Feature weight adaptive superpixel segmentation method based on sonar image

Abstract

The invention provides a feature weight adaptive super-pixel segmentation method based on a sonar image. The method comprises the following steps: Step 1, converting an original sonar image into a grayscale image; step 2, denoising processing is carried out on the image by using guide filtering; step 3, extracting specific features of the sonar image; step 4, carrying out superpixel segmentation preprocessing of automatic adjustment of feature weights; step 5, carrying out maximum flow-minimum cut superpixel aggregation; step 6, carrying out aggregation on the super pixels through a preset threshold value; and Step 7, separating the target from the background to obtain a segmentation result. According to the method, a segmentation technology suitable for the target object area of the acoustic image can be realized, and a better image foundation is laid for subsequent target object disease type judgment and volume and area measurement through analysis of the segmented area, so that segmentation is more accurate, the type of the target object is easier to judge, and the size of the disease is more accurately measured.

Description

Technical Field

[0001] The invention relates to the technical field of underwater acoustic image processing, and in particular to a feature weight adaptive superpixel segmentation method based on sonar images. Background Art

[0002] Image segmentation is to divide a digital image into regions with independent semantics or features. The goal is to divide the image into distinct parts so that pixels in the same region have similar attributes and the attributes between different regions have obvious differences. At present, due to the low resolution of sonar itself and the interference of complex acoustic field environment such as crosstalk between beams, the image grayscale distortion, irregular shape, and incomplete boundaries are caused.

[0003] Therefore, there is an urgent need for a feature weight adaptive superpixel segmentation method based on sonar images to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and propose a feature weight adaptive superpixel segmentation method based on sonar images to solve the problems raised in the background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a feature weight adaptive superpixel segmentation method based on sonar images, which comprises the following steps: Step 1, convert the original sonar image into a grayscale image; Step 2, use guided filtering to denoise the image; Step 3, extract the unique features of sonar images; Step 4: Superpixel segmentation preprocessing with automatic feature weight adjustment; Step 5, maximum flow-minimum cut superpixel aggregation; Step 6, aggregate superpixels according to a preset threshold; Step 7: Separate the target and background to obtain the segmentation result.

[0006] Preferably, the specific method of converting the original sonar image into a grayscale image in Step 1 is to use a component method, that is, to use any one channel component of the three channels of the color image, where the R channel is selected, and use it as the grayscale value of the grayscale image.

[0007] Preferably, the image denoising process in Step 2 uses a guided filter with strong filtering capability. The guided filter is a local linear image filter, which has the advantages of smoothing image details and retaining edges, and can retain more image details near the edges. A single pixel k has a linear relationship in the center of the window, and the linear function can be expressed as a complex function.

[0008] Preferably, the core of the guided filtering is to find the linear model of each pixel, that is: ; In the above formula is the pixel value of the guidance image, is the pixel value of the output image, The radius is The square window, , is the linear coefficient in the window; To solve for the linear coefficients , , it is necessary to introduce prevention Too large regularization parameter ,Right now: ; in, ,in is the noise value in the image, is the input noisy image, that is: ; ; in, and Represented as windows Middle Guide Image The mean and variance of yes Number of pixels It is in the window In The average value of As input in the filter constraint, output on the model In the process, n (noise or extra texture) is the input Unnecessary components can be removed, and the formula can be expressed as: .

[0009] Preferably, the specific method steps for extracting the unique features of the sonar image in Step 3 are: Step 3.1, extract spatial relationship features, spatial features are Indicates that and are the vertical and horizontal coordinates of pixels on the image, given two feature vectors and , and The spatial distance between is defined as follows: ; Step 3.2: Extract contour features and use the contour features generated by image gradient , the image gradient measures the directional intensity change of the image, and its amplitude is realized by the square root of the sum of the squares of the directional signal changes. The gradient of the image is expressed as , using the visual sensitivity of the human eye, calculations are performed in the brightness domain, The amplitude and size of is defined as follows: ; ; In the formula, the subscript and Represent the vertical and horizontal directions respectively. and represents the change in directional strength, is the magnitude of the amplitude, and an image filter can be used to calculate an approximate value; Step 3.3, extract texture features and use Weber local descriptor, abbreviated as WLD, to construct texture features , the DE part of WLD calculates the relative intensity difference between the current pixel and its neighboring pixels and the ratio of the intensity of the current pixel, using the forward and reverse tangent mapping to prevent Changes rapidly as the input changes, i.e.: ; in, represents the intensity of the current pixel, Indicates the number of pixels around the current pixel. The intensity of neighboring pixels, so and They are relative strength difference and current strength; A small constant Add to on, to avoid Division by zero, and The texture distance between can be transformed into the following formula: ; Step 3.4, extract grayscale features, SLIC algorithm is based on CIELab color space , , Values ​​and pixel coordinates , The pixel space is used to cluster the pixels, which uses the similarity of color and spatial features, normalized and combined to produce a unified similarity distance metric. When, as follows: ; ; ; In the formula, and are the brightness distance and the spatial distance respectively, is the maximum intra-class spatial distance, and , is the maximum color distance, which varies with different images and is generally a fixed constant; Grayscale Features It is expressed as, .

[0010] Preferably, two thresholds are used in Step 3.2 and ( ) to stratify the gradient values, In the definition, if Less than , the layered gradient value is set to ,if Greater than , reflecting the contour features, the layered gradient value is set to , the other gradient values ​​remain unchanged, and then subtracted from the layer gradient value , so that the contour feature Starting from scratch, it is found in repeated experiments that these two thresholds can be set to =5 and =15, Features It can be expressed as the following formula: ; and The difference in gradient between is set to, .

[0011] Preferably, the specific method step of the superpixel segmentation preprocessing with automatic feature weight adjustment in Step 4 is to use a discriminative metric to automatically adjust different feature weights in each iteration, with the purpose of segmenting the image into K non-overlapping superpixel regions with consistent internal attributes of each unit, which is achieved by minimizing energy E, as shown in the following formula: ; ; energy EThe optimization can be iteratively performed by solving three sub-problems: Pixel Allocation Problem ( ): Fixed cluster center and feature weights and pixel tags; Center of mass adjustment problem ): Fixed feature weights and pixel labels to identify the best cluster centers ; Weight Adaptation Problem ( ): Fixed cluster center and pixel labels, adjusting feature weights ; exist After that, the centroid adjustment problem The following formula is optimized: ; in, Represents pixels exist The value of the feature, vector Representation clustering The centroid of is set as follows: ; Among them, when hour, The smaller, The larger , the more influential the features with small normalized intra-class variance are, and this phenomenon is consistent with the principle of feature weight adaptation.

[0012] Preferably, the image needs to be reconstructed after superpixel preprocessing. The specific method of maximum flow-minimum cut superpixel aggregation in Step 5 is that each graph node corresponds to the processed superpixel, wherein the brightness difference is used as the corresponding edge weight, and the brightness feature of a single superpixel is The calculation formula is as follows: ; in, In a superpixel area, is the pixel point in the superpixel, n is the number of superpixels in the image area. The minimum cut solution can be regarded as the minimization of the energy function. The Potts model is used to solve the energy function, which can be defined as: .

[0013] Preferably, the specific method of aggregating superpixels by a preset threshold in Step 6 is: in order to distinguish the foreground object from the background, a saliency detection method is used to firstly detect the brightness feature of the superpixel and the global brightness mean feature. The solution formula is as follows: ; in, K is the total number of superpixels, B Represents the brightness feature within the superpixel area, and extracts the image contrast matrix based on global and local brightness features Q , a single superpixel The significance of is shown by the following formula: ; ; In the formula, represents the superpixel saliency, is the significant reference coefficient, , is the salient superpixel discriminant, if , superpixel It is a salient superpixel.

[0014] Preferably, in Step 7, the target and background are separated to obtain a segmentation result, each superpixel obtains a corresponding class label, and for each cluster, the total number of superpixels in the class and the total number of significant superpixels are calculated, and finally the proportion of significant superpixels in the class is obtained. , as shown below: ; in, is the total number of significant superpixels, The total number of superpixels within a class, T is the preset threshold, T = 0.5, if the proportion of significant superpixels in the cluster is greater than the preset threshold, then mark It is the target foreground, otherwise it is the background class.

[0015] The present invention has the following beneficial effects: 1. The present invention uses the principle of forward-looking sonar imaging as the research object, and through in-depth analysis of the sending, reflection and receiving processes of sound waves, realizes a super-pixel segmentation method with adaptive feature weights, which effectively improves the segmentation accuracy of sonar images; 2. The present invention uses guided filtering to denoise the sonar image, effectively removing the interference of environmental noise and self-noise, while retaining important details and edge features in the image, thereby improving the overall quality of the image; 3. The present invention extracts the spatial, grayscale, shape and texture features of the sonar image and adaptively adjusts the feature weights according to the discrimination of these features, so that the segmentation result is more consistent with the characteristics of the actual target area; 4. The present invention uses the maximum flow-minimum cut algorithm to aggregate superpixels, and combines the preset threshold to effectively aggregate and count superpixels, ensuring accurate segmentation of the target area and facilitating qualitative and quantitative analysis; 5. The present invention simplifies the image processing process by grayscale conversion and processing of a single channel (R channel) of a color image, and ensures that the key features of the image can be maintained under the grayscale image, thus providing a guarantee for the effectiveness of the segmentation algorithm; 6. The present invention adopts superpixel segmentation preprocessing with automatic feature weight adjustment, realizes dynamic adjustment of the importance of different features, enhances the adaptability of the segmentation algorithm in complex environments, and improves the robustness of image processing; 7. The technical solution of the present invention not only improves the segmentation accuracy of the target area of ​​the sonar image, but also provides higher accuracy and reliability for subsequent three-dimensional reconstruction, defect type discrimination, and qualitative and quantitative analysis, and provides advanced technical support for underwater detection and acoustic image processing; 8. The present invention segments the target area from the sonar image with strong interference through superpixel segmentation. The segmented target area can further provide more accurate basic data for the subsequent volume area measurement of the diseased area and the classification of the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method steps of the present invention; Figure 2 Schematic diagram of filter image processing of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: See also Figure 1 to Figure 2 , a feature weight adaptive superpixel segmentation method based on sonar images, which includes the following steps: Step 1, convert the original sonar image into a grayscale image; Step 2: Use guided filtering to denoise the image; Step 3, extract the unique features of sonar images; Step 4: Superpixel segmentation preprocessing with automatic feature weight adjustment; Step 5, maximum flow-minimum cut superpixel aggregation; Step 6, aggregate superpixels according to a preset threshold; Step 7: Separate the target and background to obtain the segmentation result.

[0018] Preferably, the specific method of converting the original sonar image into a grayscale image in Step 1 is to use a component method, that is, to use any one channel component of the three channels of the color image, where the R channel is selected, and use it as the grayscale value of the grayscale image.

[0019] Preferably, the image denoising process in Step 2 uses a guided filter with strong filtering capability. The guided filter is a local linear image filter, which has the advantages of smoothing image details and retaining edges, and can retain more image details near the edges. A single pixel k has a linear relationship in the center of the window, and the linear function can be expressed as a complex function.

[0020] Preferably, the core of the guided filtering is to find the linear model of each pixel, that is: ; In the above formula is the pixel value of the guidance image, is the pixel value of the output image, The radius is The square window, , is the linear coefficient in the window; To solve for the linear coefficients , , it is necessary to introduce prevention Too large regularization parameter ,Right now: ; in, ,in is the noise value in the image, is the input noisy image, that is: ; ; in, and Represented as windows Medium Guide Image The mean and variance of yes Number of pixels It is in the window In The average value of As input in the filter constraint, output on the model In the process, n (noise or extra texture) is the input Unnecessary components can be removed, and the formula can be expressed as: .

[0021] Preferably, the specific method steps for extracting the unique features of the sonar image in Step 3 are: Step 3.1, extract spatial relationship features, spatial features are Indicates that and are the vertical and horizontal coordinates of pixels on the image, given two feature vectors and , and The spatial distance between is defined as follows: ; Step 3.2: Extract contour features and use the contour features generated by image gradient , the image gradient measures the directional intensity change of the image, and its amplitude is realized by the square root of the sum of the squares of the directional signal changes. The gradient of the image is expressed as , using the visual sensitivity of the human eye, calculations are performed in the brightness domain, The amplitude and size of is defined as follows: ; ; In the formula, the subscript and Represent the vertical and horizontal directions respectively. and represents the change in directional strength, is the magnitude of the amplitude, and an image filter can be used to calculate an approximate value; Step 3.3, extract texture features and use Weber local descriptor, abbreviated as WLD, to construct texture features , the DE part of WLD calculates the relative intensity difference between the current pixel and its neighboring pixels and the ratio of the intensity of the current pixel, using the forward and reverse tangent mapping to prevent Changes rapidly as the input changes, i.e.: ; in, represents the intensity of the current pixel, Indicates the number of pixels around the current pixel. The intensity of neighboring pixels, so and They are relative strength difference and current strength; A small constant Add to on, to avoid Division by zero, and The texture distance between can be transformed into the following formula: ; Step 3.4, extract grayscale features, SLIC algorithm is based on CIELab color space , , Values ​​and pixel coordinates , The pixel space is used to cluster the pixels, which uses the similarity of color and spatial features, normalized and combined to produce a unified similarity distance metric. When, as follows: ; ; ; In the formula, and are the brightness distance and the spatial distance respectively, is the maximum intra-class spatial distance, and , is the maximum color distance, which varies with different images and is generally a fixed constant; Grayscale Features It is expressed as, .

[0022] The feature relationships between different images are different, so it is necessary to extract a variety of sonar image features to enhance the accurate extraction of the target area. The targets segmented in the image with extracted spatial relationship features have certain spatial position and direction relationships with each other. Spatial relationships are divided into connection relationships, overlapping relationships, and inclusion relationships. These relationships emphasize the distance and orientation between targets. In three-dimensional space, the absolute position of the target can obtain position information to express a simpler relative spatial position. The role of extracting spatial relationship features can have a certain ability to distinguish the image content description, but the spatial relationship features are sensitive to related changes in the image, such as rotation, inversion, scale changes, etc. If the image is to be automatically segmented and the object area contained in the image is divided, it is necessary to extract image features based on the area and establish an index. Spatial features are used to enhance the compactness of superpixels. Compactness; extract contour features. The shape features of an image mainly include contour features and regional features. Contour features mainly describe the outside of an object and focus on the positional relationship between targets. Regional features are mainly about the features of the entire region, which mainly involve the distance, size and orientation information between targets. Shape features can also be represented by a combination of contour features and regional features. This combination can provide richer shape information. If contour features cannot distinguish the edges of two similar shapes, regional features can identify shape edges by comparing different parts inside the shape, and then use local contour features to distinguish pixels that cannot be correctly clustered using grayscale and spatial features. Image texture features can be used to visually reflect the homogeneity phenomenon in the image. It is a global feature that can extract the surface organization or structural arrangement properties of slowly changing and periodically changing objects. Local texture information is often expressed in the image as the distribution of grayscale of pixel points on the surface of the object and the surrounding neighborhood. Global texture information is the repetitive embodiment of local texture information to varying degrees. To extract texture features, methods based on structure and statistics can be used. The detection logic based on the structure method requires detecting image texture for modeling and searching for patterns of repeated features in the area of ​​the image. Natural images can be textureless or textured. For texture images, texture features are more discriminative. Grayscale features are extracted. Grayscale images are composed of pixels with different grayscale values. The grayscale distribution of each pixel is an important feature of the image.

[0023] Preferably, two thresholds are used in Step 3.2 and ( ) to stratify the gradient values, In the definition, if Less than , the layered gradient value is set to ,if Greater than , reflecting the contour features, the layered gradient value is set to , the other gradient values ​​remain unchanged, and then subtracted from the layer gradient value , so that the contour feature Starting from scratch, it is found in repeated experiments that these two thresholds can be set to =5 and =15, Features It can be expressed as the following formula: ; and The difference in gradient between is set to, .

[0024] Preferably, the specific method step of the superpixel segmentation preprocessing with automatic feature weight adjustment in Step 4 is to use a discriminative metric to automatically adjust different feature weights in each iteration, with the purpose of segmenting the image into K non-overlapping superpixel regions with consistent internal attributes of each unit, which is achieved by minimizing energy E, as shown in the following formula: ; ; energy E The optimization can be iteratively performed by solving three sub-problems: Pixel Allocation Problem ( ): Fixed cluster center and feature weights and pixel tags; Center of mass adjustment problem ): Fixed feature weights and pixel labels to identify the best cluster centers ; Weight Adaptation Problem ( ): Fixed cluster center and pixel labels, adjusting feature weights ; exist After that, the centroid adjustment problem The following formula is optimized: ; in, Represents pixels exist The value of the feature, vector Representation clustering The centroid of is set as follows: ; Among them, when hour, The smaller, The larger the value, the more influence features with small normalized intra-class variance have, which is consistent with the principle of feature weight adaptation. , the parameter is set to 7 based on experimental experience. The proposed method gives the initial weights of the features, the pixels are assigned to their nearest centers, and then the distinguishability of the features is evaluated and the feature weights are adjusted accordingly. and Solution, the proposed feature partitioning measure, normalizes the intra-class variance during clustering It provides a reliable distinguishability measure for different features. Features with small normalized intra-cluster variance have similar values ​​within a superpixel, which helps group pixels with compact properties into superpixels. Feature weights Relying on the feature of small normalized intracluster variance, there are three issues to consider in each iteration: pixel allocation, centroid adjustment, and weight adaptation. ) is usually set to a fixed value. is set to 10, and once the stopping threshold is met, the pixel labels are further processed to enforce continuity within the superpixel.

[0025] Preferably, the image needs to be reconstructed after superpixel preprocessing. The specific method of maximum flow-minimum cut superpixel aggregation in Step 5 is that each graph node corresponds to the processed superpixel, wherein the brightness difference is used as the corresponding edge weight, and the brightness feature of a single superpixel is The calculation formula is as follows: ; in, In a superpixel area, is the pixel point in the superpixel, n is the number of superpixels in the image area. The minimum cut solution can be regarded as the minimization of the energy function. The Potts model is used to solve the energy function, which can be defined as: .

[0026] Preferably, the specific method of aggregating superpixels by a preset threshold in Step 6 is: in order to distinguish the foreground object from the background, a saliency detection method is used to firstly detect the brightness feature of the superpixel and the global brightness mean feature. The solution formula is as follows: ; in, K is the total number of superpixels, BRepresents the brightness feature within the superpixel area, and extracts the image contrast matrix based on global and local brightness features Q , a single superpixel The significance of is shown by the following formula: ; ; In the formula, represents the superpixel saliency, is the significant reference coefficient, , is the salient superpixel discriminant, if , superpixel It is a salient superpixel.

[0027] Preferably, in Step 7, the target and background are separated to obtain a segmentation result, each superpixel obtains a corresponding class label, and for each cluster, the total number of superpixels in the class and the total number of significant superpixels are calculated, and finally the proportion of significant superpixels in the class is obtained. , as shown below: ; in, is the total number of significant superpixels, The total number of superpixels within a class, T is the preset threshold, T = 0.5, if the proportion of significant superpixels in the cluster is greater than the preset threshold, then mark It is the target foreground, otherwise it is the background class.

[0028] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A feature weight adaptive superpixel segmentation method based on sonar images, characterized in that: It includes the following steps: Step 1, convert the original sonar image into a grayscale image; Step 2, use guided filtering to denoise the image; Step 3, extract the unique features of sonar images; Step 4: Superpixel segmentation preprocessing with automatic feature weight adjustment; Step 5, maximum flow-minimum cut superpixel aggregation; Step 6, aggregate superpixels according to a preset threshold; Step 7: Separate the target and background to obtain the segmentation result.

2. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: The specific method of converting the original sonar image into a grayscale image in Step 1 is to use a component method, that is, to use any one channel component of the three channels of the color image, and here the R channel is selected, and use it as the grayscale value of the grayscale image.

3. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: In the Step 2, the image is denoised using a guided filter with strong filtering capability. The guided filter is a local linear image filter that has the advantages of smoothing image details and preserving edges. It can preserve more image details near the edges. A single pixel k has a linear relationship in the center of the window, and the linear function can be expressed as a complex function.

4. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 3 is characterized in that: The core of the guided filtering is to find the linear model of each pixel, namely: ； In the above formula is the pixel value of the guidance image, is the pixel value of the output image, The radius is The square window, , is the linear coefficient in the window; To solve for the linear coefficients , , it is necessary to introduce prevention Too large regularization parameter ,Right now: ； in, ,in is the noise value in the image, is the input noisy image, that is: ； ； in, and Represented as windows Medium Guide Image The mean and variance of yes Number of pixels It is in the window In The average value of As input in the filter constraint, output on the model In the process, n (noise or extra texture) is the input Unnecessary components can be removed, and the formula can be expressed as: .

5. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: The specific method steps for extracting the unique features of the sonar image in Step 3 are: Step 3.1, extract spatial relationship features, spatial features are Indicates that and are the vertical and horizontal coordinates of pixels on the image, given two feature vectors and , and The spatial distance between is defined as follows: ； Step 3.2: Extract contour features and use the contour features generated by image gradient , the image gradient measures the directional intensity change of the image, and its amplitude is realized by the square root of the sum of the squares of the directional signal changes. The gradient of the image is expressed as , using the visual sensitivity of the human eye, calculations are performed in the brightness domain, The amplitude and size of is defined as follows: ； ； In the formula, the subscript and Represent the vertical and horizontal directions respectively. and represents the change in directional strength, is the magnitude of the amplitude, and an image filter can be used to calculate an approximate value; Step 3.3, extract texture features and use Weber local descriptor, abbreviated as WLD, to construct texture features , the DE part of WLD calculates the relative intensity difference between the current pixel and its neighboring pixels and the ratio of the intensity of the current pixel, using the forward and reverse tangent mapping to prevent Changes rapidly as the input changes, i.e.: ； in, represents the intensity of the current pixel, Indicates the number of pixels around the current pixel. The intensity of neighboring pixels, so and They are relative strength difference and current strength; A small constant Add to on, to avoid Division by zero, and The texture distance between can be transformed into the following formula: ； Step 3.4, extract grayscale features, SLIC algorithm is based on CIELab color space , , Values ​​and pixel coordinates , The pixel space is used to cluster the pixels, which uses the similarity of color and spatial features, normalized and combined to produce a unified similarity distance metric. When, as follows: ； ； ； In the formula, and are the brightness distance and the spatial distance respectively, is the maximum intra-class spatial distance, and , is the maximum color distance, which varies with different images and is generally a fixed constant; Grayscale Features It is expressed as, 。 6. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 5, characterized in that: Two thresholds are used in Step 3.2 and ( ) to stratify the gradient values, In the definition, if Less than , the layered gradient value is set to ,if Greater than , reflecting the contour features, the layered gradient value is set to , the other gradient values ​​remain unchanged, and then subtracted from the layer gradient value , so that the contour feature Starting from scratch, it is found in repeated experiments that these two thresholds can be set to =5 and =15, Features It can be expressed as the following formula: ； and The difference in gradient between is set to, 。 7. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: The specific method steps of the superpixel segmentation preprocessing with automatic feature weight adjustment in Step 4 are to use a discriminative metric to automatically adjust different feature weights in each iteration, with the purpose of segmenting the image into K non-overlapping superpixel regions with consistent internal attributes of each unit, which is achieved by minimizing energy E, as shown in the following formula: ； ； energy E The optimization can be iteratively performed by solving three sub-problems: Pixel Allocation Problem ( ): Fixed cluster center and feature weights and pixel tags; Center of mass adjustment problem ): Fixed feature weights and pixel labels to identify the best cluster centers ; Weight Adaptation Problem ( ): Fixed cluster center and pixel labels, adjusting feature weights ; exist Afterwards, the centroid adjustment problem The following formula is optimized: ； in, Represents pixels exist The value of the feature, vector Representation clustering The centroid of is set as follows: ； Among them, when hour, The smaller, The larger , the more influential the features with small normalized intra-class variance are, and this phenomenon is consistent with the principle of feature weight adaptation.

8. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: The image needs to be reconstructed after superpixel preprocessing. The specific method of maximum flow-minimum cut superpixel aggregation in Step 5 is as follows: each graph node corresponds to the processed superpixel, where the brightness difference is used as the corresponding edge weight, and the brightness feature of a single superpixel is The calculation formula is as follows: ； in, In a superpixel area, is the pixel point in the superpixel, n is the number of superpixels in the image area. The minimum cut solution can be regarded as the minimization of the energy function. The Potts model is used to solve the energy function, which can be defined as: 。 9. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: The specific method of aggregating superpixels by a preset threshold in Step 6 is as follows: In order to distinguish the foreground object from the background, a saliency detection method is used to firstly detect the brightness feature of the superpixel and the global brightness mean feature. The solution formula is as follows: ； in, K is the total number of superpixels, B Represents the brightness feature within the superpixel area, and extracts the image contrast matrix based on global and local brightness features Q , a single superpixel The significance of is shown by the following formula: ； ； In the formula, represents the superpixel saliency, is the significant reference coefficient, , is the salient superpixel discriminant, if , superpixel It is a salient superpixel.

10. The feature weight adaptive superpixel segmentation method based on sonar images according to claim 1, characterized in that: In Step 7, the target and background are separated to obtain the segmentation result, and each superpixel obtains the corresponding class label. For each cluster, the total number of superpixels in the class and the total number of significant superpixels are calculated, and finally the proportion of significant superpixels in the class is obtained. , as shown below: ； in, is the total number of significant superpixels, The total number of superpixels within a class, T is the preset threshold, T = 0.5, if the proportion of significant superpixels in the cluster is greater than the preset threshold, then mark It is the target foreground, otherwise it is the background class.

Images