SAR image ship wake parameter estimation method based on improved residual light network

Abstract

The application discloses a SAR image ship wake parameter estimation method based on an improved residual light network, which comprises the following steps: firstly, constructing a ship wake SAR image dataset; then, building an improved residual light network; introducing a non-local operation at the bottom of an original ResNet-18 network to calculate the correlation degree of different positions of features and enhance global feature perception; reducing the model parameter quantity by updating a 3*3 convolution process into a 3*3 channel-by-channel convolution process and a 1*1 point-by-point convolution process; and finally, increasing a convolution block attention module before a full connection layer to enhance the related feature extraction capacity of the top layer of the network. The network is used for SAR image ship wake parameter estimation and has certain precision and model parameter quantity advantages.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and particularly relates to a SAR image ship wake parameter estimation method based on an improved residual light network. BACKGROUND

[0002] The ship wake SAR image is an image obtained by detecting and imaging the weak signal generated by the ship movement on the sea surface by SAR. It can show the wake traces left by the ship on the sea surface, and through the analysis of the traces, the trajectory, speed and heading of the ship can be obtained.

[0003] For the extraction of ship parameters contained in the SAR image wake, Dong Kaiquan et al. used the sea surface ship wake features and related parameters of optical remote sensing images, extracted the wake length by two different wake length detection methods of gray level accumulation method and Radon transform method, and simply estimated the ship speed and heading by using the wake length feature; Jia Hui realized the detection of linear wake based on local Radon transform, then started from the generation mechanism of Kelvin wake, proposed a two-dimensional estimation method of ship parameters, and in the case of filtering the sea surface background noise of the wake SAR image based on Markov process, combined with this method, the estimation of ship parameters was studied; Shuofa based on the wavelength of Kelvin wake speed estimation method, extracted the wavelength based on Kelvin wake fitting, and further estimated the speed information; Fan Wen et al. simulated the full polarization SAR image features of ship wake, explored the differences of ship wake features in SAR images under different polarizations, and then combined with the ship speed inversion method based on wake, used the ship wake SAR images under different polarizations to carry out ship speed inversion, and analyzed the influence of sea conditions, ship movement speed and polarization on the inversion results.

[0004] The current estimation of ship speed, direction, ship size and other parameters from the wake is mostly based on the shape and length of the wake, and such wake estimation method has high requirements for the extraction accuracy of the wake length, and has great limitations. The method using convolutional neural network (CNN) has strong adaptive ability to different wakes, and does not need to manually extract and fit the wake.

[0005] Residual Network (ResNet) is a series of deep residual neural networks, which includes ResNet-18, ResNet-50, ResNet-101 and ResNet-152 models based on different number of layers, and the ResNet network is composed of multiple basic blocks, each basic block contains two convolutional layers and a skip connection, the skip connection directly adds the input to the output of the block, forming a residual connection, which can let the gradient directly backpropagate to the shallower network, thus avoiding the problem of gradient vanishing. Since the ResNet-18 model is relatively shallow, it has achieved a good balance between computational efficiency and accuracy, and is a classic model in the field of deep learning.

[0006] However, the ship wake needs a larger feature size for network attention, while the ResNet network has a relatively small receptive field. The basic module of the ResNet network is composed of two 3x3 convolutional layers, which results in a receptive field of only 3x3 for each basic module, and the stride of each layer of the ResNet network is 1, without using pooling layers or larger stride convolutional layers to increase the receptive field. Although on this basis, new network models such as ResNeXt, DenseNet and EfficientNet have been proposed, but they still have some problems, such as: large number of parameters, poor interpretability, difficult to train, etc. SUMMARY

[0007] The purpose of the present application is to provide a SAR image ship wake parameter estimation method based on an improved residual lightweight network.

[0008] The technical solution for achieving the purpose of the application is as follows: in a first aspect, the present application provides a SAR image ship wake parameter estimation method based on an improved residual lightweight network, the steps are as follows:

[0009] First step: modeling the sea surface ship wake model, decomposing the sea surface ship wake into turbulent wake, Kelvin wake and sea surface model, then modeling the three respectively and then linearly superimposing to obtain the total sea surface ship wake model.

[0010] Second step: set the simulation radar parameters, realize the fast imaging of SAR image by using the SBR (Shot-Bounce-Rays) method, construct the data set by simulating different ship parameters, divide the data set into a certain proportion of training set and test set, take the SAR image as the input of the training sample, take the simulation parameters of the ship wake as the training label, and label the corresponding label for each SAR image.

[0011] Third step: build the improved residual lightweight network, send the training sample and the training label into the network at the same time for training, then use the trained network to detect the unknown SAR sample of the ship parameter, and evaluate the estimation result in terms of accuracy and parameter quantity.

[0012] In a second aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of the first aspect when executing the program.

[0013] In a third aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the program is executable on a processor to implement the steps of the method of the first aspect.

[0014] Compared with the prior art, the present application has the following beneficial effects:

[0015] (1) The improved residual lightweight network is proposed in the present application, which uses non-local operation to enhance global feature perception, increases the receptive field of the network, and introduces CBAM to enhance the learning ability of the top layer features of the network architecture, and decomposes the convolution kernel to reduce the network parameter quantity, which means that the network has certain accuracy and network parameter quantity advantages compared with the mainstream network without relying on the stacking depth;

[0016] (2) The improved residual lightweight network is trained by using the ship wake SAR image as the training sample, and the trained network model can better estimate the ship parameters contained in the unknown ship wake SAR image. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a schematic diagram of a ship turbulent wake.

[0018] Figure 2 is a schematic diagram of a ship Kelvin wake.

[0019] Figure 3 is a schematic diagram of a sea surface wake obtained by linearly superimposing a ship turbulent wake, a ship Kelvin wake, and a sea surface.

[0020] Figure 4 is a schematic diagram of a sea surface wake SAR image obtained by SBR fast imaging simulation.

[0021] Figure 5 is a schematic diagram of an improved residual lightweight network. DETAILED DESCRIPTION

[0022] A SAR image ship wake parameter estimation method based on an improved residual lightweight network comprises the following steps:

[0023] Step 1: modeling the sea surface ship wake model, decomposing the sea surface ship wake into turbulent wake, Kelvin wake and sea surface model, then modeling the three respectively and then linearly superimposing to obtain the total sea surface ship wake model;

[0024] Step 2: setting the simulation radar parameters, realizing the SAR image fast imaging by the bounce ray method, constructing the data set by simulating different ship parameters, dividing the data set into a certain proportion of training set and test set, taking the SAR image as the input of the training sample, taking the simulation parameters of the ship wake as the training label, and labeling the corresponding label for each SAR image;

[0025] Step 3: building an improved residual lightweight network, training the training sample and the training label in the network at the same time, and then using the trained network to detect the unknown SAR sample ship parameters, and evaluating the accuracy and parameter amount of the estimation result.

[0026] Further, in step 2, the data set built takes different ship speeds V S , ship length L, ship width B and ship draft D as random variables to generate ship wake SAR images.

[0027] Further, in step 3, an improved residual lightweight network is built, and the steps are as follows:

[0028] For the input wake SAR image, first, a 7x7 large convolution kernel is selected to extract the bottom layer features, which include important information such as image outline, details and texture; After 7x7 convolution operation, non-local operation is introduced to calculate the correlation degree of different positions of the features in the image, and the global feature perception is enhanced; Then through 3x3 maximum pooling, non-maximum values are eliminated and the bottom layer calculation complexity is reduced; After that, four Conv_x layers are passed through, each including four 3x3 ordinary convolution layers, and the output channel numbers are 64, 128, 256 and 512 respectively, the four layers of ordinary 3x3 convolution process are updated to 3x3 channel-by-channel convolution process and 1x1 point-by-point convolution process, so that the network parameter amount is reduced; Then the convolution block attention module (CBAM) is used to enhance the related feature extraction ability of the top layer of the network, and then the average pooling is performed to output 1x1 size features with 512 channels; Finally, it is output through a fully connected layer FC to directly output the ship parameters.

[0029] The details of the non-local operation process are as follows:

[0030] Suppose the index number of the feature at any position i of the input SAR image is i, the feature is processed, and the feature at i is redefined according to the following relationship

[0031]

[0032] where x i is the input feature at i, x j represents the input feature at j, y i is the output feature at i, and a unary function g(·) is taken as a linear transformation, then

[0033] g(x j ) = W g x j

[0034] where W g is the weight matrix to be learned, which can be realized by a 1x1 convolution in the spatial domain, and f(·,·) is a function measuring the similarity between i and j, which is expressed as

[0035]

[0036] Correspondingly, the normalization factor can be expressed as

[0037]

[0038] The detailed steps will be described below in conjunction with the drawings and examples.

[0039] Example

[0040] First, a ship turbulent wake model is established. At a distance x from the wake behind the ship's travel path, the width W(x) of the turbulent wake is

[0041]

[0042] where L is the length of the ship, B is the width of the ship, and a≈5.

[0043] The turbulent height S(k) decay model is

[0044]

[0045] where k is the wave number corresponding to the energy spectrum, V S is the ship speed, L' is the velocity integral length, E(k) = ξ 2 / 3 k -5 / 3 , ξ = 9.0 × 10 -6 , and y is the position perpendicular to the ship direction.

[0046] The results of the turbulent wake modeling are shown in Figure 1 .

[0047] Then a ship Kelvin wake model is established. At a distance x from the wake behind the ship's travel path, the width W(x) of the turbulent wake is

[0048]

[0049] where, is the wake wave height, Re represents taking the real part, θ is the angle of wave propagation direction relative to the x-axis, is the phase coefficient, is the wave number of the wave component in the propagation direction, A(θ) is the ship free wave spectrum, A(θ) satisfies

[0050]

[0051]

[0052] where, H(K, θ) is the Kochin function, S H is the ship surface, z is the draft, and the flow intensity σ(x, y, z) can be expressed as

[0053]

[0054] where, f is the ship body characteristic equation

[0055]

[0056] where, d is the side wall draft.

[0057] The results of the Kelvin wake modeling are shown in reference Figure 2 .

[0058] Then the sea surface is modeled, and the PM sea spectrum model is a simple and widely used simulation sea surface model, wherein the main wave direction sea wave power spectrum model formula is as follows

[0059]

[0060] In the formula, a = 8.10 x 10 -3 , b = 0.74, k x , k y is the spatial wave number of the sea wave, U 19.5 is the wind speed at 19.5 m above the sea surface, and g0 is the gravitational acceleration.

[0061] It is assumed that the lengths of the PM sea surface represented by two-dimensional discrete points in the x and y directions are L x and L y , respectively, and the number of discrete points is M and N, respectively, and the distance between adjacent points is Δx and Δy, wherein the above parameters satisfy L x = MΔx, L y= NΔy, the sea surface height at any point (m, n) on the sea surface is

[0062]

[0063] m k , n k is a matrix element sequence, then

[0064]

[0065]

[0066] where N1, N2 are random number matrices obeying normal distribution N(0, 1). In order to make the sea surface height f(x m ,y n a real number, should satisfy

[0067]

[0068] Figure 3 The results of the combined model of turbulent wake, Kelvin wake and sea surface are given.

[0069] Then, the SBR fast imaging algorithm is used to simulate the SAR image of the combined model, and the imaging formula is

[0070]

[0071] where k is the wave number, k x and k z are the wave number components in x and z directions respectively, E0 is the amplitude of the incident wave, j is the imaginary unit, r is the distance from the observation point to the, far-field echo of θ or polarization.

[0072] Figure 4 is a schematic diagram of the SAR image of the ship wake on the sea surface.

[0073] By simulating different ship parameters, a SAR data set is constructed, and the data set is divided into a certain proportion of training set and test set.

[0074] Taking the SAR image as the input of the training sample and the simulation parameters of the ship wake as the training label, each SAR image is labeled with the corresponding label, and the training sample and the training label are sent into the network for training at the same time.

[0075] The improved residual lightweight network is built based on ResNet-18. For the input SAR image, a large convolution kernel of 7x7 is selected to extract the bottom features, including important information such as the outline, details and texture of the image.

[0076] After the 7x7 convolution operation, a non-local operation is introduced to calculate the correlation degree of different positions of the features in the image, and the global feature perception is enhanced. The detailed process is as follows. Assuming that the index number of any position of the input SAR image feature is i, the feature is processed, and the feature at i is redefined according to the following relationship

[0077]

[0078] where x i is the input feature at i, x j represents the input feature at j, y i is the output feature at i, and a unary function g(·) is regarded as a linear transformation, that is,

[0079] g(x j ) = W g x j

[0080] where W g is a weight matrix to be learned, which can be realized by a 1x1 convolution in the spatial domain, and the function f(·,·) is a similarity function between i and j, which is expressed as

[0081]

[0082] Correspondingly, the normalization factor c(x) can be expressed as

[0083]

[0084] After the non-local operation, 3x3 max pooling is used to eliminate non-maximum values and reduce the computational complexity of the bottom layer.

[0085] After that, four Conv_x layers are used, each including four 3x3 ordinary convolution layers, and the output channel numbers are 64, 128, 256 and 512 respectively. Depth separable convolution (DSC) is introduced here, which updates the four-layer ordinary 3x3 convolution process to a 3x3 channel-by-channel convolution process and a 1x1 point-by-point convolution process, reducing the network parameter quantity.

[0086] Subsequently, the convolution block attention module (CBAM) is used to enhance the related feature extraction capability of the top layer of the network. The detailed process is as follows: for a given intermediate feature F∈R C×H×WAs input, firstly, Max Pool and Avg Pool are performed on the feature map by channel, two one-dimensional vectors are sent into a shared multi-layer perception (MLP), a 1*1 convolution kernel is used instead of a full connection layer, and a one-dimensional channel attention M is generated C ∈R C×1×1 Then, the channel attention is multiplied with F element by element, and a channel attention adjusted feature map F' is obtained;Then, Max Pool and Avg Pool are performed on F' by space, two two-dimensional vectors are generated, which are spliced and then convolved, and finally a two-dimensional spatial attention M is generated S ∈R 1×H×W Then, the spatial attention is multiplied with F' element by element. The overall attention generation process of CBAM can be described as

[0087]

[0088] Wherein represents element-wise multiplication, and the intermediate adopts a broadcast mechanism to transform and match the dimensions of the features.

[0089] After CBAM, the average pooling output is a 1*1 size feature with 512 channels, and finally it is output through a fully connected layer FC.

[0090] The improved residual network structure is as shown in Figure 5 .

[0091] In the network training process, Adam is used as the gradient optimizer, and the average absolute error L1 of the estimation result is used as the loss function, which is defined as

[0092]

[0093] Wherein, f(x i ) represents the parameter estimation value of the network input image, y i represents the true label value, and n represents the number of samples input each time.

[0094] The evaluation index selects the relative root mean square error (RRMSE), which can be represented by the difference between each parameter estimation value f(x i ) and the true value y i .

[0095]

[0096] The beneficial effects of the present application can be illustrated by the following experiments.

[0097] (1) Experimental environment:

[0098] Hardware configuration: Intel Core i5 10400F 2.9GHz, 40G RAM, Windows 10, Nvidia 3090 40G;

[0099] Software configuration: operating system Windows 10, Pycharm development software based on deep learning architecture Python 3.9+Pytorch 1.12+CUDA 11.7 environment.

[0100] (2) Training results and analysis:

[0101] According to the ship wake modeling formula, the shape of the ship wake is related to the ship length L, ship width B, ship draft D, ship movement speed V S Four ship parameters.

[0102] First, the ship wake SAR image dataset is established. For the simulation process of the dataset, the differences in the wake in the SAR image caused by different parameters are used to calculate the wake model under different parameters. Then, using the SBR fast imaging algorithm, under the simulation conditions of radar center frequency 10 GHz, azimuth 0°, pitch angle θ 80°, VV polarization, 300 sampling points in azimuth and range, according to the parameter sampling range in Table 1, two-dimensional SAR image simulation is carried out. The four ship parameters are randomly selected within the simulation range given in the table, and a total of 4800 random SAR wake images are calculated.

[0103] Table 1 Four ship parameter simulation range table

[0104]

[0105] In the experiment, first, 4800 pictures are randomly divided into training set and validation set according to the ratio of 9:1, and cropped to 224x224 size. Adam is used as the gradient optimizer, Batch size is set to 16, initial learning rate is set to 0.0001, training is performed for 200 rounds, then learning rate is decayed to one tenth of the original, training is performed for 100 rounds, a total of 300 rounds, and the random dropout rate Drop out of FC is set to 0.25.

[0106] The RRMSE error result of this experiment refers to the estimation of the four parameters of the 480 randomly divided validation set wake SAR images after each round of training in a total of 300 rounds of training using the network model trained in each round, and the RRMSE average value of the four parameters estimated in each round is calculated. Thus, in the total of 300 rounds, the optimal result with the smallest RRMSE is selected, representing the optimal training result of the network.

[0107] Ablation experiments are performed on each part of the network, and the network parameter amount and the RRMSE of the network for the four to-be-estimated ship parameters are taken as evaluation indexes. In summary, the comparison results in Table 2 are obtained, wherein RRMSE VS , RRMSE L , RRMSE B , and RRMSE D respectively represent the RRMSE of V S , L, B, and D, and the smaller the RRMSE is, the higher the estimation accuracy is, and Ours represents the final improved residual lightweight network.

[0108] Table 2 Distribution table of training results under different networks

[0109]

[0110] The RRMSE of the network training result of the present application is compared with the original ResNet-18. In the case where the network parameter amount is reduced to 1 / 7 of the original, the RRMSE of the network in the present application for V S , L, B, and D is reduced by 3.91%, 0.55%, 0.51%, and 0.31% respectively, and a better ship parameter estimation effect is achieved with a smaller network parameter amount.

[0111] Under the same network training conditions, the network Ours proposed in this section is compared with the mainstream CNN network in recent years, and the average optimal results of RRMSE of 300 rounds of training are recorded to obtain Table 3.

[0112] Table 3 Comparison result distribution table of various mainstream CNN models

[0113]

[0114]

[0115] The comparison proves that the improved residual lightweight network has the advantages of higher precision and smaller network parameter amount in the aspect of tail echo SAR ship parameter estimation.

[0116] The above only describes the preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the inventive concept of the present application, a number of modifications and improvements can be made, which all belong to the protection scope of the present application.

Claims

1. A method for estimating ship wake parameters in SAR images based on an improved residual lightweight network, characterized in that, Includes the following steps: Step 1: Model the ship wake model on the sea surface. Decompose the ship wake on the sea surface into turbulent wake, Kelvin wake and sea surface model. Then, model the three separately and then superimpose them linearly to obtain the overall ship wake model on the sea surface. Step 2: Set the simulated radar parameters, use the bouncing ray method to achieve rapid SAR image imaging, construct a dataset by simulating different ship parameters, and divide the dataset into a certain proportion of training set and test set. Use SAR images as input for training samples and simulated parameters of ship wakes as training labels. Label each SAR image with the corresponding label. Step 3: Construct an improved residual lightweight network. Training samples and labels are simultaneously fed into the network for training. The trained network is then used to detect ship parameters for unknown SAR samples. The accuracy and parameter quantity of the estimation results are evaluated. The improved residual lightweight network is constructed as follows: For the input wake SAR image, a large 7×7 convolutional kernel is first used to extract the low-level features, including the image's contour, details, and texture. After the 7×7 convolution operation, nonlocal operations are introduced to calculate the correlation between different locations of features in the image, enhancing global feature perception. Then, non-maximum values ​​are eliminated through 3×3 max pooling. After that, four Conv_x layers, each consisting of four 3×3 ordinary convolutional layers with output channels of 64, 128, 256 and 512 respectively, are passed. The four ordinary 3×3 convolutional processes are updated to 3×3 channel-wise convolutional processes and 1×1 pointwise convolutional processes. Then, a convolutional block attention module is used to enhance the network's top-level feature extraction capabilities. After that, average pooling is used to output 1×1 size features with 512 channels. Finally, the ship parameters are directly output through a fully connected layer (FC).

2. The method according to claim 1, characterized in that, In the second step, the dataset built is used to generate ship wake SAR images with different ship speeds V S , ship lengths L, ship widths B, and ship draft depths D as random variables.

3. The method according to claim 1, characterized in that, The specific method for the non-local operation process is as follows: Assuming the index number of any location of a feature in the input SAR image is i, process the feature and redefine the feature at location i according to the following relationship. Where, x i The input feature at position i, x j Let y represent the input feature at position j. i If the output characteristic at point i is considered as a linear transformation of the univariate function g(·), then we have: g(x j )=W g x j Among them, W g The weight matrix to be learned is implemented through a 1×1 convolution in the spatial domain. The function f(·,·) is the similarity function between points i and j, expressed as: The normalization factor is expressed as 4. The method according to claim 1, characterized in that, In the first step, a ship turbulent wake model is established. At a distance x after the ship's path, the width W(x) of the turbulent wake is... Where L is the ship's length, B is the ship's width, and a = 5; The turbulent height S(k) attenuation model is as follows: Where k is the wavenumber corresponding to the energy spectrum, V S Let L be the ship's speed, L′ be the length of the velocity integral, and E(k) = ξ. 2 / 3 k -5 / 3 ξ=9.0×10 -6 y represents the position perpendicular to the direction of the ship's hull.

5. The method according to claim 4, characterized in that, In the first step, a Kelvin wake model of the ship is established. The wake wave height is determined at a distance x from the wake and y perpendicular to the ship's hull direction after the ship's path. for Where Re represents the real part, and θ is the angle between the wave propagation direction and the x-axis. For phase coefficients, Let A(θ) be the wave number of the propagation direction wave component, and let A(θ) be the ship's free wave spectrum. A(θ) satisfies... Where H(K,θ) is the Kochin function, S H Let z be the surface of the ship, z be the draft, and the current intensity σ(x,y,z) be expressed as: Where f is the ship's hull characteristic equation. Where d is the draft of the sidewall.

6. The method according to claim 5, characterized in that, In the first step, the sea surface is modeled. The PM sea surface power spectrum model is a simple and widely used simulation sea surface model. The formula for the wave power spectrum model in the main wave direction is as follows: In the formula, a = 8.10 × 10 -3 b = 0.74, k x k y The spatial wave number of the ocean waves. U 19.5 The wind speed is 19.5m above the sea surface, and g0 is the acceleration due to gravity. Assume that the lengths of the PM sea surface in the x and y directions, represented by two-dimensional discrete points, are L and L, respectively. x and L y Let the number of discrete points be M and N, and the distances between two adjacent points be Δx and Δy, where the above parameters satisfy L x =MΔx,L y =NΔy, then the sea level height at any point (m,n) on the PM sea surface is: m k n k If it is a sequence of matrix elements, then Where N1 and N2 are random number matrices following a normal distribution N(0,1); to make the sea surface height f(x m ,y n () is a real number. Should meet 7. The method according to claim 6, characterized in that, The SBR fast imaging algorithm is used to simulate SAR images of the merged model. The imaging formula is as follows: Where k is the wave number, k x With k z Let be the wavenumber components in the x and z directions, respectively; E0 be the amplitude of the incident wave; j be the imaginary unit; and r be the distance from the observation point. For θ or Polarized far-field echo.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images