CN110335285B - SAR image target marking method, system and device based on sparse representation - Google Patents

Abstract

The invention belongs to the field of security remote sensing detection of marine ships, airport airplanes and local automobiles, and particularly relates to a sparse representation-based SAR image target marking method, system and device, aiming at solving the problems of low target detection and marking efficiency and accuracy of the existing SAR image. The method comprises the steps of obtaining an SAR image, performing morphological processing after threshold segmentation of the image, and extracting an interested region; adopting coarse filtration, and taking the region of interest after the coarse filtration as a first region image; extracting multi-scale feature vectors of the first region image, and respectively obtaining sparse representations corresponding to the feature vectors; based on sparse representation of different scales, obtaining corresponding classification results through SVM classification models respectively, and obtaining the category of the first region image by adopting a preset decision method; the first region image is labeled in the SAR image based on the category and location information of the first region image. The invention can efficiently and accurately detect and mark the target.

Description

SAR image target labeling method, system and device based on sparse representation

technical field

The invention belongs to the field of remote sensing detection of marine ships, airport aircraft and position automobiles, and in particular relates to a sparse representation-based SAR image target marking method, system and device.

Background technique

At present, synthetic aperture radar (SAR) is widely used in the security industry. The SAR imaging system moves at a certain speed with the carrying platform, and then transmits electromagnetic wave pulses to the target area at a predetermined speed and time interval, and records the echoes. Carrying the intensity and phase information, the synthetic target image is recorded after multiple observations. Due to the characteristics of remote sensing images and airborne mobile imaging, the targets in remote sensing image scenes are usually small, and the remote sensing images themselves have serious speckle noise and rich texture information, which makes target detection difficult, and the detection efficiency and accuracy are relatively high. Difference.

So far, the methods of SAR image target detection and marking mainly include the method based on image template matching and the method based on feature template matching. Among them, the method based on image template matching needs to estimate the azimuth angle of the image and establish a template that matches the test sample. This method is simple and easy to implement, but it needs a lot of memory and has high computational complexity. The method based on feature template matching first uses algorithms such as PCA and ICA to extract features, and then uses support vector machine for classification and identification. However, this method still has the problem of too large feature dimension and long time consumption, which seriously affects the recognition rate of the target. . With the rapid development of my country's security needs, the accuracy requirements of target detection and marking are continuously improved, and the work of target detection and marking is increasing day by day. Traditional SAR image target detection and marking methods cannot meet the current rapidly developing security industry, including the detection and marking requirements of marine ships, airport aircraft and position vehicles.

How to improve the automatic target detection and marking ability of the SAR image processing system, accurately discover and identify various important strategic targets, and realize the rapid transformation of SAR images from data to intelligence information has become a new topic.

SUMMARY OF THE INVENTION

In order to solve the above problems in the prior art, that is, in order to solve the problems of low efficiency and accuracy of the existing SAR image target detection and labeling, the first aspect of the present invention proposes a method for SAR image target labeling based on sparse representation, The method includes:

Step S10, acquiring the SAR image to be marked, performing morphological processing after threshold segmentation of the SAR image, and extracting a region of interest;

Step S20, using coarse filtering based on the extracted region of interest, using the coarsely filtered region of interest as the first region image, and acquiring its position information in the SAR image;

Step S30, extracting the multi-scale feature vectors of the first area image, respectively obtaining sparse representations corresponding to the feature vectors of different scales;

Step S40, based on the sparse representation of different scales, obtain the corresponding classification results through the SVM classification model, and adopt a preset decision method to obtain the category of the first area image;

Step S50, marking the first region image in the to-be-marked SAR image according to the category and position information of the first region image;

in,

During the training process of the SVM classification model, the eigenvalues obtained by the singular value decomposition of the sample images of the SAR training set are used as attributes.

In some preferred embodiments, in step S30, "multi-scale extraction of the feature vector of the first area image", the method is: constructing a preset scale difference pyramid image set based on the first area image, extracting the Feature vectors of images of different scales in the image set.

In some preferred embodiments, in step S30 "respectively obtain sparse representations corresponding to feature vectors of different scales", the method is: obtaining through a sparse representation model; the construction method of the sparse representation model is:

Randomly extract some atoms of the training set to construct a dictionary;

The dictionary obtains an over-complete dictionary according to a preset approximation error value, and constructs an over-complete dictionary matrix;

The optimal sparse solution is obtained by decomposing the overcomplete dictionary matrix based on the MP algorithm;

An error matrix is constructed according to the optimal sparse solution, the error matrix is decomposed by a singular value decomposition method, and a dictionary is updated according to the decomposition result.

In some preferred embodiments, "decomposing the error matrix by using the singular value decomposition method", the method is: performing the position calibration of the non-zero coefficients on the error matrix to form a new error matrix, and using the singular value decomposition method to decompose the error matrix. The new error matrix is decomposed; the singular value decomposition method is constrained to optimize based on the regular term.

In some preferred embodiments, step S20 "uses coarse filtering based on the extracted region of interest, and uses the coarsely filtered region of interest as the first region image", where the coarse filtering includes geometric feature filtering, spatial location Filtering, grayscale feature filtering.

In some preferred embodiments, the scale ratio between the images of adjacent scales in the differential pyramid image set is 2, and the smoothing coefficients between the images of the adjacent layers of the same scale are different.

In some preferred embodiments, the dictionary is constructed by using partial atoms of the training set, and the number of partial atoms is 50%-80% of the atoms of the training set.

In a second aspect of the present invention, a system for SAR image target marking based on sparse representation is proposed, the system includes an extraction module, a filtering module, a sparse representation module, a classification and identification module, and a marking module;

The extraction module is configured to obtain the SAR image to be marked, perform morphological processing after threshold segmentation of the SAR image, and extract the region of interest;

The filtering module is configured to use coarse filtering based on the extracted region of interest, take the coarsely filtered region of interest as the first region image, and obtain its position information in the SAR image;

The sparse representation module is configured to extract multi-scale feature vectors of the first region image, and obtain sparse representations corresponding to the feature vectors of different scales respectively;

The classification and recognition module is configured to be based on sparse representation of different scales, obtain its corresponding classification results through the SVM classification model, and use a preset decision method to obtain the category of the first area image;

The marking module is configured to mark the first region image in the to-be-marked SAR image according to the category and position information of the first region image;

in,

During the training process of the SVM classification model, the eigenvalues obtained by the singular value decomposition of the sample images of the SAR training set are used as attributes.

In a third aspect of the present invention, a storage device is provided, in which a plurality of programs are stored, and the program applications are loaded and executed by a processor to implement the above-mentioned sparse representation-based SAR image target marking method.

In a fourth aspect of the present invention, a processing setup is proposed, including a processor and a storage device; the processor is adapted to execute various programs; the storage device is adapted to store multiple programs; the programs are adapted to be loaded by the processor And execute to realize the above-mentioned sparse representation-based SAR image target labeling method.

Beneficial effects of the present invention:

The present invention can efficiently and accurately perform target detection and marking. The invention accurately detects different targets in complex environments based on sparse representation; and based on the idea of unbalanced sampling, the number of target images that account for a minority and the number of background images that account for the majority can be balanced; further identification results are returned after target detection. Marks and gives the exact value of the locator in world coordinates. It effectively solves the technical problems of poor detection and marking efficiency and accuracy in the existing detection methods for marine ships, airport aircraft and position vehicles, thereby achieving efficient and accurate detection and marking technical effects.

Description of drawings

Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments taken with reference to the following drawings.

1 is a schematic flowchart of a sparse representation-based SAR image target labeling method according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a framework of a sparse representation-based SAR image target labeling method according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not All examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the related invention, but not to limit the invention. In addition, it should be noted that, for the convenience of description, only the parts related to the related invention are shown in the drawings.

It should be noted that the embodiments in the present application and the features of the embodiments may be combined with each other in the case of no conflict.

The sparse representation-based SAR image target marking method of the present invention includes the following steps:

Step S10, acquiring the SAR image to be marked, performing morphological processing after threshold segmentation of the SAR image, and extracting a region of interest;

Step S20, using coarse filtering based on the extracted region of interest, using the coarsely filtered region of interest as the first region image, and acquiring its position information in the SAR image;

Step S30, extracting the multi-scale feature vectors of the first area image, respectively obtaining sparse representations corresponding to the feature vectors of different scales;

Step S40, based on the sparse representation of different scales, obtain the corresponding classification results through the SVM classification model, and adopt a preset decision method to obtain the category of the first area image;

Step S50, marking the first region image in the to-be-marked SAR image according to the category and position information of the first region image;

in,

During the training process of the SVM classification model, the eigenvalues obtained by the singular value decomposition of the sample images of the SAR training set are used as attributes.

In order to more clearly describe the sparse representation-based SAR image target marking method of the present invention, each step in an embodiment of the method of the present invention will be described in detail below with reference to FIG. 1 .

In the following preferred embodiments, the sparse representation model and the SVM classification model are first described in detail, and then the sparse representation-based SAR image target labeling method for obtaining the target in the SAR image to be identified by using the sparse representation model and the SVM classification model is described in detail. .

1. Sparse representation model

1.1 Construction of training samples

(1) Acquire SAR images and extract multi-scale features

According to the preset target detection types, various types of remote sensing target image data are classified and collected with different resolutions and imaging angles, and the scales of the multi-scale images to be detected are counted to obtain the largest scale in the multi-scale image data to be detected.

Use the largest scale as the scale of the image dataset to construct the first group of images of the differential pyramid, and then use the sampling method to continue to construct scale images of the second, third, and fourth groups, so as to obtain multi-scale image features.

Image pyramid is a structure with multi-resolution images. The original image is sampled to obtain images of different scales, and the image of the same scale is filtered to obtain images with different degrees of smoothness. The largest resolution images are placed at the bottom and arranged in a pyramid form. The higher the resolution, the lower the image resolution, which constitutes an image pyramid. The steps are as follows:

Step 11: Sampling the original image to obtain the first layer image of the first group of pyramids, if the original image is larger than the set scale of the first group of pyramids, use downsampling, otherwise use upsampling. Then, the first layer images of the first group are obtained by Gaussian filtering to obtain the second layer images of the first group of pyramids. The Gaussian filter function is shown in formula (1), and the smoothing coefficient σ=1.6:

Among them, G(x, y) is a Gaussian filter function, x and y are the two-dimensional coordinates of the Gaussian distribution, and x ₀ , y ₀ are the mean values corresponding to the two-dimensional coordinates of the Gaussian distribution.

Step 12: Repeat step 11 to obtain the images of the third layer, the fourth layer, and the fifth layer of the first group, and the smoothing coefficients are 1.26×1.6, 1.26 ² ×1.6, and 1.26 ³ ×1.6, respectively. The first group of images in the pyramid has a total of five layers, and they have the same scale and different smoothing coefficients.

Step 13, down-sampling the images of the third layer of the first group to obtain the images of the first layer of the second group of the pyramid, the scale factor is 2, and adjacent groups can also be understood as adjacent scales. Repeat steps 11 and 12 to obtain five-layer images of the second group of pyramids, and the scale of the second group of images is half that of the first group. This cycle forms an image pyramid, with a total of four groups, each with five layers.

Step 14, construct a difference pyramid on the basis of the image pyramid, the first group of the first layer images of the difference pyramid is obtained by subtracting the first layer from the first group of the second layer of the image pyramid, that is, the 0th group of the first layer image of the difference pyramid. It is obtained by subtracting the Oth group I+1 level of the pyramid from the Oth group I level, where O and I are natural numbers, representing the number of groups and levels of the differential pyramid. In this way, a differential pyramid is obtained, a total of four groups, each with four layers.

(2) Unbalanced sampling of SAR images

In order to overcome the problem that the number of target images is less than the number of background images, the unbalanced sampling method SMOTE is used to construct training samples. The core idea of SMOTE is K-nearest neighbors and linear interpolation. The specific method is to add a new minority class sample through linear interpolation between two adjacent minority class samples, so that the number of minority class samples increases, so as to achieve minority class and majority class. The purpose of class sample size balance. The steps are as follows:

Step 21: For each sample _xi of the minority class, calculate the distance from _xi to all the samples of the minority class, select N nearest neighbor samples from it, and record the subscript. The K value is generally 5. Assuming that samples x _i and x _j have M-dimensional attributes, the distance is calculated according to formula (2):

Among them, x _j is a minority class sample other than sample x _i , and x _i (m) and x _j (m) are the mth attributes of samples x _i and x _j , respectively.

Step 22: Set the sampling ratio N according to the ratio of the majority class to the minority class (imbalance ratio) U in the original training set, and for each sample _xi of the minority class, randomly select t samples from the K nearest neighbor samples of _xi , denote t one of them as x _j (j=1,2,...,t), and perform random linear interpolation between x _i and x _j to obtain a new sample x _new . The calculation is shown in formula (3):

x _new = x _i +rand(0,1)×(x _j -x _i ) (3)

where rand(0,1) is a random number between 0 and 1.

Step 23, adding the obtained new samples to the original training set to obtain a new training set, in which the number of samples of the majority class and the minority class is balanced, that is, the number of background images and target images is balanced.

1.2 Building a training sparse representation model

The basic idea of sparse representation is that the signal consists of a large number of basic signals, which represent different types of features from different angles. Through the linear superposition of these fundamental signals, we can effectively represent arbitrary signals, which are called atoms, and the collection of atoms is called a dictionary. The steps to build the training are as follows:

In step S121 , a dictionary D is constructed using the atoms in the training set. The number of atoms should be smaller than the number of samples in the training set, preferably 50% to 80% of the samples in the training set.

Suppose we need to process the input signal x∈R ^m , the dictionary D={d ₁ ,d ₂ ,...,d _i ,...,d _n }∈R ^m×n , where each column d _i in the dictionary ,i∈[1,n] represents an atom in the dictionary. The basic assumption in sparse representation theory is that the signal x can be represented by a linear combination of atoms in a suitable dictionary D, as shown in equation (4):

where α _i is the expansion coefficient of the signal x on the atom d _i , α=[α ₁ ;α ₂ ;...;α _n ]∈R ⁿ .

Here, if the number of atoms in the dictionary is less than the dimension of the signal, that is, n<m, the atoms in the dictionary cannot fully represent the m-dimensional input signal, and this dictionary is called a non-complete dictionary; if n=m and ^D in D If the atoms are linearly independent, the dictionary can linearly represent any m-dimensional input signal. This kind of dictionary is called a complete dictionary. Atoms in a dictionary are called complete orthonormal basis if they are orthogonal to each other. For example, traditional Fourier transform, wavelet transform, cosine transform, etc. all use complete orthogonal basis; if n>m, it is a multi-solution problem, then the representation of the input signal x in the dictionary is not unique, that is, the solution of α is not unique. Uniquely, such a dictionary is called an overcomplete redundant dictionary.

A complete orthogonal dictionary usually uses a single type of orthogonal bases as dictionary atoms, but the signals in the real world are rich and diverse, and the signal representation needs to use different types of base signals. Therefore, in practice, the dictionary D usually contains multiple types of atoms. A complete redundant dictionary is required to solve multi-solution problems. The purpose of sparse representation theory is to find the most concise but efficient representation of the signal x in these solutions, and such solutions are called sparse solutions.

In step S122, the dictionary matrix is decomposed based on the MP algorithm iterative calculation, and the residuals are updated to obtain the optimal sparse solution.

The most intuitive measure of the sparsity of the sparse solution is the number of non-zero elements in the solution vector. If x only contains k non-zero coefficients α _i , i ∈ [1, n] in the dictionary D, then we call the signal x Under the representation of the dictionary D, it is k-sparse (K- _Sparsity ), and the lp norm of the vector α can be directly used to measure the sparsity, as shown in formula (5):

where ||α|| _p represents the _lp norm of the vector α.

The l ₀ norm can be regarded as a special case of the l _p norm, which is calculated by formula (6):

Among them, ||α|| ₀ represents the l ₀ norm of the vector α, that is, the number of non-zero elements in the vector α, and the non-zero coefficients represent the basic features and structure of the image, that is, the concentrated area of image energy.

According to the sparsity measure, the sparse representation model is obtained by formula (7):

由尽量少的字典原子表示，p＝2是信号处理领域中的均方误差度量准则，是一个平滑的凸优化问题；p＝1是总变分度量准则；当0＜p＜1时，它的数学分析是非凸的，称为拟范数；p＝0属于组合优化问题，并且是NP-hard问题。可以看出，当0≤p＜2时，都可以度量稀疏性，在这个范围内，p值越大，获得的解稀疏性越差，否则反之。因此，p＝0不仅是最直观的稀疏性度量方法，也是最合适稀疏性度量准则，缺点是求解比较困难，在多项式时间内难以获得解，其具体的求解方法是MP算法。where, the vector of samples x

Represented by as few dictionary atoms as possible, p=2 is the mean square error metric in the field of signal processing, which is a smooth convex optimization problem; p=1 is the total variation metric; when 0<p<1, it is The mathematical analysis of is non-convex, called quasi-norm; p=0 belongs to combinatorial optimization problem and is NP-hard problem. It can be seen that when 0≤p<2, the sparsity can be measured. In this range, the larger the value of p, the worse the sparsity of the obtained solution, otherwise the opposite is true. Therefore, p=0 is not only the most intuitive measure of sparsity, but also the most suitable measure of sparsity. The disadvantage is that it is difficult to solve, and it is difficult to obtain a solution in polynomial time. The specific solution method is MP algorithm.

When using the l ₀ norm to measure the sparsity, the core problem of the sparse solution is to use as few atoms in D as possible to represent the input signal x, that is, to convert the problem of solving the constraint equation, as shown in formula (8):

The above description is for an ideal noise-free image, and in practical applications, the image will inevitably have noise interference, so the model in the above formula is not necessarily exactly the same as the actual result, considering the actual error and noise, An error tolerance value δ is introduced, and the model expressed by the above formula is transformed into the following expression, that is, the signal is represented by as few atoms as possible within the allowable error range, as shown in formula (9):

表示真实样本x与其稀疏表示的平方差。in,

represents the squared difference of the true sample x and its sparse representation.

Or it can be transformed into a solution to find the best reconstruction accuracy under a certain sparsity limit, as shown in formula (10):

where T represents the error threshold of the l ₀ norm of α.

In order to calculate specific values, the constrained optimization problem described above can be transformed into an unconstrained optimization problem through Lagrange multipliers, as shown in Equation (11):

where λ represents the penalty factor.

Detailed explanation of MP algorithm: Due to the complexity and diversity of actual signal types, a single type of atom cannot better characterize the characteristics of diverse signals. Therefore, it is necessary to construct redundant dictionaries containing various basic atoms to obtain a sparse representation of the signal. At this time, it is necessary to find a solution of an ill-conditioned underdetermined equation system. It is known that the problem has multiple solutions. The sparse solution is to find the sparsest solution from the solution space of the problem. At this time, the problem is transformed by solving the ill-conditioned underdetermined equation system. To solve the minimization problem of the l0 norm, the problem is an NP-hard problem, and the solution is also very difficult. In the existing mathematical theory, there is no polynomial-time decomposition algorithm for NP problems, so only suboptimal algorithms can be used to solve sparse problems. At present, sparse solving algorithms can be divided into two categories: (1) greedy algorithms; (2) convex relaxation optimization algorithms. Next, a brief introduction to the greedy algorithm MP algorithm:

The greedy algorithm is a local optimization algorithm. It is carried out in stages. In each stage and each step, the current decision and choice are considered to be the best, and the impact of the choice at this stage on the future is not considered. , when the algorithm terminates, we hope that the local optimum is the global optimum, but in fact the algorithm often terminates on the sub-optimal solution. The MP algorithm is a typical greedy algorithm. The MP algorithm is briefly introduced below.

Assuming that the signal to be decomposed in space is x∈R ^m , the overcomplete redundancy dictionary is expressed as D={d ₁ ,d ₂ ,...,d _i ,...,d _n }∈R ^m×n , Then our goal is to select a sequence of atoms from the dictionary D and represent the signal x by their linear combination. It is assumed that the signal can be decomposed into equation (12):

x=<x,d _j0 >d _j0 +Rx (12)

Among them, Rx is the residual vector after the signal is approximated by atomic d _j0 , obviously d _j0 and Rx are mutually orthogonal, so the signal can be expressed as formula (13):

||x|| ² =|<x,d _j0 >| ² +||Rx|| ² (13)

In order to minimize the energy of the residual vector Rx, a suitable atom d _j0 must be chosen to maximize the energy of the signal x on this atom, i.e. d _j0 needs to satisfy equation (14):

The MP algorithm is an iterative algorithm. Each iterative calculation is to find the atom that best matches the residual vector Rx in the dictionary D, and then update the residual as described above, and then continue to repeat the process. Next, the MP algorithm is further explained:

First initialize the data, initialize the residual vector to the original signal, that is, R ⁰ x=x, assuming that the residual vector after the nth (n>0) iteration is expressed as R ⁿ x, then R ⁿ x will be decomposed into the formula ( 15):

R ⁿ x=|<R ⁿ x, d _jn >|d _jn +R ⁿ⁺¹ x (15)

Since R ⁿ⁺¹ x and atom d _jn are orthogonal to each other, the residual vector can be expressed as formula (16):

||R ⁿ x|| ² =|<R ⁿ x,d _jn >| ² +||R ⁿ⁺¹ x|| ² (16)

次时，将这一系列的等式相加，可以得到信号x表示为公式(17)：Similar equations can be obtained for each decomposition, when the decomposition proceeds to the first

The second time, adding this series of equations, the signal x can be obtained as equation (17):

It can also be transformed into equation (18):

Similarly, the energy of the signal x can also be obtained by a series of decomposition and summation, as shown in Equation (19):

The energy of the signal x is obtained jointly, as shown in Equation (20):

As can be seen from the above discussion, the signal x is decomposed into a dictionary D of sums of a series of atoms that best match the residual vector during each iteration. Although the decomposition process is nonlinear and non-orthogonal decomposition, the whole process preserves the energy of the signal, similar to the linear process and the orthogonal decomposition process. The specific operation steps of the MP algorithm are as follows:

Step 1221, input: the signal to be determined x, the specific dictionary D, the allowable error value δ or the maximum sparse value K;

Step 1222, initialization: initial value of residual vector R ⁰ x=x, number of iterations n=0;

Step 1223, start iterative calculation: search for atoms in dictionary D, as shown in formula (21):

R ⁿ x=|<R ⁿ x,d _jn >|d _jn +R ⁿ⁺¹ x (21)

Among them, d _jn satisfies

Step 1224, determine whether to continue the iteration:

If n>K||||R ⁿ⁺¹ x|| ² >δ, then n=n+1, go to step 1223 to continue the calculation;

If n≤K||||R ⁿ⁺¹ x|| ² ≤δ, go to step 1225;

Step 1225, the calculation ends, and the signal x to be determined is shown in formula (22):

Step S123, fixing the sparse solution, constructing an error matrix, and using singular value decomposition to update the non-zero coefficients of atoms.

Fixed α to update the dictionary. When updating the dictionary, it is updated column by column, and the non-zero coefficient corresponding to the atom is updated.

The method used is Singular Value Decomposition (SVD), and the specific calculation of dictionary update is as formula (23):

最小。Formula (23) expresses the fixed α to find the dictionary D such that the squared difference between the sample X and its sparse representation

minimum.

中的非零系数。此时，目标函数可以转换为公式(24)：Suppose D={d ₁ ,d ₂ ,...,d _i ,...,d _K }, where d _i is the ith column of the dictionary, when we need to update the atom d _k , we also need to update its corresponding The kth row of coefficients

nonzero coefficients in . At this point, the objective function can be transformed into formula (24):

表示原子d_k对应的第k行非零系数，E_k表示字典中未包含第k个原子时，此时所有训练样本在在该字典上的误差矩阵。where F is the F norm,

Represents the non-zero coefficient of the kth row corresponding to the atom _dk , and _Ek represents the error matrix of all training samples on the dictionary when the dictionary does not contain the kth atom.

Two regular terms are added to constrain and prevent overfitting. The objective function above is rewritten as formula (25):

为字典D正则项的权重因子，τ为α正则项的权重因子。in,

is the weight factor of the regular term of dictionary D, and τ is the weight factor of the regular term of α.

中的非零系数的位置，形成向量index^k，如公式(26)所示：When the atom d _k is updated, the remaining atoms are fixed. If the error matrix can be approximately decomposed into a column multiplied by a row vector, and the decomposed form closest to the error matrix is selected, the column can be used to replace the kth atom that is not included in the dictionary, and this row is the training sample. The projection coefficients on this atom, corresponding to this assumption, correspond to the well-known singular value decomposition. During the decomposition, we also need to maintain the corresponding sparsity. If the error matrix is directly decomposed into singular value, it may cause coefficient diffusion and cannot maintain sparsity, so the error matrix is not decomposed directly. To solve this problem, we calibrate the coefficient of the kth row

The positions of the non-zero coefficients in , form the vector index ^k , as shown in equation (26):

该矩阵在位置(index^k(i),i)上的元素为1，其余元素全都为0。那么设由缺失第k个原子引起的误差矩阵

可以通过公式(27)求得：These vectors are where the training samples are concentrated: the position of the projected sample set on the kth atom. First define the matrix

The element of this matrix at position (index ^k (i), i) is 1, and the rest of the elements are all 0. Then let the error matrix caused by the missing kth atom

It can be obtained by formula (27):

根据分解结果更新字典原子及相应位置系数，

其中，误差矩阵

是一个m×n矩阵，对误差矩阵进行奇异值分解，U是一个m×m的左奇异矩阵，Δ代表一个m×n的奇异值矩阵，除了主对角线元素以外其余元素为0，V代表一个n×n右奇异矩阵。将U中的第一列用于更新字典第k个原子d_k，将V中的第一列与Δ_(1,1)的乘积用于更新原第k行系数

中的非零系数，其中，Δ_(1,1)代表矩阵第一行第一列的元素。通过对字典中的原子逐个执行上述操作，即完成K-SVD字典学习。Decompose the new error matrix directly using SVD

Update dictionary atoms and corresponding position coefficients according to the decomposition result,

Among them, the error matrix

is an m×n matrix, which performs singular value decomposition on the error matrix, U is an m×m left singular matrix, Δ represents an m×n singular value matrix, and the remaining elements except the main diagonal elements are 0, V represents an n×n right singular matrix. Use the first column in U to update the kth atom d _k of the dictionary, and use the product of the first column in V and Δ _(1,1) to update the original kth row coefficients

The non-zero coefficients in , where Δ _(1,1) represents the element in the first row and first column of the matrix. By performing the above operations on the atoms in the dictionary one by one, the K-SVD dictionary learning is completed.

2. SVM classification model

Support vector machine (SVM) is a new type of machine learning method, which has the characteristics of strong classification ability, generalization ability, flexible classification method and insensitive feature parameters, and has been widely used in the field of pattern recognition.

The combination of linear identifiers presented by the dictionary constructed based on the sparse representation model, the linear identifiers are divided into linearly separable and inseparable. The SVM classification model can map linearly inseparable data points to a new space and convert them into linearly separable data in the new space for classification. If you go back to the space of the original data, it is actually non-linearly separated data.

The dictionary learning of singular value decomposition is adopted, and certain eigenvalues are obtained from the sparse representation of the sample images of the SAR training set as attributes, and the classification results are obtained through SVM training. The specific training process of the SVM classification model is as follows:

(1) Input training sample set T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _N ,y _N )}, where x=Dα, y∈{-1 ,+1};

其中，

时求得最优解

i，j为参数值；(2) Select the penalty parameter C and solve the convex quadratic programming

in,

find the optimal solution

i, j are parameter values;

选取α^*的一个分量

满足

计算

其中，w^*为法向量，b^*为截距；(3) Calculation

pick a component of α ^*

Satisfy

calculate

where w ^* is the normal vector and b ^* is the intercept;

(4) Obtain the separation hyperplane w ^* x+b ^* =0, and obtain the binary classification result according to the classification decision function f(x)=sign(w ^* x+b ^* ).

Based on the classification model of the SVM method, the classification results of the four groups of image sets of the differential pyramid are obtained respectively, and the image categories are obtained by using a preset decision method. decision), you can also choose other picture decision methods.

3. SAR image target labeling method based on sparse representation

A sparse representation-based SAR image target marking method according to an embodiment of the present invention includes the following steps:

Step S10: Obtain the SAR image to be marked, perform morphological processing after threshold segmentation of the SAR image, and extract a region of interest.

Threshold segmentation refers to the process of extracting objects of interest by setting corresponding thresholds according to the characteristics of the image. Threshold segmentation based on the grayscale information of an image is the most widely used image segmentation method. The threshold segmentation of the image is mainly divided into two steps: first, determine the optimal segmentation threshold; secondly, compare the pixel gray value with the segmentation threshold to realize the attribution of the region.

In actual situations, there is often a lot of noise interference, so that only one-dimensional grayscale histogram cannot observe obvious peaks and troughs, so a better segmentation effect cannot be obtained. The OTSU method based on two-dimensional grayscale histogram utilizes the grayscale information of the image, and considers the relevant information of each pixel in the image and its neighborhood space, and has a good anti-noise effect.

After threshold segmentation and morphological processing, the region of interest is obtained. The region of interest is the image of the region that is smaller than the set threshold after OTSU threshold segmentation and morphological processing.

In this embodiment, the SAR image to be marked is acquired, the SAR image is subjected to morphological processing after threshold segmentation, and the region of interest is extracted.

In step S20, based on the extracted region of interest, coarse filtering is adopted, and the coarsely filtered region of interest is used as a first region image, and its position information in the SAR image is obtained.

In this embodiment, based on the size of the ship, the grayscale, and the grayscale characteristics of its neighborhood, the primary selected region of interest is filtered. This method first uses prior knowledge to constrain the size, grayscale of the target and the grayscale of its neighborhood within a given range, and then uses a threshold method to filter it. This method utilizes the characteristics of the target as much as possible to ensure a low missed detection rate.

Based on the extracted region of interest, coarse filtering is adopted, and the coarsely filtered region of interest is used as the first region image. The first area image is an area image of which the area of interest is smaller than the set threshold after the scale (geometric) feature, the spatial position feature, and the grayscale feature are applied. and obtain its position information in the SAR image.

Step S30: Extract multi-scale feature vectors of the first region image, and obtain sparse representations corresponding to feature vectors of different scales respectively.

The essential idea of sparse representation is template matching, and the dictionary constructed based on the sparse representation model is a linear combination of training samples. Therefore, if the image patch we extract contains a target, its representation in the dictionary will show a certain sparse characteristic. And if our input target block does not contain the target, the lexicographic representation of the image will take on a stationary and random nature.

In this example, a differential pyramid image set with a preset scale is constructed based on the first region image, feature vectors of images of different scales in the image set are respectively extracted, and sparse representations corresponding to the feature vectors of different scales are obtained by using a sparse representation model.

Step S40 , based on the sparse representation of different scales, obtain the corresponding classification results through the SVM classification model, and use a preset decision method to obtain the category of the first area image.

In this example, based on sparse representations of different scales, the corresponding classification results are obtained through the SVM classification model, and a preset decision method is used to select the optimal classification result as the category of the first region image. The preset decision method is preferably joint voting decision (minority obeys majority decision) in this implementation, and other picture decision methods can also be selected.

Step S50: Mark the first area image in the to-be-marked SAR image according to the category and location information of the first area image.

The invention belongs to the detection of marine ships, airport aircraft and position vehicles, and needs to give the precise value of the SAR image to be located in the world coordinate system. In this example, returning to the original resolution SAR image block to be discriminated, the first region image is marked with a rectangular frame or other signs.

A sparse representation-based SAR image target labeling system according to the second embodiment of the present invention, as shown in FIG. 2 , includes: an extraction module 100, a filtering module 200, a sparse representation module 300, a classification and identification module 400, and a labeling module 500;

The extraction module 100 is configured to obtain a SAR image to be marked, perform morphological processing after threshold segmentation of the SAR image, and extract a region of interest;

The filtering module 200 is configured to use coarse filtering based on the extracted region of interest, and use the coarsely filtered region of interest as a first region image, and obtain its position information in the SAR image;

The sparse representation module 300 is configured to extract multi-scale feature vectors of the first region image, and obtain sparse representations corresponding to the feature vectors of different scales respectively;

The classification and identification module 400 is configured to obtain the corresponding classification results through the SVM classification model based on sparse representations of different scales, and adopts a preset decision method to obtain the category of the first area image;

The marking module 500 is configured to mark the first area image in the to-be-marked SAR image according to the category and position information of the first area image;

in,

During the training process of the SVM classification model, the eigenvalues obtained by the singular value decomposition of the sample images of the SAR training set are used as attributes.

Those skilled in the technical field can clearly understand that, for the convenience and brevity of description, for the specific working process and related description of the system described above, reference may be made to the corresponding process in the embodiment of the signing method, which will not be repeated here.

It should be noted that the sparse representation-based SAR image target labeling system provided by the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be allocated to different functional modules as required. To complete, that is, the modules or steps in the embodiments of the present invention are decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or can be further split into multiple sub-modules to complete all or part of the functions described above. . The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing each module or step, and should not be regarded as an improper limitation of the present invention.

A storage device according to the third embodiment of the present invention stores a plurality of programs, and the programs are suitable for being loaded by a processor and implementing the above-mentioned sparse representation-based SAR image target marking method.

A processing device according to a fourth embodiment of the present invention includes a processor and a storage device; the processor is adapted to execute various programs; the storage device is adapted to store multiple programs; the programs are adapted to be loaded and executed by the processor In order to realize the above-mentioned sparse representation-based SAR image target labeling method.

Those skilled in the technical field can clearly understand that the undescribed convenience and brevity, the specific working process and related description of the storage device and the processing device described above, can refer to the corresponding process in the example of the signing method, and will not be repeated here. Repeat.

Those skilled in the art should be able to realize that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two, and the programs corresponding to the software modules and method steps Can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or as known in the art in any other form of storage medium. In order to clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been described generally in terms of functionality in the foregoing description. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may use different methods of implementing the described functionality for each particular application, but such implementations should not be considered beyond the scope of the present invention.

The terms "first," "second," etc. are used to distinguish between similar objects, and are not used to describe or indicate a particular order or sequence.

The term "comprising" or any other similar term is intended to encompass a non-exclusive inclusion such that a process, method, article or device/means comprising a list of elements includes not only those elements but also other elements not expressly listed, or Also included are elements inherent to these processes, methods, articles or devices/devices.

So far, the technical solutions of the present invention have been described with reference to the preferred embodiments shown in the accompanying drawings, however, those skilled in the art can easily understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims (8)

1. A SAR image target marking method based on sparse representation is characterized by comprising the following steps:

step S10, acquiring an SAR image to be marked, performing morphological processing after threshold segmentation of the SAR image, and extracting an interested region;

step S20, based on the extracted region of interest, adopting coarse filtering, taking the region of interest after the coarse filtering as a first region image, and acquiring the position information of the region of interest in the SAR image;

step S30, extracting multi-scale feature vectors of the first region image, and respectively acquiring sparse representations corresponding to the feature vectors of different scales;

step S40, based on sparse representation of different scales, obtaining corresponding classification results through SVM classification models respectively, and obtaining the category of the first region image by adopting a preset decision method;

step S50, according to the category and the position information of a first area image, marking the first area image in the SAR image to be marked;

wherein,

in the training process of the SVM classification model, a characteristic value obtained by singular value decomposition of an SAR training set sample image is used as an attribute;

"respectively obtain sparse representations corresponding to feature vectors of different scales", the method is as follows: obtaining through a sparse representation model; the construction method of the sparse representation model comprises the following steps:

randomly extracting part of atoms in a training set to construct a dictionary;

the dictionary obtains an over-complete dictionary according to a preset approximation error value, and an over-complete dictionary matrix is constructed;

decomposing the overcomplete dictionary matrix based on an MP algorithm to obtain an optimal sparse solution;

constructing an error matrix according to the optimal sparse solution, decomposing the error matrix by adopting a singular value decomposition method, and updating a dictionary according to a decomposition result;

wherein, the method for decomposing the error matrix by adopting a singular value decomposition method comprises the following steps: carrying out non-zero coefficient position calibration on the error matrix to form a new error matrix, and decomposing the new error matrix by adopting a singular value decomposition method;

the singular value decomposition method carries out constraint optimization based on a regular term, and the corresponding objective function is as follows:

wherein, X represents a sample,

represents the squared difference of sample X and its sparse representation,

is the weight factor of dictionary D regular term, tau is the weight factor of alpha regular term, F is expressed as F norm,

represents an atom d_kCorresponding non-zero coefficient of k-th row, E_kWhen the representation dictionary does not contain the kth atom, all training samples are on the dictionaryThe error matrix of (a) is calculated,

k-th row coefficient, d, representing an error matrix_kAtom representing the k-th column in the dictionary, d_jRepresenting atoms in the dictionary except for column k.

2. The sparse representation-based SAR image target labeling method of claim 1, wherein in step S30, "extracting multi-scale feature vectors of the first region image" is performed by: and constructing a difference pyramid image set with a preset scale based on the first region image, and respectively extracting the feature vectors of the images with different scales in the image set.

3. The SAR image target marking method based on sparse representation as claimed in claim 1, wherein step S20 "based on the extracted region of interest, using coarse filtering to take the region of interest after coarse filtering as the first region image", wherein the coarse filtering includes geometric feature filtering, spatial position filtering, and gray feature filtering.

4. The SAR image target marking method based on sparse representation as claimed in claim 2, wherein the scale ratio between the images of adjacent scales in the differential pyramid image set is 2, and the smoothing coefficients between the images of adjacent layers of the same scale are different.

5. The SAR image target labeling method based on sparse representation as claimed in claim 1, wherein a dictionary is constructed by using training set partial atoms, and the number of the partial atoms is 50% -80% of the training set atoms.

6. A SAR image target marking system based on sparse representation is characterized by comprising an extraction module, a filtering module, a sparse representation module, a classification identification module and a marking module;

the extraction module is configured to acquire an SAR image to be marked, perform morphological processing after threshold segmentation of the SAR image, and extract an interested region;

the filtering module is configured to adopt coarse filtering based on the extracted region of interest, take the region of interest after the coarse filtering as a first region image, and acquire position information of the region of interest in the SAR image;

the sparse representation module is configured to extract multi-scale feature vectors of the first region image and respectively obtain sparse representations corresponding to the feature vectors of different scales;

the classification recognition module is configured to obtain corresponding classification results through SVM classification models based on sparse representations of different scales, and acquire the category of the first region image by adopting a preset decision method;

the marking module is configured to mark a first region image in the SAR image to be marked according to the category and the position information of the first region image;

wherein,

in the training process of the SVM classification model, a characteristic value obtained by singular value decomposition of an SAR training set sample image is used as an attribute;

"respectively obtain sparse representations corresponding to feature vectors of different scales", the method is as follows: obtaining through a sparse representation model; the construction method of the sparse representation model comprises the following steps:

randomly extracting part of atoms in a training set to construct a dictionary;

the dictionary obtains an over-complete dictionary according to a preset approximation error value, and an over-complete dictionary matrix is constructed;

decomposing the overcomplete dictionary matrix based on an MP algorithm to obtain an optimal sparse solution;

constructing an error matrix according to the optimal sparse solution, decomposing the error matrix by adopting a singular value decomposition method, and updating a dictionary according to a decomposition result; wherein, the method for decomposing the error matrix by adopting a singular value decomposition method comprises the following steps: carrying out non-zero coefficient position calibration on the error matrix to form a new error matrix, and decomposing the new error matrix by adopting a singular value decomposition method;

the singular value decomposition method carries out constraint optimization based on a regular term, and the corresponding objective function is as follows:

wherein, X represents a sample,

represents the squared difference of sample X and its sparse representation,

is the weight factor of dictionary D regular term, tau is the weight factor of alpha regular term, F is expressed as F norm,

represents an atom d_kCorresponding non-zero coefficient of k-th row, E_kRepresenting the error matrix of all training samples on the dictionary when the kth atom is not contained in the dictionary,

k-th row coefficient, d, representing an error matrix_kAtom representing the k-th column in the dictionary, d_jRepresenting atoms in the dictionary except for column k.

7. A storage device having stored therein a plurality of programs, wherein said program applications are loaded and executed by a processor to implement the sparse representation based SAR image target labeling method of any of claims 1-5.

8. A processing arrangement comprising a processor, a storage device; a processor adapted to execute various programs; a storage device adapted to store a plurality of programs; characterized in that the program is adapted to be loaded and executed by a processor to implement the sparse representation based SAR image target labeling method of any of claims 1-5.

Images