperture radar SAR test by the MLPH classification method of the present invention and the prior art, wherein fig. 3(a) is an original TerraSAR-X polarized synthetic aperture radar SAR test image used in the simulation, the size is 256 × 256, and the resolution is 1 m. Fig. 3(b) is a diagram showing a simulation result of classification of the test image of fig. 3(a) by the MLPH classification method of the related art, and fig. 3(c) is a diagram showing a simulation result of classification of the test image of fig. 3(a) by the present invention.
3. Simulation effect analysis:
fig. 2(b) is a diagram showing a classification simulation result of the test image of fig. 2(a) by using the MLPH classification method of the prior art. As can be seen from the simulation results of fig. 2(b), the clarity and detail integrity of the edge boundaries remain relatively poor for the homogeneous region. Fig. 2(c) is a diagram of a classification simulation result of the test image of fig. 2(a) by using the present invention, and it can be seen from the simulation result of fig. 2(c) that the classification result of the present invention can effectively distinguish various regions, the regions in the classification diagram are more uniform and complete, the classified edges between different regions are smoother and clearer, and the detail information is complete.
Fig. 3(b) is a diagram showing a classification simulation result of the test image of fig. 3(a) by using the MLPH classification method of the prior art. As can be seen from the simulation results of fig. 3(b), there are distinct noise regions and the edges are not smooth between different regions. Fig. 3(c) is a diagram of the classification simulation result of the test image of fig. 3(a) by using the present invention, and it can be seen from the simulation result of fig. 3(c) that the result of the present invention is significantly better than the result of the existing MLPH classification, and different regions in the classification diagram are more uniform and accurate, the edges are clearer, and the image visual effect is better.
Claims (4)
1. A SAR image classification method based on a contraction self-encoder comprises the following steps:
(1) inputting an image:
inputting a synthetic aperture radar SAR image to be classified;
(2) performing stationary wavelet decomposition:
performing 1-layer stationary wavelet decomposition on the synthetic aperture radar SAR image to be classified, and respectively acquiring 1 low-frequency component and 3 high-frequency components of the synthetic aperture radar SAR image to be classified;
(3) selecting a training sample:
(3a) 3 high-frequency components of the synthetic aperture radar SAR image stationary wavelet decomposition to be classified are connected end to end in the column direction and combined into a total high-frequency component;
(3b) respectively and randomly selecting 10% of samples from low-frequency components and total high-frequency components of the stable wavelet decomposition of the synthetic aperture radar SAR image to be classified as low-frequency training samples and high-frequency training samples;
(4) a parallel two-level punctured self-encoder is constructed:
(4a) generating output characteristics of the first layer of the punctured self-encoder by puncturing the encoding of the first layer of the self-encoder according to:
y=f(W(1)x+b(1))
where y represents the output characteristics of the first layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation, W(1)Weight coefficients representing a matrix of all 1 values between the input unit of the punctured self-encoder and the first layer of the punctured self-encoder, x representing the original input data of the punctured self-encoder, b(1)A bias coefficient representing a matrix of all 1's values between an input unit of the punctured self-encoder and a first layer of the punctured self-encoder;
(4b) inputting output characteristics of a first layer of a punctured self-encoder to a second layer of the punctured self-encoder;
(4c) generating output characteristics of the second layer of the punctured self-encoder by decoding the second layer of the punctured self-encoder according to the following formula, and composing a two-level punctured self-encoder:
z=f(W(2)y+b(2))
where z represents the output characteristics of the second layer of the punctured autocoder, f (-) represents Sigmod nonlinear activation operation, W(2)Weight coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, y represents an output characteristic of the first layer of the punctured self-encoder, b(2)Bias coefficients representing a matrix of values all 1 between the second layer of the punctured self-encoder and the output unit of the punctured self-encoder, b(2)The initial value of (1) is a full 1 matrix;
(4d) constructing an identical two-level shrinking self-encoder by adopting the same method of the step (4a), the step (4b) and the step (4 c);
(4e) paralleling two-level contraction self-encoders to obtain parallel two-level contraction self-encoders;
(5) training a parallel two-level punctured self-encoder:
(5a) inputting a low-frequency training sample into a two-level contraction self-encoder by adopting a contraction self-encoder network training method to obtain high-level low-frequency characteristics of the training sample;
(5b) inputting the high-frequency training sample into another two-layer level self-contraction encoder by adopting a self-contraction encoder network training method to obtain high-level high-frequency characteristics of the training sample;
(6) constructing a sample feature set:
carrying out head-to-tail joint on the high-level low-frequency features and the high-level high-frequency features of the training sample in the column direction, and combining the high-level low-frequency features and the high-level high-frequency features into a total high-level feature;
(7) training a Softmax classifier:
(7a) setting the training step length of the Softmax classifier to be 1000;
(7b) inputting the total high-grade characteristics into a Softmax classifier, and training the Softmax classifier by using a random gradient descent (SGD) method;
(7c) performing parameter fine adjustment on the whole composed of the parallel two-level contraction self-encoder and the Softmax classifier until a set step length is reached;
(8) and (4) classification:
and respectively inputting all low-frequency components and total high-frequency components acquired by the stable wavelet decomposition of the synthetic aperture radar SAR image to be classified into a trained two-level shrinking self-encoder network, and finally outputting the classification result of the synthetic aperture radar SAR image to be classified.
2. The method of claim 1, wherein the SAR image classification based on systolic self-encoder is as follows: the specific steps of the 1-layer stationary wavelet decomposition in the step (2) are as follows:
firstly, calculating the low-frequency component of the synthetic aperture radar SAR image to be classified according to the following formula:
wherein,representing synthetic aperture radars to be classifiedMth layer of SAR image smooth wavelet decomposition1Line n1The low frequency component of the column, ∑ denotes the summing operation,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column;
secondly, calculating the high-frequency component of the synthetic aperture radar SAR image to be classified in the horizontal direction according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency component of the horizontal direction of the column, ∑ denotes the summing operation,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column;
thirdly, calculating the high-frequency component of the synthetic aperture radar SAR image to be classified in the vertical direction according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency component of the vertical direction of the column, ∑ denotes the summing operation,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,filter coefficients H represented in an orthogonal wavelet filter HjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column;
fourthly, calculating high-frequency components in the diagonal direction of the synthetic aperture radar SAR image to be classified according to the following formula:
wherein,representing the mth layer of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified1Line n1The high frequency components in the diagonal direction of the column, ∑ denotes the summing operation,filter coefficient G expressed in orthogonal wavelet filter GjInterject 2j-an operation of 1 zero,representing the m < th > layer j-1 < th > of the steady wavelet decomposition of the synthetic aperture radar SAR image to be classified2Line n2Low frequency components of the column.
3. The method of claim 1, wherein the SAR image classification based on systolic self-encoder is as follows: the method for training the network of the self-encoder in the step (5a) specifically comprises the following steps of obtaining high-level low-frequency characteristics of a training sample:
first, the reconstruction error of the punctured autocoder is calculated as follows:
where L (x, z) represents the reconstruction error of the punctured self-encoder, x represents the low frequency training samples, z represents the output characteristics of the second layer of the punctured self-encoder, ∑ represents the summing operation, dxRepresenting the total dimension of the feature, i representing the specific feature dimension, xiFeatures representing the ith dimension of the low frequency training sample, log representing the log taking operation, ziFeatures representing an ith dimension of an output of a second layer punctured from the encoder;
secondly, calculating the penalty term of the contraction self-encoder according to the following formula:
wherein,a penalty term representing a punctured self-encoder, J (x) a Jacobian matrix representing output values of the punctured self-encoder with respect to weights, x representing low frequency training samples,denotes the squaring operation of the F norm, dhRepresenting the total dimension of the output features of the first layer of the punctured self-encoder, j representing the specific feature dimension, yjRepresents the output characteristics of the jth dimension of the first layer of the shrinking self-encoder, W represents the weight coefficient of the shrinking self-encoder, | | | | | sweet wind2A squaring operation representing a norm;
thirdly, calculating a loss function of the punctured self-encoder according to the following formula:
wherein, Jcae(theta) represents shrinkage fromLoss function of encoder, theta denotes parameter of punctured self-encoder, ∑ denotes summation operation, Tn={x(1),x(2),...,x(n)Denotes a training sample set containing n samples, L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a low frequency training sample, z denotes an output characteristic of a second layer of the punctured self-encoder, λ denotes a weight adjustment parameter,a penalty term representing a punctured self-encoder, J (x) a Jacobian matrix representing output values of the punctured self-encoder with respect to weights, x representing low frequency training samples,a squaring operation representing the F norm;
and fourthly, training the self-contraction encoder, optimizing a loss function of the self-contraction encoder through a random echelon descent SGD method, and adjusting a weight coefficient and a bias coefficient between layers in the self-contraction encoder to enable a reconstruction error of the self-contraction encoder to be minimum and obtain high-level low-frequency characteristics of the training sample.
4. The method of claim 1, wherein the SAR image classification based on systolic self-encoder is as follows: the method for training the network of the self-encoder in the step (5b) obtains the high-level and high-frequency characteristics of the training sample by the following specific steps:
first, the reconstruction error of the punctured autocoder is calculated as follows:
where L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a high frequency training sample, z denotes an output characteristic of the second layer of the punctured self-encoder, ∑ denotes a summing operation, dxRepresenting the total dimension of the feature, i representing the specific feature dimension, xiFeatures representing the ith dimension of the high frequency training sample, log representing the log taking operation, ziFeatures representing an ith dimension of an output of a second layer punctured from the encoder;
secondly, calculating the penalty term of the contraction self-encoder according to the following formula:
wherein,represents a penalty term for the punctured autocoder, J (x) represents a Jacobian matrix of output values of the punctured autocoder for the first layer with respect to weights, x represents high frequency training samples,denotes the squaring operation of the F norm, dhRepresenting the total dimension of the output features of the first layer of the punctured self-encoder, j representing the specific feature dimension, yjRepresents the output characteristics of the jth dimension of the first layer of the shrinking self-encoder, W represents the weight coefficient of the shrinking self-encoder, | | | | | sweet wind2A squaring operation representing a norm;
thirdly, calculating a loss function of the punctured self-encoder according to the following formula:
wherein, Jcae(theta) represents shrinkage fromLoss function of encoder, theta denotes parameter of punctured self-encoder, ∑ denotes summation operation, Tn={x(1),x(2),...,x(n)Denotes a training sample set containing n samples, L (x, z) denotes a reconstruction error of the punctured self-encoder, x denotes a high frequency training sample, z denotes an output characteristic of a second layer of the punctured self-encoder, λ denotes a weight adjustment parameter,represents a penalty term for the punctured autocoder, J (x) represents a Jacobian matrix of output values of the punctured autocoder for the first layer with respect to weights, x represents high frequency training samples,a squaring operation representing the F norm;
and fourthly, training the self-contraction encoder, optimizing a loss function of the self-contraction encoder through a random echelon descent SGD method, and adjusting a weight coefficient and a bias coefficient between layers in the self-contraction encoder to enable a reconstruction error of the self-contraction encoder to be minimum and obtain high-level high-frequency characteristics of the training sample.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610407324.6A CN106096650B (en) | 2016-06-12 | 2016-06-12 | Based on the SAR image classification method for shrinking self-encoding encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610407324.6A CN106096650B (en) | 2016-06-12 | 2016-06-12 | Based on the SAR image classification method for shrinking self-encoding encoder |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106096650A true CN106096650A (en) | 2016-11-09 |
CN106096650B CN106096650B (en) | 2019-08-13 |
Family
ID=57228565
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610407324.6A Active CN106096650B (en) | 2016-06-12 | 2016-06-12 | Based on the SAR image classification method for shrinking self-encoding encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106096650B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109214401A (en) * | 2017-06-30 | 2019-01-15 | 清华大学 | SAR image classification method and device based on stratification autocoder |
CN110411566A (en) * | 2019-08-01 | 2019-11-05 | 四川长虹电器股份有限公司 | A kind of Intelligent light spectrum signal denoising method |
CN112418267A (en) * | 2020-10-16 | 2021-02-26 | 江苏金智科技股份有限公司 | Motor fault diagnosis method based on multi-scale visual and deep learning |
-
2016
- 2016-06-12 CN CN201610407324.6A patent/CN106096650B/en active Active
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109214401A (en) * | 2017-06-30 | 2019-01-15 | 清华大学 | SAR image classification method and device based on stratification autocoder |
CN109214401B (en) * | 2017-06-30 | 2020-10-16 | 清华大学 | SAR image classification method and device based on hierarchical automatic encoder |
CN110411566A (en) * | 2019-08-01 | 2019-11-05 | 四川长虹电器股份有限公司 | A kind of Intelligent light spectrum signal denoising method |
CN112418267A (en) * | 2020-10-16 | 2021-02-26 | 江苏金智科技股份有限公司 | Motor fault diagnosis method based on multi-scale visual and deep learning |
CN112418267B (en) * | 2020-10-16 | 2023-10-24 | 江苏金智科技股份有限公司 | Motor fault diagnosis method based on multi-scale visual view and deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN106096650B (en) | 2019-08-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
CN107239751B (en) | High-resolution SAR image classification method based on non-subsampled contourlet full convolution network | |
CN103456018B (en) | Remote sensing image change detection method based on fusion and PCA kernel fuzzy clustering | |
CN104915676B (en) | SAR image sorting technique based on further feature study and watershed | |
CN106874889A (en) | Multiple features fusion SAR target discrimination methods based on convolutional neural networks | |
CN102609701B (en) | Remote sensing detection method based on optimal scale for high-resolution SAR (synthetic aperture radar) | |
CN104091151A (en) | Vehicle identification method based on Gabor feature extraction and sparse representation | |
CN105930772A (en) | City impervious surface extraction method based on fusion of SAR image and optical remote sensing image | |
CN105809198A (en) | SAR image target recognition method based on deep belief network | |
CN104299232B (en) | SAR image segmentation method based on self-adaptive window directionlet domain and improved FCM | |
CN107358260A (en) | A kind of Classification of Multispectral Images method based on surface wave CNN | |
CN112083422A (en) | Single-voyage InSAR system end-to-end classification method based on multistage deep learning network | |
Lefebvre et al. | Object-oriented approach and texture analysis for change detection in very high resolution images | |
CN104680184B (en) | Polarization SAR terrain classification method based on depth RPCA | |
CN104657717A (en) | Pedestrian detection method based on layered kernel sparse representation | |
Liang et al. | Maximum likelihood classification of soil remote sensing image based on deep learning | |
CN107133653B (en) | High-resolution SAR image classification method based on deep ladder network | |
CN106096650B (en) | Based on the SAR image classification method for shrinking self-encoding encoder | |
CN108256557B (en) | Hyperspectral image classification method combining deep learning and neighborhood integration | |
CN105138966A (en) | Quick density peak value clustering based polarimetric SAR image classification method | |
Mahrooghy et al. | A machine learning framework for detecting landslides on earthen levees using spaceborne SAR imagery | |
CN106846363A (en) | A kind of scale adaptability compression tracking for improving sparse matrix | |
CN106446854A (en) | High-resolution optical remote sensing image target detection method based on rotation invariant HOG feature | |
CN104102900A (en) | Vehicle identification system | |
CN109284752A (en) | A kind of rapid detection method of vehicle |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |