CN113093164A - Translation-invariant and noise-robust radar image target identification method - Google Patents
Translation-invariant and noise-robust radar image target identification method Download PDFInfo
- Publication number
- CN113093164A CN113093164A CN202110349429.1A CN202110349429A CN113093164A CN 113093164 A CN113093164 A CN 113093164A CN 202110349429 A CN202110349429 A CN 202110349429A CN 113093164 A CN113093164 A CN 113093164A
- Authority
- CN
- China
- Prior art keywords
- frequency domain
- feature description
- domain feature
- sample
- unknown sample
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/02—Systems using reflection of radio waves, e.g. primary radar systems; Analogous systems
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
- G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
- G01S13/9004—SAR image acquisition techniques
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/418—Theoretical aspects
Landscapes
- Engineering & Computer Science (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Physics & Mathematics (AREA)
- Computer Networks & Wireless Communication (AREA)
- General Physics & Mathematics (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a translation invariant and noise robust radar image target identification method. The technical problems that a traditional SAR image target identification method is affected by local micro-disturbance such as target space offset and noise pollution and the identification effect is poor are solved, and the method comprises the following steps: carrying out frequency domain transformation on the selected reference sample; constructing a reference sample frequency domain feature description vector; constructing an over-complete redundant dictionary; processing an unknown sample to be identified; calculating the sparsest representation of the frequency domain feature description vector of the unknown sample; and reconstructing an unknown sample and carrying out class judgment. According to the invention, the frequency domain feature description of a low-frequency component design signal is utilized, an over-complete redundant dictionary is constructed to realize the sparsest representation of the frequency domain feature of an unknown sample, and the technical problems that in the SAR image target recognition field, similar targets have spatial offset and cannot be aligned accurately and the technical problem that image disturbance in an actual scene affects the recognition rate are solved. The method can be used for target classification and identification under complex imaging conditions of actual scenes of the radar.
Description
Technical Field
The invention belongs to the technical field of signal and information processing, mainly relates to SAR image target identification, and particularly relates to a radar image target identification method with constant translation and stable noise. The method can be used for realizing target classification and identification under the complex imaging condition of the actual scene of the radar.
Background
The target identification is an important ring in radar image interpretation, and plays an important role in strategic early warning, air defense guiding, situation perception and the like. The classical target identification method usually needs to meet the requirement of accurate registration of a training sample and a test sample, however, the precondition is too harsh, and an interested region extracted under the complex imaging condition of an actual scene cannot be accurately aligned with a reference sample in the target space position, so that the practical performance is greatly reduced.
John Wright et al in 2009 proposed a face recognition method based on sparse representation in the text "Robust face recognition via sparse representation" (IEEE Transactions on Pattern. analysis and Machine significance, vol.31, No.2, pp.210-227, Feb.2009), successfully solving the problem that a signal sparse representation classification method is difficult to obtain through a complete redundant dictionary; thereafter, Thiagarajan et al applied the method to target classification recognition of SAR images, "Sparse representation for automatic target classification in SAR images" (20104th International Symposium on Communications, Control and Signal Processing), explaining the physical meaning of Sparse representation recognition from the perspective of manifold. However, the classical method performs dimensionality reduction on the image in a space domain, constructs sparse representation of an unknown image, and is applied on the premise that an ideal data set, namely a test sample and a training sample, performs strict registration in the space domain. Although the classical method can obtain a better experimental result on an ideal data set, the precondition of the classical method cannot be directly applied to an actual scene, and an unknown region of interest transmitted to a classification stage after detection and identification is located in the center of an image but cannot be strictly registered with a training image, namely, the spatial offset problem exists. In addition, the classical method has poor classification effect due to factors such as target structure change, sensor measurement error and noise pollution in actual scenes, and cannot effectively cope with target identification tasks under extended working conditions.
In the SAR image target recognition, the assumption of ideal conditions, including the image center, cannot be strictly registered with a training image, the problem of spatial offset is generated, and the classification effect in the target recognition process is poor due to system errors and noise pollution.
Disclosure of Invention
Aiming at the problems and the defects of the existing method, the invention provides the SAR image frequency domain sparse representation target identification method with more stable judgment, which can effectively solve the problems of target space offset and local micro-disturbance such as noise pollution.
The invention relates to a radar image target identification method with constant translation and stable noise, which is characterized by comprising the following steps:
(1) selecting a reference sample: selecting n actually-measured SAR ground target images with category marks as reference samples, wherein each image has one and only one category mark;
(2) starting off-line calculation, and performing frequency domain transformation on the selected reference sample: respectively carrying out discrete Fourier transform on all selected reference samples, and carrying out subsequent analysis based on the amplitude spectrum characteristics of the reference samples after frequency domain transform;
(3) constructing a reference sample frequency domain feature description vector: acquiring the low-frequency component of a reference sample signal for the amplitude spectrum of each reference sample, and performing energy normalization on the extracted low-frequency component; assume that only m sets of low frequency components are extracted per reference sample signal
The process of (2) frequency domain transformation and (3) frequency domain feature description vector construction is marked as mapping
M represents the number of rows of the space domain signals, and N represents the number of columns of the space domain signals;
(4) constructing an overcomplete redundant dictionary, and finishing off-line calculation: constructing by using m-dimensional frequency domain feature description vectors corresponding to the n reference samples respectivelyStacking the frequency domain feature description vectors of all n reference samples according to the category sequence on the other dimension to obtain an over-complete redundant dictionary
(5) Starting on-line prediction, processing unknown samples to be identified: for any unknown sample y to be detected, firstly carrying out discrete Fourier transform, then obtaining low-frequency components and carrying out energy normalization to generate a frequency domain feature description vector of the unknown sample y to be detected
The specific method is consistent with the steps (2) and (3);
(6) with overcomplete redundant dictionaries
And (3) calculating to obtain the most sparse representation of the frequency domain feature description vector of the unknown sample: overcomplete redundant dictionary constructed using offline computation
Frequency domain feature description vector of unknown sample y to be measured
Performing linear representation, constructing an objective function according to a method for solving the most sparse representation coefficient, and solving by using an optimization algorithm to obtain a frequency domain feature description vector
Most sparsely represented coefficient of
(7) Reconstructing an unknown sample to be detected: reconstructing the frequency domain feature description vector of the unknown sample to be detected by utilizing the sparsest representation result,
each representation coefficient item in the vector corresponds to one category, a plurality of representation coefficient items are not 0, and if k items are not 0, the k items are respectively taken
Term of the expression coefficient other than 0
And are of the same kind as
Middle corresponding sub-dictionary
Frequency domain feature vector of unknown sample to be measured
Reconstructing to obtain k reconstructed results of the frequency domain characteristics of the unknown sample to be detected
(8) Judging the category of the unknown sample to be detected to obtain an online prediction result: according to the minimum reconstruction error criterion, category attribution judgment is carried out, and k reconstruction results P are respectively calculatediAnd selecting the category corresponding to the reconstruction result with the minimum error as the category prediction result of the sample to be detected.
The method solves the technical problems that in the field of SAR image target identification, space deviation exists between targets of the same category and accurate alignment cannot be realized, and the technical problem that local micro-disturbance such as noise pollution in an actual scene influences identification accuracy.
The method converts the space domain signals into frequency domains, designs frequency domain feature description vectors of the signals by using low-frequency components, constructs an over-complete redundant dictionary, realizes the most sparse representation of the frequency domain feature description vectors of unknown samples, and carries out category judgment according to reconstruction errors.
Compared with the prior art, the invention has the remarkable advantages of two aspects:
effectively deal with the space deviation of target, improve the recognition accuracy: in the prior art, sparse representation of SAR image signals is realized in an airspace, the sparse representation of signals constructed in a frequency domain is selected, the problem of space offset of a target can be effectively solved, the high dependence of a classical method on accurate alignment of a reference sample and an unknown sample is broken, and the translation invariance of the target is realized;
filtering out noise, a more robust decision is achieved: the method and the device have the advantages that the noise pollution of the SAR image is not processed in the prior art, the high-frequency component containing noise is filtered through a frequency domain mask technology, the local micro-disturbance problems such as the noise pollution can be effectively solved, and more stable judgment is realized.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a frequency domain feature mask diagram;
FIG. 3 is a schematic diagram of a spatial shift of a target;
FIG. 4 is a graph of the recognition rate of different methods when spatial offset occurs;
FIG. 5 is a schematic diagram of random noise pollution;
FIG. 6 is a graph of recognition rates for different methods when contaminated with noise.
Detailed Description
The present invention will be described in detail with reference to the accompanying drawings.
Example 1
In a traditional SAR image target recognition algorithm, feature extraction usually needs to estimate an attitude angle and a plurality of fussy preprocessing links, such as speckle suppression, image segmentation, mathematical morphology processing and the like. The SAR image target classification method based on sparse representation can avoid the preprocessing link.
The invention relates to a radar image target identification method with constant translation and stable noise, which is shown in figure 1 and comprises the following steps:
(1) selecting a reference sample: selecting n actually-measured SAR ground target images with class marks as reference samples, wherein each image has one and only one class mark.
(2) Starting off-line calculation, and performing frequency domain transformation on the selected reference sample: and respectively carrying out discrete Fourier transform on all the selected reference samples, projecting and transforming all the reference samples to a frequency domain, and then carrying out subsequent analysis based on a harmonic analysis theory according to the amplitude spectrum characteristics of the reference samples. The invention uses a transform domain processing method to realize the steady target identification in a complex scene.
(3) Constructing a reference sample frequency domain feature description vector: and (3) acquiring the low-frequency component of the reference sample signal according to the discrete Fourier transform result, namely the reference sample amplitude spectrum obtained based on the frequency domain transform in the step (2), and performing energy normalization on the extracted low-frequency component. Assume that only m sets of low frequency components are extracted per reference sample signal
The frequency domain transformation in the step (2) and the frequency domain feature description vector construction process in the step (3) are marked as mapping
M denotes the number of rows of spatial signals, and N denotes the number of columns of spatial signals. And performing frequency domain transformation on each reference sample, extracting low-frequency components, and performing energy normalization to obtain an m-dimensional vector, namely the frequency domain feature description vector of the sample.
(4) Constructing an overcomplete redundant dictionary, and finishing off-line calculation: constructing an over-complete redundant dictionary by using the m-dimensional frequency domain feature description vectors of all the reference samples, and stacking the frequency domain feature description vectors of all the n reference samples on the other dimension according to the category sequence to obtain the over-complete redundant dictionary
(5) Starting on-line prediction, processing unknown samples to be identified: for any unknown sample y to be detected, firstly carrying out discrete Fourier transform, then obtaining low-frequency components and carrying out energy normalization to generate a frequency domain feature description vector of the unknown sample y to be detected
The specific method is consistent with the steps (2) and (3).
(6) With overcomplete redundant dictionaries
The calculation yields the sparsest representation of the unknown sample: overcomplete redundant dictionary constructed using offline computation
Frequency domain feature description vector of unknown sample y to be measured
Performing linear representation, constructing an objective function according to a method for solving the sparsest representation coefficient, and solving the frequency domain feature description vector of the unknown sample by using an optimization algorithm
Redundant dictionary over-complete
Most sparsely represented coefficient of
(7) Reconstructing an unknown sample to be detected: utilizing the result of the step (6) to carry out frequency domain characteristic vector on the unknown sample to be detected
Is reconstructed due to
There may be multiple values in the vector different from 0, assuming that k values are different from 0, each value corresponds to a category, and for each category, the values are respectively taken
The expression coefficient term related thereto
And
middle corresponding sub-dictionary
Reconstructing to obtain k reconstructed results of the unknown sample to be detected
And finishing the reconstruction of the frequency domain characteristic vector of the unknown sample to be detected.
(8) Judging the category of the unknown sample to be detected to obtain an online prediction result: according to the minimum reconstruction error criterion, category attribution judgment is carried out, and k reconstruction results P are respectively calculatediThe method comprises the steps of obtaining a category judgment result of a sample to be detected according to the minimum reconstruction error, completing the identification of the SAR image frequency domain sparse representation target, and effectively solving the problems of target space offset and noise disturbance.
The invention mainly develops and researches aiming at the technical problem of poor classification effect caused by space offset and system error noise pollution in the existing SAR image target identification task, and the basic thought of the invention is as follows: firstly, a reference sample is projected to a frequency domain by means of a harmonic analysis theory, then an over-complete dictionary is constructed by using a small amount of low-frequency components, sparse representation of the low-frequency components of the sample to be detected is realized by using dictionary atoms, the most sparse representation is obtained by constraining a feasible set of the representation coefficients, and finally the most attributive type judgment is carried out according to the reconstruction error of each type of sparse representation coefficients, so that the disturbance such as noise, space offset and the like can be effectively dealt with.
In the invention, the image signals are subjected to two-dimensional discrete Fourier transform, and the frequency domain energy is mainly concentrated in a few frequency bands, so that the frequency bands with concentrated energy have stronger discrimination capability than other redundant frequency bands and contain main discrimination information of the signals. The amplitude spectrum energy in the SAR image frequency spectrum after the shifting of the ftshift frequency spectrum is concentrated in the frequency spectrum wave band of the central area, and the amplitude spectrum has the characteristic of unchanged space translation and rotation, so that the frequency spectrum of the central area is only taken out, the characteristic vectors are rearranged to generate the characteristic vectors, the dimension of the characteristic vectors is reduced, and the identification capability is improved.
Example 2
The method for identifying the radar image target with unchanged translation and stable noise is the same as the embodiment 1, and the method for acquiring the low-frequency component in the steps (3) and (5) comprises the following steps:
assuming that m groups of low-frequency components of a sample signal are extracted as frequency domain feature description, extracting the low-frequency components of the sample signal by using a frequency domain mask method on a frequency spectrum image obtained by two-dimensional discrete Fourier transform of the sample signal, and filtering out the high-frequency components: and performing fftshift shift on the original frequency spectrum, selecting a wave band of the center part of the shifted frequency spectrum, wherein the number m of the selected low-frequency component groups depends on the area size of the frequency domain mask. Specifically, as shown in fig. 2, fig. 2 is a frequency domain feature mask diagram, and the present invention uses a square mask to only reserve the high energy spectrum in the central area on the shifted two-dimensional spectrum image and remove the interference of the peripheral high frequency noise. The method and the device have the advantages that the noise pollution of the SAR image is not processed in the prior art, the high-frequency component containing noise is filtered through a frequency domain mask technology, the local micro-disturbance problems such as the noise pollution can be effectively solved, and more stable judgment is realized.
Example 3
A method for identifying a radar image target with invariant translation and robust noise as in embodiment 1-2, wherein the method for constructing an overcomplete redundant dictionary described in step (4) constructs an overcomplete redundant dictionary by using m-dimensional frequency domain feature description vectors of n reference samples, and specifically comprises the following steps: stacking m-dimensional frequency domain feature description vectors of all n reference samples on the other dimension to obtain a dictionary
Reference samples of the same class form a sub-dictionary, wherein the sub-dictionary formed by the reference samples of the ith class is
niIs the number of reference samples belonging to the i-th class, K is the total number of classes; to satisfy the condition of overcomplete redundancy, the total number of reference samples should be larger than the dimension n of the frequency domain feature description vector>m。
Example 4
A radar image target identification method with translation invariance and stable noise is the same as that of the embodiments 1-3, and the step (6) is described by means of an over-complete redundant dictionary
Calculating to obtain the sparsest expression of the unknown sample, which specifically comprises the following steps:
linear representation of unknown samples: overcomplete redundant dictionary constructed using offline computation
Frequency domain feature description vector of unknown sample y to be measured
The linear representation is a representation of the linear,
wherein α ═ α1,α2,…,αK]∈RnTo represent the coefficients.
Constructing a coefficient optimization objective function of the unknown sample: according to the method for solving the sparsest expression coefficient, a coefficient optimization objective function of the frequency domain feature description vector of the unknown sample to be detected is constructed,
solving the sparsest representation coefficients: by using
Frequency domain characteristic tracing obtained by solving norm optimization algorithmThe vector
Most sparsely represented coefficient of
And gradually updating the representation coefficients by adopting an LARS minimum angle regression algorithm.
In the prior art, sparse representation of SAR image signals is realized in a space domain, the sparse representation of signals constructed in a frequency domain is selected, the problem of space offset of a target can be effectively solved, the high dependence of a classical method on accurate alignment of a reference sample and an unknown sample is broken, and the translation invariance of the target is realized.
Example 5
A method for identifying a radar image target with constant translation and robust noise, which is the same as that in embodiments 1 to 4, and the unknown sample class decision in step (8), specifically includes:
calculating reconstruction error of each reconstruction result obtained in the step (7) respectively for use
Measuring the reconstruction result P of the unknown sample to be measured by the square of the normiAnd true frequency domain feature description vectors
The error between the two-dimensional data of the two-dimensional data,
and selecting the corresponding category i when the reconstruction error is minimum as a final prediction result of the unknown sample to be detected.
An example of a concise expression is given below
Example 6
A method for identifying radar image targets with unchanged translation and stable noise, which is the same as the embodiments 1-5,
referring to fig. 1, the specific steps of the present invention are summarized as follows:
step 1: transforming all the reference sample projections to a frequency domain according to a harmonic analysis theory;
step 2: extracting frequency domain low-frequency components of all reference samples, normalizing energy, and constructing a frequency domain feature description vector;
and step 3: constructing an over-complete redundant dictionary by using the reference sample frequency domain feature description vector;
and 4, step 4: performing linear representation on the frequency domain feature description vector of the unknown sample by using the constructed dictionary;
and 5: constraining a feasible set of the linear representation model representation coefficients and solving the sparsest representation;
step 6: respectively reconstructing unknown samples by using the representation coefficient items of each category;
and 7: and judging the category attribution according to the minimum reconstruction error criterion.
The invention converts the space domain signal into the frequency domain, designs the feature description vector by using the low-frequency component, constructs the over-complete dictionary, realizes the most sparse representation of the unknown sample frequency domain description, and judges according to the reconstruction error.
A more detailed example is given below to further illustrate the invention
Example 7
A method for identifying radar image targets with unchanged translation and stable noise, which is the same as the embodiments 1-6,
fig. 1 depicts a brief flow of the present invention, with specific target identification steps as follows:
assume that the i-th class of reference samples has niIs marked as
All K classes are common
Reference samples, each denoted X ═ X1,X2,…,Xk]The pixel size of the reference sample is M × N, and for a given unknown sample y to be detected, the main task of target identification is to accurately infer the type of the unknown sample to be detected according to the given reference sample and the type thereof.
Step 1: all n sets of reference samples are separately discrete fourier transformed. For a reference sample f of pixel size M N, its discrete Fourier transform
Step 2: and constructing a frequency domain low-frequency description according to the discrete Fourier transform result. Calculating the amplitude spectrum of the image signal, extracting the low frequency component of the signal by means of a frequency domain mask as shown in fig. 2, energy normalizing the extracted low frequency component, the calculation process being marked as mapping
Assuming that m sets of low frequency components of the signal are extracted as frequency domain feature descriptions,
and step 3: combining the frequency domain features of all reference samples to construct an over-complete redundant dictionary,
wherein
Is a sub-dictionary of the i-th class reference sample construction. To satisfy the condition of overcomplete redundancy, the number of reference samples should be larger than the dimension n of the frequency domain feature>m；
And 4, step 4: for any unknown sample y, carrying out frequency domain change in the same way, obtaining low-frequency components and energy normalization, and obtaining a frequency domain feature description vector of the unknown sample y
Overcomplete redundant dictionary with construction
The frequency domain feature description vector of the unknown sample is linearly represented,
wherein α ═ α1,α2,…,αK]∈RnIs a representative weight coefficient;
and 5: according to the idea of solving the sparsest expression coefficient, an objective function is constructed,
by using
Norm optimization algorithm for solving sparsest representation
Step 6: respectively using k sub-dictionaries and corresponding optimal representation coefficients
The reconstruction is performed on the unknown sample,
and 7: calculating the reconstruction error of each category, making a decision according to the minimum reconstruction error,
the invention relates to a target identification method in radar image interpretation, which can effectively cope with the adverse effect of local disturbance such as random noise, space translation and the like on target identification, and the target identification is more accurate.
The effect of the present invention will be further described below by simulation experiments in combination with data.
Example 8
A method for identifying radar image targets with unchanged translation and stable noise, which is the same as the embodiments 1-7,
the experimental conditions are as follows:
the invention utilizes MSTAR SAR actual measurement data to carry out verification, the radar sensor parameters of the collected data are shown in the following table, an experiment operation system is an Intel (R) core (TM) i7-8565 CPU @1.80GHz and 64-bit Windows10 operation system, and simulation software adopts MATLAB (R2016 b). Table 1 lists the SAR image imaging parameters used for the experiments.
Table 1 experimental SAR image imaging parameters
| Center frequency | 9.6GHz |
| Bandwidth of signal | 0.591GHz |
| Mode of operation | Strip imaging |
| Polarization mode | HH |
| Multiplicative noise | -10dB |
| Additive noise | -32～34dB |
| Dynamic range | 64dB |
| Azimuth beam width | 8.8 degree |
| Tilt angle beamwidth | 6.8 degree |
| Resolution ratio | 0.3X 0.3 m |
| Pixel pitch | 0.2X 0.2 m |
Table 2 lists the reference algorithms in the experiments for comparison
TABLE 2 other algorithms involved in experimental comparisons
| Abbreviations | Description of the method |
| SVM | Support vector machine classification using spatial domain features as input |
| SVMFT | Support vector machine classification using frequency domain features as input |
| MINACE | Minimum noise and correlation energy filter classification |
| OTSDF | Optimally compromised synthetic discriminant function classification |
| SRC | Sparse representation classification with spatial domain features as input |
| SRCHa | Sparse representation classification with Haar wavelets as input |
| SRCGb | Sparse representation classification with Gabor filtering as input |
| SRC of the inventionFT | Sparse representation classification with frequency domain features as input |
Table 3 shows the training data and the test data in the experiment. SAR image data of four targets, namely BMP2, BTR70, T72 and BRDM2, in the MSTAR data set are selected for experiment, wherein data acquired with a radar pitch angle of 17 degrees are used as reference samples for training, and data acquired with a radar pitch angle of 15 degrees are used as unknown samples for testing.
TABLE 3 training data and test data in the experiment
The experimental contents are as follows:
the method proposed by the invention is verified by constructing a spatial migration experiment of the target, and the advantages of the invention are illustrated by comparing with the classical method.
All the reference samples are kept unchanged, the unknown samples are manually subjected to spatial offset along the horizontal direction (distance direction) and the vertical direction (azimuth direction) respectively, the problem of probability registration between the reference samples of the actual scene and the observed target is simulated, and the method is shown in figure 3, and figure 3 is a schematic diagram of the spatial offset of the target. From left to right in the figure are-3 pixels, -1 pixel, 0 pixel, 1 pixel, 3 pixels, 5 pixels, 10 pixels shifted images, respectively, where 0 pixel shift is the original registered image. In addition, the offset in the lower right-hand direction is defined as a negative offset, and the offset in the upper left-hand direction is defined as a positive offset. The white box in the figure is the center of the object of the original registered image. And respectively testing the recognition accuracy of the various algorithms under the condition of different degrees of spatial offset.
Analysis of Experimental results
The experimental result is shown in fig. 4, and fig. 4 is a graph of the recognition rate of different methods when spatial offset occurs; the horizontal coordinate represents the degree of spatial offset of the unknown sample, the unit is a pixel value, and the vertical coordinate is the accuracy of model identification. The curve with dots in the figure represents the MINACE method, the curve with upper triangular dots represents the OTSDF method, the curve with lower triangular dots represents the SRC method, the curve with five-pointed points represents the SVM method, and the curve with star-shaped dots represents the SRC methodHaMethod, curve with diamond points representing SRCGbMethod, curve with hexagonal points representing SVMFTMethod, the last curve with multiplication points represents the SRC of the inventionFTA method.
It can be seen that the eight methods involved in the comparison performed differently by gradually increasing the unknown sample from-3 pixel offset to 10 pixel offset. Two space domain feature classification methods SVM and SRC and two SRC method variants adopting different input featuresHa、SRCGbThe identification accuracy rate is greatly reduced under the condition that the target has large spatial offset, and the performance is poor. Frequency domain feature classification method SRCFTAnd the two related filters MINAE and OTSDF methods have good effects, but the accuracy rate still has a remarkable descending trend under the condition that the target has serious spatial deviation. Method SRC proposed by the inventionFTThe effect is optimal, the accuracy rates are optimal values under different pixel offset conditions, the identification performance is not reduced along with spatial offset, and the accuracy rates are kept stable.
Example 9
The method for identifying the radar image target with unchanged translation and stable noise is the same as the embodiments 1-7, and the experimental conditions are the same as the embodiment 8.
The experimental contents are as follows:
the method provided by the invention is verified by respectively constructing random noise pollution experiments of the target, and the advantages of the method are explained by comparing with the classical method.
All reference samples are kept unchanged, noise with a certain level is randomly added to unknown samples, the noise pollution positions are randomly selected, fig. 5 is a schematic diagram of adding random noise with different levels, and the diagram shows the situation that 0%, 1%, 5%, 10%, 15%, 20% and 30% of pixels in the SAR image are polluted by the random noise from left to right. And respectively testing the recognition accuracy of the various algorithms under the conditions of different degrees of noise pollution.
Analysis of Experimental results
The experimental result is shown in fig. 6, fig. 6 is a graph of the recognition rate of different methods when the unknown sample is polluted by noise, the abscissa is the degree of the unknown sample polluted by noise and represents the percentage of pixels in the unknown sample image polluted by noise, the ordinate is the accuracy rate of model recognition, and the corresponding relationship between the curve and the method in the graph is the same as that in example 8.
It can be seen that, as the noise pollution is increased, all the methods involved in the comparison have performance degradation of different degrees, and when the noise pollution reaches 30%, the recognition rate of the comparison reference method is generally reduced to below 70%, wherein the SVM method and the SRC method are adoptedHaThe accuracy of the method is greatly reduced, the accuracy is lower than 50 percent, and the effect is the worst; SVMFTMINCE, SRCGbThe method has general effect; the OTSDF method has good effect, but the accuracy rate is slightly lower than 70%. In contrast, the method SRC proposed by the present inventionFTThe effect is best, the accuracy rate of about 90% can be still kept at the moment, and the reduction range of the identification accuracy rate is small. The experimental results show that the method SRC provided by the inventionFTAnd the judgment is more reliable and more robust.
In conclusion, the invention provides a radar image target identification method with unchanged translation and stable noise. The technical problems that a traditional SAR image target identification method is affected by local micro-disturbance such as target space offset and noise pollution and the identification effect is poor are solved, and the method comprises the following steps: selecting a reference sample; performing frequency domain transformation on the reference sample; constructing a frequency domain feature description vector of a reference sample; constructing an over-complete redundant dictionary; processing an unknown sample to be identified; calculating the sparsest representation of the frequency domain feature description vector of the unknown sample; reconstructing an unknown sample; and carrying out class judgment on the unknown sample. According to the method, a space domain signal is converted into a frequency domain, a frequency domain feature description vector of the signal is designed by using a low-frequency component, an over-complete redundant dictionary is constructed, the most sparse representation of the frequency domain feature description vector of an unknown sample is realized, category judgment is carried out according to a reconstruction error, and the technical problems that in the field of SAR image target identification, space deviation exists among targets of the same category and accurate alignment cannot be achieved, and the technical problem that noise pollution and other local image micro-disturbance in an actual scene influence identification accuracy are solved. The method can be used for target classification and identification under complex imaging conditions of actual scenes of the radar.
Claims (5)
1. A radar image target identification method with unchanged translation and stable noise is characterized by comprising the following steps:
(1) selecting a reference sample: selecting n actually-measured SAR ground target images with category marks as reference samples, wherein each image has one and only one category mark;
(2) starting off-line calculation, and performing frequency domain transformation on the selected reference sample: respectively performing discrete Fourier transform on all selected reference samples, projecting and transforming all the reference samples to a frequency domain, wherein the reference samples after frequency domain transformation have amplitude spectrum characteristics and phase spectrum characteristics, and performing subsequent analysis according to a harmonic analysis theory based on the amplitude spectrum characteristics;
(3) constructing a reference sample frequency domain feature description vector: according to the discrete Fourier transform result, namely based on the amplitude frequency spectrum of each reference sample obtained after the frequency domain transformation in the step (2), acquiring the low-frequency component of each reference sample signal, and dividing the extracted low frequencyCarrying out energy normalization on the quantity; assume that only m sets of low frequency components are extracted per reference sample signal
The process of (2) frequency domain transformation and (3) frequency domain feature description vector construction is marked as mapping
M represents the number of rows of the space domain signals, and N represents the number of columns of the space domain signals; carrying out frequency domain transformation on each reference sample, extracting low-frequency components and carrying out energy normalization to obtain an m-dimensional vector, namely the frequency domain feature description vector of the sample;
(4) constructing an overcomplete redundant dictionary: constructing an over-complete redundant dictionary by using the frequency domain feature description vectors of n reference samples, and stacking the frequency domain feature description vectors of all the n reference samples on the other dimension according to the category sequence to obtain the over-complete redundant dictionary
Finishing off-line calculation;
(5) starting on-line prediction, processing unknown samples to be identified: for any unknown sample y to be detected, firstly carrying out discrete Fourier transform, then obtaining low-frequency components and carrying out energy normalization to generate a frequency domain feature description vector of the unknown sample y to be detected
The specific method is consistent with the steps (2) and (3);
(6) with overcomplete redundant dictionaries
And (3) calculating to obtain the most sparse representation of the frequency domain feature description vector of the unknown sample: overcomplete redundant dictionary constructed using offline computation
Frequency domain feature description vector of unknown sample y to be measured
Performing linear representation, constructing an objective function according to a method for solving the most sparse representation coefficient, and solving by using an optimization algorithm to obtain a frequency domain feature description vector
Most sparsely represented coefficient of
(7) Reconstructing an unknown sample to be detected: frequency domain feature description vector of unknown sample to be detected by using the sparsest representation result obtained in the step (6)
The reconstruction is carried out and the reconstruction is carried out,
each representation coefficient item in the vector corresponds to one category, a plurality of representation coefficient items are not 0, and if k items are not 0, the k items are respectively taken
Term of the expression coefficient other than 0
And are of the same kind as
Middle corresponding sub-dictionary
Frequency domain feature description vector of unknown sample to be measured
Reconstructing to obtain k reconstructed results of the frequency domain feature description vector of the unknown sample to be detected
(8) Judging the category of the unknown sample to be detected to obtain an online prediction result: and judging category attribution according to a minimum reconstruction error criterion, respectively calculating errors of k reconstruction results, selecting the reconstruction result with the minimum reconstruction error, and taking the corresponding category as a prediction result of the sample to be detected to finish the identification of the SAR image frequency domain sparse representation target.
2. The method for radar image target identification with translation invariance and noise robustness according to claim 1, wherein the obtaining of the low frequency component in the steps (3) and (5) specifically includes:
assuming that m groups of low-frequency components of a sample signal are extracted as frequency domain feature description vectors, extracting the low-frequency components of the sample signal on a frequency spectrum image obtained by two-dimensional discrete Fourier transform of the sample signal by means of a frequency domain mask method, filtering out the high-frequency components, performing fftshift shifting on an original frequency spectrum, selecting a wave band of a center region of the shifted frequency spectrum, wherein the number m of the selected low-frequency component groups depends on the area size of a frequency domain mask.
3. The method for radar image target recognition with translation invariance and noise robustness as claimed in claim 1, wherein the constructing of the overcomplete redundant dictionary in step (4) utilizes m-dimensional frequency domain feature description vectors corresponding to n reference samples to construct the overcomplete redundant dictionary, which specifically includes:
stacking the frequency domain feature description vectors of all n reference samples according to the category order on the other dimension to obtainTo dictionary
Wherein
Is a sub-dictionary constructed from the totality of the i-th class reference samples, niIs the number of reference samples belonging to the i-th class, K is the total number of classes; in order to satisfy the condition of overcomplete redundancy, the total number of reference samples should be larger than the dimension n > m of the frequency-domain feature description vector.
4. The method for identifying a radar image target with translation invariance and noise robustness as claimed in claim 1, wherein the step (6) is performed by means of an overcomplete redundant dictionary
Calculating to obtain the most sparse representation of the frequency domain feature description vector of the unknown sample, specifically comprising:
linear representation of unknown samples: overcomplete redundant dictionary constructed using offline computation
Frequency domain feature description vector for unknown sample y
The linear representation is carried out in such a way that,
wherein α ═ α1,α2,…,αK]∈RnIs to represent a coefficient;
constructing a coefficient optimization objective function of the unknown sample: according to the method for solving the sparsest expression coefficient, a coefficient optimization objective function of the frequency domain feature description vector of the unknown sample to be detected is constructed,
solving the sparsest representation coefficients: by means of1Solving by norm optimization algorithm to obtain frequency domain feature description vector
In a dictionary
Most sparsely represented coefficient of
Specifically, the representation coefficients are gradually updated by adopting an LARS minimum angle regression algorithm.
5. The method for identifying a radar image target with constant translation and robust noise according to claim 1, wherein the step (8) of determining the unknown sample class specifically comprises:
respectively calculating the reconstruction error of each reconstruction result obtained in the step (7) by using l2Measuring the reconstruction result P of the unknown sample to be measured by the square of the normiAnd true frequency domain feature description vectors
Inter error according to the minimum reconstruction error formula
And selecting the corresponding category i when the reconstruction error is minimum as a final prediction result of the unknown sample to be detected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110349429.1A CN113093164B (en) | 2021-03-31 | 2021-03-31 | Translation-invariant and noise-robust radar image target identification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110349429.1A CN113093164B (en) | 2021-03-31 | 2021-03-31 | Translation-invariant and noise-robust radar image target identification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113093164A true CN113093164A (en) | 2021-07-09 |
| CN113093164B CN113093164B (en) | 2023-02-10 |
Family
ID=76672139
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110349429.1A Active CN113093164B (en) | 2021-03-31 | 2021-03-31 | Translation-invariant and noise-robust radar image target identification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113093164B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113887278A (en) * | 2021-08-23 | 2022-01-04 | 西安电子科技大学 | SAR image target identification method based on nonlinear manifold modeling |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824088A (en) * | 2014-01-23 | 2014-05-28 | 西安电子科技大学 | SAR target variant recognition method based on multi-information joint dynamic sparse representation |
| 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 |
| CN103984966A (en) * | 2014-05-29 | 2014-08-13 | 西安电子科技大学 | SAR image target recognition method based on sparse representation |
| CN104008373A (en) * | 2014-06-03 | 2014-08-27 | 西安电子科技大学 | SAR target recognition method based on multi-information dictionary learning |
| CN106096505A (en) * | 2016-05-28 | 2016-11-09 | 重庆大学 | The SAR target identification method of expression is worked in coordination with based on Analysis On Multi-scale Features |
| CN108122008A (en) * | 2017-12-22 | 2018-06-05 | 杭州电子科技大学 | SAR image recognition methods based on rarefaction representation and multiple features decision level fusion |
| WO2019006835A1 (en) * | 2017-07-06 | 2019-01-10 | 清华大学深圳研究生院 | Target recognition method based on compressed sensing |
| CN112001257A (en) * | 2020-07-27 | 2020-11-27 | 南京信息职业技术学院 | SAR image target recognition method and device based on sparse representation and cascade dictionary |
-
2021
- 2021-03-31 CN CN202110349429.1A patent/CN113093164B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824088A (en) * | 2014-01-23 | 2014-05-28 | 西安电子科技大学 | SAR target variant recognition method based on multi-information joint dynamic sparse representation |
| 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 |
| CN103984966A (en) * | 2014-05-29 | 2014-08-13 | 西安电子科技大学 | SAR image target recognition method based on sparse representation |
| CN104008373A (en) * | 2014-06-03 | 2014-08-27 | 西安电子科技大学 | SAR target recognition method based on multi-information dictionary learning |
| CN106096505A (en) * | 2016-05-28 | 2016-11-09 | 重庆大学 | The SAR target identification method of expression is worked in coordination with based on Analysis On Multi-scale Features |
| WO2019006835A1 (en) * | 2017-07-06 | 2019-01-10 | 清华大学深圳研究生院 | Target recognition method based on compressed sensing |
| CN108122008A (en) * | 2017-12-22 | 2018-06-05 | 杭州电子科技大学 | SAR image recognition methods based on rarefaction representation and multiple features decision level fusion |
| CN112001257A (en) * | 2020-07-27 | 2020-11-27 | 南京信息职业技术学院 | SAR image target recognition method and device based on sparse representation and cascade dictionary |
Non-Patent Citations (6)
| Title |
|---|
| GANGGANG DONG: "TARGET RECOGNITION IN SAR IMAGE VIA SPARSE REPRESENTATION IN TRANSFORMED DOMAIN", 《IGARSS 2019》 * |
| GANGGANG DONG: "Target recognition in SAR images via sparse representation in the frequency domain", 《PATTERN RECOGNITION》 * |
| 刘宏伟等: "雷达高分辨距离像目标识别研究进展", 《电子与信息学报》 * |
| 王燕霞等: "基于特征参数稀疏表示的SAR图像目标识别", 《重庆邮电大学学报(自然科学版)》 * |
| 胡允侃: "基于稀疏表示和字典学习的雷达目标识别方法研究", 《中国优秀硕士学位论文 信息科技辑》 * |
| 雷磊等: "基于多尺度稀疏字典的SAR图像目标识别方法", 《火力与指挥控制》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113887278A (en) * | 2021-08-23 | 2022-01-04 | 西安电子科技大学 | SAR image target identification method based on nonlinear manifold modeling |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113093164B (en) | 2023-02-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107301661B (en) | High-resolution remote sensing image registration method based on edge point features | |
| CN108389188B (en) | Sparse hyperspectral abnormal target detection method | |
| CN108122008B (en) | SAR image recognition method based on sparse representation and multi-feature decision-level fusion | |
| CN104299232B (en) | SAR image segmentation method based on self-adaptive window directionlet domain and improved FCM | |
| CN110458192B (en) | Hyperspectral remote sensing image classification method and system based on visual saliency | |
| CN110135438B (en) | Improved SURF algorithm based on gradient amplitude precomputation | |
| CN110516525B (en) | SAR image target recognition method based on GAN and SVM | |
| CN110837768A (en) | Rare animal protection oriented online detection and identification method | |
| CN106991661B (en) | Non-local mean denoising method fusing KL (karhunen-Loeve) transformation and grey correlation degree | |
| CN104834931A (en) | Improved SIFT algorithm based on wavelet transformation | |
| CN109712149B (en) | Image segmentation method based on wavelet energy and fuzzy C-means | |
| CN107895139A (en) | A kind of SAR image target recognition method based on multi-feature fusion | |
| CN109117880A (en) | Tile image sorting algorithm based on WAVELET PACKET DECOMPOSITION selection coefficient weighting reconstruct | |
| CN110222661B (en) | Feature extraction method for moving target identification and tracking | |
| CN102651132A (en) | Medical image registration method based on intersecting cortical model | |
| CN112990314B (en) | Hyperspectral image anomaly detection method and device based on improved isolated forest algorithm | |
| CN110598613A (en) | Expressway agglomerate fog monitoring method | |
| CN107590785A (en) | A kind of Brillouin spectrum image-recognizing method based on sobel operators | |
| CN112308873A (en) | Edge detection method for multi-scale Gabor wavelet PCA fusion image | |
| CN104680536A (en) | Method for detecting SAR image change by utilizing improved non-local average algorithm | |
| CN112598708A (en) | Hyperspectral target tracking method based on four-feature fusion and weight coefficient | |
| CN107944497A (en) | Image block method for measuring similarity based on principal component analysis | |
| CN113093164B (en) | Translation-invariant and noise-robust radar image target identification method | |
| CN111460966B (en) | Hyperspectral remote sensing image classification method based on metric learning and neighbor enhancement | |
| CN112418241A (en) | Power equipment identification method based on infrared imaging technology |
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 |