A method for generating saliency feature maps of SAR images based on background awareness and multi-scale superpixel fusion

Abstract

This invention discloses a method for generating saliency feature maps (SMs) in SAR images based on background awareness and multi-scale superpixel fusion, belonging to the field of image processing technology. First, this invention identifies and constructs background reference superpixel sets at various scales. Then, based on the reference sets, it calculates global and local contrast to generate SMs at that scale. Finally, by fusing SMs from all scales, a comprehensive saliency feature map is formed. This invention introduces a multi-scale fusion strategy, making it more flexible in adapting to targets of different sizes in the scene and enhancing scale robustness; and by measuring the differences between regions and the background, it highlights the target and suppresses the background.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and particularly relates to the field of synthetic aperture radar image detection. Background Technology

[0002] Synthetic Aperture Radar (SAR) image interpretation typically involves three key stages: detection, identification, and classification. Among these, detection, as the foundation of the interpretation process, has a decisive impact on the accuracy and reliability of the entire SAR information extraction process. Currently, mainstream SAR image target detection methods can be mainly divided into three categories: methods based on the Constant False Alarm Rate (CFAR) strategy, methods based on deep learning, and methods based on visual saliency.

[0003] Deep learning-based methods demonstrate superior detection performance, but they typically rely on large-scale labeled data for model training, which limits their application in practical engineering. CFAR-like methods, as classic SAR detection techniques, are highly dependent on the matching degree of clutter modeling. With the improvement of imaging resolution and the increasing complexity of monitoring scenarios, accurate modeling of background clutter has become increasingly difficult, gradually limiting the applicability of CFAR methods. In contrast, saliency detection methods only require a single SAR image to achieve target detection, better meeting the flexibility requirements of practical applications. Furthermore, these methods do not rely on clutter modeling and can fully exploit the rich semantic information contained in SAR images based on the "feature integration" theory, demonstrating good adaptability and potential for data mining.

[0004] However, existing SAR image saliency detection methods mainly follow the field of computer vision, and their model designs often fail to take into account the unique imaging mechanisms and data attributes of SAR images. Due to the fundamental differences between optical images and SAR images in terms of imaging principles and information representation, this direct transfer strategy results in many saliency models performing worse than expected in real-world SAR scenarios.

[0005] In conclusion, to further improve the detection performance of SAR images, it is necessary to start from visual saliency, comprehensively consider the characteristics of SAR images themselves, and propose a more superior and applicable detection scheme. Summary of the Invention

[0006] In SAR monitoring scenarios, ground targets often exhibit significant scale differences, and single-scale processing is prone to performance loss due to scale mismatch. To address this, this invention introduces a multi-scale superpixel segmentation strategy. By generating superpixels at different scales, it more flexibly adapts to targets of different sizes in the scene, enhancing scale robustness. Secondly, targets typically possess high saliency due to the difference in scattering characteristics between themselves and the background. However, traditional methods often measure saliency by calculating the grayscale difference between a specific region and all other sub-regions of the entire image. This approach has two problems: first, it fails to exclude potential target components in other sub-regions, thus weakening the contrast of the true target; second, it deviates conceptually from the fundamental starting point of saliency—the difference between the target and the background.

[0007] To address the aforementioned issues, this invention proposes a background-aware mechanism. First, a set of background reference superpixels is identified and constructed at each scale. Then, global and local contrast is calculated based on the reference set to generate a Saliency Map (SM) for that scale. Finally, the SMs from all scales are fused to form a comprehensive saliency feature map.

[0008] Based on the above ideas, this invention adopts the following technical solution: a method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion. Here, scale represents the size of a single superpixel. Specifically, this invention includes the following steps:

[0009] Step 1: Filtering Preprocessing: Input the original SAR image I, denoted as I. The noise is smoothed by using Gaussian filtering to obtain the filtered image I. gau ;

[0010] Step 2: Use linear normalization to transform I gau Normalizing to the range [0, 1], we get As shown in the following formula, where, This represents the minimum value operator. This represents the maximum value operator; (1);

[0011] Step 3: Multi-scale superpixel segmentation: Select k scales {S1, S2, ..., S...} k}, for I at each scale gau Perform superpixel segmentation; where, at scale S i Next, 1≤i≤k, I gau Divided into N i Superpixels { , ,···, };

[0012] Step 4: Based on scale S i Frequency statistics are performed on the superpixel segmentation results, where 1 ≤ i ≤ k, and superpixels are extracted. frequency characteristics , , The total number of superpixels is represented by the following process:

[0013] 4-1) Divide the 256 gray levels from 0 to 255 into 64 non-overlapping gray level intervals, each with a length of 4, and denoted as: Itvl1, Itvl2, ..., Itvl 64 ;

[0014] 4-2) Based on I gau Statistical grayscale falls into Itvl m The number of pixels, denoted as frequency h. m , 1≤m≤64;

[0015] 4-3) Denote superpixels The number of pixels is , , pixels The frequency corresponding to the gray range that falls into the range is denoted as , Then the frequency corresponding to this superpixel is Calculate using the following formula (2);

[0016] (2);

[0017] 4-4) Normalized to (0, 1], , to obtain frequency features As shown in equation (3);

[0018] (3);

[0019] in, For sequence The maximum value in;

[0020] Step 5: Based on scale S i Background bias estimation is performed on the superpixel segmentation results, 1≤i≤k, and superpixels are extracted. Background deviation characteristics , The specific process is as follows:

[0021] 5-1) From the 64 grayscale intervals, select the T reference intervals with the highest frequency characteristics to estimate the background grayscale level, denoted as TT. The corresponding frequencies are denoted as follows: The median of each interval is denoted as follows: ;

[0022] 5-2) Calculate the estimated background gray level based on T reference intervals. As shown in equation (4);

[0023] (4);

[0024] 5-3) Calculate each superpixel The average gray level, , in order as ;as well as exist Corresponding area The average grayscale values ​​are denoted as follows: ;

[0025] 5-4) For each superpixel Calculate the difference between its grayscale mean and the estimated background grayscale level. As shown in equation (5); , (5); Step 5-5: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Normalizing to (0, 1] yields the background bias. As shown in equation (6); , (6);

[0026] in, For sequence The maximum value in;

[0027] Step 6: For scale S i The superpixel segmentation results are based on frequency features. and background deviation features Calculate superpixels Background score As shown below; , (7);

[0028] Step 7: Select the background score with the highest score. Each superpixel serves as a background superpixel, constituting scale S. i Background superpixel set ; exist The corresponding area is , Reference The elements are divided into There are 1 superpixel, denoted as follows: , and One-to-one correspondence, ;

[0029] Step 8: Calculation The average gray level of all pixels within the range is denoted as . ;

[0030] Step 9: Calculate superpixels Global contrast based on background distance The specific process is as follows:

[0031] 9-1) Calculation The average grayscale value of each superpixel in the image is denoted as { , , , , , }; express The grayscale value of the b-th superpixel within the range; calculate The average grayscale value of each superpixel in the image is denoted as { , , , , , }, express The grayscale value of the b-th superpixel within the range;

[0032] 9-2) Calculate superpixels The difference between the superpixel set and the background superpixel set is denoted as . As shown in equation (8) below,

[0033] (8);

[0034] in, The function is used to determine Is it positive? For positive, Take 1; It is not positive. Set to 0; These are empirical parameters; It is obtained in step 5;

[0035] 9-3) Calculate superpixels Corresponding background distance weight As shown in equation (9);

[0036] (9);

[0037] 9-4) Corresponding global contrast based on background distance weighting It can be obtained by the following formula (10);

[0038] (10);

[0039] Step 10: Determine superpixels Local reference superpixel set and The specific process is as follows:

[0040] 10-1) Calculation geometric center ,in, for The mean of all pixel row coordinates in the image. for The mean of all column coordinates of pixels in the image;

[0041] 10-2) with Construct a rectangle centered at the top left corner, with coordinates of the top left corner as follows: The coordinates of the top right vertex are The coordinates of the lower left vertex are The coordinates of the bottom right vertex are For S i Length of empirical reference window at scale;

[0042] 10-3) The superpixels that the four sides of the rectangle mentioned above pass through constitute the local reference superpixel set. , The elements in the middle are denoted as follows: , express The number of superpixels in the local reference superpixel set;

[0043] 10-4) exist The corresponding area is denoted as ,Will Reference The elements are divided into There are 1 superpixel, denoted as follows: ;

[0044] Step 11: Calculate superpixels Local contrast based on background distance weighting The specific process is as follows:

[0045] 11-1) Calculation The average gray value of each element in the array is denoted as follows: , express The Middle The average gray level of each element;

[0046] 11-2) Based on the local reference superpixel set, calculate Local contrast As shown in equation (11), where, It is obtained in step 5; The same as the empirical parameters in step 9;

[0047] (11); These are empirical parameters;

[0048] 11-3) Corresponding local contrast based on background distance weighting It can be obtained by the following formula (12);

[0049] (12);

[0050] Step 12: Calculate superpixels Corresponding significance value As shown in equation (13);

[0051] (13)

[0052] Step 13: Traverse Repeat steps 9-12 to calculate scale S. i The saliency values ​​of all superpixels are used to obtain the scale S. i The saliency feature map below ;

[0053] Step 14: Traverse all scales S i Repeat steps 4-13 to obtain the saliency feature map at each scale. The saliency feature maps at all scales are fused to obtain the final saliency feature map. As shown in equation (14).

[0054] , (14);

[0055] Given the current limitations of saliency models in SAR image detection performance, this invention proposes a method for generating superpixels (SMs) to enhance real targets, based on the inherent characteristics of SAR images and the principle of saliency. It introduces a multi-scale processing strategy, generating superpixels at different scales to more flexibly adapt to targets of varying sizes in the scene, thus enhancing scale robustness. Secondly, since targets typically possess high saliency due to differences in scattering characteristics between themselves and the background, this invention employs a contrast strategy based on adaptive background component perception, effectively highlighting targets while suppressing background noise. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the structure of the present invention.

[0057] Figure 2 These are the original SAR images from the SAR image set used in this experiment.

[0058] Figure 3 It corresponds to the target ground truth label.

[0059] Figure 4 This invention is aimed at Figure 2 The generated salient feature map.

[0060] Figure 5 The PR curves are generated for the saliency maps of this invention and four other classic models. Detailed Implementation

[0061] The core idea of ​​this invention is multi-scale fusion and background awareness. First, multi-scale superpixel segmentation is performed. Then, at a single scale, global and local contrast is calculated based on background awareness to generate a single-scale superpixel sensor (SM). Finally, the SMs from each scale are fused to output the final SM. To more clearly illustrate this invention, a detailed description is provided below with reference to the accompanying drawings.

[0062] The following points are worth noting during the implementation process:

[0063] 1) Gaussian filtering was used in step 1, and its formula is shown below.

[0064] (14);

[0065] in, This represents the coefficients at (x, y) on the Gaussian filter template; in practice, With a value of 1.5 and a template size of 7... 7, that is ;Will Substitute the value into the Gaussian filter formula and iterate through... Take the value and generate 7 Using a template of size 7, convolving the template with the input image I yields I. gau ;

[0066] 2) In step 1, when convolving the filter template with the input image I, to ensure that the result after convolution has the same size as the input image... The image needs to be expanded at the image boundary; here, the mirror fill method is used to expand I to... Size; thus, I gau The size will be consistent with I;

[0067] 3) In step 2, this implementation sets three scales: {20, 40, 80}, which correspond to small, medium, and large targets that may exist in the SAR scene, respectively; the superpixel segmentation method used is the classic SLIC (Simple Linear Iterative Clustering) algorithm.

[0068] 4) In step 2, after setting a certain scale, superpixel segmentation is performed at this scale. The actual size of the superpixel obtained will not be strictly equal to the preset scale, but will fluctuate around the scale. This is the result of the algorithm adaptive clustering based on local image features.

[0069] 5) In step 10, when constructing the rectangle, if the coordinates of the top-left vertex fall outside the image, then replace the point with the top-left vertex of the image, and use the coordinates of the bottom-right vertex and the top-left vertex of the image to determine the rectangle; if the coordinates of the top-right vertex fall outside the image, then replace the point with the top-right vertex of the image, and use the bottom-left vertex and the top-right vertex of the image to determine the rectangle; if the coordinates of the bottom-left vertex of the image fall outside the image, then replace it with the bottom-left vertex of the image, and use the coordinates of the top-right vertex and the bottom-left vertex of the image to determine the rectangle; if the coordinates of the bottom-right vertex fall outside the image, then replace it with the bottom-right vertex of the image, and use the coordinates of the top-left vertex and the bottom-right vertex of the image to construct the rectangle.

[0070] 6) In step 5, the number of reference intervals T is set to 10; in step 7, the number of background reference superpixels... Set as That is, 30% of the number of superpixels at the current scale; in steps 9 and 11, empirical parameters Set to 25; in step 10, empirical parameter The value range is usually the current scale S. i One to two times, in this implementation, make and scale S i Equivalent.

[0071] Based on the above precautions, refer to Figure 1 The present invention is implemented through the following steps:

[0072] Step 1: Filtering Preprocessing: Input the original SAR image I, denoted as I. The noise is smoothed by using Gaussian filtering to obtain the filtered image I. gau ;

[0073] Step 2: Use linear normalization to transform I gau Normalizing to the range [0, 1], we get As shown in the following formula, where, This represents the minimum value operator. This represents the maximum value operator;

[0074] ;

[0075] Step 3: Multi-scale superpixel segmentation: Select three scales {20, 40, 80}, and use the SLIC (Simple Linear Iterative Clustering) algorithm at each scale to segment I... gau Perform superpixel segmentation; where, at scale S i (1≤i≤3), I gau Divided into N i Superpixels { , ,···, };

[0076] Step 4: Based on scale S i Frequency statistics are performed on the superpixel segmentation results under (1≤i≤3) conditions to extract superpixels. frequency characteristics The specific process is as follows:

[0077] 4-1) Divide the 256 gray levels from 0 to 255 into 64 non-overlapping gray level intervals, each with a length of 4, and denoted as: Itvl1, Itvl2, ..., Itvl 64 ;

[0078] 4-2) Based on I gau Statistical grayscale falls into Itvl m The number of pixels, denoted as frequency h. m (1≤m≤64);

[0079] 4-3) Denote superpixels The number of pixels is , pixels The frequency corresponding to the gray range that falls into the range is denoted as Then the frequency corresponding to this superpixel is As shown in the following formula (2);

[0080] , ;

[0081] 4-4) Normalizing to (0, 1] yields the frequency features. As shown in the following formula, where, For sequence The maximum value in;

[0082] , ;

[0083] Step 5: Based on scale S i Background bias estimation is performed on the superpixel segmentation results under (1≤i≤3) conditions, and superpixels are extracted. Background deviation characteristics The specific process is as follows:

[0084] 5-1) From the 64 grayscale intervals, select the T reference intervals with the highest frequency characteristics to estimate the background grayscale level, denoted as TT. The corresponding frequencies are denoted as follows: The median of each interval is denoted as follows: ;

[0085] 5-2) Calculate the estimated background gray level based on T reference intervals. As shown in the following formula;

[0086] ;

[0087] 5-3) Calculate each superpixel The average grayscale values ​​are denoted as follows: ;as well as exist Corresponding area The average grayscale values ​​are denoted as follows: ;

[0088] 5-4) For each superpixel Calculate the difference between its grayscale mean and the estimated background grayscale level. As shown in the following formula;

[0089] , ;

[0090] 5-5) Normalizing to (0, 1] yields the background bias estimate. As shown in the following formula, where, For sequence The maximum value in;

[0091] , ;

[0092] Step 6: For scale S i Superpixel segmentation results under (1≤i≤3) based on frequency features and background deviation features Calculate superpixels Background score As shown below;

[0093] , ;

[0094] Step 7: Select the background score with the highest score. Each superpixel serves as a background superpixel, constituting scale S. i Background superpixel set under (1≤i≤3) ; exist The corresponding area is , Reference The elements are divided into There are 1 superpixel, denoted as follows: , and One-to-one correspondence;

[0095] Step 8: Calculation The average gray level of all pixels within the range is denoted as . ;

[0096] Step 9: Calculate superpixels Global contrast based on background distance The specific process is as follows:

[0097] 9-1) Calculation The average grayscale value of each superpixel in the image is denoted as { , , , };

[0098] 9-2) Calculate superpixels The difference between the superpixel set and the background superpixel set is denoted as . As shown in the following formula, where, The function is used to determine Is it positive? For positive, Take 1; It is not positive. Set to 0; These are empirical parameters; It is obtained in step 5;

[0099] ;

[0100] 9-3) Calculate superpixels Corresponding background distance weight As shown in the following formula;

[0101] ;

[0102] 9-4) Corresponding global contrast based on background distance weighting It can be obtained by the following formula;

[0103] ;

[0104] Step 10: Determine superpixels Local reference superpixel set and The specific process is as follows:

[0105] 10-1) Calculation geometric center ,in, for The mean of all pixel row coordinates in the image. for The mean of all column coordinates of pixels in the image;

[0106] 10-2) with Construct a rectangle centered at the top left corner, with coordinates of the top left corner as follows: The coordinates of the top right vertex are The coordinates of the lower left vertex are The coordinates of the bottom right vertex are For S i Length of empirical reference window at scale;

[0107] 10-3) The superpixels that the four sides of the rectangle mentioned above pass through constitute the local reference superpixel set. , The elements in the middle are denoted as follows: , express The number of superpixels in the local reference superpixel set;

[0108] 10-4) exist The corresponding area is denoted as ,Will Reference The elements are divided into There are 1 superpixel, denoted as follows: ;

[0109] Step 11: Calculate superpixels Local contrast based on background distance weighting The specific process is as follows:

[0110] 11-1) Calculation The average gray value of each element in the array is denoted as follows: ;

[0111] 11-2) Based on the local reference superpixel set, calculate Local contrast As shown in the following formula, where, It is obtained in step 5; The same as the empirical parameters in step 9;

[0112] ;

[0113] 11-3) Corresponding local contrast based on background distance weighting It can be obtained by the following formula;

[0114] ;

[0115] Step 12: Calculate superpixels Corresponding significance value As shown in the following formula;

[0116] ;

[0117] Step 13: Traverse Repeat steps 9-12 to calculate scale S. i The significance values ​​of all superpixels under the condition (1≤i≤3) are used to obtain the scale S. i Significance feature map ;

[0118] Step 14: Traverse all scales S i (1≤i≤3), repeat steps 4~13 to obtain the saliency feature map at each scale. The saliency feature maps at all scales are fused to obtain the final saliency feature map. As shown in the following formula.

[0119] , ;

[0120] The salient target highlighting effect of the present invention is further illustrated by the following simulation experiments.

[0121] Experimental Scenario: The experiment uses real SAR images from the publicly available MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset, specifically the BTR-60 (test) subset. This subset contains 192 images, each with a size of [missing information - likely a unit of measurement]. The resolution is 0.3m. All of these are actual scene images containing ground features. Figure 2 This example SAR image shows the target in the central highlighted area, with the remaining portion consisting of background and the target's shadow. Its corresponding target label (Ground Truth, GT) is as follows: Figure 3 As shown.

[0122] Experimental comparison method: The experiment introduced four classic significance models as comparison and reference, namely ITTI model, Spectral Residual (SR) model, Region Contrast (RC) model and Contour-Guided VisualSearch (CGVS) model.

[0123] Experimental results: The experimental results are as follows Figure 4 and Figure 5 As shown. Figure 4 This invention is based on Figure 2 The generated saliency feature map shows that the target region exhibits a high saliency value, while other regions are relatively thoroughly suppressed. This demonstrates that the present invention effectively highlights the target and suppresses the background. On the other hand, Figure 5 The Precision-Recall (PR) curves corresponding to the saliency map generation results of this invention and four classical models are described. The thickest solid line ('Ours') corresponds to this invention. The more the PR curve leans towards the upper right, the better the performance of its corresponding saliency feature map. The curve corresponding to this invention is significantly closer to the upper right corner compared to the other four classical models. Therefore, this invention exhibits superior performance compared to classical methods.

[0124] The above experiments show that the present invention can effectively enhance the target and suppress background clutter in real SAR scenarios. Its overall performance is better than many classic saliency models, demonstrating good advancement.

Claims

1. A method for generating saliency feature maps of SAR images based on background awareness and multi-scale superpixel fusion, characterized in that, The generation of salient feature maps (SMs) utilizes multi-scale superpixel segmentation and a background-aware contrastive strategy, specifically including the following steps: Step 1: Filtering Preprocessing: Input the original SAR image I, and use Gaussian filtering to smooth the noise in I, obtaining the filtered image I. gau ; Step 2: Use linear normalization to transform image I gau Normalizing to the range [0,1], we get As shown in the following formula, where, This represents the minimum value operator. This represents the maximum value operator; （1）； Step 3: Multi-scale superpixel segmentation: Select k scales {S1, S2, ..., S...} k }, for I at each scale gau Perform superpixel segmentation; where, at scale S i Next, 1≤i≤k, I gau Divided into N i Superpixels { , ,···, }; Step 4: Based on scale S i Frequency statistics are performed on the superpixel segmentation results, where 1 ≤ i ≤ k, and superpixels are extracted. frequency characteristics , ; Step 5: Based on scale S i Background bias estimation is performed on the superpixel segmentation results, and superpixels are extracted. Background deviation characteristics ; Step 6: For scale S i The superpixel segmentation results are based on frequency features. and background deviation features Calculate superpixels Background score ; Step 7: Select the background score with the highest score. Each superpixel serves as a background superpixel, constituting scale S. i Background superpixel set : ； turn up exist The corresponding region is defined as the normalized superpixel set. ;reference The method of dividing elements in the middle will Divided into There are 1 superpixel, denoted as follows: , and One-to-one correspondence, ; Step 8: Calculation The average gray level of all superpixels within the range is denoted as . ; Step 9: Calculate superpixels Global contrast based on background distance ; Step 10: Determine superpixels Locally referenced background superpixel set and normalized superpixel set ; Step 11: Calculate superpixels Local contrast based on background distance weighting ; Step 12: Calculate superpixels Corresponding significance value , ; Step 13: Traverse Repeat steps 9-12 to calculate scale S. i The saliency values ​​of all superpixels are used to obtain the scale S. i Significance feature map ; Step 14: Traverse all scales S i Repeat steps 4-13 to obtain the saliency feature map at each scale. , scale k The saliency feature maps at all scales are merged to obtain the final saliency feature map. .

2. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 1, characterized in that, The specific method for step 4 is as follows: Step 4-1: Divide the 256 gray levels (0-255) into 64 non-overlapping gray level intervals, each with a length of 4, and label them as Itvl1, Itvl2, ..., Itvl... 64 ; Step 4-2: Based on the filtered image The number of pixels whose grayscale values ​​fall within a specified area is counted and recorded as the frequency h. m , 1≤m≤64; Step 4-3: Record superpixels The number of pixels is , , Represents the total number of superpixels, pixels The frequency corresponding to the gray range that falls into the range is denoted as , Then the frequency corresponding to this superpixel is Calculate using the following formula (2); （2）； in These are the length and width of the original image, respectively; Step 4-4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Normalized to (0, 1], , to obtain frequency features As shown in equation (3); （3）； in, For sequence The maximum value in.

3. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 2, characterized in that, The specific method for step 5 is as follows: Step 5-1: From the 64 grayscale intervals, select the T reference intervals with the highest frequency features to estimate the background grayscale level, denoted as TT. Each element represents the background grayscale value of the corresponding reference interval; the corresponding frequencies are denoted as follows: ,in Indicates the corresponding reference interval The number of times the background grayscale appears; the median of each interval is denoted as follows. ,in Indicates the corresponding reference interval The median gray level; Step 5-2: Calculate the background gray level based on T reference intervals. As shown in equation (4); （4）； Step 5-3: Calculate each superpixel grayscale mean , , in order as ;as well as exist Corresponding area grayscale mean , in order as ; Step 5-4: For each superpixel Calculate its mean gray level and background gray level. Valuation differences As shown in equation (5); ， （5）； Step 5-5: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Normalizing to (0, 1] yields the background bias. As shown in equation (6); ， （6）； in, For sequence The maximum value in.

4. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 3, characterized in that, Background score in step 6 The calculation method is as follows: ， （7）； Indicates frequency characteristics.

5. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 4, characterized in that, The specific method for step 9 is as follows: Step 9-1: Calculation The average grayscale value of each background superpixel in the image is denoted as { , , , , , }, express The grayscale value of the b-th superpixel within the range; calculate The average grayscale value of each superpixel in the image is denoted as { , , , , , }, express The grayscale value of the b-th superpixel within the range; Step 9-2: Calculate superpixels The difference between the superpixel set and the background superpixel set is denoted as . As shown in equation (8) below, （8）； in, These are empirical parameters. This represents the total number of superpixels in the partition. The result obtained in step 5 exist Corresponding area The average gray level; The function is used to determine Is it positive? For positive, This indicates that the value is 1; It is not positive. Set to 0; Step 9-3: Calculate superpixels Corresponding background distance weight As shown in equation (9); （9）； 9-4) Corresponding global contrast based on background distance weighting It can be obtained by the following formula (10); （10）。 6. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 5, characterized in that, The specific method for step 10 is as follows: Step 10-1: Calculation geometric center ,in, for The mean of all pixel row coordinates in the image. for The mean of all column coordinates of pixels in the image; Step 10-2: with Construct a rectangle centered at the top left corner, with coordinates of the top left corner as follows: The coordinates of the top right vertex are The coordinates of the lower left vertex are The coordinates of the bottom right vertex are For S i Length of empirical reference window at scale; Step 10-3: The superpixels crossed by the four sides of the rectangle in Step 10-2 constitute a local reference superpixel set. , The elements in the middle are denoted as follows: , express The number of superpixels in the local reference superpixel set; Step 10-4: exist The corresponding area is denoted as ,Will Reference The elements are divided into There are 1 superpixel, denoted as follows: .

7. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 6, characterized in that, The specific method for step 11 is as follows: Step 11-1: Calculation The average gray value of each element in the array is denoted as follows: , express The Middle The average gray level of each element; Step 11-2: Calculate based on the local reference superpixel set. Local contrast As shown in equation (11), where, It is obtained in step 5; The same as the empirical parameters in step 9; （11）； Step 11-3: Corresponding local contrast based on background distance weighting It can be obtained by the following formula (12); （12）； The background distance weights are obtained in step 9.

8. The method for generating SAR image saliency feature maps based on background awareness and multi-scale superpixel fusion according to claim 7, characterized in that, In step 14, the final saliency feature map is calculated. The specific method is as follows: ， ； Here, min() means taking the smallest element in the parentheses, and max() means taking the largest element in the parentheses.

Images