Ship recognition method in SAR images based on non-stationary feature extraction and deep learning

Abstract

The present invention discloses a SAR image ship recognition method based on non-stationary feature extraction and deep learning, which belongs to the field of synthetic aperture radar image processing technology. Firstly, the SAR image of the ship is preprocessed, and the second-order synchronous compression transform (2 nd Fourier Transform‑based Synchrosqueezing Transform, FSST2) performs time-frequency analysis on the preprocessed ship SAR image to generate a time-frequency feature space diagram describing the non-stationary characteristics of the ship; secondly, extract the one-dimensional HOG feature value from the time-frequency feature space diagram of the ship obtained in step 1, and then reshape the extracted one-dimensional HOG feature value to obtain the final two-dimensional HOG feature; then input the two-dimensional HOG feature into the deep learning classifier for classification to realize the recognition of the ship type; finally, the classification algorithm of step 3 is evaluated for performance. The method provided by the present invention not only improves the accuracy of SAR ship classification by integrating the advantages of non-stationary features and deep learning, but also enhances the robustness and generalization ability of the model.

Description

Technical Field

[0001] The invention belongs to the technical field of synthetic aperture radar (SAR) image processing, and in particular relates to a SAR image ship recognition method based on non-stationary feature extraction and deep learning. Background Art

[0002] As an advanced remote sensing imaging technology, synthetic aperture radar (SAR) is widely used in ocean surveillance, environmental monitoring, military reconnaissance and other fields because it can provide high-quality images under various weather conditions. SAR transmits electromagnetic waves and receives the signals reflected by the target, and synthesizes high-resolution images using the principle of synthetic aperture. However, the interpretation and analysis of SAR images face many challenges, especially in the task of ship classification.

[0003] In the field of marine surveillance, ship classification is crucial for maritime traffic management, fishery resource assessment, environmental monitoring, and safety assurance. Traditional SAR ship classification methods rely on hand-crafted features such as ship shape, size, texture, etc., which are extracted from SAR images using image processing techniques. Subsequently, these features are used to train machine learning models such as support vector machines and decision trees to achieve automatic recognition and classification of ships. Although these methods have achieved some success in some cases, they usually require a lot of domain knowledge and human intervention, and their performance and robustness are limited when dealing with complex and changing marine scenes.

[0004] With the rapid development of deep learning technology, especially the breakthrough of convolutional neural network in the field of image recognition, researchers have begun to explore the application of deep learning to automatic feature extraction and classification of SAR images. Convolutional neural network can automatically learn complex feature representations from a large amount of data, thereby improving the accuracy of classification. However, most existing deep learning-based SAR ship classification methods rely too much on abstract feature representations, which may lead to a lack of interpretability of the model and limit its further application in practical applications.

[0005] SAR images have unique non-stationary characteristics, which are the result of the combined effect of the SAR imaging mechanism and the physical properties of the target. These non-stationary characteristics are often ignored in traditional image processing methods, but may contain important information in the ship classification task. For example, the reflection characteristics of electromagnetic waves from different parts of a ship (such as the hull, deck, and mast) may change with time and observation angle, and these changes appear as non-stationary signals in SAR images. Therefore, how to effectively extract and utilize these non-stationary characteristics is crucial to improving the performance of SAR ship classification. Summary of the invention

[0006] The purpose of the present invention is to provide a SAR image ship recognition method based on non-stationary feature extraction and deep learning, so as to solve the problems existing in the prior art that traditional SAR ship classification mainly relies on manually designed features and traditional machine learning algorithms, resulting in limitations in processing complex ocean scenes, and the SAR image target recognition method based on deep learning method makes less use of the non-stationary characteristics of the target.

[0007] To achieve the above object, the present invention provides a SAR image ship recognition method based on non-stationary feature extraction and deep learning, comprising the following steps:

[0008] Step 1: Preprocess the SAR image of the ship using the second-order synchronous compression transform based on Fourier transform (2 nd Fourier Transform-based Synchrosqueezing Transform, FSST2) performs time-frequency analysis on the preprocessed ship SAR image to generate a time-frequency feature space diagram that describes the non-stationary characteristics of the ship;

[0009] Step 2: extracting one-dimensional HOG feature values ​​from the time-frequency feature space diagram of the ship obtained in step 1, and then reshaping the extracted one-dimensional HOG feature values ​​to obtain the final two-dimensional HOG features;

[0010] Step 3: Input the two-dimensional HOG features into the deep learning classifier for classification to realize the identification of ship types;

[0011] Step 4: Evaluate the performance of the classification algorithm in step 3.

[0012] Preferably, the preprocessing of the SAR image of the ship in step 1 includes using a median filter to remove salt and pepper noise, using a filter such as a multiscale wavelet filter to remove multiplicative noise, and unifying the size of the SAR image by a reshape function.

[0013] Preferably, in step 1, the specific process of using FSST2 to perform time-frequency analysis on the preprocessed ship SAR image to generate a time-frequency feature space diagram describing the non-stationary characteristics of the ship is as follows:

[0014] S11, for each line signal {x m (n)} to perform FSST2 to capture the instantaneous frequency change of the signal and obtain the corresponding time-frequency spectrum. The specific expression is as follows:

[0015]

[0016] Where w∈{1,2,…,N} is the frequency, m represents the row number of the SAR image matrix, n represents the column number of the SAR image matrix, δ(·) represents the unit impulse function, and the signal {x m (n)}Short-time Fourier Transform (STFT) for g(·) represents the window function, N represents the number of columns of the SAR image matrix, and the second-order instantaneous frequency estimation operator ω2(n,h) is:

[0017]

[0018] as well as:

[0019]

[0020] Among them, ω1(n,h) represents the first-order instantaneous frequency estimation operator, q(n,h) and are new variable symbols introduced in the calculation of the second-order instantaneous frequency estimation operator. Ro{.} represents the rounding operator, Re{Z} and Im{Z} represent the real part and imaginary part of the complex number Z, respectively. and Respectively represent the signal {x m (n)} STFT when the window functions are g′(x), g″(x), xg(x) and xg′(x);

[0021] S12. Extract key time-frequency features, including instantaneous frequency and amplitude, from the results of FSST2 and map them onto a two-dimensional plane to form a time-frequency feature space diagram.

[0022] Preferably, the specific process of step 2 is as follows:

[0023] S21, image standardization: perform global contrast normalization on the time-frequency feature space map of the non-stationary characteristics generated by FSST2 to eliminate the contrast differences between different images and use contrast-limited adaptive histogram equalization to highlight the local features in the image;

[0024] S22. Perform gradient calculation on the slice of the standardized non-stationary characteristic time-frequency feature space graph, including calculating the gradient amplitude G(m,n) and direction θ(m,n) of each pixel point. The specific expression is as follows:

[0025]

[0026] In the formula, G x (m,n) represents the difference of the pixel along the x-axis, G y (m,n) represents the difference of the pixel along the y-axis;

[0027] S23, dividing the time-frequency feature space map into cells of equal size, each cell containing a sufficient number of pixels to capture local texture information, grouping the pixels in each cell according to the gradient direction, and calculating the gradient magnitude of each group to construct a gradient histogram;

[0028] S24, combining adjacent cells into blocks, each block containing a number of cells, extracting texture features in a large area, and normalizing the cell histogram in each block to reduce the impact of changes in illumination and shadow;

[0029] S25, concatenate the descriptors of all blocks into a long vector to form a one-dimensional HOG feature value;

[0030] S26. Reshape the one-dimensional HOG feature value to obtain the final two-dimensional HOG feature.

[0031] Preferably, the specific process of step 3 is as follows:

[0032] S31. Structural design of deep learning classifier: Design a deep learning network model, including multiple convolutional layers, pooling layers and fully connected layers, to automatically learn the hierarchical features of the image, and in the final stage of the network, add a fully connected layer to predict the ship category;

[0033] S32, use the two-dimensional HOG features obtained in S26 to perform end-to-end training on the deep learning network model, wherein the cross entropy loss function is used as the training objective to minimize the difference between the predicted category and the actual category. The specific expression is as follows:

[0034]

[0035] In the formula, L represents the number of samples, C represents the total number of categories, and y i,j represents the label of the i-th sample on the j-th category (usually takes a value of 0 or 1, used to indicate whether the sample belongs to the j-th category), Represents the predicted label of the i-th sample for the j-th category;

[0036] S33. Use the Adam optimizer to perform gradient descent, quickly converge and find the optimal solution to identify the type of ship.

[0037] Preferably, the specific process of step 4 is as follows:

[0038] S41. Evaluate the performance of the classification algorithm on an independent test set. Use the evaluation indicators accuracy, recall rate, precision, F1 measure, and LR+ index to comprehensively evaluate the classification effect of the classification algorithm. The specific expressions are as follows:

[0039] The evaluation index Accuracy is defined by the following formula:

[0040]

[0041] Among them, #TP represents the number of true positive samples, #TN represents the number of false positive samples, #FN represents the number of false negative samples, #TN represents the number of true negative samples, the number of correct ship classifications is #TP+#TN, and the total number of all ships is #TP+#TN+#FP+#FN;

[0042] The recall rate Recall is defined as follows:

[0043]

[0044] Precision is defined as follows:

[0045]

[0046] The F1 measure is defined as follows:

[0047]

[0048] LR+ is defined as follows:

[0049]

[0050] S42, draw the ROC curve and calculate the AUC value to evaluate the performance of the model at different thresholds;

[0051] S43. Perform confusion matrix analysis to identify the performance of the model on specific categories and perform targeted optimization.

[0052] Therefore, the present invention adopts the above-mentioned SAR image ship recognition method based on non-stationary feature extraction and deep learning. First, FSST2 is used to perform time-frequency analysis on SAR images to capture the non-stationary characteristics of ships; then, the gradient histogram feature extraction is used to further enhance the description ability of ship shape and texture; finally, deep learning networks such as VGG, Mobilenet, Densenet, Efficientnet and Resnet are combined to achieve efficient classification of ships in SAR images; the proposed method not only improves the accuracy of SAR ship classification, but also enhances the robustness, generalization ability and interpretability of the model by integrating the advantages of non-stationary features and deep learning, providing a new perspective for SAR image analysis.

[0053] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1It is an overall flow chart of the SAR image ship recognition method based on non-stationary feature extraction and deep learning of the present invention;

[0055] Figure 2 are the original SAR images of the same target in the embodiment of the present invention and their corresponding non-stationary characteristic space spectrograms; wherein (a) is the original SAR image of the fishing boat, (b) is the characteristic space spectrogram of the fishing boat, (c) is the original SAR image of the passenger ship, and (d) is the characteristic space spectrogram of the passenger ship;

[0056] Figure 3 1 is a comparison diagram of ROC curves of deep learning ship classification and identification with and without non-stationary feature extraction on different data sets in the embodiment of the present invention; wherein (a) is a curve comparison diagram on the OpenSARShip data set, and (b) is a curve comparison diagram on the FuSARShip data set;

[0057] Figure 4 1 is a ROC curve diagram of different time-frequency analysis methods on the OpenSARShip data set of an embodiment of the present invention; wherein, (a) is a ROC curve diagram of the STFT time-frequency analysis method, (b) is a ROC curve diagram of the CWT time-frequency analysis method, (c) is a ROC curve diagram of the MSST time-frequency analysis method, and (d) is a ROC curve diagram of the FSST2 time-frequency analysis method;

[0058] Figure 5 1 is a ROC curve diagram of different time-frequency analysis methods on the FUSARShip data set of an embodiment of the present invention; wherein, (a) is a ROC curve diagram of the STFT time-frequency analysis method, (b) is a ROC curve diagram of the CWT time-frequency analysis method, (c) is a ROC curve diagram of the MSST time-frequency analysis method, and (d) is a ROC curve diagram of the FSST2 time-frequency analysis method. DETAILED DESCRIPTION

[0059] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] See also Figure 1-5 , a SAR image ship recognition method based on non-stationary feature extraction and deep learning includes the following steps:

[0061] Step 1: Preprocess the SAR image of the ship, where the preprocessing includes using filters such as median filters and multiscale wavelet filters to remove multiplicative noise and unifying the size of the SAR image through the reshape function. Use FSST2 to perform time-frequency analysis on the preprocessed ship SAR image to generate a time-frequency feature space diagram that describes the non-stationary characteristics of the ship. The specific process is as follows:

[0062] S11, for each line signal {x m (n)} to perform FSST2 to capture the instantaneous frequency change of the signal and obtain the corresponding time-frequency spectrum. The specific expression is as follows:

[0063]

[0064] Where w∈{1,2,…,N} is the frequency, m represents the row number of the SAR image matrix, n represents the column number of the SAR image matrix, δ(·) represents the unit impulse function, and the signal {x m (n)}Short-time Fourier transform for g(·) represents the window function, N represents the number of columns of the SAR image matrix, and the second-order instantaneous frequency estimation operator ω2(n,h) is:

[0065]

[0066] as well as:

[0067]

[0068] Among them, ω1(n,h) represents the first-order instantaneous frequency estimation operator, q(n,h) and are new variable symbols introduced in the calculation of the second-order instantaneous frequency estimation operator. Ro{.} represents the rounding operator, Re{Z} and Im{Z} represent the real part and imaginary part of the complex number Z, respectively. and Respectively represent the signal {x m (n)} STFT when the window functions are g′(x), g″(x), xg(x) and xg′(x);

[0069] S12. Extract key time-frequency features, including instantaneous frequency and amplitude, from the results of FSST2 and map them onto a two-dimensional plane to form a time-frequency feature space diagram.

[0070] Step 2: Extract one-dimensional HOG feature values ​​from the time-frequency feature space diagram of the ship obtained in step 1, and then reshape the extracted one-dimensional HOG feature values ​​to obtain the final two-dimensional HOG features; the specific process is as follows:

[0071] S21, image standardization: perform global contrast normalization on the time-frequency feature space map of the non-stationary characteristics generated by FSST2 to eliminate the contrast differences between different images and use contrast-limited adaptive histogram equalization to highlight the local features in the image;

[0072] S22. Perform gradient calculation on the slice of the standardized non-stationary characteristic time-frequency feature space graph, including calculating the gradient amplitude G(m,n) and direction θ(m,n) of each pixel point. The specific expression is as follows:

[0073]

[0074] In the formula, G x (m,n) represents the difference of the pixel along the x-axis, G y (m,n) represents the difference of the pixel along the y-axis;

[0075] S23, dividing the time-frequency feature space map into cells of equal size, each cell containing a sufficient number of pixels to capture local texture information, grouping the pixels in each cell according to the gradient direction, and calculating the gradient magnitude of each group to construct a gradient histogram;

[0076] S24, combining adjacent cells into blocks, each block containing a number of cells, extracting texture features in a large area, and normalizing the cell histogram in each block to reduce the impact of changes in illumination and shadow;

[0077] S25, concatenate the descriptors of all blocks into a long vector to form a one-dimensional HOG feature value;

[0078] S26. Reshape the one-dimensional HOG feature value to obtain the final two-dimensional HOG feature.

[0079] Step 3: Input the two-dimensional HOG features into the deep learning classifier for classification to realize the identification of ship types; the specific process is as follows:

[0080] S31. Structural design of deep learning classifier: Design a deep learning network model, including multiple convolutional layers, pooling layers and fully connected layers, to automatically learn the hierarchical features of the image, and in the final stage of the network, add a fully connected layer to predict the ship category;

[0081] S32, use the two-dimensional HOG features obtained in S26 to perform end-to-end training on the deep learning network model, wherein the cross entropy loss function is used as the training objective to minimize the difference between the predicted category and the actual category. The specific expression is as follows:

[0082]

[0083] In the formula, L represents the number of samples, C represents the total number of categories, and y i,j represents the label of the i-th sample on the j-th category (usually takes a value of 0 or 1, used to indicate whether the sample belongs to the j-th category), Represents the predicted label of the i-th sample for the j-th category;

[0084] S33. Use the Adam optimizer to perform gradient descent, quickly converge and find the optimal solution to identify the type of ship.

[0085] Step 4: Evaluate the performance of the classification algorithm in step 3. The specific process is as follows:

[0086] S41. Evaluate the performance of the classification algorithm on an independent test set. Use the evaluation indicators accuracy, recall rate, precision, F1 measure, and LR+ index to comprehensively evaluate the classification effect of the classification algorithm. The specific expressions are as follows:

[0087] The evaluation index Accuracy is defined by the following formula:

[0088]

[0089] Among them, #TP represents the number of true positive samples, #TN represents the number of false positive samples, #FN represents the number of false negative samples, #TN represents the number of true negative samples, the number of correct ship classifications is #TP+#TN, and the total number of all ships is #TP+#TN+#FP+#FN;

[0090] The recall rate Recall is defined as follows:

[0091]

[0092] Precision is defined as follows:

[0093]

[0094] The F1 measure is defined as follows:

[0095]

[0096] LR+ is defined as follows:

[0097]

[0098] S42, draw the ROC curve and calculate the AUC value to evaluate the performance of the model at different thresholds;

[0099] S43. Perform confusion matrix analysis to identify the performance of the model on specific categories and perform targeted optimization.

[0100] Example

[0101] The overall process of the SAR image ship recognition method based on non-stationary feature extraction and deep learning is as follows:

[0102] (1) Data preparation

[0103] (1.1) Download and prepare SAR image datasets, including the OpenSARShip and FUSAR-Ship public datasets. These datasets should contain SAR images of various types of ships under different conditions, providing basic data for subsequent feature extraction and classification.

[0104] (1.2) Preprocess the dataset, including image cropping, size standardization, and denoising, to ensure data quality and lay a solid foundation for subsequent analysis.

[0105] (2) FSST2 implementation steps

[0106] according to Figure 1 After the SAR image is preprocessed in the flowchart shown, the specific steps of step (2) are:

[0107] (2.1) The preprocessed SAR image is processed by FSST2 to generate a time-frequency feature space diagram, which provides a time-frequency domain representation for the feature extraction of the ship.

[0108] (2.2) Select appropriate window functions and frequency resolution to optimize the quality of time-frequency representation and computational efficiency, ensuring the accuracy of time-frequency features and the feasibility of computation.

[0109] (2.3) Extract key time-frequency features from the time-frequency feature space diagram. These features will be used for subsequent gradient histogram feature extraction to provide important time-frequency information for ship classification.

[0110] according to Figure 1 The flowchart shown selects the time-frequency slice in the time-frequency feature space diagram. Figure 2 The original SAR images of fishing boats and passenger ships and the corresponding non-stationary feature spatial spectrum results. It can be seen from the figure that compared with the original SAR images, the spatial spectrum obtained after non-stationary feature extraction can describe the shape and texture of the target more finely and clearly, which verifies the effectiveness of the proposed method.

[0111] (3) Implementation of gradient histogram feature extraction steps

[0112] according to Figure 1 After obtaining the non-stationary feature space spectrum in the flowchart shown, the specific steps of step (3) are:

[0113] (3.1) The time-frequency feature space map generated by FSST2 is normalized to eliminate the influence of illumination and contrast and improve the stability and comparability of the features.

[0114] (3.2) The gradient magnitude and direction of each pixel are calculated to capture local texture information and provide a basis for texture feature extraction of the ship.

[0115] (3.3) The image is divided into cells and a gradient histogram is constructed in each cell to capture the local texture features of the image.

[0116] (3.4) Adjacent cells are grouped into blocks and block descriptors are calculated to extract texture features within a larger area.

[0117] (3.5) Generate the final gradient histogram feature vector and convert it into a two-dimensional feature through two-dimensional reshaping as the input of the ship classifier, providing rich feature information for the classifier.

[0118] (4) Deep Learning Classifier Implementation

[0119] according to Figure 1 After the gradient histogram features are extracted in the flowchart shown, the specific steps of step (4) are:

[0120] (4.1) Use the two-dimensional feature dataset obtained in the previous steps to train the ResNet model so that the model can learn effective features from the data.

[0121] (4.2) Select a suitable loss function (cross entropy) and optimizer (Adam), apply the CosineAnnealingLR learning rate scheduler, and use 0.0005 as the benchmark to smoothly reduce the learning rate during training to improve the generalization ability of the model.

[0122] (4.3) During the training process, in order to further optimize the model performance, the training batch size was set to a more appropriate 32 based on experience. The validation set was used to evaluate the model performance and make necessary hyperparameter adjustments to improve the generalization ability and classification performance of the model.

[0123] (5) Model evaluation and optimization

[0124] After comparing and analyzing the experimental results, the specific steps of step (5) are:

[0125] (5.1) The performance of the trained ResNet model is evaluated on the OpenSARship and FuSARship test sets, and the classification effect of the model is comprehensively evaluated using indicators such as accuracy, recall, precision, and F1 measure.

[0126] (5.2) Based on the evaluation results, the model structure or training strategy is adjusted to further improve the classification performance and ensure the accuracy and robustness of the model.

[0127] (5.3) Figure 3 Table 1 and Table 2 respectively give the ROC curve comparison and performance index comparison results of the ship classification experiment results with and without the non-stationary feature extraction step on different data sets.

[0128] Table 1 Comparison of performance indicators of different methods

[0129]

[0130] from Figure 3 As can be seen from Table 1, the classification effect of the input deep learning network model using FSST2 combined with gradient histogram for non-stationary feature extraction is significantly improved compared with the model without non-stationary feature extraction. On the OpenSARShip and FuSARShip datasets, the accuracy is increased by 15% to 20% and 17% to 21%, respectively, which verifies the effectiveness of non-stationary feature extraction in the field of SAR image ship recognition, that is, the effectiveness of the proposed method.

[0131] (5.4) To further verify the superiority of FSST2 in non-stationary feature extraction, Figure 4 , Figure 5 Table 2 and Table 3 give the experimental results comparison of STFT, Continuous Wavelet Transform (CWT), Multisynchrosqueezing Transform (MSST) and FSST2 on the OpenSARShip and FuSARShip datasets respectively.

[0132] Table 2 Comparison of performance indicators obtained by extracting non-stationary features using different time-frequency analysis methods

[0133]

[0134] Depend on Figure 4 , Figure 5As shown in Table 1 and Table 2, on the OpenSARShip and FuSARShip datasets, the average accuracy of no time-frequency analysis, CWT, STFT, MSST and FSST2 are 72% / 76%, 74% / 77%, 76% / 81%, 76% / 80% and 82% / 90% respectively. On the OpenSARShip and FuSARShip datasets, the classification accuracy of using FSST2 for time-frequency analysis, that is, non-stationary feature extraction, is 2%~10% / 1%~14% higher than that of not performing non-stationary feature extraction and other time-frequency analysis methods, which verifies that the model has good generalization ability and provides reliable support for practical applications.

[0135] Therefore, the present invention adopts the above-mentioned SAR image ship recognition method based on non-stationary feature extraction and deep learning, which not only improves the accuracy of SAR ship classification, but also enhances the robustness, generalization ability and interpretability of the model by integrating the advantages of non-stationary features and deep learning, providing a new perspective for SAR image analysis.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A SAR image ship recognition method based on non-stationary feature extraction and deep learning, characterized in that: The following steps are involved: Step 1: Preprocess the SAR image of the ship, and use the second-order synchronous compression transform based on Fourier transform to perform time-frequency analysis on the preprocessed ship SAR image to generate a time-frequency feature space diagram describing the non-stationary characteristics of the ship; the specific process is as follows: S11. For each line of signal in the SAR image Perform FSST2 to capture the instantaneous frequency change of the signal and obtain the corresponding time-frequency spectrum. The specific expression is as follows: ； in is the frequency, Indicates the row number of the SAR image matrix, Represents the column number of the SAR image matrix, represents the unit impulse function, signal The short-time Fourier transform for , represents the window function, Represents the number of columns of the SAR image matrix, the second-order instantaneous frequency estimation operator for: ； as well as: ； ； ； in, represents the first-order instantaneous frequency estimation operator, Ro{.} represents the rounding operator, Re{} and Im{} represent the real part and imaginary part of the complex number respectively, , , and Respectively represent signals The window functions are , , as well as STFT in the case; S12, extract key time-frequency features from the results of FSST2, including instantaneous frequency and amplitude, and map them onto a two-dimensional plane to form a time-frequency feature space graph; Step 2: extracting one-dimensional HOG feature values ​​from the time-frequency feature space diagram of the ship obtained in step 1, and then reshaping the extracted one-dimensional HOG feature values ​​to obtain the final two-dimensional HOG features; Step 3: Input the two-dimensional HOG features into the deep learning classifier for classification to realize the identification of ship types; Step 4: Evaluate the performance of the classification algorithm in step 3.

2. The SAR image ship recognition method based on non-stationary feature extraction and deep learning according to claim 1 is characterized in that: In step 1, the SAR image of the ship is preprocessed by using a median filter and a multiscale wavelet filter to remove multiplicative noise and unifying the size of the SAR image by a reshape function.

3. The SAR image ship recognition method based on non-stationary feature extraction and deep learning according to claim 2 is characterized in that: The specific process of step 2 is as follows: S21, image standardization: perform global contrast normalization on the time-frequency feature space map of the non-stationary characteristics generated by FSST2 to eliminate the contrast differences between different images and use contrast-limited adaptive histogram equalization to highlight the local features in the image; S22, perform gradient calculation on the section of the time-frequency feature space graph of the standardized non-stationary characteristics, including calculating the gradient amplitude of each pixel point and direction , the specific expression is as follows: ； ； In the formula, Indicates that the pixel is along The difference of the axes, Indicates that the pixel is along Differentiation of axes; S23, dividing the time-frequency feature space map into cells of equal size, each cell containing a sufficient number of pixels to capture local texture information, grouping the pixels in each cell according to the gradient direction, and calculating the gradient magnitude of each group to construct a gradient histogram; S24, combining adjacent cells into blocks, each block containing a number of cells, extracting texture features in a large area, and normalizing the cell histogram in each block to reduce the impact of changes in illumination and shadow; S25, concatenate the descriptors of all blocks into a long vector to form a one-dimensional HOG feature value; S26. Reshape the one-dimensional HOG feature value to obtain the final two-dimensional HOG feature.

4. The SAR image ship recognition method based on non-stationary feature extraction and deep learning according to claim 3 is characterized in that: The specific process of step 3 is as follows: S31. Structural design of deep learning classifier: Design a deep learning network model, including multiple convolutional layers, pooling layers and fully connected layers, to automatically learn the hierarchical features of the image, and in the final stage of the network, add a fully connected layer to predict the ship category; S32, use the two-dimensional HOG features obtained in S26 to perform end-to-end training on the deep learning network model, wherein the cross entropy loss function is used as the training objective to minimize the difference between the predicted category and the actual category. The specific expression is as follows: ； In the formula, represents the number of samples, C represents the total number of categories, Indicates Sample pair The labels on the categories, Indicates Sample pair The predicted labels for each category; S33. Use the Adam optimizer to perform gradient descent, quickly converge and find the optimal solution to identify the type of ship.

5. The SAR image ship recognition method based on non-stationary feature extraction and deep learning according to claim 4 is characterized in that: The specific process of step 4 is as follows: S41. Evaluate the performance of the classification algorithm on an independent test set. Use the evaluation indicators accuracy, recall rate, precision, F1 measure, and LR+ index to comprehensively evaluate the classification effect of the classification algorithm. The specific expressions are as follows: Evaluation index accuracy Defined by: ； Among them, #TP represents the number of true positive samples, #TN represents the number of false positive samples, #FN represents the number of false negative samples, #TN represents the number of true negative samples, the number of correct ship classifications is #TP + #TN, and the total number of all ships is #TP+#TN+#FP+#FN; Recall The definition is as follows: ； Accuracy The definition is as follows: ； The F1 measure is defined as follows: ； LR+ is defined as follows: ； S42, draw the ROC curve and calculate the AUC value to evaluate the performance of the model at different thresholds; S43. Perform confusion matrix analysis to identify the performance of the model on specific categories and perform targeted optimization.

Images