CN110232362B - Ship size estimation method based on convolutional neural network and multi-feature fusion - Google Patents

Abstract

The invention discloses a method for estimating ship size based on convolutional neural network and multi-feature fusion, which mainly solves the problem in the prior art that the target size estimation error is relatively large under the condition of medium and low resolution. The implementation plan is: 1) Obtain training samples and test samples, and perform preprocessing; 2) Obtain the magnitude image of the target area, and calculate multi-dimensional features; 3) Construct the target size estimation network framework, and train it with training samples to obtain the trained model; 4) use the trained model to estimate the preliminary size feature of the test sample; 5) form the multi-dimensional feature and the preliminary size feature into a new multi-dimensional feature; 6) use the new multi-dimensional feature to train the gradient boosting decision tree GBDT; 7) use the trained The GBDT model estimates the final size characteristics of the test sample. The invention uses the CNN network to independently learn the target features of the SAR image, improves the estimation accuracy of the target size, and can be used for the identification and classification of the ship target in the SAR image.

Description

Ship size estimation method based on convolutional neural network and multi-feature fusion

Technical Field

The invention belongs to the technical field of radar image processing, and mainly relates to a method for estimating the size of a ship in a SAR image, and can be used for identifying and classifying ship targets in a SAR image.

Background Art

Synthetic aperture radar SAR can observe the marine environment over a large area without being affected by weather and environmental factors. It has become an effective means of marine management. It is also widely used in military reconnaissance, disaster forecasting and other fields, and has broad research and application prospects. In recent years, based on the monitoring of marine targets, research has been carried out in the fields of ship detection and classification, iceberg detection, wind inversion, oil spill detection, sea ice monitoring, ship wake detection, etc., and its practicality has been proved. Among them, the size estimation of ship targets is the key and basis for ship target detection and classification. The precise geometric parameter estimation is also the key to SAR image interpretation, which has important research significance.

A lot of research has been done on ship target size estimation. However, due to practical limitations, it is difficult to obtain a large number of high-resolution SAR images. For medium and low-resolution SAR images, the ship target details are not rich enough, which affects the accuracy of target size estimation. To solve this problem, Bjorn Tings et al. proposed a dynamic adaptive ship parameter estimation method, which uses cross entropy and multivariate linear regression to optimize the algorithm parameters. This method has achieved high estimation accuracy on TerraSAR-X data, but the algorithm is greatly affected by environmental factors such as sea clutter.

In addition, for medium and low resolution SAR images, Lui Bedini et al. also proposed a ship target size extraction method based on SAR images. This method performs morphological analysis on SAR images, aiming to find the outline of the target from the cluttered SAR images, thereby extracting the size characteristics of the target. In addition, machine learning has also begun to be applied to the size estimation of ship targets, and has achieved good results. Boying Li et al. proposed a ship size estimation method based on dual polarization fusion and nonlinear regression, using gradient boosting decision tree GBDT to reduce the impact of imaging quality on the accuracy of ship size estimation, but these algorithms all rely on artificially designed features and are not robust.

Summary of the invention

The purpose of this invention is to address the shortcomings of existing SAR target size estimation methods and propose a ship target size estimation method based on convolutional neural network and multi-feature fusion to improve the performance of target size estimation in medium and low resolution situations, thereby improving the accuracy of target size estimation.

The technical idea of the present invention is: by performing image segmentation on training samples and test samples, the target area amplitude image of each sample is obtained, and the feature extraction of the target area amplitude image is performed to obtain the multi-dimensional features of the target; the training samples are input into the target size estimation network framework based on the convolutional neural network CNN for training to obtain a trained network model, and the test samples are input into the trained network model to obtain a preliminary target size estimation result. Then, the gradient boosting decision tree GBDT is used to further correct the preliminary target size estimation result to obtain the final target size feature. The implementation steps include the following:

(1) Five types of ship targets were selected from the public OpenSARShip dataset as the experimental dataset, and 70% of the experimental dataset was randomly selected as training samples and 30% as test samples;

(2) Perform histogram equalization and mean filtering on the original image G in the data set to remove noise interference and obtain the filtered image P;

(3) Performing threshold segmentation on the filtered image P, and performing morphological filtering and clustering processing on the segmentation results in turn to obtain a binary image Q of the target area, and then performing sidelobe removal processing to obtain a binary image Q' after sidelobe removal;

(4) Masking the binary image Q' after removing the side lobes and the original image G to obtain the target area amplitude image, calculate the multi-dimensional features of the target area amplitude image, and obtain the multi-dimensional features h of the target;

(5) Construct a target size estimation network framework Ψ based on convolutional neural network CNN:

5a) setting six convolutional layers, namely, a first convolutional layer L ₁ , a second convolutional layer L ₂ , a third convolutional layer L ₃ , a fourth convolutional layer L ₄ , a fifth convolutional layer L ₅ , and a sixth convolutional layer L ₆ ;

5b) setting four maximum pooling layers, namely, a first maximum pooling layer P ₁ , a second maximum pooling layer P ₂ , a third maximum pooling layer P ₃ , and a fourth maximum pooling layer P ₄ ;

5c) cross-arranging the six convolutional layers in 5a) and the four maximum pooling layers in 5b), that is, the target size estimation network framework Ψ is formed by sequentially connecting in series the first convolutional layer L ₁ , the first maximum pooling layer P ₁ , the second convolutional layer L ₂ , the second maximum pooling layer P ₂ , the third convolutional layer L ₃ , the third maximum pooling layer P ₃ , the fourth convolutional layer L ₄ , the fourth maximum pooling layer P ₄ , the fifth convolutional layer L ₅ , and the sixth convolutional layer L ₆ ;

(6) Input the training samples into the constructed target size estimation network framework Ψ for training, obtain the trained network model Ψ′, and obtain the preliminary size estimation results of the training samples;

(7) Input the test sample into the trained network model Ψ′ to obtain a preliminary size estimation result of the test sample;

(8) Combining the training sample part of the multidimensional feature vector h with the preliminary size estimation result of the training sample to form a new multidimensional feature vector h ₁ of the training sample, and combining the test sample part of the multidimensional feature vector h with the preliminary size estimation result of the test sample to form a new multidimensional feature vector h ₂ of the test sample;

(9) Using the multi-dimensional feature vector _h1 of the training sample to train the gradient boosting decision tree GBDT classifier, a trained gradient boosting decision tree GBDT model is obtained;

(10) The multidimensional feature vector _h2 of the test sample is sent to the trained gradient boosting decision tree (GBDT) model to obtain the final size feature of the test sample.

Compared with the prior art, the present invention has the following advantages:

1) The present invention uses a convolutional neural network to make a preliminary size estimation of the ship target in the SAR image. It does not need to manually design and extract features based on the SAR image, which greatly reduces the burden of manpower and improves the robustness of the size estimation algorithm.

2) This paper draws on the idea of deep learning to extract the size features of ship targets in SAR images. At the same time, it combines traditional algorithms to extract the scattering information of SAR images and performs multi-feature fusion to improve the accuracy of ship target size estimation.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the implementation of the present invention;

FIG2 is a diagram of a target size estimation network framework in the present invention;

FIG3 is a diagram showing the result of segmenting a portion of a sample image using the present invention;

FIG4 is a diagram showing a comparison between the target size estimation result and the target real size using the present invention;

DETAILED DESCRIPTION

The implementation scheme and effects of the present invention are described in detail below with reference to the accompanying drawings:

Referring to Figure 1, the implementation steps of this embodiment are as follows:

Step 1: Obtain training samples and test samples.

Five types of ship targets are selected from the public OpenSARShip dataset as the experimental dataset, and 70% of the experimental dataset is randomly selected as training samples and 30% as test samples.

Step 2: preprocess the original image G in the experimental data set.

The original image G in the experimental data set is first histogram equalized to improve the light and dark contrast between the target area and the background area, and then mean filtering is performed to remove noise interference to obtain the filtered image P.

Step 3: Perform threshold segmentation on the filtered image P.

3a) Obtaining the maximum amplitude value I _max and the minimum amplitude value I _min of the image P;

3b) Set the first scale parameter γ and calculate the segmentation threshold thr of the target-background area:

thr＝γ(I _max −I _min )+I _min

3c) Perform binarization operation on image P:

Wherein, I(x,y) is the amplitude value of the pixel at the position (x,y) in the image P, I _mask (x,y) is the encoding value of the pixel at the position (x,y) after binarization, and the encoded image is the binary image after segmentation;

Step 4: Correct the segmented binary image.

4a) Perform morphological filtering and clustering processing on the segmented binary image in turn to obtain a binary image Q of the target area;

4b) extracting the principal axis of the target area in the binary image Q by principal component analysis, and obtaining the length L _PA of the principal axis;

4c) extracting all straight lines parallel to the main axis in the target area and calculating the length L _line of the straight lines;

4d) Select the second scale parameter ρ, compare the lengths of the main axis and all other straight lines in the target area, if L _line <ρ×L _PA , erase all points on the straight line L _line , that is, set the pixels to 0, and obtain the binary image Q' after removing the side lobes.

Step 5: Calculate the multi-dimensional feature h of the target.

4a) Masking the binary image Q' after removing the side lobes and the original image G to obtain the amplitude image of the target area;

4b) Use principal component analysis to extract the principal axis of the target area amplitude image, obtain the target azimuth angle θ, and calculate the cosine value θ _cos of the azimuth angle value θ.

4c) Calculate the pixel amplitude and s ₁ , mean μ ₁ , variance σ ₁ of the target area, and let the first target feature f ₁ =(s ₁ ,μ ₁ ,σ ₁ );

4d) Use edge operator to extract edge pixels of the target area, calculate the amplitude and s ₂ , mean μ ₂ , variance σ ₂ of the edge pixels, and let the second target feature f ₂ =(s ₂ ,μ ₂ ,σ ₂ );

4e) Divide the target area into three equal parts along the main axis direction, namely, the bow, the midship, and the stern. Calculate the amplitude sum s ₃ , mean μ ₃ , and variance σ ₃ of the bow part; calculate the amplitude sum s ₄ , mean μ ₄ , and variance σ ₄ of the midship part; calculate the amplitude sum s ₅ , mean μ ₅ , and variance σ ₅ of the stern part. Let the third target feature f ₃ =(s ₃ ,μ ₃ ,σ ₃ ,s ₄ ,μ ₄ ,σ ₄ ,s ₅ ,μ ₅ ,σ ₅ );

4f) Combine the features obtained above into a multi-dimensional feature h = (θ _cos , f ₁ , f ₂ , f ₃ ).

The calculation formulas for the above amplitude, mean, and variance are:

Amplitude and:

Mean:

variance:

Wherein, N is the number of pixels in the area to be counted, I( _xi , _yi ) is the amplitude value of the pixel located at ( _xi , _yi ), and j represents the jth group of statistics, j=1,…,5.

Step 6: Construct the target size estimation network framework Ψ and parameters based on the convolutional neural network CNN.

6a) Referring to Figure 2, construct the target size estimation network framework Ψ:

6a1) setting six convolutional layers, namely, a first convolutional layer L ₁ , a second convolutional layer L ₂ , a third convolutional layer L ₃ , a fourth convolutional layer L ₄ , a fifth convolutional layer L ₅ , and a sixth convolutional layer L ₆ ;

6a2) setting four maximum pooling layers, namely, a first maximum pooling layer P ₁ , a second maximum pooling layer P ₂ , a third maximum pooling layer P ₃ , and a fourth maximum pooling layer P ₄ ;

6a2) Cross-arrange the six convolutional layers in 6a1) and the four maximum pooling layers in 6a2).

That is, the target size estimation network framework Ψ is composed of the first convolution layer L ₁ , the first maximum pooling layer P ₁ , the second convolution layer L ₂ , the second maximum pooling layer P ₂ , the third convolution layer L ₃ , the third maximum pooling layer P ₃ , the fourth convolution layer L ₄ , the fourth maximum pooling layer P ₄ , the fifth convolution layer L ₅ , and the sixth convolution layer L ₆ which are sequentially connected in series;

6b) Set the parameters of each layer of the network framework Ψ:

The first convolution layer L ₁ takes as input image data x ₁ , with a size of 128×128×1, a window size of 5×5, a sliding step size S ^L1 of ¹ , and a padding parameter P=0. It is used to perform convolution operations on the input data and output 16 feature maps Y ^{1 ,} with a size of 124×124×16.

The first maximum pooling layer P ₁ , the input data is Y ¹ , the padding parameter P = 0, the window size of the pooling kernel U ¹ is 2×2, the sliding step size S ^P1 is 2, which is used to downsample the input data, and the size of the output feature map Y ² is 62×62×16;

The second convolution layer L ₂ , the input data is Y ² , the window size of its convolution kernel K ² is 5×5, the sliding step size S ^L2 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 32 feature maps Y ³ , the size of Y ³ is 58×58×32;

The second maximum pooling layer P ₂ , the input data is Y ³ , the padding parameter P = 0, the window size of the pooling kernel U ² is 2×2, the sliding step size S ^P2 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁴ is 29×29×32;

The third convolution layer L ₃ , the input data is Y ⁴ , the window size of its convolution kernel K ³ is 6×6, the sliding step size S ^L3 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 64 feature maps Y ⁵ , the size of Y ⁵ is 24×24×64;

The third maximum pooling layer P ₃ , the input data is Y ⁵ , the padding parameter P = 0, the window size of the pooling kernel U ³ is 2×2, the sliding step size S ^P3 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁶ is 12×12×64;

The fourth convolution layer L ₄ , the input data is Y ⁶ , the window size of its convolution kernel K ⁴ is 3×3, the sliding step size S ^L4 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 128 feature maps Y ⁷ , the size of Y ⁷ is 10×10×128;

The fourth maximum pooling layer P ₄ , the input data is Y ⁷ , the padding parameter P = 0, the window size of the pooling kernel U ⁴ is 2×2, the sliding step size S ^P4 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁸ is 5×5×128;

The fifth convolution layer L ₅ , the input data is Y ⁸ , the window size of its convolution kernel K ⁵ is 3×3, the sliding step size S ^L5 is 1, the padding parameter P=0, it is used to perform convolution operation on the input data, and output 64 feature maps Y ⁹ , the size of Y ⁹ is 3×3×64;

The sixth convolution layer L ₆ , the input data is Y ⁹ , the window size of its convolution kernel K ⁶ is 3×3, the sliding step size S ^L6 is 1, the padding parameter P=0, it is used to perform a convolution operation on the input data and output a feature map Y ¹⁰ , the size of Y ¹⁰ is 1×1×1, and the preliminary target size is obtained.

Step 7: Train the target size estimation network framework Ψ to obtain the trained network model Ψ′ and obtain the preliminary size estimation result of the training sample.

7a) Input the training sample into the target size estimation network framework Ψ and calculate the loss of the network output layer:

loss=(y _true -y _prediction ) ² ,

Where y _true is the true size of the sample, and y _prediction is the estimated size of the sample target;

7b) Use the back-propagation algorithm to propagate the loss of the output layer forward, and use the stochastic gradient descent algorithm to calculate the gradient vector of the loss function loss, and update the parameters of each layer in the network;

7c) Repeat 7b), iteratively update the parameters continuously until the loss function converges, and obtain the trained network model Ψ′;

7d) Input the training sample into the trained network model Ψ′ to obtain a preliminary size estimation result of the training sample.

Step 8: Input the test sample into the trained network model Ψ′ to obtain the preliminary size estimation result of the test sample.

Step 9: Fusion the target preliminary size estimation result with the multi-dimensional feature vector.

The training sample part of the multidimensional feature vector h is combined with the preliminary size estimation result of the training sample to form a new multidimensional feature vector h ₁ of the training sample, and the test sample part of the multidimensional feature vector h is combined with the preliminary size estimation result of the test sample to form a new multidimensional feature vector h ₂ of the test sample.

Step 10: Train the gradient boosted decision tree GBDT to obtain a trained gradient boosted decision tree GBDT model.

10a) Initialize the gradient boosting regression tree model f ₀ = 0;

10b) Train the gradient boosting regression tree model, assuming the number of cycles t = 1,…,10:

10b1) Taking the multi-dimensional feature h ₁ of the training sample as input, a trained regression tree model f _t is obtained by fitting the residual loss r _t-1 :

r _t-1 = y _true -y _pre

Where y _true is the true size of the sample, and y _pre is the estimated size of the target output by the regression tree model f _t-1 ;

10b2) Update the gradient boosting regression tree model _ft = ft _-1 + _vft , where v is the learning rate;

10b3) Repeatedly iterate and continuously update the gradient boosting regression tree model until the number of iterations reaches t = 10, and the trained gradient boosting regression tree model ζ is obtained.

Step 11, the multidimensional feature vector _h2 of the test sample is sent to the trained gradient boosting decision tree GBDT model ζ to obtain the final size feature of the test sample.

The effect of the present invention can be further illustrated by the following experimental data:

1. Experimental conditions:

1) Experimental data:

The data used in the experiment is the OpenSARShip dataset compiled by Shanghai Jiao Tong University. The dataset can be downloaded from the OpenSAR platform of Shanghai Jiao Tong University at https://opensar.sjtu.edu.cn/. The experimental dataset comes from the VH polarization interferometric wide-band IW mode, with a resolution of 20m×20m and a pixel spacing of 10m. The size truth value of the ship image used in this experiment is provided by the ship automatic identification system AIS. The dataset used contains five types of targets: Tanker, Cargo, Bulk carrier, General cargo, and Container.

The dataset selected for the experiment contains a total of 2467 target images with an image size of 128×128. 70% of the dataset is randomly selected as training samples, including 1727 target images, and 30% of the dataset is test samples, including 740 target images.

The targets used in the experiment ranged from 42m to 399m in length and from 6m to 65m in width.

This experiment selected two evaluation criteria, relative error and absolute error, to evaluate the error of the estimation results.

2. Experimental content:

Experiment 1: The present invention is used to conduct experiments on the above experimental data, setting the first scale parameter γ = 0.67, the second scale parameter ρ = 0.59, and the learning rate v = 0.5, and visualizing the image segmentation results. The results are shown in Figure 3, where:

Figure 3(a) shows the original images of some samples;

Figure 3(b) shows the target area segmentation image of some samples.

As can be seen from Figure 3, the target region segmentation result is not very accurate, so directly using the target region segmentation result to estimate the target size will produce a large error.

Experiment 2: The above experimental data were compared using the method of the present invention and the existing method. In order to verify the size extraction effect of the present invention, the size extraction results were compared with other methods. The comparison results are shown in Table 1.

Table 1 Comparison results of size estimation accuracy between the method of the present invention and the existing method

In Table 1: The existing method is a ship size estimation method based on dual polarization fusion and nonlinear regression, which mainly includes two stages: image preprocessing and nonlinear regression. The experimental reproduction method may differ from the original method in details.

As can be seen from Table 1, the error of ship size estimation using the present invention is smaller than the estimation error of the existing method, achieving better target size estimation performance and higher robustness. Comparing the preliminary target size estimation results with the final target size estimation results, it is found that the use of gradient boosting regression tree to correct the preliminary size estimation results can improve the accuracy of target size estimation, indicating that the multi-dimensional features obtained from the binary image of the target area are helpful for target size regression.

Experiment 3: The above experimental data were tested using the method of the present invention, and the comparison results between the estimated target size and the actual target size were visualized. The results are shown in Figure 4, where:

FIG4( a ) is a comparison result between the estimated target length and the actual target length obtained by using the present invention;

FIG4( b ) is a comparison result between the estimated width of the target obtained by using the present invention and the actual width of the target.

As can be seen from FIG4 , there is a high correlation between the target size estimation value obtained by using the present invention and the real size value of the target.

The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. It is obvious that for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims (7)

1. A ship target size estimation method based on convolutional neural network and multi-feature fusion, characterized by comprising: (1) Five types of ship targets were selected from the public OpenSARShip dataset as the experimental dataset, and 70% were randomly selected as training samples and 30% as test samples; (2) Perform histogram equalization and mean filtering on the original image G in the data set to remove noise interference and obtain the filtered image P; (3) Performing threshold segmentation on the filtered image P, and performing morphological filtering and clustering processing on the segmentation results in turn to obtain a binary image Q of the target area, and then performing sidelobe removal processing to obtain a binary image Q' after sidelobe removal; (4) Masking the binary image Q' after removing the side lobes and the original image G to obtain the target area amplitude image, calculate the multi-dimensional features of the target area amplitude image, and obtain the multi-dimensional features h of the target; (5) Construct a target size estimation network framework Ψ based on convolutional neural network CNN: 5a) setting six convolutional layers, namely, a first convolutional layer L ₁ , a second convolutional layer L ₂ , a third convolutional layer L ₃ , a fourth convolutional layer L ₄ , a fifth convolutional layer L ₅ , and a sixth convolutional layer L ₆ ; 5b) setting four maximum pooling layers, namely, a first maximum pooling layer P ₁ , a second maximum pooling layer P ₂ , a third maximum pooling layer P ₃ , and a fourth maximum pooling layer P ₄ ; 5c) cross-arranging the six convolutional layers of 5a) and the four maximum pooling layers of 5b), that is, the target size estimation network framework Ψ is formed by sequentially connecting in series the first convolutional layer L ₁ , the first maximum pooling layer P ₁ , the second convolutional layer L ₂ , the second maximum pooling layer P ₂ , the third convolutional layer L ₃ , the third maximum pooling layer P ₃ , the fourth convolutional layer L ₄ , the fourth maximum pooling layer P ₄ , the fifth convolutional layer L ₅ , and the sixth convolutional layer L ₆ ; (6) Inputting the training samples into the constructed target size estimation network framework Ψ for training, obtaining the trained network model Ψ′, and obtaining the preliminary size estimation results of the training samples; (7) Input the test sample into the trained network model Ψ′ to obtain a preliminary size estimation result of the test sample; (8) Combining the training sample part of the multidimensional feature vector h with the preliminary size estimation result of the training sample to form a new multidimensional feature vector h ₁ of the training sample, and combining the test sample part of the multidimensional feature vector h with the preliminary size estimation result of the test sample to form a new multidimensional feature vector h ₂ of the test sample; (9) Using the multi-dimensional feature vector _h1 of the training sample to train the gradient boosting decision tree GBDT classifier, a trained gradient boosting decision tree GBDT model is obtained; (10) The multidimensional feature vector _h2 of the test sample is sent to the trained gradient boosting decision tree GBDT model to obtain the final size feature of the test sample. 2. The method according to claim 1, characterized in that the threshold segmentation of the image P in (3) is performed according to the following steps: 3a) Obtaining the maximum amplitude value I _max and the minimum amplitude value I _min of the image P; 3b) Set the first scale parameter γ and calculate the segmentation threshold thr of the target-background area: thr＝γ(I _max −I _min )+I _min 3c) Perform binarization operation on image P:

Among them, I(x,y) is the amplitude value of the pixel at the position (x,y) in the image P, I _mask (x,y) is the encoding value of the pixel at the position (x,y) after binarization, and the encoded image is the binary image after segmentation. 3. The method according to claim 1, characterized in that the sidelobe removal process in (3) is performed according to the following steps: 3d) extracting the principal axis of the target area in the binary image Q by principal component analysis, and obtaining the length L _PA of the principal axis; 3e) extracting all straight lines parallel to the main axis in the target area and calculating the length L _line of the straight lines; 3f) Select the second scale parameter ρ, compare the lengths of the main axis and all other straight lines in the target area, if L _line <ρ×L _PA , erase all points on the straight line L _line , that is, set the pixels to 0, and obtain the binary image Q' after removing the side lobes. 4. The method according to claim 1 is characterized in that (4) calculating the multidimensional features of the target area amplitude image to obtain the multidimensional features h of the target is performed according to the following steps: 4a) Use principal component analysis to extract the principal axis of the target area amplitude image, obtain the target azimuth θ, and calculate the cosine value of the azimuth θ

4b) Calculate the pixel amplitude and s ₁ , mean μ ₁ , variance σ ₁ of the target area, and let the first target feature f ₁ =(s ₁ ,μ ₁ ,σ ₁ ); 4c) Use edge operator to extract edge pixels of the target area, calculate the amplitude and s ₂ , mean μ ₂ , variance σ ₂ of the edge pixels, and let the second target feature f ₂ =(s ₂ ,μ ₂ ,σ ₂ ); 4d) Divide the target area into three equal parts along the main axis direction, namely, the bow, the midship, and the stern. Calculate the amplitude sum s ₃ , mean μ ₃ , and variance σ ₃ of the bow part; calculate the amplitude sum s ₄ , mean μ ₄ , and variance σ ₄ of the midship part; calculate the amplitude sum s ₅ , mean μ ₅ , and variance σ ₅ of the stern part. Let the third target feature f ₃ =(s ₃ ,μ ₃ ,σ ₃ ,s ₄ ,μ ₄ ,σ ₄ ,s ₅ ,μ ₅ ,σ ₅ ); 4e) Combine the features obtained above into a multi-dimensional feature h = (θ _cos , f ₁ , f ₂ , f ₃ ). 5. The method according to claim 1 is characterized in that, in 5c), a ship size estimation network framework Ψ based on a convolutional neural network is constructed, and the parameters of each layer are as follows: The first convolution layer L ₁ takes as input image data x ₁ , with a size of 128×128×1, a window size of 5×5 for the convolution kernel K ¹ , a sliding step size S ^L1 of 1, and a padding parameter P=0. It is used to perform convolution operations on the input data and output 16 feature maps Y ¹ , with ^a size of 124×124×16. The first maximum pooling layer P ₁ , the input data is Y ¹ , the padding parameter P = 0, the window size of the pooling kernel U ¹ is 2×2, the sliding step size S ^P1 is 2, which is used to downsample the input data, and the size of the output feature map Y ² is 62×62×16; The second convolution layer L ₂ , the input data is Y ² , the window size of its convolution kernel K ² is 5×5, the sliding step size S ^L2 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 32 feature maps Y ³ , the size of Y ³ is 58×58×32; The second maximum pooling layer P ₂ , the input data is Y ³ , the padding parameter P = 0, the window size of the pooling kernel U ² is 2×2, the sliding step size S ^P2 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁴ is 29×29×32; The third convolution layer L ₃ , the input data is Y ⁴ , the window size of its convolution kernel K ³ is 6×6, the sliding step size S ^L3 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 64 feature maps Y ⁵ , the size of Y ⁵ is 24×24×64; The third maximum pooling layer P ₃ , the input data is Y ⁵ , the padding parameter P = 0, the window size of the pooling kernel U ³ is 2×2, the sliding step size S ^P3 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁶ is 12×12×64; The fourth convolution layer L ₄ , the input data is Y ⁶ , the window size of its convolution kernel K ⁴ is 3×3, the sliding step size S ^L4 is 1, the padding parameter P=0, which is used to perform convolution operation on the input data and output 128 feature maps Y ⁷ , the size of Y ⁷ is 10×10×128; The fourth maximum pooling layer P ₄ , the input data is Y ⁷ , the padding parameter P = 0, the window size of the pooling kernel U ⁴ is 2×2, the sliding step size S ^P4 is 2, which is used to downsample the input data, and the size of the output feature map Y ⁸ is 5×5×128; The fifth convolution layer L ₅ , the input data is Y ⁸ , the window size of its convolution kernel K ⁵ is 3×3, the sliding step size S ^L5 is 1, the padding parameter P=0, it is used to perform convolution operation on the input data, and output 64 feature maps Y ⁹ , the size of Y ⁹ is 3×3×64; The sixth convolution layer L ₆ , the input data is Y ⁹ , the window size of its convolution kernel K ⁶ is 3×3, the sliding step size S ^L6 is 1, the padding parameter P=0, it is used to perform a convolution operation on the input data and output a feature map Y ¹⁰ , the size of Y ¹⁰ is 1×1×1, and the preliminary target size is obtained. 6. The method according to claim 1 is characterized in that in (6), the training samples are input into the constructed target size estimation network framework Ψ for training, which is implemented as follows: 6a) Input the training sample into the target size estimation network framework Ψ and calculate the loss of the network output layer: loss = (y _true -y _pre ) ² Where y _true is the true size of the sample, and y _pre is the estimated size of the sample target; 6b) Use the back-propagation algorithm to propagate the loss of the output layer forward, and calculate the gradient vector of the loss function loss through the stochastic gradient descent algorithm to update the weights of each layer in the network; 6c) Repeat 6b) and iterate to continuously update the weights until the loss function converges to obtain a trained network model. 7. The method according to claim 1, characterized in that in (9), a gradient boosting decision tree (GBDT) classifier is trained using the multidimensional feature vector _h1 of the training sample, which is implemented as follows: 9a) Initialize the gradient boosting regression tree model f ₀ = 0; 9b) Train the gradient boosting regression tree model, assuming the number of cycles t = 1,…,10: 9b1) Taking the multi-dimensional feature h ₁ of the training sample as input, a trained regression tree model f _t is obtained by fitting the residual loss r _t-1 : r _t-1 = y _true -y _pre Where _ytrue is the true size of the sample, and _ypre is the estimated size of the target output by the regression tree model ft _-1 ; 9b2) Update the gradient boosted regression tree model _ft = ft _-1 + _νft , where ν is the learning rate; 9b3) Iterate repeatedly to continuously update the gradient boosting regression tree model until the number of iterations reaches t=10, and obtain the trained gradient boosting regression tree model ζ.

Images