SAR (Synthetic Aperture Radar) seaborne ship identification device based on variational mode decomposition

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) seaborne ship identification device based on variational mode decomposition. The SAR seaborne ship identification device comprises an SAR radar, a database and an upper computer, the SAR radar, the database and the upper computer are sequentially connected, the SAR radar monitors a sea area in real time and stores image data obtained by the SAR radar into the database, and the upper computer comprises an image preprocessing module, a variational mode decomposition module, a classification module, a decision fusion module and a result display module. According to the method, a high-resolution image is obtained by using a synthetic aperture technology of an SAR radar, multi-mode representation of the image is obtained through two-dimensional variational mode decomposition, each mode is classified by using a convolutional neural network and probability distribution is output, and then the probability distribution of each mode is fused based on a Bayesian theory and a target is determined. And finally, high-precision ship target identification is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of SAR maritime radar target recognition, two-dimensional variational mode decomposition, and deep learning algorithms, and in particular to a SAR maritime ship recognition device based on variational mode decomposition. Background Art

[0002] As an important means of large-scale early warning and monitoring at sea, SAR radar ship identification modeling technology is widely used in the fields of marine security, maritime search and rescue, illegal smuggling monitoring, etc. SAR radar can generate high-resolution images and provide clear and detailed target information. This is crucial for the accurate identification and monitoring of targets such as ships at sea. Unlike optical sensors, SAR radar is not affected by weather conditions such as clouds, rain and fog, and can maintain efficient monitoring capabilities in harsh meteorological environments.

[0003] At present, ship target detection based on SAR images has been widely studied, but there are many shortcomings of traditional methods. Traditional ship target recognition methods based on manual feature extraction and classification models rely on complex artificial feature design, which is often difficult to maintain a high accuracy rate when facing the diversity of ship target variants and complex sea surface background. In addition, the existing deep learning methods require a large number of labeled samples to improve the recognition effect, but due to the scarcity and incompleteness of marine ship data sets, this method is greatly limited in practical applications. Therefore, when faced with complex SAR image data, existing models often find it difficult to cope with the high-precision requirements of target recognition, resulting in unsatisfactory recognition results. The low accuracy of ship target recognition based on SAR images has become a key problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a SAR marine ship recognition device based on variational mode decomposition to solve the problem that the current marine ship target recognition accuracy based on SAR images is not high.

[0005] The technical solution adopted by the present invention to solve its technical problems is: a SAR maritime ship recognition device based on variational mode decomposition, including a SAR radar, a database, and a host computer, which are connected in sequence. The SAR radar monitors the sea area in real time and stores the image data obtained by the SAR radar in the database; the host computer includes an image preprocessing module, a variational mode decomposition module, a classification module, a decision fusion module, and a result display module; the image preprocessing module, the variational mode decomposition module, the classification module, the decision fusion module, and the result display module are connected in sequence; the image preprocessing module is used to perform equalization processing on the SAR radar image data; the variational mode decomposition module is used to decompose the equalized image into multiple modes; the classification module is used to train a classifier based on the CNN network and establish a classification model; the decision fusion module is used to perform decision fusion based on the Bayesian network.

[0006] Further, the image preprocessing module is used to preprocess the SAR radar image data and is completed by the following process:

[0007] 2.1 Use the N noisy SAR radar grayscale images S * ={x 1 * ,x 2 * ,...,x N *} stored in the database as training samples, where S * represents the SAR radar grayscale image matrix, and x 1 * ,x 2 * ,...,x N * represent the 1st, 2nd,..., Nth SAR radar grayscale images respectively;

[0008] 2.2 Perform equalization processing on the grayscale histogram of the training samples:

[0009] 2.2.1 Calculate the probability mass function of the grayscale histogram, and the formula is as follows:

[0010]

[0011] Among them, p r (r) represents the probability mass function of the grayscale image, n r represents the number of pixels with grayscale level r in the image, and n represents the total number of pixels;

[0012] 2.2.2 Calculate the equalized grayscale level and map it back to the image pixels, and the formula is as follows:

[0013]

[0014] wherein, x ig represents the result of the g-th gray level equalization of the i-th sample, T(·) represents the histogram equalization transformation, r represents the gray level, and L represents the number of gray levels of the image.

[0015] The image after the gray histogram is equalized is S = {x 1 , x 2 ,..., x N}, where S represents the SAR radar gray image matrix after the gray histogram is equalized, and x 1 , x 2 ,..., x N respectively represent the images after the 1st, 2nd,..., Nth gray histograms are equalized.

[0016] Furthermore, the variational mode decomposition module is used to obtain the multi-modal representation of the equalized image, and is completed by the following process:

[0017] 3.1 Establish a two-dimensional constrained variational model, the formula is as follows:

[0018]

[0019] wherein, u k = {u 1 , u 2 ,... u k} represents the decomposed modal functions; ω k = {ω 1 , ω 2 ,..., ω k} represents the center frequencies of the decomposed modes, x 1 , x 2 ,..., x N represents the image obtained after the image preprocessing module, α k represents the quadratic penalty factor, and u As,k (x) represents the two-dimensional analytical signal, represents the corrected signal to modulate the spectrum of each analytical signal to the corresponding baseband, and x i represents the i-th image in the image set S = {x 1 , x 2 ,..., x N}, k represents the index of the modal function, representing the number or serial number of a specific mode in the decomposition process;

[0020] 3.2 For this constrained variational problem, the Lagrange multiplier operator λ and the quadratic penalty factor α are used to find the optimal solution, and it is transformed into an unconstrained variational problem, resulting in:

[0021]

[0022] Among them, L(·) represents the Lagrangian equation, λ represents the Lagrange multiplier operator, λ(x) represents the operator term, x represents the position of the current pixel, and ω k represents the central frequency of the mode function u k .

[0023] 3.3 Based on the multiplicative operator alternating direction method, parameters such as and are iterated. represents the mode function after the (n + 1)-th iteration, represents the central frequency of the above mode function, represents the Lagrangian operator of the above mode function, and the update process of obtaining is described as:

[0024]

[0025] In the formula: ω k is equivalent to ; ∑ k u k (x) is equivalent to .

[0026] 3.4 According to the Parseval Fourier isometric transformation principle, the frequency-domain expressions of each mode can be obtained through time-frequency transformation as:

[0027]

[0028] Among them, represents the form of the k-th mode function in the frequency domain after the (n + 1)-th iteration, represents the frequency-domain signal of the original decomposed signal, represents the form of the i-th mode function in the current iteration in the frequency domain, represents the Lagrange multiplier operator in the current frequency domain, and α is the quadratic penalty term used to control the smoothness in mode decomposition.

[0029] Similarly, the update expression of obtaining is:

[0030]

[0031] Among them, is the central frequency of the k-th mode function; Equivalent to the current residual: Wiener filtering;

[0032] Furthermore, the classification module trains the classifier using a CNN network and completes the process as follows:

[0033] 4.1 The input of the first layer of the classifier is the decomposed image matrix U. After passing through the convolutional layer, pooling layer, and fully connected layer, the feature extraction and mapping are completed. The ReLU activation function is used for all layers, and batch normalization is used to process the output of each layer;

[0034] 4.2 The last layer uses the classical regression classifier Softmax, and the formula is as follows:

[0035]

[0036] where z l represents the original output of the previous layer of the network, c represents the number of ship image categories, represents performing an exponential operation on each output signal of the previous layer of the network. The classifier independently trains and classifies each modality and outputs the decision variables in each modality, that is, the posterior probability vector of the current modality belonging to each category;

[0037] Furthermore, the decision fusion module is implemented using a Bayesian network. This module only acts on test samples and completes the process as follows:

[0038] 5.1 Use T = {T 1 , T 2 ,..., T C} and Y = {y 1 , y 2 ,..., y K} to represent C categories and K decisions respectively. The probability that the decision y k comes from the category Tc is denoted as:

[0039]

[0040] where T c represents the Cth category, y k represents the kth decision, represents the probability that the decision comes from the category;

[0041] 5.2 Under the condition that each decision is independent of each other, the joint probability distribution formula is:

[0042] P(T c |Y) = P(T c |y 1 )P(T c |y 2 )…P(Tc |y K ) (23);

[0043] 5.3 Based on decision fusion, determine the target category of the current test sample according to the maximum a posteriori probability, and the formula is:

[0044] identity(Y) = arg max c (P(T c |Y)) (24);

[0045] Among them, identity(Y) means that the input test sample selects the most likely category Y among all possible categories.

[0046] Furthermore, the host computer further includes a result display module for displaying the types of ships in the input SAR image on the screen.

[0047] As another preferred solution: The variational mode decomposition module included in the host computer can decompose the image at multiple scales, thereby providing the ability to model different hierarchical structures of the image. This helps to capture the information of the image more comprehensively at different scales.

[0048] The technical concept of the present invention is: In view of the characteristics that the marine SAR radar can work all-weather and can collect high-resolution images, the present invention first performs image preprocessing on the marine pictures taken by the SAR radar, then performs variational mode decomposition to obtain multi-modal representations of single pictures, and finally establishes a classification model through the training process of the convolutional neural network, and uses the Bayesian network to achieve decision fusion, thereby completing the ship target recognition task of the marine SAR radar.

[0049] The present invention overcomes the drawbacks of traditional methods and achieves the effect of improving the recognition accuracy through the following technical details: First, the SAR radar image data is equalized by the image preprocessing module to enhance the image quality and reduce the influence of noise; second, the variational mode decomposition module decomposes the processed image into multiple modes, and each mode provides different information, which helps to analyze the image content more carefully; then, the classification module trains the classifier based on the CNN network to establish a classification model, effectively improving the generalization ability and recognition accuracy of the model; finally, the decision fusion module uses the Bayesian network for decision fusion to optimize the classification result, thereby realizing high-accuracy ship recognition.

[0050] The beneficial effects of the present invention are mainly manifested in three aspects:

[0051] 1) Innovatively introduce the variational mode decomposition technology, which can effectively extract valuable information from the complex sea surface environment and significantly improve the classification accuracy;

[0052] 2) Compared with traditional deep learning models, the present invention requires fewer training samples and is more suitable for practical applications with limited samples.

[0053] 3) The proposed classification results have high reliability against noise interference, ensuring the stability of the system under various environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is the overall structural diagram of the system proposed by the present invention;

[0055] Figure 2 is the image target classification flowchart based on variational mode decomposition and Bayesian decision fusion in the host computer of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The present invention will be further described below with reference to the drawings.

[0057] Embodiment 1

[0058] Refer to Figure 1 and Figure 2 , a SAR maritime ship recognition device based on variational mode decomposition, comprising a SAR radar 1, a database 2 and a host computer 3. The SAR radar 1, the database 2 and the host computer 3 are connected in sequence. The SAR radar 1 irradiates the monitored sea area and stores the SAR radar image in the database 2. The host computer 3 includes:

[0059] An image preprocessing module 4 for preprocessing SAR radar image data, which is completed by the following process:

[0060] 2.1 First, read N noisy SAR radar grayscale images S * ={x 1 * , x 2 * ,..., x N *} from the database as training samples, where S * represents the SAR radar grayscale image matrix, and x 1 * , x 2 * ,..., x N * respectively represent the 1st, 2nd,..., Nth SAR radar grayscale images;

[0061] 2.2 Perform equalization processing on the grayscale histogram of the training samples.

[0062] 2.2.1 Calculate the probability mass function of the grayscale histogram, and the formula is as follows:

[0063]

[0064] Among them, p r (r) represents the probability mass function of the grayscale image, and n r represents the number of pixels with grayscale level r in the image, and n represents the total number of pixels.

[0065] 2.2.2 After obtaining the probability mass function, calculate the grayscale level and map it back to the image pixels. The formula is as follows:

[0066]

[0067] Among them, x ig represents the result of the g-th grayscale equalization of the i-th sample, T(·) represents the histogram equalization transformation, r represents the grayscale level, and L represents the number of grayscales of the image.

[0068] After completing the above steps, the equalized image S = {x 1 , x 2 ,..., x N} can be obtained, where S represents the SAR radar grayscale image matrix after the grayscale histogram is equalized, and x 1 , x 2 ,..., x N respectively represent the images after the 1st, 2nd,..., Nth grayscale histograms are equalized;

[0069] The variational mode decomposition module 5 is used to decompose the equalized image to obtain a multi-modal representation, which is completed by the following process:

[0070] 3.1 The essence of variational mode decomposition is an optimization model based on variational problems. Its innovative core idea is to determine the frequency and amplitude of each mode through variational optimization. Therefore, to solve each mode, first establish a two-dimensional constrained variational model. The formula is as follows:

[0071]

[0072] Among them, u k = {u 1 , u 2 ,... u k} represents the decomposed modal functions; ω k = {ω 1 , ω 2 ,..., ω k} represents the central frequencies of the decomposed modes, and x 1 , x 2 ,..., x N represent the image obtained after passing through the image preprocessing module 4, and αk denotes the quadratic penalty factor, denotes the correction signal, which modulates the spectrum of each analytic signal to the corresponding base frequency band. k represents the index of the mode function, indicating the number or sequence number of a specific mode in the decomposition process; u AS,k (x) represents the two-dimensional analytic signal, and its calculation formula is:

[0073]

[0074] where, u k (x) represents the one-dimensional analytic signal of the k-th mode function, δ(<x, ω k ) This is a Dirac δ function on x and frequency ω k , and is usually used to represent the impulse at the points of x and ω k , representing a function dependent on x and ω k , used to adjust the effect of the δ function. * represents the convolution operation.

[0075] 3.2 To solve the above constrained variational optimization problem, it is considered to transform it into an unconstrained variational problem, and then use the Lagrange multiplier operator λ and the quadratic penalty factor α to find the optimal solution. The obtained unconstrained variational problem is:

[0076]

[0077] where, L(·) represents the Lagrange equation, λ represents the Lagrange multiplier operator, λ(x) represents the operator term, x represents the position of the current pixel point, ω k represents the center frequency of the mode function u k .

[0078] 3.3 Use the multiplicative operator alternating direction method to iterate on and and other parameters. represents the mode function after the (n + 1)-th iteration, represents the center frequency of the above mode function, represents the Lagrange operator of the above mode function, and the update process of obtaining is described as:

[0079]

[0080] In the formula: ω k is equivalent to ; ∑kuk(x) is equivalent to .

[0081] 3.4 According to the Parseval Fourier isometric transformation principle, the frequency-domain expressions of each mode can be obtained through time-frequency transformation as:

[0082]

[0083] Among them, represents the form of the k-th modal function in the frequency domain after the (n + 1)-th iteration, is expressed as the frequency-domain signal of the original decomposed signal, represents the form of the i-th modal function in the current iteration in the frequency domain, represents the Lagrange multiplier operator in the current frequency domain, and α is the quadratic penalty term used to control the smoothness in modal decomposition.

[0084] Similarly, the updated expression of is:

[0085]

[0086] Among them, is the central frequency of the k-th modal function; is equivalent to the current residual: Wiener filtering of

[0087] 3.5 Each image x 1 , x 2 ,..., x N} in the image set S = {x i decomposed by the variational mode decomposition module is decomposed into multiple modes, satisfying

[0088] For the classification module 6, a classifier is built with a convolutional neural network for training, and the following process is adopted to complete it:

[0089] 4.1 The input of the first layer of the classifier is the decomposed image matrix U. Through the convolutional layer, pooling layer and fully connected layer, the extraction and mapping of features are completed, and the ReLU activation function is used for each layer, and batch normalization is used to process the output of each layer.

[0090] 4.2 Convolutional layer: There are 5 layers in total, separated by maxpool layer by layer. The first two layers contain 2 3x3 convolutional kernels, and the last three layers contain 3 3x3 convolutional kernels. When inputting a two-dimensional image, the convolution operation can be expressed by the following formula:

[0091] C(i, j) = ∑ m ∑ n I(i + m, j + n) × Kernel(m, n) + b (34);

[0092] Where C(i, j) is the output after convolution, I(i, j) is the pixel value of the input image, Kernel(m, n) is the weight of the convolution kernel, and b is the bias term. m is the row index in the convolution kernel, and n is the column index in the convolution kernel.

[0093] For multi-channel input images and convolution kernels, the convolution operation is performed separately on each channel and then added together to obtain the final output.

[0094] Fully connected layer: The fully connected layer connects all nodes in the previous layer to each node in the current layer. The calculation of the fully connected layer can be represented by matrix multiplication.

[0095] The last layer uses the classical regression classifier Softmax, and the formula is as follows:

[0096]

[0097] Where z l represents the raw output of the previous layer of the network, denotes performing an exponential operation on each output signal of the previous layer of the network, c represents the number of ship picture categories. The classifier independently trains and classifies each modality, and outputs the decision variables in each modality, that is, the posterior probability vector of the current modality belonging to each category. So far, the training link of the classification module in the host computer has been built.

[0098] The decision fusion module 7 is implemented using a Bayesian network. In the host computer, this module only acts on the test samples. The test samples must go through the same module operations as the training samples. After obtaining the probability distributions of different modalities of the test samples belonging to each category, the following process is used to complete the decision fusion:

[0099] 5.1 Use {T 1 , T 2 ,..., T C} and Y = {y 1 , y 2 ,..., y K} to represent C categories and K decisions respectively. The probability that the decision y k comes from the category T c is denoted as

[0100]

[0101] Where, T c represents the Cth category, y k represents the kth decision, represents the probability that the decision y k comes from the category T c ;

[0102] 5.2 Under the condition that each decision is independent of each other, the joint probability distribution formula is:

[0103] P(T c |Y) = P(T c |y 1 )P(T c |y 2 )…P(T c |y K ) (37);

[0104] 5.3 Based on decision fusion, determine the target category of the current test sample according to the maximum a posteriori probability:

[0105] identity(Y) = arg max c (P(T c |Y)) (38);

[0106] 5.4 The output of the decision fusion module serves as the final recognition result of the SAR maritime ship recognition device based on variational mode decomposition.

[0107] The upper computer 3 further includes: a result display module 8 for displaying the recognition result of the classification module on the upper computer.

[0108] The hardware part of the upper computer 3 includes: I / O components for data acquisition and information transfer; a data memory for storing data samples and operation parameters required for operation, etc.; a program memory for storing software programs for implementing functional modules; an arithmetic unit for executing programs to implement specified functions; and a display module for displaying set parameters and detection results.

[0109] The above embodiments are used to explain the present invention, rather than limiting the present invention. Any modifications and changes made to the present invention within the spirit and scope of the protection of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A SAR marine ship identification device based on variational mode decomposition, characterized in that: The system comprises a SAR radar, a database and a host computer, which are connected in sequence. The SAR radar performs real-time monitoring of the sea area and stores the image data obtained by the SAR radar in the database; the host computer comprises an image preprocessing module, a variational mode decomposition module, a classification module, a decision fusion module and a result display module; the image preprocessing module, the variational mode decomposition module, the classification module, the decision fusion module and the result display module are connected in sequence; the image preprocessing module is used for performing equalization processing on the SAR radar image data; the variational mode decomposition module is used for decomposing the image after the equalization processing into multiple modes; the classification module is used for performing classifier training based on a CNN network and establishing a classification model; the decision fusion module is used for performing decision fusion based on a Bayesian network.

2. The SAR marine ship identification device based on variational modal decomposition according to claim 1 is characterized in that: The image preprocessing module is used to preprocess the SAR radar image data, which is completed by the following process: (2.1) Using the N noisy SAR radar grayscale images S stored in the database * ={x1 * ,x2 * ,...,x N * } as training samples, where S * Represents the SAR radar grayscale image matrix, x1 * ,x2 * ,...,x N * Respectively represent the 1st, 2nd, ..., Nth SAR radar grayscale images; (2.2) Equalize the grayscale histogram of the training sample: (2.2.1) Calculate the probability mass function of the grayscale histogram, the formula is as follows: Among them, p r (r) represents the probability mass function of the grayscale image, n r represents the number of pixels with gray level r in the image, and n represents the total number of pixels; (2.2.2) Calculate the equalized gray level and map it back to the image pixel. The formula is as follows: i=1,2,...,N g=0,1,2,...,L-1 Among them, x ig represents the result of grayscale equalization of the gth sample of the ith sample, T(·) represents the histogram equalization transformation, r represents the grayscale level, and L represents the number of grayscales of the image.

3. The SAR marine ship identification device based on variational modal decomposition according to claim 2 is characterized in that: The image after the grayscale histogram is equalized is S = {x1, x2, ..., x N }, where S represents the SAR radar grayscale image matrix after grayscale histogram equalization, x1, x2, ..., x N They respectively represent the images after the 1st, 2nd, ..., Nth grayscale histograms are equalized.

4. The SAR marine ship identification device based on variational modal decomposition according to claim 1 is characterized in that: The variational modal decomposition module is used to obtain a multimodal representation of the image after equalization processing, and is completed by the following process: (3.1) Establish a two-dimensional constrained variational model, the formula is as follows: Among them, u k = {u1, u2, ... u k } represents each mode function after decomposition; ω k ={ω1, ω2, ..., ω k } represents the center frequency of each mode after decomposition, x1, x2, ..., x N represents the image obtained after the image preprocessing module, α k represents the quadratic penalty factor, u AS，k (x) represents the two-dimensional analytical signal, represents the correction signal, which modulates the spectrum of each analytical signal to the corresponding baseband; x i Denotes an image set S = {x1, x2, ..., x N }; k represents the index of the mode function, which indicates the number or sequence of a specific mode in the decomposition process; (3.2) For this constrained variational problem, we use the Lagrange multiplication operator λ and the quadratic penalty factor α to find the optimal solution and transform it into an unconstrained variational problem, and we get: Where L(·) represents the Lagrange equation, λ represents the Lagrange multiplication operator, λ(x) represents the operator term, x represents the position of the current pixel, ω k represents the modal function u k The center frequency of (3.3) Based on the alternating direction method of multiplication operator and Iterate with the same parameters, represents the modal function after the n+1th iteration, represents the center frequency of the above mode function, Representing the Lagrangian operator of the above modal function, we obtain The update process is described as: Where: k and Equivalence;∑ k u k (x) and equivalence; (3.4) According to the Parseval Fourier isometric transform principle, the frequency domain expression of each mode can be obtained by time-frequency transformation: in, represents the form of the kth mode function in the frequency domain after the n+1th iteration, Represented as the frequency domain signal of the original decomposed signal, represents the current iteration form of the ith mode function in the frequency domain, represents the Lagrange multiplication operator in the current frequency domain, α is the quadratic penalty term used to control the smoothness in the modal decomposition; Similarly, obtain The update expression is: in, is the center frequency of the kth mode function; Equivalent to the current residual: Wiener filtering.

5. The SAR marine ship identification device based on variational modal decomposition according to claim 1, characterized in that: The classification module uses a CNN network to perform classifier training, which is completed using the following process: (4.1) The first layer input of the classifier is the decomposed image matrix U, which is extracted and mapped through convolutional layers, pooling layers, and fully connected layers. The activation function uses ReLU and batch normalization is used to process the output of each layer. (4.2) The last layer uses the classic regression classifier Softmax, the formula is as follows: Among them, z l represents the original output of the previous layer of the network, c represents the number of ship image categories, It means that an exponential operation is performed on each output signal of the previous layer of the network. The classifier trains and classifies each mode independently and outputs the decision variables under each mode, that is, the posterior probability vector of the current mode belonging to each category.

6. The SAR marine ship identification device based on variational modal decomposition according to claim 1 is characterized in that: The decision fusion module is implemented using a Bayesian network. This module only acts on test samples and is completed using the following process: (5.1) Using T = {T1, T2, ..., T C } and Y = {y1, y2, ..., y K } represent C categories and K decisions respectively, and decision y k From category T c The probability of is recorded as: Among them, T c represents the Cth category, y k represents the kth decision, Represents decision y k From category T c The probability of (5.2) Under the condition that each decision is independent of each other, the joint probability distribution formula is: P(T c |Y)=P(T c |y1)P(T c |y2)…P(T c |y K ) (11); (5.3) Based on decision fusion, the target category of the current test sample is determined according to the maximum posterior probability. The formula is: identity(Y)=arg max c (P(T c |Y)) (12); Among them, identity(Y) means that the input test sample is most likely to be category Y among all possible categories.

7. The SAR marine ship identification device based on variational modal decomposition according to claim 1 is characterized in that: The host computer also includes a result display module, which is used to display the type of ship in the input SAR image on the screen.

Images