CN110414494A - SAR image classification method with ASPP deconvolution network - Google Patents
SAR image classification method with ASPP deconvolution network Download PDFInfo
- Publication number
- CN110414494A CN110414494A CN201910829626.6A CN201910829626A CN110414494A CN 110414494 A CN110414494 A CN 110414494A CN 201910829626 A CN201910829626 A CN 201910829626A CN 110414494 A CN110414494 A CN 110414494A
- Authority
- CN
- China
- Prior art keywords
- image
- aspp
- training
- test
- pixels
- 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
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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of SAR image classification method with ASPP deconvolution network, mainly solve the problems, such as that classification accuracy is not high, classification effectiveness is low, complicated for operation in SAR image assorting process.The specific steps of the present invention are as follows: (1) reading training image and test image；(2) training set and test set are generated；(3) building has the deconvolution network of void space pyramid pond ASPP；(4) training band has the deconvolution network of empty spatial pyramid pond ASPP；(5) classify to test image.The present invention overcomes being unable to fully in existing SAR image sorting algorithm using multiscale target information, cannot achieve the problem of end-to-end classification, have the advantages that classification accuracy is high, classification is quick, easy to operate.
Description
Technical field
The invention belongs to technical field of image processing, and it is free to further relate to one of Image Classfication Technology field band
The synthetic aperture radar of hole spatial pyramid pond ASPP (Atrous Spatial Pyramid Pooling) deconvolution network
SAR (Synthetic Aperture Radar) image classification method.The present invention can be used for in synthetic aperture radar SAR image
Ground object target type classify, ground object target type such as construction area, farmland region, airfield runway region.
Background technique
Synthetic aperture radar SAR because its round-the-clock, it is round-the-clock, penetration power is strong the features such as, SAR imaging technique has obtained extensively
Research.Thus SAR imaging technique is constantly mature, and the image resolution ratio of acquisition is higher and higher.SAR image classification is radar image
One important branch of processing technology field.SAR image classification be to ground object target different in SAR image, as construction area,
Classification is realized using different textural characteristics in farmland region, airfield runway region.Selected feature is determined in SAR image classification
The quality of classification results is determined, common textural characteristics include mean value, variance, entropy, energy and gray level co-occurrence matrixes etc..Depth
Learning method can automatically extract validity feature, solve the problems, such as that artificial selected characteristic is difficult in conventional method, wide at present
The general SAR image that is applied to is classified.Common deep learning classification method includes classification method based on convolutional neural networks, is based on
Classification method, classification method based on deconvolution neural network of depth confidence network etc..
Xian Electronics Science and Technology University its application patent document " based on super-pixel segmentation and convolution deconvolution network
It is disclosed in SAR image classification method " (number of patent application: 201810512092.X, application publication number: CN 108764330A)
A kind of SAR image classification method based on super-pixel segmentation and convolution deconvolution network.This method first send SAR image data
Enter convolution deconvolution network, obtains the preliminary classification of image as a result, by the super-pixel segmentation result of SAR image data and initial point
Class result combines to obtain final classification result.This method using the smooth SAR image of super-pixel segmentation preliminary classification as a result, gram
The problem of time-consuming of classifying in the SAR image classification method based on deep learning has been taken, the nicety of grading of SAR image is improved.
But the shortcoming that this method still has is, SAR image data is sent into convolution deconvolution neural network, without abundant
Using the multiscale target information of SAR image, SAR image classification accuracy is still not high enough.
HeFei University of Technology is in a kind of patent document " SAR image classification side based on textural characteristics and DBN of its application
It is disclosed in method " (number of patent application: 201710652513.4, application publication number: CN 107506699A) a kind of special based on texture
It seeks peace the SAR image classification method of DBN.This method is mainly by the GLCM feature of gray level image block and initial data image block
The strength information of GMRF feature and image is sent into depth confidence network DBN, obtains the classification results of image.This method introduces image
Textural characteristics assist DBN network class, overcome the SAR image classification method based on deep learning and only image intensity utilized to believe
The problem of breath, improves the nicety of grading of SAR image.But the shortcoming that this method still has is, utilizes the line of image
It manages feature and assists DBN network class, do not directly obtain the classification results of SAR image to be sorted, cannot achieve and divide end to end
Class, causes complicated for operation, and test is only capable of obtaining the classification results of a pixel every time, is unfavorable for Project Realization, SAR figure
As classification effectiveness is low.
Summary of the invention
It is a kind of with ASPP deconvolution network it is an object of the invention in view of the above shortcomings of the prior art, propose
SAR image classification method, mainly solve the classification accuracy in SAR image assorting process it is still not high enough, it is complicated for operation, classification
The problem of low efficiency.
Realizing the thinking of the object of the invention is: reading training image and test image, is mentioned from training image using sliding window
It takes training dataset and carries out data extending and generate training set, slided using sliding window from the different initial positions of test image
Generating test set, building have the deconvolution network of void space pyramid pond ASPP, and training set is sent into deconvolution network
It is trained to network parameter and restrains, obtain trained deconvolution network, be respectively sent into the different test set in initial position
Trained deconvolution network is classified to obtain final classification result.
Specific steps of the invention include the following:
(1) training image and test image are read:
Reading two width includes the SAR image data of identical ground object target respectively as training image and test image, reading
Any width SAR image data in include at least three classes ground object target, and pixel in any two classes ground object target of training image
The more one kind of number a kind of pixel number Ratio control less with pixel number is in the range of 1 to 1.5；
(2) training set and test set are generated:
(2a) constructs the sliding window that a size is 90 × 90 pixels；
Sliding window is placed on the upper left corner of training image by (2b), is successively slided by sequence from left to right, from top to bottom, sliding
Step-length is 10 pixels, will slide the training image region being overlapped in result with sliding window as a training image block of pixels；
Every other number of type in all pixels point, big by the number of classification in each training image block of pixels of (2c) statistics
In being equal to classification of 0.5 times of the classification of sliding window pixel sum as the block of pixels, a sample is constituted；
The other all samples of every type are constituted such training dataset by (2d)；
(2e) upsets sequence to every class training data set at random and randomly selects equal number of sample, forms initial instruction
Practice sample set, wherein the number of training that every class is chosen is not less than 2500；
(2f) rotates clockwise 90 for centered on the central pixel point of each training sample in initial training sample set
Degree, obtains the first expansion sets；
(2g) is 0 to each training sample addition mean value in initial training sample set, and the Gauss that standard deviation is 0.1 makes an uproar
Sound obtains the second expansion sets；
Initial training sample set, the first expansion sets, the second expansion sets are formed training set by (2h)；
Sliding window is placed on 5 different initial positions of test image by (2i), successively by sequence from left to right, from top to bottom
Sliding, sliding step are 90 pixels, will slide the test image region being overlapped in result with sliding window as a test image
Block of pixels separately constitutes first, second, third and fourth, five test sets；
(3) building has the deconvolution network of void space pyramid pond ASPP:
(3a) builds a deconvolution network, and structure is successively are as follows: input layer → 1 → pond of convolutional layer layer → nested type mould
Block → 1 → convolutional layer of splicing layer 2 → up-sampling layer → output layer；
The nested type module is made of first void space pyramid pond ASPP with branch 1 parallel, branch 1
Structure is followed successively by 3 → nested type of convolutional layer submodule 1 → splicing 2 → convolutional layer of layer, 4 → warp lamination 1；
The nested type submodule 1 is made of second void space pyramid pond ASPP with branch 2 parallel；Branch
2 structure is followed successively by 5 → nested type of convolutional layer submodule 2 → splicing 3 → convolutional layer of layer, 6 → warp lamination 2；
The nested type submodule 2 is made of third void space pyramid pond ASPP with branch 3 parallel；Branch
3 structure is followed successively by convolutional layer 7 → four void space pyramid pond ASPP → warp lamination 3；
The parameter of (3b) setting deconvolution network；
(4) training band has the deconvolution network of empty spatial pyramid pond ASPP:
Training set is input in the deconvolution network with void space pyramid pond ASPP, training band has empty sky
Between pyramid pond ASPP deconvolution network, until network parameter restrain, obtain trained deconvolution network；
(5) classify to test image:
First, second, third and fourth, five test sets are successively input in trained deconvolution network by (5a) respectively, are obtained every
The preliminary classification block of pixels of a test set；
(5b) finds the most label of each pixel corresponding position frequency of occurrence from all preliminary classification block of pixels and makees
For the classification of the pixel, final classification block of pixels is obtained, using the final classification block of pixels as the final classification of test image
As a result.
Compared with the prior art, the present invention has the following advantages:
First, since present invention uses four void space pyramid pond ASPP, cavity is carried out to characteristic pattern respectively
The operation of spatial pyramid pondization overcomes caused by algorithm existing in the prior art underuses because of multiscale target information
The not high enough problem of classification accuracy, so that the present invention improves the accuracy rate of SAR image classification.
Second, since present invention uses up-sampling layers to carry out up-sampling operation to characteristic pattern, it can once obtain one
The classification results of block of image pixels overcome algorithm existing in the prior art and are only capable of obtaining a pixel because testing every time
The low problem of SAR image classification effectiveness caused by classification results, so that the present invention improves classification effect in SAR image classification
Rate.
Third obtains the first of test set since test set is directly input in trained deconvolution network by the present invention
Beginning classified pixels block, overcoming cannot achieve the problem of classifying end to end existing for algorithm in the prior art, so that this hair
It is bright simple in SAR image sort operation.
Detailed description of the invention
Attached drawing 1 is flow chart of the invention；
Attached drawing 2 is schematic network structure of the invention；
Attached drawing 3 is the SAR image of actual measurement used in emulation experiment of the present invention；
Attached drawing 4 is the authentic signature figure of the SAR image of actual measurement used in emulation experiment of the present invention；
Attached drawing 5 is the classification results figure of emulation experiment of the present invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawing.
Referring to Fig.1, the specific steps realized to the present invention are further described.
Step 1, training image and test image are read.
Reading two width includes the SAR image data of identical ground object target respectively as training image and test image, reading
Any width SAR image data in include at least three classes ground object target, and pixel in any two classes ground object target of training image
The more one kind of number a kind of pixel number Ratio control less with pixel number is in the range of 1 to 1.5.
Step 2, training set and test set are generated.
Construct the sliding window that a size is 90 × 90 pixels.
Sliding window is placed on to the upper left corner of training image, is successively slided by sequence from left to right, from top to bottom, sliding step
For 10 pixels, the training image region being overlapped in result with sliding window will be slided as a training image block of pixels.
Every other number of type in all pixels point is counted in each training image block of pixels, the number of classification is greater than etc.
In classification of 0.5 times of the classification as the block of pixels of sliding window pixel sum, a sample is constituted.
The other all samples of every type are constituted to such training dataset.
Sequence is upset to every class training data set at random and randomly selects equal number of sample, forms initial training sample
This collection, wherein the number of training that every class is chosen is not less than 2500.
Centered on the central pixel point of each training sample in initial training sample set, 90 degree are rotated clockwise,
Obtain the first expansion sets.
It is 0 to each training sample addition mean value in initial training sample set, the Gaussian noise that standard deviation is 0.1 obtains
To the second expansion sets.
Initial training sample set, the first expansion sets, the second expansion sets are formed into training set.
Sliding window is placed on 5 different initial positions of test image, is successively slided by sequence from left to right, from top to bottom
Dynamic, sliding step is 90 pixels, will slide the test image region being overlapped in result with sliding window as a test image picture
Plain block separately constitutes first, second, third and fourth, five test sets.
Sliding window is placed on 5 different initial positions of test image and refers to, using the upper left corner of test image as initial position
Obtained test set is the first test set；It is obtained using the test image upper left corner to the position of 30 pixels of right translation as initial position
The test set arrived is the second test set；It is obtained using the test image upper left corner to the position of 60 pixels of right translation as initial position
Test set be third test set；What the position for translating 30 pixels downwards using the test image upper left corner was obtained as initial position
Test set is the 4th test set；The survey that the position for translating 60 pixels downwards using the test image upper left corner is obtained as initial position
Examination collection is the 5th test set.
Step 3, building has the deconvolution network of void space pyramid pond ASPP.
Referring to Fig. 2, network structure of the invention is further described.
A deconvolution network is built, structure is successively are as follows: input layer → 1 → pond of convolutional layer layer → nested type module →
Splice 1 → convolutional layer of layer 2 → up-sampling layer → output layer.
The nested type module is made of first void space pyramid pond ASPP with branch 1 parallel, branch 1
Structure is followed successively by 3 → nested type of convolutional layer submodule 1 → splicing 2 → convolutional layer of layer, 4 → warp lamination 1.
The nested type submodule 1 is made of second void space pyramid pond ASPP with branch 2 parallel；Branch
2 structure is followed successively by 5 → nested type of convolutional layer submodule 2 → splicing 3 → convolutional layer of layer, 6 → warp lamination 2.
The nested type submodule 2 is made of third void space pyramid pond ASPP with branch 3 parallel；Branch
3 structure is followed successively by convolutional layer 7 → four void space pyramid pond ASPP → warp lamination 3.
First, second, third, each structure of the 4th void space pyramid pond ASPP successively are as follows: simultaneously
4 capable branches → splicing layer → convolutional layer, the 1st in 4 parallel branches, the 2nd, the 3rd branch be sky
Hole convolutional layer, the structure of the 4th branch are followed successively by pond layer → up-sampling layer.
The parameter setting of void space pyramid pond ASPP is as follows:
The spreading rate of the empty convolutional layer of 1st, the 2nd, the 3rd branch is respectively set to 2,4,8 pixels, each
Empty convolution kernel size is 3 × 3, and each sliding step is 1 pixel.
Global average pond is set by the pond mode of the pond layer of the 4th branch.
Bilinear interpolation up-sampling is set by the up-sampling layer of the 4th branch.
The convolution kernel of convolutional layer is dimensioned to 1 × 1, sliding step is 1 pixel.
A matrix splicing function is set by splicing layer.
The parameter of deconvolution network is set.
The parameter setting of the deconvolution network is as follows:
32 are set by the number of the Feature Mapping figure of convolutional layer 1, convolution kernel size is 5 × 5, and sliding step is 1
Pixel.
Set the number of the Feature Mapping figure of convolutional layer 2 to total classification of the included ground object target of SAR image to be sorted
Number, convolution kernel size are 1 × 1, and sliding step is 1 pixel.
The number of the Feature Mapping figure of convolutional layer 3,5,7 is respectively set to 64,128,256, each convolution kernel size
It is 5 × 5, each sliding step is 2 pixels.
The number of the Feature Mapping figure of convolutional layer 4,6 is respectively set to 64,128, each convolution kernel size is 1 ×
1, each sliding step is 1 pixel.
Maximum pond is set by the pond mode of pond layer, the number of Feature Mapping figure is 32, pond window size
It is 2 × 2, sliding step is 2 pixels.
The number of the Feature Mapping figure of warp lamination 1,2,3 is respectively set to 64,128,256, each convolution kernel is big
Small is 5 × 5, and each sliding step is 2 pixels.
By first, second, third, the 4th void space pyramid pond ASPP Feature Mapping figure number
It is respectively set to 32,64,128,256.
Splicing layer 1,2,3 is respectively set to a matrix splicing function.
Bilinear interpolation up-sampling is set by up-sampling layer.
Step 4, training band has the deconvolution network of empty spatial pyramid pond ASPP.
Training set is input in the deconvolution network with void space pyramid pond ASPP, training band has empty sky
Between pyramid pond ASPP deconvolution network, until network parameter restrain, obtain trained deconvolution network.
Step 5, classify to test image.
Successively first, second, third and fourth, five test sets are input in trained deconvolution network respectively, obtain each survey
Try the preliminary classification block of pixels of collection.
The most label conduct of each pixel corresponding position frequency of occurrence is found from all preliminary classification block of pixels should
The classification of pixel obtains final classification block of pixels, using the final classification block of pixels as the final classification result of test image.
Effect of the invention can be illustrated by following emulation experiment.
1. emulation experiment condition:
Emulation experiment of the invention be dominant frequency 2.9GHz Intel Xeon (R) CPU E3-1535M v5, interior save as
It is carried out under the hardware environment and MATLAB R2017a of 31.1GB, the software environment of Python2.7.
2. emulation content and interpretation of result:
Emulation experiment of the invention is that the present invention is respectively adopted with two kinds of prior arts (based on multistage local mode histogram
The support vector machines classification method of MLPH feature, the classification method based on super-pixel segmentation and convolution deconvolution network) respectively
Classify to the width test image in the SAR image of actual measurement.
Fig. 3 is the SAR image of actual measurement used in emulation experiment of the present invention, and wherein Fig. 3 (a) is the width training figure of actual measurement
Picture, Fig. 3 (b) are a width test image of actual measurement.
Fig. 4 is the authentic signature figure of the SAR image of actual measurement used in emulation experiment of the present invention, and wherein Fig. 4 (a) is Fig. 3 (a)
The authentic signature figure of one width training image of actual measurement, Fig. 4 (b) are the authentic signature figure of a width test image of Fig. 3 (b) actual measurement.
Fig. 5 is the classification results figure of emulation experiment of the present invention, and wherein Fig. 5 (a) is a width of the present invention to Fig. 3 (b) actual measurement
The classification results figure of test image；Fig. 5 (b) is the svm classifier method using the prior art based on MLPH feature, real to Fig. 3 (b)
The classification results figure for the width test image surveyed；Fig. 5 (c) is to be based on super-pixel segmentation and convolution deconvolution net using the prior art
The classification method of network, to the classification results figure of a width test image of Fig. 3 (b) actual measurement.Comparison diagram 5 (a), Fig. 5 (b), Fig. 5 (c),
The method of the present invention is smoother for the boundary classification of Fig. 3 (b) width test image surveyed, and surveys referring to a width of Fig. 4 (b) actual measurement
Attempt the authentic signature figure of picture, it is seen that classification results Fig. 5 (a) of the invention is compared to now there are two types of technology classification result figures 5
(b), Fig. 5 (c) illustrates that classification of the invention is more acurrate, classifying quality is more preferable closer to authentic signature figure.
Table 1 is to be utilized respectively the method for the present invention and now divide there are two types of technology the width test image that Fig. 3 (b) is surveyed
The performance parameter of class compares, and the F1 in table indicates the svm classifier method based on MLPH feature, F2 indicate based on super-pixel segmentation and
The classification method of convolution deconvolution network, F3 indicate the method for the present invention.Classification accuracy is the width surveyed to Fig. 3 (b) in table
The correct pixel number of the classification that test image is classified, with the total pixel number purpose ratio of the test image, classification is just
Really refer to that the classification results classified to test image authentic signature corresponding with preceding test image of classifying is consistent.Average mark
Class accuracy rate is to classify to the width test image that Fig. 3 (b) is surveyed, the mean value of the classification accuracy of all kinds of ground object targets；
Runing time is to carry out runing time used of classifying to the width test image that Fig. 3 (b) is surveyed.
1. 3 kinds of classification methods of table are to the sorted performance parameter contrast table of test image
As it can be seen that the method for the present invention tests a width of Fig. 3 (b) actual measurement compared to now there are two types of technologies from contrast table 1
The average classification accuracy of image is higher, and runing time is also most short.Illustrate the present invention by utilizing four void space pyramids
Pond ASPP extracts the multiscale target information of SAR image, improves classification accuracy, by utilizing warp lamination and up-sampling
Layer, which to test every time, obtains the classification results of SAR image block of pixels, realizes and classifies end to end, shortens runing time,
Convenient for Project Realization.
Claims (5)
1. a kind of SAR image classification method with ASPP deconvolution network, which is characterized in that read training image and test chart
Picture generates training set and test set, and building has the deconvolution network of void space pyramid pond ASPP, and training band has cavity
The step of deconvolution network of spatial pyramid pond ASPP, classifies to test image, this method include the following:
(1) training image and test image are read:
Reading two width includes the SAR image data of identical ground object target respectively as training image and test image, times of reading
It include at least three classes ground object target, and pixel number in any two classes ground object target of training image in one width SAR image data
More one kind a kind of pixel number Ratio control less with pixel number is in the range of 1 to 1.5；
(2) training set and test set are generated:
(2a) constructs the sliding window that a size is 90 × 90 pixels；
Sliding window is placed on the upper left corner of training image by (2b), is successively slided by sequence from left to right, from top to bottom, sliding step
For 10 pixels, the training image region being overlapped in result with sliding window will be slided as a training image block of pixels；
Every other number of type in all pixels point in each training image block of pixels of (2c) statistics, the number of classification is greater than etc.
In classification of 0.5 times of the classification as the block of pixels of sliding window pixel sum, a sample is constituted；
The other all samples of every type are constituted such training dataset by (2d)；
(2e) upsets sequence to every class training data set at random and randomly selects equal number of sample, forms initial training sample
This collection, wherein the number of training that every class is chosen is not less than 2500；
(2f) rotates clockwise 90 degree for centered on the central pixel point of each training sample in initial training sample set,
Obtain the first expansion sets；
(2g) is 0 to each training sample addition mean value in initial training sample set, and the Gaussian noise that standard deviation is 0.1 obtains
To the second expansion sets；
Initial training sample set, the first expansion sets, the second expansion sets are formed training set by (2h)；
Sliding window is placed on 5 different initial positions of test image by (2i), is successively slided by sequence from left to right, from top to bottom
Dynamic, sliding step is 90 pixels, will slide the test image region being overlapped in result with sliding window as a test image picture
Plain block separately constitutes first, second, third and fourth, five test sets；
(3) building has the deconvolution network of void space pyramid pond ASPP:
(3a) builds a deconvolution network, and structure is successively are as follows: and input layer → 1 → pond of convolutional layer layer → nested type module →
Splice 1 → convolutional layer of layer 2 → up-sampling layer → output layer；
The nested type module is made of first void space pyramid pond ASPP with branch 1 parallel, the structure of branch 1
It is followed successively by 3 → nested type of convolutional layer submodule 1 → splicing 2 → convolutional layer of layer, 4 → warp lamination 1；
The nested type submodule 1 is made of second void space pyramid pond ASPP with branch 2 parallel；Branch 2
Structure is followed successively by 5 → nested type of convolutional layer submodule 2 → splicing 3 → convolutional layer of layer, 6 → warp lamination 2；
The nested type submodule 2 is made of third void space pyramid pond ASPP with branch 3 parallel；Branch 3
Structure is followed successively by convolutional layer 7 → four void space pyramid pond ASPP → warp lamination 3；
The parameter of (3b) setting deconvolution network；
(4) training band has the deconvolution network of empty spatial pyramid pond ASPP:
Training set is input in the deconvolution network with void space pyramid pond ASPP, training band has void space golden
The deconvolution network of word tower basin ASPP obtains trained deconvolution network until network parameter convergence；
(5) classify to test image:
First, second, third and fourth, five test sets are successively input in trained deconvolution network by (5a) respectively, obtain each survey
Try the preliminary classification block of pixels of collection；
(5b) finds the most label conduct of each pixel corresponding position frequency of occurrence from all preliminary classification block of pixels should
The classification of pixel obtains final classification block of pixels, using the final classification block of pixels as the final classification result of test image.
2. the SAR image classification method according to claim 1 with ASPP deconvolution network, which is characterized in that step
Sliding window described in (2i) is placed on 5 different initial positions of test image and refers to, using the upper left corner of test image as starting
The test set that position obtains is the first test set；Using the test image upper left corner to the position of 30 pixels of right translation as start bit
The test set set is the second test set；Using the test image upper left corner to the position of 60 pixels of right translation as initial position
Obtained test set is third test set；The position for translating 30 pixels downwards using the test image upper left corner is obtained as initial position
The test set arrived is the 4th test set；The position for translating 60 pixels downwards using the test image upper left corner is obtained as initial position
Test set be the 5th test set.
3. the SAR image classification method according to claim 1 with ASPP deconvolution network, which is characterized in that step
First described in (3a), second, third, each structure of the 4th void space pyramid pond ASPP successively
Are as follows: parallel 4 branches → splicing layer → convolutional layer, the 1st in 4 parallel branches, the 2nd, the 3rd branch it is equal
For empty convolutional layer, the structure of the 4th branch is followed successively by pond layer → up-sampling layer.
4. the SAR image classification method according to claim 3 with ASPP deconvolution network, which is characterized in that described
Void space pyramid pond ASPP parameter setting it is as follows:
The spreading rate of the empty convolutional layer of 1st, the 2nd, the 3rd branch is respectively set to 2,4,8 pixels, each cavity
Convolution kernel size is 3 × 3, and each sliding step is 1 pixel；
Global average pond is set by the pond mode of the pond layer of the 4th branch；
Bilinear interpolation up-sampling is set by the up-sampling layer of the 4th branch；
The convolution kernel of convolutional layer is dimensioned to 1 × 1, sliding step is 1 pixel；
A matrix splicing function is set by splicing layer.
5. the SAR image classification method according to claim 1 with ASPP deconvolution network, which is characterized in that step
The parameter setting of deconvolution network described in (3b) is as follows:
32 are set by the number of the Feature Mapping figure of convolutional layer 1, convolution kernel size is 5 × 5, and sliding step is 1 pixel；
It sets the number of the Feature Mapping figure of convolutional layer 2 to total classification number of the included ground object target of SAR image to be sorted, rolls up
Product core size is 1 × 1, and sliding step is 1 pixel；
The number of the Feature Mapping figure of convolutional layer 3,5,7 is respectively set to 64,128,256, each convolution kernel size is 5
× 5, each sliding step is 2 pixels；
The number of the Feature Mapping figure of convolutional layer 4,6 is respectively set to 64,128, each convolution kernel size is 1 × 1, often
A sliding step is 1 pixel；
Set maximum pond for the pond mode of pond layer, the number of Feature Mapping figure is 32, pond window size for 2 ×
2, sliding step is 2 pixels；
The number of the Feature Mapping figure of warp lamination 1,2,3 is respectively set to 64,128,256, each convolution kernel size is equal
It is 5 × 5, each sliding step is 2 pixels；
First, second, third, the number of Feature Mapping figure of the 4th void space pyramid pond ASPP are distinguished
It is set as 32,64,128,256；
Splicing layer 1,2,3 is respectively set to a matrix splicing function；
Bilinear interpolation up-sampling is set by up-sampling layer.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2019100710516 | 2019-01-25 | ||
| CN201910071051 | 2019-01-25 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110414494A true CN110414494A (en) | 2019-11-05 |
| CN110414494B CN110414494B (en) | 2022-12-02 |
Family
ID=68370301
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910829626.6A Active CN110414494B (en) | 2019-01-25 | 2019-09-03 | SAR image classification method with ASPP deconvolution network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110414494B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110827275A (en) * | 2019-11-22 | 2020-02-21 | 吉林大学第一医院 | Liver nuclear magnetic artery phase image quality grading method based on raspberry group and deep learning |
| CN111563528A (en) * | 2020-03-31 | 2020-08-21 | 西北工业大学 | SAR image classification method based on multi-scale feature learning network and bilateral filtering |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160104058A1 (en) * | 2014-10-09 | 2016-04-14 | Microsoft Technology Licensing, Llc | Generic object detection in images |
| WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN107944353A (en) * | 2017-11-10 | 2018-04-20 | 西安电子科技大学 | SAR image change detection based on profile ripple BSPP networks |
| CN108764330A (en) * | 2018-05-25 | 2018-11-06 | 西安电子科技大学 | SAR image sorting technique based on super-pixel segmentation and convolution deconvolution network |
-
2019
- 2019-09-03 CN CN201910829626.6A patent/CN110414494B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160104058A1 (en) * | 2014-10-09 | 2016-04-14 | Microsoft Technology Licensing, Llc | Generic object detection in images |
| WO2018028255A1 (en) * | 2016-08-11 | 2018-02-15 | 深圳市未来媒体技术研究院 | Image saliency detection method based on adversarial network |
| CN107944353A (en) * | 2017-11-10 | 2018-04-20 | 西安电子科技大学 | SAR image change detection based on profile ripple BSPP networks |
| CN108764330A (en) * | 2018-05-25 | 2018-11-06 | 西安电子科技大学 | SAR image sorting technique based on super-pixel segmentation and convolution deconvolution network |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110827275A (en) * | 2019-11-22 | 2020-02-21 | 吉林大学第一医院 | Liver nuclear magnetic artery phase image quality grading method based on raspberry group and deep learning |
| CN110827275B (en) * | 2019-11-22 | 2023-12-22 | 吉林大学第一医院 | Liver nuclear magnetic artery image quality grading method based on raspberry pie and deep learning |
| CN111563528A (en) * | 2020-03-31 | 2020-08-21 | 西北工业大学 | SAR image classification method based on multi-scale feature learning network and bilateral filtering |
| CN111563528B (en) * | 2020-03-31 | 2022-03-11 | 西北工业大学 | SAR image classification method based on multi-scale feature learning network and bilateral filtering |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110414494B (en) | 2022-12-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108564006B (en) | Polarized SAR terrain classification method based on self-learning convolutional neural network | |
| CN104077599B (en) | Polarization SAR image classification method based on deep neural network | |
| CN107229918A (en) | A kind of SAR image object detection method based on full convolutional neural networks | |
| CN108830330B (en) | Multispectral image classification method based on self-adaptive feature fusion residual error network | |
| CN110084159A (en) | Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint | |
| CN110263705A (en) | Towards two phase of remote sensing technology field high-resolution remote sensing image change detecting method | |
| CN108921030B (en) | SAR automatic target recognition method | |
| CN110110596B (en) | Hyperspectral image feature extraction, classification model construction and classification method | |
| CN103927551B (en) | Polarimetric SAR semi-supervised classification method based on superpixel correlation matrix | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
| CN109389080A (en) | Hyperspectral image classification method based on semi-supervised WGAN-GP | |
| CN107358203B (en) | A kind of High Resolution SAR image classification method based on depth convolution ladder network | |
| CN103761526B (en) | Urban area detecting method based on feature position optimization and integration | |
| CN103578110A (en) | Multi-band high-resolution remote sensing image segmentation method based on gray scale co-occurrence matrix | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN108960404B (en) | Image-based crowd counting method and device | |
| CN107944370A (en) | Classification of Polarimetric SAR Image method based on DCCGAN models | |
| CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
| CN110309781A (en) | Damage remote sensing recognition method in house based on the fusion of multi-scale spectrum texture self-adaption | |
| CN104299232B (en) | SAR image segmentation method based on self-adaptive window directionlet domain and improved FCM | |
| CN103247059A (en) | Remote sensing image region of interest detection method based on integer wavelets and visual features | |
| CN113657326A (en) | Weed detection method based on multi-scale fusion module and feature enhancement | |
| CN113128335B (en) | Method, system and application for detecting, classifying and finding micro-living ancient fossil image | |
| CN112232328A (en) | Remote sensing image building area extraction method and device based on convolutional neural network |
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 |