Image object extraction method based on LRW and density clustering

Abstract

The present invention discloses an image target extraction method based on LRW and density clustering, comprising the following steps: S1, LRW superpixel algorithm pre-segmentation; S2, superpixel texture feature extraction; S3, density similarity measurement: constructing a multi-feature similarity matrix M between different superpixels ij ; S4, select parameters ε and Minpts; S5, complete image target extraction. The goal of the present invention is to extract the image target based on the differences in texture, color and structure, and use a clustering method to effectively combine the differences between the various microscopic components to achieve color image segmentation.

Description

Technical Field

[0001] The invention relates to the technical field of image processing and target extraction, and in particular to an image target extraction method based on LRW and density clustering. Background Art

[0002] Target region extraction is a key step in automatic image analysis. To accurately identify image targets, it is necessary to analyze the relevant features of the image. The target region and background region can be segmented according to the color, texture and other features in different target regions. The segmented regions will not overlap but have certain similarities. Image segmentation is an effective image processing and analysis method, which is widely used in various fields.

[0003] The research on image segmentation algorithms has a history of several decades. According to the needs of different situations, the current image segmentation methods at home and abroad can be roughly divided into traditional methods and methods based on specific theories. Among them, traditional methods include: threshold-based, edge detection-based and region-based segmentation methods. The development of image segmentation technology is closely related to new methods and new theories in other disciplines and fields. Most scholars use these new theories and methods to improve image segmentation algorithms, resulting in the emergence of many segmentation algorithms based on specific theories, including fuzzy theory, wavelet analysis, multifractal, ant colony algorithm, clustering, graph theory, neural network and granular computing. However, with the increasing complexity of segmentation scenes and the continuous improvement of segmentation requirements, it is difficult for single-theory segmentation algorithms to meet the needs of researchers for image segmentation results. Therefore, algorithms that combine a large number of theories have become the main research direction in the field of image segmentation. Therefore, how to provide an image target extraction method based on LRW and density clustering is a problem that technicians in this field need to solve urgently. Summary of the invention

[0004] One purpose of the present invention is to propose an image target extraction method based on LRW and density clustering. The goal of the present invention is to extract parts of image targets based on differences in texture, color and structure. The clustering method can effectively combine the differences between the various microscopic components to achieve color image segmentation.

[0005] According to an embodiment of the present invention, an image target extraction method based on LRW and density clustering includes the following steps:

[0006] S1. LRW superpixel algorithm pre-segmentation: The LRW superpixel algorithm is used to pre-segment the gas ash color image to obtain a superpixel image;

[0007] S2. Superpixel texture feature extraction: an improved method of extracting texture features is selected to count the superpixel texture feature quantities in the gray-level co-occurrence matrix of different carbon-containing materials;

[0008] S3. Density similarity measure: construct a multi-feature similarity matrix M between different superpixels ij ;

[0009] S4. Select parameters ε and Minpts: Get the sequential matrix KNN based on the image superpixel points n×(n-1) , KNN n×(n-1) The kth (k = 1, 2, ..., n) column in the matrix is ​​the k-nearest neighbor distance value set Dist of all pixels in the sample set. k , in the k-nearest neighbor distance graph, the distance corresponding to the point with the largest slope change is the ε value;

[0010] S5, complete the image target extraction: according to the similarity matrix M ij Select any superpixel p, and the set of pixels in the hypersphere area with p as the center and ε as the radius in the ε-neighborhood point set of p is:

[0011] N ε (p) = {q∈D|M pq ≥τ};

[0012] Among them, |N ε (p)| is the number of sample points in the ε-neighborhood of p;

[0013] If |N ε (p)| is greater than or equal to the specified parameter Minpts value, then p is a core point, and all points that are density-reachable from p are found to form a class C, and p is marked as processed. Otherwise, p is marked as a noise point;

[0014] After class C is determined, all sample points in class C are repeatedly judged for expansion, and then sample points outside class C are visited until all sample points are searched, and the final target extraction result is obtained.

[0015] Optionally, the super-pixel image includes a plurality of super-pixel blocks of different and irregular shapes.

[0016] Optionally, the S1 specifically includes:

[0017] S12, for each pixel x in the image i Using linear representation:

[0018]

[0019] Among them, N i is the pixel x i The neighborhood pixel set, w ij and b i are the weights and bias terms determined through the optimization process;

[0020] S12, optimize the initial LRW segmentation result, using the optimization goal:

[0021]

[0022] Among them, C represents a super pixel block, μ C It is the average LRW value of all pixels in the block;

[0023] S13. Based on the optimization, the edge of the super pixel block is refined To evaluate each pixel x i Is it at the edge of the superpixel block? If it exceeds the set threshold, then x i At the edge of the superpixel block.

[0024] Optionally, the edge refinement includes using a Gaussian pyramid to represent the image at multiple scales, and calculating the gradient at each scale separately:

[0025]

[0026] Among them, Gσ represents the Gaussian smoothing kernel, σ represents the scale, and * represents the convolution operation;

[0027] Based on edge refinement, the second-order derivative of the calculated gradient is used to further enhance the edge clarity:

[0028]

[0029] When the absolute value of the second-order derivative exceeds the set threshold, the point is considered to be at the edge;

[0030] A convolutional neural network is used to learn and identify edge features in an image by training a network model, wherein the network model input is an LRW-processed image and the output is an edge detection result.

[0031] Optionally, the S2 specifically includes:

[0032] S21, extract the coordinate set of the pixel points in the superpixel (X * ,Y * ), extract the selected super pixel area separately onto a black background image of the same size as the original image, and the position and color of the super pixel block should be consistent with the original image;

[0033] S22, according to the coordinate method, the minimum and maximum values ​​of the horizontal and vertical coordinates in the superpixel are represented as min(X * )、max(X * )、min(Y * )、max(Y *), two by two are combined to form a rectangular area, and the coordinates are calibrated with 4 red points;

[0034] S23, counting the eigenvalues ​​of the gray-level co-occurrence matrix of the superpixels in the calibration area, and completing the extraction of superpixel texture feature quantities of the superpixels.

[0035] Optionally, S2 further includes extracting superpixel texture features, combining superpixel color and position information, and using a 10-dimensional vector to represent the i-th superpixel in the superpixel image from three aspects of color, position and texture:

[0036] C i =(l i ,a i ,b i ,x i ,y i ,MP i ,CON i ,IDM i ,ASM i ,ENT i );

[0037] Among them, the color feature is the average value of the internal pixel color components in the super pixel block, and the position feature is the center of gravity of each super pixel block, represented by coordinates x and y.

[0038] Optionally, the multi-feature similarity matrix M between different superpixels is constructed ij include:

[0039]

[0040] in, They are respectively represented as the color, spatial and texture feature vectors of pixel i, σ c , σ p and σ t They represent the color feature sensitivity coefficient, position distance feature sensitivity coefficient and texture feature sensitivity coefficient respectively.

[0041] Optionally, the ε describes a neighborhood distance threshold of a certain sample, and Minpts describes a threshold of the number of samples in a neighborhood of a certain sample whose distance is ε.

[0042] The beneficial effects of the present invention are:

[0043] (1) The present invention uses superpixels instead of pixels, which speeds up the existing pixel-based algorithms, reduces the running time, improves the efficiency of the algorithms, groups pixels by using the similarity between them to obtain redundant information of the image, and greatly reduces the complexity of subsequent image processing tasks;

[0044] (2) The present invention uses a superpixel method to preprocess image target extraction, replaces a large number of pixels with a small number of superpixels, and searches for density-connected areas through multi-feature similarity between superpixels. While ensuring that the algorithm time is greatly reduced, the quality of image extraction can also be improved. Therefore, the scheme proposed by the present invention has important application potential in the field of image processing and target extraction, and can provide a more accurate and stable solution for research and application in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0046] Figure 1 A flowchart of an image target extraction method based on LRW and density clustering proposed by the present invention;

[0047] Figure 2 A flowchart of extracting superpixel texture feature in an image target extraction method based on LRW and density clustering proposed by the present invention;

[0048] Figure 3 The k-nearest neighbor distance graph used by the parameters ε and Minpts selected in the improved density clustering algorithm in the image target extraction method based on LRW and density clustering proposed by the present invention;

[0049] Figure 4 A comparison diagram of the pre-segmentation results of LRW superpixel and SLIC superpixel in the solution of the present invention in an image target extraction method based on LRW and density clustering proposed by the present invention;

[0050] Figure 5 This is a comparison chart of image target extraction results under different algorithms in the image target extraction method based on LRW and density clustering proposed in the present invention. DETAILED DESCRIPTION

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0052] refer to Figure 1-2 , an image target extraction method based on LRW and density clustering, comprising the following steps:

[0053] S1. LRW superpixel algorithm pre-segmentation: The LRW superpixel algorithm is used to pre-segment the gas ash color image to obtain a superpixel image;

[0054] In this embodiment, according to the analysis of color and texture characteristics of different carbon-containing structures in the gas ash color image, it can be known that the three color component values ​​​​L, a, and b in the CIELab color space can better highlight the color differences between different carbon-containing materials, and the five texture feature quantities of maximum probability, contrast, inverse moment, angular second moment, and entropy based on the grayscale co-occurrence matrix can better characterize the texture details and differences of each carbon-containing structure. Therefore, these five parameters are selected to represent the texture feature values ​​​​of different carbon-containing areas.

[0055] After preprocessing, the image is divided into many superpixel blocks of different shapes and irregularities. Therefore, it is impossible to extract the texture features of each superpixel block based on the traditional method. The present invention selects an improved method of extracting texture features to count the feature quantities in the grayscale co-occurrence matrix of different carbon-containing substances.

[0056] In this implementation, S1 specifically includes:

[0057] S12, for each pixel x in the image i Using linear representation:

[0058]

[0059] Among them, N i is the pixel x i The neighborhood pixel set, w ij and b i are weights and biases determined by the optimization process. The purpose of this step is to find a linear combination such that the pixel x i It can be as much as possible from its neighboring pixel point x j Linear representation. ;

[0060] S12, optimize the initial LRW segmentation result, the goal is to minimize the pixel difference within the super pixel block and enhance the distinction between blocks, using the optimization goal:

[0061]

[0062] Among them, C represents a super pixel block, μ C It is the average LRW value of all pixels in the block;

[0063] S13. Based on the optimization, the edge of the super pixel block is refined to make the edge clearer and more precise. To evaluate each pixel x i Is it at the edge of the superpixel block? If it exceeds the set threshold, then x i At the edge of the superpixel block.

[0064] In this implementation, the edge refinement includes using a Gaussian pyramid to represent the image at multiple scales, and calculating the gradient at each scale:

[0065]

[0066] Among them, Gσ represents the Gaussian smoothing kernel, σ represents the scale, and * represents the convolution operation;

[0067] Based on edge refinement, the second-order derivative of the calculated gradient is used to further enhance the edge clarity:

[0068]

[0069] When the absolute value of the second-order derivative exceeds the set threshold, the point is considered to be at the edge;

[0070] A convolutional neural network is used to learn and identify edge features in an image by training a network model, wherein the network model input is an LRW-processed image and the output is an edge detection result.

[0071] By precisely adjusting the parameters of the LRW algorithm and optimizing the edges of superpixel blocks through S11-S13, efficient and accurate image segmentation is achieved, providing a better foundation for subsequent density clustering.

[0072] S2. Superpixel texture feature extraction: an improved method of extracting texture features is selected to count the superpixel texture feature quantities in the gray-level co-occurrence matrix of different carbon-containing materials;

[0073] In this embodiment, S2 specifically includes:

[0074] S21, extract the coordinate set of the pixel points in the superpixel (X * ,Y * ), extract the selected super pixel area separately onto a black background image of the same size as the original image, and the position and color of the super pixel block should be consistent with the original image;

[0075] S22, according to the coordinate method, the minimum and maximum values ​​of the horizontal and vertical coordinates in the superpixel are represented as min(X * )、max(X * )、min(Y * )、max(Y * ), two by two are combined to form a rectangular area, and the coordinates are calibrated with 4 red points;

[0076] S23, counting the eigenvalues ​​of the gray-level co-occurrence matrix of the superpixels in the calibration area, and completing the extraction of superpixel texture feature quantities of the superpixels.

[0077] In this embodiment, S2 also includes extracting superpixel texture features, combining the color and position information of the superpixel, and using a 10-dimensional vector to represent the i-th superpixel in the superpixel image from three aspects of color, position and texture:

[0078] C i =(l i ,a i ,b i ,x i ,y i ,MP i ,CON i ,IDM i ,ASM i ,ENT i );

[0079] Among them, the color feature is the average value of the internal pixel color components in the super pixel block, and the position feature is the center of gravity of each super pixel block, represented by coordinates x and y.

[0080] The traditional DBSCAN algorithm uses a single feature distance to describe the similarity between superpixels, which makes it difficult to describe the potential relationship between complex structures. In order to better characterize the local structural features between superpixels, the present invention proposes a new multi-feature similarity matrix.

[0081] S3. Density similarity measure: construct a multi-feature similarity matrix M between different superpixels ij ;

[0082] In this embodiment, a multi-feature similarity matrix M between different superpixels is constructed. ij include:

[0083]

[0084] in, They are respectively represented as the color, spatial and texture feature vectors of pixel i, σ c , σ p and σ t They represent the color feature sensitivity coefficient, position distance feature sensitivity coefficient and texture feature sensitivity coefficient respectively.

[0085] In this embodiment, the DBSCAN algorithm mainly analyzes the density distribution between sample points through a set of neighborhoods, and needs to define two parameters (ε, Minpts) to describe the density of sample distribution in the neighborhood. ε describes the neighborhood distance threshold of a certain sample, and Minpts describes the threshold of the number of samples in the neighborhood with a distance of ε of a certain sample. Sample points in space can be divided into three forms: core points, boundary points, and noise points. Core points refer to points that meet the requirement of containing at least Minpts number of samples in an area with a set neighborhood radius of ε; boundary points refer to points with less than Minpts samples in an area with a neighborhood radius of ε, but are sample points within the neighborhood radius of the core point; sample points that do not meet the requirements of core points and boundary points are noise points.

[0086] For the entire sample set, select a pixel point p arbitrarily from the sample set, and determine whether the pixel point is a core point by using the above core point definition. If the pixel point is a core point, then search for all pixel points that are density-reachable starting from point p in turn. These pixel points form a new cluster, and then use the pixel points in the new cluster as new pixel points, and continuously obtain points that are density-reachable with these pixel points to achieve cluster expansion until no new pixel points are added, and finally it becomes a complete cluster; if the pixel point p is not a core point, it is classified as a noise point. Select the next undivided pixel point in the sample set and repeat the judgment to complete the clustering of the next cluster until all pixel points in the sample set are divided, and then the clustering ends. Therefore, the method for determining the parameters ε and Minpts is as follows:

[0087] S4. Select parameters ε and Minpts: Get the sequential matrix KNN based on the image superpixel points n×(n-1) , KNN n×(n-1) The kth (k = 1, 2, ..., n) column in the matrix is ​​the k-nearest neighbor distance value set Dist of all pixels in the sample set. k , in the k-nearest neighbor distance graph, the distance corresponding to the point with the largest slope change is the ε value;

[0088] The present invention selects the k-nearest neighbor distance graph when k=4 to determine the ε value. Generally, the Minpts value is selected as n / 25 (n is the number of samples). The k-nearest neighbor distance graph when k=4 is as follows: Figure 3 As shown, point A is the point where the slope changes the most.

[0089] S5, complete the image target extraction: according to the similarity matrix M ij Select any superpixel p, and the set of pixels in the hypersphere area with p as the center and ε as the radius in the ε-neighborhood point set of p is:

[0090] N ε (p) = {q∈D|M pq ≥τ};

[0091] Among them, |N ε (p)| is the number of sample points in the ε-neighborhood of p;

[0092] If |N ε (p)| is greater than or equal to the specified parameter Minpts value, then p is a core point, and all points that are density-reachable from p are found to form a class C, and p is marked as processed. Otherwise, p is marked as a noise point;

[0093] After class C is determined, all sample points in class C are repeatedly judged for expansion, and then sample points outside class C are visited until all sample points are searched, and the final target extraction result is obtained.

[0094] Embodiment 1:

[0095] Example 1 A plurality of gas ash color images (481×321) containing different target areas are selected as experimental objects. The experiment is implemented in the Matlab 2016b programming environment, and the hardware environment is: Intel (R) Core (TM) i5-4200H CPU @ 2.80GHz.

[0096] In an implementation manner, the overall framework of the solution of the present invention includes two main components: LRW superpixel pre-segmentation and improved density clustering algorithm secondary segmentation. The LRW superpixel segmentation algorithm is improved based on the RW algorithm. The RW algorithm calculates the first arrival probability, that is, the random walk starts from a pixel and first reaches one of each labeled seeds, and then represents the pixel as having the same label with the maximum probability of the same seed point. Since the random walk algorithm first reaches the position of the seed point marked in advance after starting from the pixel point, it mainly analyzes the partial relationship between the seed point and the current pixel point, which may cause the shape of the superpixel block finally generated to be irregular. The LRW superpixel segmentation algorithm proposed in the present invention takes into account the overall relationship between all pixels and seed points, adds self-loops to effectively solve the problems of weak boundaries and segmentation of complex texture areas, and introduces an energy optimization framework to eliminate the sensitivity of the initial seed point. Therefore, the LRW algorithm can effectively perform initial segmentation of the image. The traditional density clustering algorithm continuously searches for pixels in the image and clusters them by setting parameters, and the setting of parameters is mainly selected based on single features such as the distance or color between the pixels of the image. Because the algorithm forms clusters by continuously searching for density-connected sets between pixels, if all pixels on the image are traversed point by point, not only will the time complexity of the algorithm increase due to the huge data processing, but also the segmentation effect will be poor. Therefore, the present invention uses the superpixel method to perform the initial clustering of the image, replaces a large number of pixels with a small number of superpixels, and searches for density-connected areas through the multi-feature similarity between superpixels. While ensuring that the algorithm time is greatly reduced, the accuracy of the image after segmentation can also be improved.

[0097] Example 1 uses commonly used evaluation indicators in image target extraction to perform quantitative performance evaluation on the gas ash microscopic image target extraction results, including:

[0098] (1) Precision: It indicates the ratio of true positive pixels to all positive pixels in the segmentation result, and is used to measure the quality of the edge of the segmented target area.

[0099] (2) Recall rate: It indicates the proportion of all true positive pixels in the segmentation result that are judged as positive examples, which is used to reflect the completeness of the edge of the segmented target.

[0100] (3) Intersection-Union Ratio: It represents the ratio of the intersection and union of the real pixel value and the predicted pixel value, which is used to reflect the overlap between the segmented area and the real area.

[0101] (4) Pixel accuracy: It indicates the ratio of pixels correctly judged as positive examples to the total pixels in the segmentation target area. It is the most basic indicator in image segmentation evaluation.

[0102] (5) Algorithm running time: represents the execution time of the entire algorithm.

[0103] The experimental results are shown in Table 1 below:

[0104] Table 1 Comparison of performance evaluation results of different segmentation algorithms in the present invention under color image set

[0105] Precision(%) Recall(%) IOU(%) PA(%) RunningTime(s) Watershed 74.20 71.94 68.14 78.45 6.553 Kmeans 72.47 75.54 70.48 82.48 25.872 DBSCAN 85.62 82.52 83.12 88.77 32.471 This algorithm 94.99 96.12 92.87 95.75 8.691

[0106] Refer to Table 1 and Figure 4-Figure 5 It can be seen that the segmentation effect of the watershed algorithm and the Kmeans algorithm is relatively poor, and the segmentation effect of the traditional DBSCAN algorithm is relatively good, but there is still a gap compared with the algorithm in this paper; in addition, from the analysis of the running time of the algorithm, the running time of the Kmeans clustering algorithm and the traditional DBSCAN algorithm is relatively long. This is because the algorithm is based on the clustering of each pixel point, so it will consume a lot of time. Although the running speed of the watershed algorithm is faster, its segmentation accuracy is lower than that of other algorithms. It is not advisable to increase the running speed of the algorithm at the cost of sacrificing segmentation accuracy. On the contrary, the algorithm in this paper uses superpixels for initial clustering, and uses a small number of superpixels instead of a large number of pixels for clustering. On the basis of the traditional DBSCAN algorithm, the running time is greatly reduced, and better segmentation accuracy is also guaranteed. In summary, the image segmentation method combining LRW and DBSCAN algorithms proposed in this paper has the best overall segmentation performance.

[0107] Through experimental verification on the gas ash color image set in Example 1, the solution of the present invention shows superior performance in image target extraction, surpassing the prior art. Through the technical solution of the present invention, effective extraction of the target area of ​​the color image can be achieved. Compared with the prior art, the present invention provides a more efficient and accurate method, and provides a new solution for the further development of color image target extraction technology.

[0108] The present invention uses superpixels instead of pixels, which speeds up the existing pixel-based algorithms, reduces the running time, improves the efficiency of the algorithms, groups pixels by utilizing the similarity between pixels to obtain redundant information of the image, and also greatly reduces the complexity of subsequent image processing tasks.

[0109] In summary, the scheme of the present invention has significant advantages in object extraction of color images. The present invention preprocesses the image object extraction by using a superpixel method, replaces a large number of pixels with a small number of superpixels, searches for density-connected areas through the multi-feature similarity between superpixels, and improves the quality of image extraction while ensuring that the algorithm time is greatly reduced. Therefore, the scheme proposed by the present invention has important application potential in the field of image processing and object extraction, and can provide a more accurate and stable solution for research and application in related fields.

[0110] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An image target extraction method based on LRW and density clustering, characterized in that: The steps include: S1. LRW superpixel algorithm pre-segmentation: The LRW superpixel algorithm is used to pre-segment the gas ash color image to obtain a superpixel image; S2. Superpixel texture feature extraction: an improved method of extracting texture features is selected to count the superpixel texture feature quantities in the gray-level co-occurrence matrix of different carbon-containing materials; S3. Density similarity measure: construct a multi-feature similarity matrix M between different superpixels ij ; The multi-feature similarity matrix M between different superpixels is constructed ij include: in, They are respectively represented as the color, spatial and texture feature vectors of pixel i, σ c , σ p and σ t They represent the color feature sensitivity coefficient, the position distance feature sensitivity coefficient, and the texture feature sensitivity coefficient respectively; S4. Select parameters ε and Minpts: Get the sequential matrix KNN based on the image superpixel points n×(n-1) , KNN n×(n-1) The kth column in the matrix, k = 1, 2, ..., n, is the set of k-nearest neighbor distance values ​​of all pixels in the sample set Dist k , in the k-nearest neighbor distance graph, the distance corresponding to the point with the largest slope change is the ε value; S5, complete the image target extraction: according to the similarity matrix M ij Select any superpixel p, and the set of pixels in the hypersphere area with p as the center and ε as the radius in the ε-neighborhood point set of p is: N ε (p)={q∈D|M pq ≥τ}; Among them, |N ε (p)| is the number of sample points in the ε-neighborhood of p; If |N ε (p)| is greater than or equal to the specified parameter Minpts value, then p is a core point, and all points that are density-reachable from p are found to form a class C, and p is marked as processed. Otherwise, p is marked as a noise point; After class C is determined, all sample points in class C are repeatedly judged for expansion, and then sample points outside class C are visited until all sample points are searched, and the final target extraction result is obtained.

2. The image target extraction method based on LRW and density clustering according to claim 1 is characterized in that: The super-pixel image includes a plurality of super-pixel blocks of different and irregular shapes.

3. The image target extraction method based on LRW and density clustering according to claim 1 is characterized in that: The S1 specifically includes: S12, for each pixel x in the image i Using linear representation: Among them, N i is the pixel x i The neighborhood pixel set, w ij and b i are the weights and bias terms determined through the optimization process; S12, optimize the initial LRW segmentation result, using the optimization goal: Among them, C represents a super pixel block, μ C It is the average LRW value of all pixels in the block; S13. Based on the optimization, the edge of the super pixel block is refined To evaluate each pixel x i Is it at the edge of the superpixel block? If it exceeds the set threshold, then x i At the edge of the superpixel block.

4. The image target extraction method based on LRW and density clustering according to claim 3 is characterized in that: Edge refinement involves representing the image at multiple scales using a Gaussian pyramid and computing the gradient at each scale separately: Among them, Gσ represents the Gaussian smoothing kernel, σ represents the scale, and * represents the convolution operation; Based on edge refinement, the second-order derivative of the calculated gradient is used to further enhance the edge clarity: When the absolute value of the second-order derivative exceeds the set threshold, the point is considered to be at the edge; A convolutional neural network is used to learn and identify edge features in an image by training a network model, wherein the network model input is an LRW-processed image and the output is an edge detection result.

5. The image target extraction method based on LRW and density clustering according to claim 1, characterized in that: The S2 specifically includes: S21, extract the coordinate set of the pixel points in the superpixel (X * ,Y * ), extract the selected super pixel area separately onto a black background image of the same size as the original image, and the position and color of the super pixel block should be consistent with the original image; S22, according to the coordinate method, the minimum and maximum values ​​of the horizontal and vertical coordinates in the superpixel are represented as min(X * )、max(X * )、min(Y * )、max(Y * ), two by two are combined to form a rectangular area, and the coordinates are calibrated with 4 red points; S23, counting the eigenvalues ​​of the gray-level co-occurrence matrix of the superpixels in the calibration area, and completing the extraction of superpixel texture feature quantities of the superpixels.

6. The image target extraction method based on LRW and density clustering according to claim 5 is characterized in that: The S2 also includes extracting the superpixel texture features, combining the color and position information of the superpixel, and using a 10-dimensional vector to represent the i-th superpixel in the superpixel image from the three aspects of color, position and texture: C i =(l i ,a i ,b i ,x i ,y i ,MP i ,CON i ,IDM i ,ASM i ,ENT i ); Among them, the color feature is the average value of the internal pixel color components in the super pixel block, and the position feature is the center of gravity of each super pixel block, represented by coordinates x and y.

7. The image target extraction method based on LRW and density clustering according to claim 1 is characterized in that: The ε describes the neighborhood distance threshold of a certain sample, and Minpts describes the threshold of the number of samples in the neighborhood of a certain sample whose distance is ε.

Images