CN108122008A - SAR image recognition methods based on rarefaction representation and multiple features decision level fusion - Google Patents
SAR image recognition methods based on rarefaction representation and multiple features decision level fusion Download PDFInfo
- Publication number
- CN108122008A CN108122008A CN201711405236.3A CN201711405236A CN108122008A CN 108122008 A CN108122008 A CN 108122008A CN 201711405236 A CN201711405236 A CN 201711405236A CN 108122008 A CN108122008 A CN 108122008A
- Authority
- CN
- China
- Prior art keywords
- convolution
- target
- feature
- random
- formula
- 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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
Abstract
The present invention relates to the SAR image recognition methods based on rarefaction representation and multiple features decision level fusion.To improve the discrimination and recognition speed of SAR Target Recognition Algorithms, SAR sectioning images are extracted gray feature and the random convolution feature vector of dimensionality reduction by the present invention, then the dictionary formed using dictionary learning algorithm to the feature vector extracted from each classification training sample is optimized, dictionary is formed, recovering sample sparse coefficient finally by dictionary obtains classification results.Method proposed by the present invention improves accuracy of identification while recognition speed is drastically increased, and has better application prospect.
Description
Technical field
The invention belongs to SAR (Synthetic Aperture Radar) Image Automatic Targets to identify field, is related to one kind
SAR image recognition methods based on rarefaction representation and multiple features decision level fusion.
Background technology
SAR image automatic target detection is that SAR image interpretation needs one of key problem urgently solved, the course of work
To find out area-of-interest in SAR image first, definite target classification of classifying then is carried out to it.SAR image target identification
It is widely applied in national economy and national defense construction, detected including marine monitoring system, mineral reserve etc..
Feature extraction and classifier design are to influence two key factors of SAR image target identification precision.From SAR image
The feature of middle extraction mainly includes the feature based on mathematic(al) manipulation, computer vision feature and electromagnetic signature etc., Computer
Visual signature mainly includes texture, attitude angle, shape etc..Current main SAR image Target Recognition Algorithms are included based on template
Matched method, the method based on support vector machines, the method based on Boosting, method based on rarefaction representation etc..It is based on
The Target Recognition Algorithms of rarefaction representation are applied to recognition of face earliest, have started to be widely used in SAR image target in recent years
Identification, and achieve higher accuracy of identification.To improve the SAR image target identification precision based on rarefaction representation, from SAR image
The effective target signature of middle extraction, and carry out dictionary optimization be frequently with means.The main feature used at present includes gray scale
Feature singly drills signal characteristic, improves SIFT feature etc., and main dictionary learning algorithm includes KSVD, LCKSVD, Online
Learning etc..
Depth learning technology can be based on magnanimity big data and carry out feature learning, and the feature that study obtains is typically superior to people
The well-designed feature of work is based especially on the depth model of convolutional neural networks, by carrying out convolution, Chi Hua to two dimensional image
Deng operation, the multiple dimensioned two-dimentional local feature of target can be effectively extracted, the classification performance in terms of recognition of face is special better than HOG
Sign.But problems with need to be solved when carrying out feature learning based on depth convolutional neural networks：(1) training sample lacks；(2) it is deep
Degree model needs optimization design；(3) training time is long.Even if literature research shows using the convolution collecting image generated at random
It is filtered, by reasonable design grader, the random convolution feature of extraction can also obtain excellent classifying quality.
When carrying out SAR image information fusion using various features, the complementation of various features how is effectively utilized
Advantage, reasonable design grader and fusion rule are the key elements for influencing SAR image target identification precision.When using dilute
When dredging method for expressing progress target identification, to improve the target identification precision and recognition speed of SAR image, the present invention is from image
Gray feature and random convolution feature are extracted respectively, are carried out using decision level fusion strategy comprehensive based on two kinds of tagsort results
It closes and differentiates.Due to the calculation amount of the Target Recognition Algorithms based on rarefaction representation and the intrinsic dimensionality used and the element of sparse dictionary
Number is closely related, to improve the recognition speed of the SAR Target Recognition Algorithms based on rarefaction representation, on the one hand using sparse random
Projecting method carries out dimensionality reduction to the random convolution feature of the higher-dimension of extraction, on the other hand using Online dictionary learning algorithms to being used for
The dictionary of target classification optimizes, and in the case where ensureing target classification precision, is formed using less dictionary element number
Dictionary reduces the calculation amount of the SAR Target Recognition Algorithms based on rarefaction representation.It, will when the recognition result of two kinds of features of fusion
It is converted into based on the signal reconstruction error after rarefaction representation sparse coefficient Optimization Solution and belongs to each classification target identification probability, so
The fusion of classification results is carried out based on Bayesian Fusion rule afterwards.The experimental results showed that using the mapping of sparse accidental projection and word
Two kinds of means of allusion quotation Optimization Learning can improve the speed of target classification, and the method that the present invention designs can be at multiple operating conditions
Document is obtained and is currently known close to the recognition performance of even more high, algorithm strong applicability.
The content of the invention
SAR magnitude images directly reflect the retroreflection coefficient of target and atural object, therefore can directly utilize SAR image
Gray feature carries out target identification.Local feature based on every pixel neighborhoods information extraction usually has the description of stronger feature
Ability, but it is affected by the variation of platform depression angle, compared to gray feature, due to fully utilizing local region information, therefore
The influence of picture noise can be weakened.The present invention considers that depth convolution feature has stronger target local message descriptive power,
To improve the discrimination and real-time of the SAR image Target Recognition Algorithms based on rarefaction representation, devise a kind of based on sparse table
Show the SAR image recognition methods with multiple features decision level fusion, using two kinds of the mapping of sparse accidental projection and dictionary Optimization Learning
Means improve the real-time of the SAR image identification based on rarefaction representation, by merging the gray feature extracted from SAR image
With the recognition result of the random convolution feature of dimensionality reduction, SAR image target identification precision is improved.
To solve the problems, such as the SAR image multiple features decision level fusion identification based on rarefaction representation, and improve the identification of algorithm
Precision and real-time, the technical solution adopted by the present invention comprise the following steps：
Step (1) pre-processes original SAR image, obtains target slice image.
Step (2) extracts target gray feature vector.
Step (3) generates the convolution kernel of multiple and different sizes at random, carries out convolutional filtering to target slice image, passes through
Multiple dimensioned random convolution feature vector is extracted in equal Fang Chihua operations；It is special to the random convolution of extraction using the mapping of sparse accidental projection
Sign vector carries out dimensionality reduction, obtains the random convolution feature vector of dimensionality reduction.
Step (4) is formed two feature vectors extracted from each classification training sample using dictionary learning algorithm
Dictionary optimizes respectively, then the dictionary after optimization is merged into the dictionary for target identification.
When step (5) carries out test sample, sparse coefficient optimization is carried out based on gray feature and the random convolution feature of dimensionality reduction
It solves, reconstructed error is converted into target classification probability and carries out decision level fusion, differentiate so as to fulfill target classification.
Compared with prior art, the present invention its remarkable advantage is：(1) when using framework of sparse representation progress target identification
When, excessively high intrinsic dimensionality can influence the real-time of target identification, be carried out by the random convolution feature of higher-dimension to extraction random
It projects dimensionality reduction and using dictionary learning means to being optimized by the dictionary that convolution feature vector is formed after dimensionality reduction, is ensureing
The real-time of algorithm is improved while SAR target identification performances.(2) two kinds of features with complementary advantage have been merged, by base
It is converted into the reconstructed error that rarefaction representation progress sparse coefficient is recovered and belongs to each classification target probability, using Bayes
Fusion rule has carried out decision level fusion, improves the discrimination of SAR image target.(3) in standard test condition and a variety of expansions
Many experiments have been carried out under exhibition test condition, the experimental results showed that, the method for invention achieves higher nicety of grading, and algorithm is fitted
It is strong with property.
Description of the drawings
Fig. 1 is the flow chart of the method for the present invention.
Specific embodiment
Below in conjunction with attached drawing, the invention will be further described.
As shown in Figure 1, specific implementation step of the present invention is as follows：
Step (1) pre-processes original SAR image, obtains target slice image I.Concrete operations are：
Original SAR image is filtered using Mean Filtering Algorithm, filtering core size is 3 × 3.With plane of delineation two dimension
Central point is coordinate origin, the SAR sectioning images I divided by 255.0 that extraction size is 64 × 64 so that gradation of image etc.
Grade is located at section [0,1].
Step (2) extracts target gray feature vector.Concrete operations are：
Target slice image I by row is arranged, it is made to be converted to vector f1.By vector f1It is normalized, removes first
With vector f12 norms, then subtract to obtain the average of vector, obtain the gray feature vector of target(i.e. N-dimensional to
Amount is added N number of element, last divided by N)
Step (3) generates the convolution kernel of multiple and different sizes at random, carries out convolutional filtering to target slice image, passes through
Multiple dimensioned random convolution feature vector is extracted in equal Fang Chihua operations；It is special to the random convolution of extraction using sparse accidental projection matrix
Sign vector carries out dimensionality reduction, obtains the random convolution feature vector of dimensionality reduction.Concrete operations are：
For the random convolution feature of extraction, target slice image I is zoomed in and out first, its resolution ratio is made to become 32 × 32,
Obtained SAR target slice images are denoted as P.Since SAR image has stronger coherent speckle noise, larger convolution need to be used
Core extracts random convolution feature.The convolution kernel width set of use is set as { 5,7,9,11,13 }, is first sampled according to certain probability
It first determines convolution kernel width, is then based on formula (1) generation two-dimensional convolution core Kernelm, the convolution nuclear volume used is N.
Kernelm- 1 formulas of (i, j)=2 × rand () (1)
Wherein i, j represent the row, column coordinate of two-dimensional convolution core respectively, and m=1,2 ..., N represent the index of two-dimensional convolution core.
Rand (), which is generated, is located at section [0 1] equally distributed random number.
Since the convolution kernel width of use is inconsistent, for convenience of subsequent processing, target slice figure is checked using two-dimensional convolution
As when being filtered, carrying out zero padding operation to original image so that using different convolution kernels to the feature that is generated after image filtering
Picture size is consistent, specifically as shown in formula (2).
Fm=Kernelm* padding (P) formula (2)
Wherein, FmRepresent the two-dimensional convolution feature being calculated, * represents convolution operation, and padding (P) is represented according to volume
The width of product core is filled target slice image P edges with 0 so that FmIt is consistent with the size of P.
Two-dimensional random convolution feature to make acquisition has certain consistency to target translation, is operated using equal Fang Chihua
Eigentransformation is carried out, as shown in formula (3).
Wherein i, j=1,2 ..., 32+1-g, g are the peak width that pondization operation carries out.
After N number of convolution kernel is based on to SAR sectioning images by convolution and pondization operation progress feature extraction, by the two of generation
Tie up convolution characteristic pattern PmStretch feature vector f in columnm, and by { fm}M=1 ... NMerge, form a bigger dimension feature to
Measure f2, dimension is [N × (32+1-g)2 1]。
Due to feature vector f2Dimension it is larger, when using sparse representation method design grader when, carry out sparse coefficient
It will be taken very much during recovery, it is general that dimensionality reduction is carried out to the feature of extraction using accidental projection mapping method, for example with Gauss point
Cloth generates accidental projection mapping matrix.To make the feature after dimensionality reduction more sparse, sparse accidental projection is generated using formula (4) and is mapped
Matrix Rab。
Wherein, a is through sparse accidental projection mapping matrix RabDimension after dimensionality reduction, b=N × (32+1-g)2, be characterized to
Measure f2Dimension, ρ determines matrix RabSparse degree.As ρ=b/c, ideally matrix RabEvery a line in only c
A element is non-zero, and c is a constant, and experimental study shows when c takes the numerical value of very little, the mapping matrix R of generationab
It is sparse.
Former feature vector f2After formula (5) dimensionality reduction, the random convolution feature vector of dimensionality reduction is obtained
Step (4) is formed two feature vectors extracted from each classification training sample using dictionary learning algorithm
Dictionary optimizes respectively, then the dictionary after optimization is merged into the dictionary for target identification.Concrete operations are：
Target Recognition Algorithms based on rarefaction representation are taken except related with the target signature dimension of use, also with forming word
Each classification target dictionary element number of allusion quotation is related, can generally pass through dictionary optimization means so that is ensureing nicety of grading
Under the premise of, use the real-time of less dictionary element number raising algorithm.
If adding up to the target of T classification, when carrying out dictionary optimization, if all trained samples that same classification t will be belonged to
The feature vector of this extractionDictionary is arranged to make up by rowWherein ntRepresent all training sample numbers of classification t, i
=1 ..., nt.The dictionary after optimization is obtained using Online dictionary learning methods1 expression is the word of gray feature
Allusion quotation, 2 are expressed as random convolution feature.
Shown in the object function such as formula (6) of Online dictionary learnings method optimization.Wherein, XtFor point to be optimized of input
Category dictionary, sparse coding sparse coefficient matrix are set toSparse coding and dictionary learning mistake are carried out by interaction
Journey, iteration optimization obtain dictionary after sparse coding sparse coefficient matrix A and optimizationFinally to belonging to the dictionaries of all categories
It is combined, obtains the dictionary for target classification
Wherein, l is the dimension of the feature vector of input, and d is the number of the dictionary element of classification t after optimizing, and λ is regularization
Coefficient.
The dictionary formed for the random convolution feature of gray feature vector sum dimensionality reduction extracted from training sample
It optimizes and is combined respectively based on formula (6), the dictionary D after being optimized1And D2。
When step (5) carries out test sample, sparse coefficient optimization is carried out based on gray feature and the random convolution feature of dimensionality reduction
It solves, reconstructed error is converted into target classification probability and carries out decision level fusion, differentiate so as to fulfill target classification.Specific behaviour
As：
Extract gray feature vector respectively from sample to be tested xWith the random convolution feature vector of dimensionality reductionBased on l1Model
Number solves the convex optimization problem as shown in formula (7) respectively, obtains corresponding sparse coefficient α1And α2.Notation index is denoted as p；Wherein,
ε is a threshold value, and general value is 0.01.
The sparse coefficient α obtained according to solution1And α2, target classification differentiation can be carried out.Since SAR image has strongly
Coherent speckle noise, method of the generally use based on minimal reconstruction error differentiate target classification.For kth class target, defined function：ForIt is coefficient vector, in wherein α at index only corresponding with kth class target
Value remains unchanged, and value is set to zero at remaining index.Define residual error rk(x) it is
It is obtained by formula (9), the n of residual error minimum1, n2Value is corresponded to test sample x and is rolled up at random based on gray feature and dimensionality reduction respectively
The recognition result of product feature.For ease of being merged to the recognition result based on two kinds of features, reconstructed error is obtained according to formula (8)Afterwards, the reconstructed error that sparse coefficient is obtained based on recovery is converted into based on SoftMax thoughts and belongs to each classification target
Identification probability, as shown in formula (10) and formula (11).
The fusion of two kinds of feature recognition results is carried out using Bayesian Fusion rule, as shown in formula (12).The classification of target
Class formula (13) Suo Shi by obtaining.
To verify effectiveness of the invention, using MSTAR databases come the performance of testing algorithm.In MSTAR databases
SAR image be polarized by HH, the SAR sensors of 0.3 × 0.3m resolution ratio, X-band collect.Including compound target
10 class targets are amounted to, the number of training and test sample number of use are as shown in table 1.Using picture centre as origin extraction 64 × 64
Section, the gray feature dimension directly extracted be 4096 × 1.To extract two-dimensional random convolution feature, by the SAR targets of extraction
Sectioning image zooms to 32 × 32.Partial parameters are as follows in experiment：N=48, g=3, ε=0.01, λ=0.01, c=4, it is each
Class target word allusion quotation element number is set as d=75.B=N × (32+1-g) at this time2=43200, random convolution is special after setting dimensionality reduction
The dimension a=4000 of sign.
1 MSTAR database objects of table describe
When carrying out classification experiments using 10 class Morph Targets in table 1, SOC1 is denoted as, to wherein BMP2, BTR70 and T72 tri-
When class target carries out classification experiments, SOC2 is denoted as.First using two kinds of configurations of SOC1 and SOC2, compare and carried from SAR image
The discriminating power of the random convolution feature of gray feature vector sum dimensionality reduction taken, at the same analyze dictionary learning for nicety of grading and
The influence of algorithm real-time.Since random convolution feature extraction in the method for design and dimensionality reduction have certain randomness, therefore carry out
5 Monte Carlo simulations experiments, experimental result comparison are as shown in table 2.It can be seen that from the extraction formula of convolution feature such as formula
(2) shown in, the feature of enhancing is generated it makes use of the neighborhood information of every pixel, therefore in (instruction similar in depression angle
It is 17 ° to practice sample, and test sample is 15 °), the back scattering distribution of target is close, the accuracy of identification based on random convolution feature
Higher even if employing sparse accidental projection method carries out Feature Dimension Reduction, does not significantly reduce the random convolution feature of extraction yet
Discriminating power.
Different characteristic classification performance comparison (5 Monte Carlo simulation) of the table 2 based on rarefaction representation
From Table 2, it can be seen that when classifying for 10 class Morph Targets (SOC1), using the random convolution feature of dimensionality reduction
The classification results of gray feature vector are slightly above directly used, nicety of grading is improved by 94.75% to 96.05%, when using word
During allusion quotation optimization algorithm, the accuracy of identification based on the random convolution feature of dimensionality reduction is promoted to 96.05% by 94.18%.It is deformed for 3 classes
When target is classified (SOC2), it is better than the classification results of gray feature using the classification results of the random convolution feature of dimensionality reduction, point
Class performance is promoted to 95.62% by 92.16% highest, if do not use the dictionary to optimize, point based on the random convolution feature of dimensionality reduction
Class result reaches 96.35%.Compare the configuration of SOC1 and SOC2 it can be found that deformation is mainly reflected in BMP2 and T72 two
Class target, therefore when target classification increases but deformation does not increase, cause based on the recognition result of gray feature close to base
In the recognition result of the random convolution feature of dimensionality reduction.
The real-time of more several algorithms can be seen that based on dictionary Optimization Learning method, is based on due to reducing composition
The dictionary element number of the target identification method of rarefaction representation, therefore algorithm real-time gets a promotion.Using dictionary optimization method pair
It is not quite similar in the recognition performance influence of target, when convolution feature random using dimensionality reduction, for SOC1 scenes, target identification essence
Degree increases, for SOC2 scenes, then target identification precision reduction.When using gray feature, situation is just opposite.
Further to verify the validity of the mapping of the sparse accidental projection of use and dictionary Optimization Learning method, dimensionality reduction is set
The dimension a=1000 of random convolution feature afterwards, as can be seen from Table 2, in the case where overall classification accuracy slightly reduces,
The real-time of sorting algorithm is significantly improved, at this point, in the case where equal SOC2 scenes have dictionary learning, the drop of use
The overall classification accuracy for tieing up random convolution feature remains above the nicety of grading based on gray feature.This illustrates the random volume used
The validity of product feature.
It is identified for the SAR image in MSTAR databases, in addition to above two Standard Operating Conditions, using in addition
Four kinds of extended operation conditions have carried out SAR experiment for target identification, are denoted as EOC1, EOC2, EOC3, EOC4 respectively.In EOC1 conditions
Under, it trains to obtain grader using the four classification target training samples of BMP2, BTR70, BRDM2 and T72 in table 1, then to T72
The test samples of 5 kinds of different models classify, the main testing classification device of EOC1 classification experiments is to the classification energy of Morph Target
Power.Under the conditions of EOC2, grader is trained using tetra- classification target training sample of 2S1, BRDM2, T72 and ZSU234 in table 1, is surveyed
Sample is originally the target sample image obtained when depression angle is 30 °, the model A64 that T72 is used.EOC2 classification experiments are mainly surveyed
Try target classification ability of the grader under larger depression angle difference condition.It is different from EOC2, EOC3 only with 2S1, BRDM2 and
ZSU234 tertiary targets, there is no using T72, training sample uses the data in table 1, and test sample is respectively adopted depression angle and is
There are target distortion situations by target sample image under 30 ° and 45 °, wherein BRMD2 and ZSU234.Under the conditions of EOC4, use
Tetra- kinds of targets of BMP2, T72, BTR60 and T62 in table 1 are classified, and wherein BMP2 and T72 are instructed there are Morph Target situation
Using SN9563 and SN132 models when practicing, the model of test sample is respectively SN9566, SNC21, SN812, SNS7, EOC4 master
Want classification capacity of the testing classification device for Morph Target.
Table 3 lists the characteristic for several typical SAR image Target Recognition Algorithms for participating in comparison, the feature including use,
Classifier methods and whether need pose estimation and dictionary learning etc..Table 4 is given in 6 kinds of behaviour such as SOC1, SOC2, EOC1
Target classification result under the conditions of work.In addition to the method for design, other experimental results arise directly from corresponding document.From table 4
As can be seen that in addition to EOC1 situations, inventive algorithm is suitable with the optimal classification precision obtained in each case.In EOC1 feelings
Under condition, due to identical with k nearest neighbor grader in the grader mechanism based on rarefaction representation of use, when target distortion situation compared with
Greatly, when having larger difference with the target in training sample, while influenced be subject to coherent speckle noise, therefore may result in identification essence
Degree declines.The equally examination of EOC4 scenes is for the recognition capability of Morph Target, and the method that the present invention designs is with highest identification
Precision.Therefore, although carrying out feature extraction and dimensionality reduction using method of randomization, dictionary optimization means is combined, pass through fusion
Two kinds of features with complementary characteristic, the sorting algorithm based on rarefaction representation of use remain to preferably progress target classification and sentence
It is fixed, it can obtain excellent classification results in several cases.
The experimental result of consolidated statement 2 and table 4 can be seen that after gray feature and dimensionality reduction random convolution feature is merged, though
So it is declined slightly for the accuracy of identification of SOC2 scenes, but under the conditions of extended operation, since two kinds of features of use have
Complementarity by reasonably selection sort device and design fusion rule, improves the adaptable of design method, achieves higher
Nicety of grading.
The specificity analysis of the typical SAR image Target Recognition Algorithms of table 3
Classification performance comparison (5 Monte Carlo simulation) of the table 4 with typical algorithm under multiple scenes
Claims (3)
1. the SAR image recognition methods based on rarefaction representation and multiple features decision level fusion, it is characterised in that this method is specifically：
Step (1) pre-processes original SAR image, obtains target slice image I；
Step (2), extraction target gray feature vector
Step (3), the convolution kernel for generating multiple and different sizes at random carry out convolutional filtering to target slice image, by square
Multiple dimensioned random convolution feature vector f is extracted in pondization operation2；Using the mapping of sparse accidental projection to the random convolution feature of extraction
Vector carries out dimensionality reduction, obtains the random convolution feature vector of dimensionality reduction
Step (4), the dictionary that two feature vectors extracted from each classification training sample are formed using dictionary learning algorithm
It optimizes respectively, then the dictionary after optimization is merged into the dictionary for target identification；
When step (5), progress test sample, sparse coefficient optimization is carried out based on gray feature and the random convolution feature of dimensionality reduction and is asked
Solution, is converted into target classification probability by reconstructed error and carries out decision level fusion, differentiates so as to fulfill target classification.
2. the SAR image recognition methods according to claim 1 based on rarefaction representation and multiple features decision level fusion, special
Sign is that step (3) is specifically：
For the random convolution feature of extraction, target slice image I is zoomed in and out first, its resolution ratio is made to become 32 × 32, will
To SAR target slice images be denoted as P；Since SAR image has stronger coherent speckle noise, need to be carried using larger convolution kernel
Take random convolution feature；The convolution kernel width set of use is set as { 5,7,9,11,13 }, first first true according to the sampling of certain probability
Determine convolution kernel width, be then based on formula (1) generation two-dimensional convolution core Kernelm, the convolution nuclear volume used is N；
Kernelm- 1 formulas of (i, j)=2 × rand () (1)
Wherein i, j represent the row, column coordinate of two-dimensional convolution core respectively, and m=1,2 ..., N represent the index of two-dimensional convolution core；rand
(), which generates, is located at section [0 1] equally distributed random number；
Since the convolution kernel width of use is inconsistent, for convenience of subsequent processing, using two-dimensional convolution check target slice image into
During row filtering, zero padding operation is carried out to original image so that using different convolution kernels to the characteristic image that is generated after image filtering
Size is consistent, specifically as shown in formula (2)；
Fm=Kernelm* padding (P) formula (2)
Wherein, FmRepresent the two-dimensional convolution feature being calculated, * represents convolution operation, and padding (P) is represented according to convolution kernel
Width is filled target slice image P edges with 0 so that FmIt is consistent with the size of P；
Two-dimensional random convolution feature to make acquisition has certain consistency to target translation, is operated and carried out using equal Fang Chihua
Eigentransformation, as shown in formula (3)；
Wherein i, j=1,2 ..., 32+1-g, g are the peak width that pondization operation carries out；
After N number of convolution kernel is based on to SAR sectioning images by convolution and pondization operation progress feature extraction, the two dimension of generation is rolled up
Product characteristic pattern PmStretch feature vector f in columnm, and by { fm}M=1 ... NIt merges, forms the feature vector of a bigger dimension
f2, dimension is [N × (32+1-g)21]；
Due to feature vector f2Dimension it is larger, when using sparse representation method design grader when, carry out sparse coefficient recovery when
It will take very much, it is general that dimensionality reduction is carried out to the feature of extraction using accidental projection mapping method, it is generated for example with Gaussian Profile
Accidental projection mapping matrix；To make the feature after dimensionality reduction more sparse, sparse accidental projection mapping matrix is generated using formula (4)
Rab；
Wherein, a is through Random Maps matrix RabDimension after dimensionality reduction, b=N × (32+1-r)2, it is characterized the dimension of vector f, ρ
Determine matrix RabSparse degree；As ρ=b/c, ideally matrix RabEvery a line in only c element be non-zero,
Show when c takes the numerical value of very little, the mapping matrix R of generationabIt is sparse；
Former feature vector f2After formula (5) dimensionality reduction, the random convolution feature vector of dimensionality reduction is obtained
3. the SAR image recognition methods according to claim 1 based on rarefaction representation and multiple features decision level fusion, special
Sign is that step (5) is specifically：
Extract gray feature vector respectively from sample to be tested xWith the random convolution feature vector of dimensionality reductionBased on l1Norm point
The convex optimization problem as shown in formula (7) is not solved, obtains corresponding sparse coefficient α1And α2；Notation index is denoted as p；Wherein, ε is
One threshold value, general value are 0.01；
The sparse coefficient α obtained according to solution1And α2, target classification differentiation can be carried out；Since SAR image has strong be concerned with
Spot noise, method of the generally use based on minimal reconstruction error differentiate target classification；For kth class target, defined function：ForIt is coefficient vector, in wherein α at index only corresponding with kth class target
Value remains unchanged, and value is set to zero at remaining index；Define residual error rk(x) it is
It is obtained by formula (9), the n of residual error minimum1, n2Value corresponds to test sample x and is based on gray feature and the random convolution spy of dimensionality reduction respectively
The recognition result of sign；For ease of being merged to the recognition result based on two kinds of features, reconstructed error is obtained according to formula (8)Afterwards, the reconstructed error that sparse coefficient is obtained based on recovery is converted into based on SoftMax thoughts and belongs to each classification target
Identification probability, as shown in formula (10) and formula (11)；
The fusion of two kinds of feature recognition results is carried out using Bayesian Fusion rule, as shown in formula (12)；The classification class of target
By being obtained formula (13) Suo Shi；
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711405236.3A CN108122008B (en) | 2017-12-22 | 2017-12-22 | SAR image recognition method based on sparse representation and multi-feature decision-level fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711405236.3A CN108122008B (en) | 2017-12-22 | 2017-12-22 | SAR image recognition method based on sparse representation and multi-feature decision-level fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108122008A true CN108122008A (en) | 2018-06-05 |
| CN108122008B CN108122008B (en) | 2020-09-08 |
Family
ID=62231135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711405236.3A Active CN108122008B (en) | 2017-12-22 | 2017-12-22 | SAR image recognition method based on sparse representation and multi-feature decision-level fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108122008B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109086704A (en) * | 2018-07-23 | 2018-12-25 | 杭州电子科技大学 | A kind of physical activity recognition methods classified based on rarefaction representation and Softmax |
| CN109598218A (en) * | 2018-11-23 | 2019-04-09 | 南通大学 | A kind of method for quickly identifying of vehicle |
| CN109658380A (en) * | 2018-11-09 | 2019-04-19 | 广西壮族自治区遥感信息测绘院 | Forest road hierarchy detection method based on forest land vector data early period |
| CN109815357A (en) * | 2019-01-28 | 2019-05-28 | 辽宁工程技术大学 | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation |
| CN109919242A (en) * | 2019-03-18 | 2019-06-21 | 长沙理工大学 | A kind of images steganalysis method based on depth characteristic and joint sparse |
| CN110147403A (en) * | 2019-05-23 | 2019-08-20 | 中国农业科学院农业信息研究所 | Agriculture big data fusion method, device, equipment and storage medium |
| CN110378415A (en) * | 2019-07-19 | 2019-10-25 | 浙江理工大学 | A kind of SAR image sorting algorithm |
| CN110751201A (en) * | 2019-10-16 | 2020-02-04 | 电子科技大学 | SAR equipment task failure cause reasoning method based on textural feature transformation |
| CN110837801A (en) * | 2019-11-06 | 2020-02-25 | 中国人民解放军国防科技大学 | SAR image fusion shielding target identification method based on segmentation image sparse representation |
| CN112990225A (en) * | 2021-05-17 | 2021-06-18 | 深圳市维度数据科技股份有限公司 | Image target identification method and device in complex environment |
| CN113093164A (en) * | 2021-03-31 | 2021-07-09 | 西安电子科技大学 | Translation-invariant and noise-robust radar image target identification method |
| CN115223050A (en) * | 2022-04-28 | 2022-10-21 | 湖北工程学院 | Polarized satellite image analysis method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103886337A (en) * | 2014-04-10 | 2014-06-25 | 西安电子科技大学 | Nearest neighbor subspace SAR target identification method based on multiple sparse descriptions |
| CN107103338A (en) * | 2017-05-19 | 2017-08-29 | 杭州电子科技大学 | Merge the SAR target identification methods of convolution feature and the integrated learning machine that transfinites |
-
2017
- 2017-12-22 CN CN201711405236.3A patent/CN108122008B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103886337A (en) * | 2014-04-10 | 2014-06-25 | 西安电子科技大学 | Nearest neighbor subspace SAR target identification method based on multiple sparse descriptions |
| CN107103338A (en) * | 2017-05-19 | 2017-08-29 | 杭州电子科技大学 | Merge the SAR target identification methods of convolution feature and the integrated learning machine that transfinites |
Non-Patent Citations (2)
| Title |
|---|
| SALMAN H. KHAN等: "A Discriminative Representation of Convolutional Features for Indoor Scene Recognition", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| 桑成伟等: "基于可区分性字典学习模型的极化SAR 图像分类", 《信号处理》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109086704A (en) * | 2018-07-23 | 2018-12-25 | 杭州电子科技大学 | A kind of physical activity recognition methods classified based on rarefaction representation and Softmax |
| CN109658380A (en) * | 2018-11-09 | 2019-04-19 | 广西壮族自治区遥感信息测绘院 | Forest road hierarchy detection method based on forest land vector data early period |
| CN109598218A (en) * | 2018-11-23 | 2019-04-09 | 南通大学 | A kind of method for quickly identifying of vehicle |
| CN109598218B (en) * | 2018-11-23 | 2023-04-18 | 南通大学 | Method for quickly identifying vehicle type |
| CN109815357A (en) * | 2019-01-28 | 2019-05-28 | 辽宁工程技术大学 | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation |
| CN109815357B (en) * | 2019-01-28 | 2022-12-06 | 辽宁工程技术大学 | Remote sensing image retrieval method based on nonlinear dimension reduction and sparse representation |
| CN109919242A (en) * | 2019-03-18 | 2019-06-21 | 长沙理工大学 | A kind of images steganalysis method based on depth characteristic and joint sparse |
| CN110147403A (en) * | 2019-05-23 | 2019-08-20 | 中国农业科学院农业信息研究所 | Agriculture big data fusion method, device, equipment and storage medium |
| CN110378415A (en) * | 2019-07-19 | 2019-10-25 | 浙江理工大学 | A kind of SAR image sorting algorithm |
| CN110751201B (en) * | 2019-10-16 | 2022-03-25 | 电子科技大学 | SAR equipment task failure cause reasoning method based on textural feature transformation |
| CN110751201A (en) * | 2019-10-16 | 2020-02-04 | 电子科技大学 | SAR equipment task failure cause reasoning method based on textural feature transformation |
| CN110837801B (en) * | 2019-11-06 | 2022-08-02 | 中国人民解放军国防科技大学 | SAR image fusion shielding target identification method based on segmentation image sparse representation |
| CN110837801A (en) * | 2019-11-06 | 2020-02-25 | 中国人民解放军国防科技大学 | SAR image fusion shielding target identification method based on segmentation image sparse representation |
| CN113093164A (en) * | 2021-03-31 | 2021-07-09 | 西安电子科技大学 | Translation-invariant and noise-robust radar image target identification method |
| CN112990225B (en) * | 2021-05-17 | 2021-08-27 | 深圳市维度数据科技股份有限公司 | Image target identification method and device in complex environment |
| CN112990225A (en) * | 2021-05-17 | 2021-06-18 | 深圳市维度数据科技股份有限公司 | Image target identification method and device in complex environment |
| CN115223050A (en) * | 2022-04-28 | 2022-10-21 | 湖北工程学院 | Polarized satellite image analysis method |
| CN115223050B (en) * | 2022-04-28 | 2023-08-18 | 湖北工程学院 | Polarized satellite image analysis method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108122008B (en) | 2020-09-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108122008A (en) | SAR image recognition methods based on rarefaction representation and multiple features decision level fusion | |
| CN108038476B (en) | A kind of facial expression recognition feature extracting method based on edge detection and SIFT | |
| CN101520894B (en) | Method for extracting significant object based on region significance | |
| CN111242174A (en) | Liver cancer image feature extraction and pathological classification method and device based on imaging omics | |
| Gao et al. | On combining morphological component analysis and concentric morphology model for mammographic mass detection | |
| Xiao et al. | Development of a CNN edge detection model of noised X-ray images for enhanced performance of non-destructive testing | |
| CN108460400A (en) | A kind of hyperspectral image classification method of combination various features information | |
| Nateghi et al. | Automatic detection of mitosis cell in breast cancer histopathology images using genetic algorithm | |
| CN109284781A (en) | Image classification algorithms and system based on manifold learning | |
| CN111639697B (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
| CN107563393A (en) | A kind of extraction of inscriptions on bones or tortoise shells picture Local textural feature and matching process and system | |
| Li et al. | Self-attention random forest for breast cancer image classification | |
| CN106548195A (en) | A kind of object detection method based on modified model HOG ULBP feature operators | |
| Asodekar et al. | Brain tumor classification using shape analysis of MRI images | |
| Huang et al. | Superpixel-based change detection in high resolution sar images using region covariance features | |
| Guo et al. | Is local dominant orientation necessary for the classification of rotation invariant texture? | |
| Dash et al. | Wavelet based features of circular scan lines for mammographic mass classification | |
| Retmin Raj C et al. | Distance based genetic algorithm for feature selection in computer aided diagnosis systems | |
| Wang et al. | Sonar objective detection based on dilated separable densely connected CNNs and quantum-behaved PSO algorithm | |
| Pan et al. | Leukocyte image segmentation using novel saliency detection based on positive feedback of visual perception | |
| Sinthia et al. | An effective two way classification of breast cancer images: A detailed review | |
| CN111832463A (en) | Deep learning-based traffic sign detection method | |
| Lakshmi et al. | Robust algorithm for Telugu word image retrieval and recognition | |
| Spratling | Comprehensive assessment of the performance of deep learning classifiers reveals a surprising lack of robustness | |
| Loka et al. | Classification of normal, benign and malignant tissues using fuzzy texton and support vector machine in mammographic images |
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 |