A polarization image segmentation method

Abstract

The present invention provides a method for polarization image segmentation, which can perform image enhancement according to the different characteristics of polarization images and can effectively improve the ability to extract polarization image features. The present invention obtains the S0 image, S1 image, S2 image and DoP image in the original polarization image, and performs data enhancement on the S0 image, S1 image, S2 image and DoP image according to the different polarization image characteristics; stacks the four enhanced images together as input data for image segmentation; converts the input data into a graph structure, and uses a greedy algorithm to perform graph structure segmentation according to the judgment criteria of the region boundary, thereby realizing the segmentation of the polarization image. Image enhancement combined with polarization characteristics strengthens the polarization information to a certain extent, effectively improves the image segmentation accuracy, and at the same time, combined with the image segmentation ideas of graph theory, detailed segmentation results can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of polarization imaging and image processing, and in particular to a polarization image segmentation method based on graph theory. Background Art

[0002] Polarization refers to the phenomenon in which the vibration vector of a shear wave (perpendicular to the direction of wave propagation) deviates in certain directions. It is the most obvious characteristic that distinguishes shear waves from longitudinal waves. Optical polarization imaging can obtain polarization information in multiple dimensions. In various natural environments, the polarization images obtained by polarization imaging systems are rich in detailed information. Unlike visible light, infrared, and spectral images, polarization images have important application value in complex backgrounds, low-light environments, and strong scattering environments. Furthermore, factors such as the target material, surface roughness, and observation angle can also affect the polarization characteristics of polarization images, which is why polarization images are widely used in target detection.

[0003] Image segmentation is the process and method of dividing an image into specific regions and extracting objects of interest. It is the foundation of computer vision and a crucial component of understanding image information. It is also one of the most challenging problems in image processing. Image segmentation can divide an image into distinct regions based on characteristics such as grayscale, color, texture, and edge shape, while also allowing regions to share similar features. Image segmentation can also be understood as the process of extracting regions of interest.

[0004] Image segmentation methods can be divided into two categories: traditional methods and deep learning-based methods. Traditional segmentation methods include threshold-based segmentation methods, edge-based segmentation methods, region-based segmentation methods, wavelet analysis and wavelet transform-based segmentation methods, graph theory-based segmentation methods, and genetic algorithm-based classification algorithms. These methods mainly utilize knowledge in digital image processing, topology, mathematics, and other fields. They are highly interpretable and can be used in conjunction with specific image segmentation scenarios. Deep learning-based image segmentation methods have developed rapidly in recent years due to their high accuracy and are widely used in unmanned driving, augmented reality, face detection, biomedicine, and other fields. However, compared with traditional methods, they are more complex and require large data sets for training. Training for specific scenarios also takes a lot of time.

[0005] Although many methods have been developed for image segmentation technology, there are not many examples of their specific application to polarization images. Moreover, due to the complexity and diversity of image information, image segmentation is relatively troublesome. Different segmentation methods also produce different results. At the same time, the noise and blur of the image itself will also affect the segmentation results.

[0006] In summary, it is necessary to propose an image segmentation method for polarization images, which can enhance the image according to the different characteristics of the polarization image. This can improve the accuracy of image segmentation while improving the polarization characteristics in the polarization image. Summary of the Invention

[0007] To solve the above problems, the present invention provides a polarization image segmentation method, which is an image segmentation method that combines the characteristics of polarization images. This method can effectively improve the ability to extract polarization image features, and at the same time, the image segmentation idea based on graph theory can obtain better image segmentation results.

[0008] A polarization image segmentation method comprises the following steps:

[0009] S1: Acquire the S0, S1, S2, and DoP images from the original polarization image, and perform data enhancement on the S0, S1, S2, and DoP images from the original image based on different polarization image characteristics;

[0010] S2: Stack the four images S0, S1, S2, and DoP after data enhancement in step S1 together to obtain a stacked image, which is used as input data for image segmentation;

[0011] S3: Filter the stacked image obtained in step S2, and then convert the filtered image into a graph structure. Segment the converted graph structure using a greedy algorithm according to the judgment criteria of the region boundary to complete the segmentation of the polarization image.

[0012] In step S1, the method for obtaining the S0 image, S1 image, S2 image, and DoP image in the original polarization image is:

[0013] S101: The polarization state of light can be described by the Stokes vector, which has four parameters S0, S1, S2 and S3. The theoretical definition of these four parameters is

[0014]

[0015] Where, I x Represents the polarization diagram corresponding to the x polarization direction, x = 0°, 45°, 90°, 135°; I l and I r The distribution represents left-handed polarization and right-handed polarization. S0 represents the intensity diagram, S1 and S2 represent two linear polarization diagrams with orthogonal polarization directions, and S3 represents the circular polarization diagram. In natural environments, the circular polarization characteristics are not obvious and can be ignored. The degree of polarization (DoP) and angle of polarization (AoP) are calculated using the four Stokes parameters.

[0016]

[0017] Different polarization images have different characteristics. Since the polarization angle image has relatively large noise, the four images S0, S1, S2 and DoP images are used as the original input data for image segmentation.

[0018] S102: By observing these images, we can find that the S0 image mainly contains intensity information, while the S1, S2 and DoP images mainly contain the contour information and edge information of the object. Therefore, the present invention uses different image processing methods to enhance the initial data;

[0019] In step S102, the method for performing data enhancement on the S0 image, S1 image, S2 image, and DoP image in the acquired original image is:

[0020] S103: Use histogram equalization to process the S0 image to enhance the image contrast and highlight the target object;

[0021] S104: For contour edge information, mainly some high-frequency signals, the exponential high-pass filter (EHPF) method is used to sharpen the S1 image, S2 image, and DoP image so that subsequent image segmentation can better extract image features;

[0022] The specific process of sharpening is to highlight the edges and contours of the image by high-pass filtering the Fourier spectrum of the image. The exponential high-pass filter has a smoother transfer function than the Butterworth high-pass filter, so the resulting image ringing phenomenon is weaker and there is no ringing phenomenon in the high-frequency image. The transfer function of the exponential high-pass filter is as follows:

[0023]

[0024] Where D0 is the cutoff frequency from the origin of the frequency domain plane, n is the filter order.

[0025] Specifically, the first step is to perform a center transformation on the polarization image; perform a Fourier transform on the result image of the first step, that is, to calculate F(u,v); use the exponential high-pass filter transfer function H(u,v) to multiply F(u,v); calculate the inverse Fourier transform and obtain the real part of the result; add the result and the original image to obtain the result image after image sharpening.

[0026] S301: In step S3, the stacked image is filtered using a Gaussian filter function.

[0027] Gaussian filtering can preserve the overall grayscale distribution characteristics of the image while smoothing the image. The present invention uses a sliding window to implement convolution. The Gaussian kernel is used when discretizing the window convolution. The size of the Gaussian kernel is an odd number because the Gaussian convolution will output the result at the center of the coverage area. Specifically, the Gaussian template is calculated by the Gaussian function, and the formula is as follows:

[0028]

[0029] Where σ is the standard deviation. The Gaussian filter template is sampled with the center of the template as the coordinate origin. The template has corresponding coordinates at each position, with the x-axis horizontally to the right and the y-axis vertically upward. The coordinates x and y of each position are substituted into the above Gaussian function G to obtain the value of each position. The Gaussian template is obtained, and the Gaussian template is used to perform Gaussian filtering on each channel image of the input data.

[0030] S302: The method of converting to a graph structure is: taking pixels in space as vertices of the graph, and taking the differences between different pixels as the edge weights between the vertices of the graph, thereby converting the image segmentation problem into a segmentation problem of a graph structure; a graph is a data structure consisting of a vertex set and a set of binary relationships between vertices, usually represented by G = (V, E), where the vertex set is represented by V, the elements in V are called vertices, represented by v, and the number of vertices is also called the order of the graph, represented by n; the edge set is represented by E, the elements in E are called edges, represented by e, and the number of edges is represented by m. In the edge set, a pair of vertices constitutes an unordered pair (v i ,v j ), representing the vertex v i and vertex v j If all edges in a graph do not have a definite direction, it is called an undirected graph. i ,v j ) and (v j ,v i ) is the same edge. The image is expressed as a graph in graph theory, that is, each pixel in the image is regarded as a vertex v∈V, and each pixel and its adjacent pixel form an edge e∈E of the graph. The graph is represented as G=(V,E). The weight w of each edge of the graph is the relationship between the pixel and the adjacent pixel, which expresses the similarity between adjacent pixels. The meaning of the weight is the dissimilarity between vertices. The graph used is an undirected graph;

[0031] A tree is a special kind of graph, that is, any two vertices in the graph are connected by a path, but there is no loop. The minimum spanning tree is a tree with the smallest sum of edge weights given the vertices to be connected. Each pixel is initialized as a vertex, and the process of gradually merging into a region is a minimum spanning tree connecting the pixels in this region, which corresponds to a segmented region.

[0032] The four-channel grayscale distance is used to measure the similarity between two points, namely

[0033]

[0034] Among them, S 01 is any pixel value in the S0 image, S 02 For S 01 Eight neighborhood pixel values; S 11 is any pixel value in the S0 image, S 12 For S 11 Eight neighborhood pixel values; S 21 is any pixel value in the S0 image, S 22 For S 21 The eight-neighborhood pixel values ​​of DOP1; DOP1 is any pixel value in the DOP image, and DOP2 is the eight-neighborhood pixel values ​​of DOP1;

[0035] Each pixel is equivalent to a one-dimensional vector. The Euclidean distance between two pixels is used as the edge weight. For the similarity between two regions or a region and a pixel, an adaptive threshold is used instead of a global threshold, so that the segmentation result is neither too coarse nor too fine.

[0036] S303: The criteria for determining the region boundary are:

[0037] For two regions (components), define the intra-class difference Int(C) of a region:

[0038]

[0039] The maximum gray value difference within a region is defined as the edge with the largest dissimilarity in the minimum spanning tree;

[0040] The difference between the two regions Diff(C1, C2):

[0041]

[0042] That is, the dissimilarity of the edge with the smallest dissimilarity among all the edges connecting the two regions, that is, the dissimilarity of the most similar parts of the two regions. If there is no edge connecting the two segments, define Diff(C1, C2) = ∞;

[0043] A more intuitive standard for judging whether to merge:

[0044] Diff(C1,C2)≤min(Int(C1),Int(C2))

[0045] Among them, the intra-class difference of the minimum split is defined as follows:

[0046] min(Int(C1),Int(C2))=min(Int(C1)+τ(C1)),Int(C2)+τ(C2)))

[0047] The threshold function τ here is mainly used to better control the boundary definition of the segmented area. Intuitively, the boundary definition of the small segmented area is stronger than that of the large segmented area. Otherwise, the small segmented areas can be merged to form a large area. The threshold function is related to the area size:

[0048]

[0049] C refers to the number of vertices in the segmented part, and k is a parameter that can be adjusted according to different needs, such as the image size. k can control the size of the formed area. When k is 0, almost every pixel is an independent area. When k approaches infinity, the entire image will be gathered into one area.

[0050] Beneficial effects

[0051] This paper applies graph theory to polarization image segmentation for the first time. By transforming polarization image segmentation into graph-structured segmentation and incorporating the polarization properties of polarization images, it effectively improves the accuracy of image segmentation in reflective environments. The graph-theoretic segmentation method incorporates an adaptive threshold, allowing the precision of the segmentation results to be adjusted based on image size. In addition to the grayscale distance used in this method, other methods, such as texture feature filtering and distance calculation, can also be used to determine regional similarity.

[0052] 2. According to the image characteristics of different polarization images, different image enhancement methods are proposed. Histogram equalization is used to perform contrast enhancement on the S0 image, and exponential high-pass filtering is used to sharpen the S1 image, S2 image and DoP image, which enhances the polarization characteristics of the image and effectively improves the image segmentation accuracy.

[0053] 3. The present invention uses four-channel images, namely, S0 image, S1 image, S2 image and DoP image, as the original data for image segmentation, effectively obtaining different features of the polarization image, and can achieve better segmentation effect than other schemes. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A flow chart of a polarization image segmentation method implemented in the present invention;

[0055] Figure 2 Comparison chart of the original S0 image, S1 image, S2 image, DoP image and the image enhancement results;

[0056] Figure 3 is a 5×5 Gaussian template;

[0057] Figure 4 is the coordinate representation of the Gaussian function template;

[0058] Figure 5 This is the segmentation result before image enhancement;

[0059] Figure 6 The segmentation result after image enhancement. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] like Figure 1 As shown, a polarization image segmentation method specifically includes the following steps:

[0062] S1: If Figure 2 As shown, the S0 image, S1 image, S2 image and DoP image in the original polarization image are obtained, and data enhancement is performed on the S0 image, S1 image, S2 image and DoP image in the obtained original image according to different polarization image characteristics;

[0063] In step S1, the method for obtaining the S0 image, S1 image, S2 image, and DoP image in the original polarization image is:

[0064] S101: The polarization state of light can be described by the Stokes vector, which has four parameters S0, S1, S2 and S3. The theoretical definition of these four parameters is

[0065]

[0066] Where, I x Represents the polarization diagram corresponding to the x polarization direction, x = 0°, 45°, 90°, 135°; I l and I r The distribution represents left-handed polarization and right-handed polarization. S0 represents the intensity diagram, S1 and S2 represent two linear polarization diagrams with orthogonal polarization directions, and S3 represents the circular polarization diagram. In natural environments, the circular polarization characteristics are not obvious and can be ignored. The degree of polarization (DoP) and angle of polarization (AoP) are calculated using the four Stokes parameters.

[0067]

[0068] Different polarization images have different characteristics. Since the polarization angle image has relatively large noise, the S0 image, S1 image, S2 image, and DoP image are used as the original input data for image segmentation.

[0069] S102: By observing these images, we can find that the intensity map S0 mainly contains intensity information, while S1, S2 and DoP images mainly contain object contour information and edge information. Therefore, the present invention uses different image processing methods to enhance the initial data.

[0070] In step S102, the method for performing data enhancement on the S0 image, S1 image, S2 image, and DoP image in the acquired original image is:

[0071] S103: Using histogram equalization to process the intensity map, the image contrast is enhanced and the target object is highlighted. Histogram equalization is a simple and effective image enhancement technique that changes the grayscale of each pixel in the image by changing the image histogram. It is mainly used to enhance the contrast of images with a small dynamic range.

[0072] Furthermore, for the S0 image, its histogram reflects the statistical situation of the different gray levels of the image. Strictly speaking, the gray level histogram of an image is a one-dimensional discrete function, which can be written as:

[0073] h(k)=n k k=0,1,...,L-1

[0074] Where n k is the number of pixels with gray level k in the image. The height of each column of the histogram corresponds to n k The histogram provides the distribution of various grayscale values ​​in the original image. Based on the histogram, the relative frequency P of the normalized histogram grayscale is further defined. r (k). That is:

[0075] P r (k) = n k / N

[0076] Where N represents the total number of image pixels, n k is the number of pixels with gray level k in the image.

[0077] Let r and s represent the grayscale of the original image and the grayscale of the image after histogram equalization, respectively. For the convenience of discussion, r and s are normalized so that 0 ≤ r, s ≤ 1. For a given image, the grayscale distribution after normalization is within the range of 0 ≤ r ≤ 1. For any value of r within this range, the following transformation is performed:

[0078] The transformation should satisfy the condition: s=T(r)

[0079] (1) In the range 0≤r≤1, T(r) is a monotonically increasing function.

[0080] (2) In the range 0≤r≤1, 0≤T(r)≤1.

[0081] The inverse transformation from s to r is expressed as follows:

[0082] r=T -1 (s)

[0083] The probability density function of r is P r (r), the probability density of s can be expressed by P r (r) Find

[0084]

[0085] Assume that the transformation function is

[0086]

[0087] Where ω is the integration variable, and It is the cumulative distribution function of r.

[0088] Taking the derivative of r in the formula, we get

[0089]

[0090] Substitute the result into the previous formula

[0091]

[0092] After the transformation, the probability density of the variable s within its domain is uniformly distributed. Therefore, using the cumulative distribution function of r as the transformation function can produce an image with a uniform probability distribution density of grayscale levels. As a result, the dynamic range of pixel values ​​is expanded.

[0093] S104: For contour edge information, mainly some high-frequency signals, the S1 image, S2 image and DoP image are sharpened using the exponential high-pass filtering (EHPF) method so that subsequent image segmentation can better extract image features.

[0094] Sharpening involves applying a high-pass filter to the Fourier spectrum of an image to emphasize edges and contours. Exponential high-pass filtering, with its smoother transfer function than the Butterworth high-pass filter, produces less ringing in the image and eliminates ringing in high-frequency images. The transfer function of an exponential high-pass filter is shown below.

[0095]

[0096] Where D0 is the cutoff frequency from the origin of the frequency domain plane, and D(u,v) is n is the filter order.

[0097] Specifically, the first step is to perform a center transformation on the polarization image; perform a Fourier transform on the result image of the first step, that is, to calculate F(u,v); use the exponential high-pass filter transfer function H(u,v) to multiply F(u,v); calculate the inverse Fourier transform and obtain the real part of the result; add the result and the original image to obtain the result image after image sharpening.

[0098] S2: Stack the four images S0, S1, S2, and DoP after data enhancement in step S1 together to obtain a stacked image, which is used as input data for image segmentation;

[0099] S3: Filter the stacked image obtained in step S2, and then convert the filtered image into a graph structure. Segment the converted graph structure using a greedy algorithm according to the judgment criteria of the region boundary to complete the segmentation of the polarization image.

[0100] S301: In step S3, the stacked image is filtered using a Gaussian filter function.

[0101] Gaussian filtering can smooth an image while preserving its overall grayscale distribution. In image processing, there are two main methods for implementing Gaussian filtering: one is discretized window sliding window convolution, and the other is Fourier transform. Fourier transform is only considered when the discretized window is very large and the sliding window calculation is very large. Therefore, the present invention uses a sliding window to implement convolution.

[0102] The Gaussian kernel is used when discretizing the window convolution. The size of the Gaussian kernel is an odd number because the Gaussian convolution will output the result at the center of the coverage area. The Gaussian kernel size used in this invention is 5×5, and its form is as follows Figure 3 shown.

[0103] Specifically, the Gaussian template is calculated by the Gaussian function, and the formula is as follows:

[0104]

[0105] Where σ is the standard deviation, and the 5×5 Gaussian filter template is sampled with the center of the template as the coordinate origin. The coordinates of the template at each position are as follows: Figure 4 As shown, the x-axis is horizontally to the right, and the y-axis is vertically upward. Substituting the coordinates of each position into the Gaussian function G above, we obtain the value at each position and thus obtain the Gaussian template. This Gaussian template is used to perform Gaussian filtering on each channel image of the input data.

[0106] S302: The method of converting to a graph structure is: taking pixels in space as vertices of the graph, and taking the differences between different pixels as the edge weights between the vertices of the graph, thereby converting the image segmentation problem into a segmentation problem of a graph structure; a graph is a data structure consisting of a vertex set and a set of binary relationships between vertices, usually represented by G = (V, E), where the vertex set is represented by V, the elements in V are called vertices, represented by v, and the number of vertices is also called the order of the graph, represented by n; the edge set is represented by E, the elements in E are called edges, represented by e, and the number of edges is represented by m. In the edge set, a pair of vertices constitutes an unordered pair (v i ,v j ), representing the vertex v i and vertex v j If all edges in a graph do not have a definite direction, it is called an undirected graph. i ,v j ) and (v j ,v i ) is the same edge. The image is expressed as a graph in graph theory, that is, each pixel in the image is regarded as a vertex v∈V, and each pixel and its adjacent pixel form an edge e∈E of the graph. The graph is represented as G=(V,E). The weight w of each edge of the graph is the relationship between the pixel and the adjacent pixel, which expresses the similarity between adjacent pixels. The meaning of the weight is the dissimilarity between vertices. The graph used is an undirected graph;

[0107] A tree is a special kind of graph, that is, any two vertices in the graph are connected by a path, but there is no loop. The minimum spanning tree is a tree with the smallest sum of edge weights given the vertices to be connected. Each pixel is initialized as a vertex, and the process of gradually merging into a region is a minimum spanning tree connecting the pixels in this region, which corresponds to a segmented region.

[0108] The four-channel grayscale distance is used to measure the similarity between two points, namely

[0109]

[0110] Among them, S 01 is any pixel value in the S0 image, S 02 For S 01 Eight neighborhood pixel values; S 11 is any pixel value in the S0 image, S 12 For S 11 Eight neighborhood pixel values; S 21 is any pixel value in the S0 image, S 22 For S 21 The eight-neighborhood pixel values ​​of DOP1; DOP1 is any pixel value in the DOP image, and DOP2 is the eight-neighborhood pixel values ​​of DOP1;

[0111] Each pixel is equivalent to a one-dimensional vector. The Euclidean distance between two pixels is used as the edge weight. For the similarity between two regions or a region and a pixel, an adaptive threshold is used instead of a global threshold, so that the segmentation result is neither too coarse nor too fine.

[0112] S303: The criteria for determining the region boundary are:

[0113] For two regions (components), define the intra-class difference Int(C) of a region:

[0114]

[0115] The maximum gray value difference within a region is defined as the edge with the largest dissimilarity in the minimum spanning tree;

[0116] The difference between the two regions Diff(C1, C2):

[0117]

[0118] That is, the dissimilarity of the edge with the smallest dissimilarity among all the edges connecting the two regions, that is, the dissimilarity of the most similar parts of the two regions. If there is no edge connecting the two segments, define Diff(C1, C2) = ∞;

[0119] A more intuitive standard for judging whether to merge:

[0120] Diff(C1,C2)≤min(Int(C1),Int(C2))

[0121] Among them, the intra-class difference of the minimum split is defined as follows:

[0122] min(Int(C1),Int(C2))=min(Int(C1)+τ(C1)),Int(C2)+τ(C2)))

[0123] The threshold function τ here is mainly used to better control the boundary definition of the segmented area. Intuitively, the boundary definition of the small segmented area is stronger than that of the large segmented area. Otherwise, the small segmented areas can be merged to form a large area. The threshold function is related to the area size:

[0124]

[0125] C refers to the number of vertices in the segmented part, and k is a parameter that can be adjusted according to different needs, such as the image size. k can control the size of the formed area. When k is 0, almost every pixel is an independent area. When k approaches infinity, the entire image will be gathered into one area.

[0126] Specifically, the segmentation algorithm steps are:

[0127] The first step is to sort the elements in the edge set E of the graph structure in ascending order according to the edge weight, and obtain the edge set E0 = (e1,...,e m ), m is the number of edges in the graph structure;

[0128] The second step is to set the edge e q The corresponding edge weight is w q , where q = 1, 2, 3, ..., m. Edge e q The two vertices of v i and v j , edge weight w q =(v i ,v j ). If vertex v i and vertex v j In two regions respectively, and the edge weight w q is smaller than the minimum intra-class difference, i.e. Then the vertex v i and vertex v j The two regions where the two vertices are located are merged to form one region, otherwise, the two regions where the two vertices are located are not merged;

[0129] In the third step, the second step is repeated according to the order of the edges in the edge set E0, the two areas connected by the edges are merged and judged, and the segmentation result is updated q times to obtain the final segmentation result.

[0130] in, For vertex v i Intra-class differences in the region, For vertex v j The intra-class difference of the region. The intra-class difference of the region Int(C) is the maximum edge weight in region C.

[0131] Figure 5 This is the segmentation result without image enhancement. The car's region of interest is segmented into the following components: the window glass, the plastic window frame, the tires, and the entire body sheet metal. The glass has the highest F1 value of 0.9796, followed by the body sheet metal at 0.9795. The plastic window frame has an F1 value of 0.8974, and the tires have an F1 value of 0.8872.

[0132] Figure 6This is the segmentation result after image enhancement. Calculations show a certain improvement in segmentation accuracy, with the F1 value for the car body sheet reaching 0.9782 and the glass 0.9715. The segmentation results for the plastic midframe and tires also show some improvement, with the F1 value for the plastic midframe reaching 0.9194 and the F1 value for the tire reaching 0.9809.

[0133] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A polarization image segmentation method, characterized in that The following steps are involved: S1: Acquire the S0, S1, S2, and DoP images from the original polarization image, and perform data enhancement on the S0, S1, S2, and DoP images from the original image based on different polarization image characteristics; S2: Stack the four images S0, S1, S2, and DoP after data enhancement in step S1 together to obtain a stacked image, which is used as input data for image segmentation; S3: filtering the stacked image obtained in step S2, and then converting the filtered image into a graph structure. Segmenting the converted graph structure using a greedy algorithm according to a region boundary judgment criterion to complete the segmentation of the polarization image. In the step S1, the S0 image, S1 image, S2 image and DoP image in the original polarization image are obtained as follows: The S0 image represents the intensity image, S0=I 0° +I 90° ; The S1 image and the S2 image represent two linear polarization images with polarization directions orthogonal to each other. S1 = I 0° -I 90° , S2=I 45° +I 135° ; Where, I x represents the polarization diagram corresponding to the x polarization direction, x = 0°, 45°, 90°, 135°; DoP image stands for degree of polarization image. In step S3, the method for converting the filtered image into a graph structure is as follows: the pixels of the filtered image are used as vertices v of the graph structure, all pixels constitute the vertex set V of the graph structure, the differences between the eight neighborhoods of the pixels of the filtered image are used as the edge weights w of the graph structure, the edge represented by the edge weight is e, the set of all edges is E, and the generated graph structure is represented as G = (V, E). The edge weight w of the graph structure is: Among them, S 01 is any pixel value in the S0 image, S 02 For S 01 Eight neighborhood pixel values; S 11 is any pixel value in the S0 image, S 12 For S 11 Eight neighborhood pixel values; S 21 is any pixel value in the S0 image, S 22 For S 21 The eight-neighborhood pixel values ​​of DOP1; DOP1 is any pixel value in the DOP image, and DOP2 is the eight-neighborhood pixel values ​​of DOP1; The first step is to sort the elements in the edge set E of the graph structure in ascending order according to the edge weight, and obtain the edge set E0 = (e1,...,e m ), m is the number of edges in the graph structure; The second step is to set the edge e q The corresponding edge weight is w q , where q = 1, 2, 3, ..., m, and edge e q The two vertices of v i and v j , edge weight w q =(v i ,v j ), if vertex v i and vertex v j In two regions respectively, and the edge weight w q is smaller than the minimum intra-class difference, i.e. Then the vertex v i and vertex v j The two regions where the two vertices are located are merged to form one region, otherwise, the two regions where the two vertices are located are not merged; In the third step, the second step is repeated according to the order of the edges in the edge set E0, the two areas connected by the edge are merged and judged, and the segmentation result is updated q times to obtain the final segmentation result; in, For vertex v i Intra-class differences in the region, For vertex v j The intra-class difference of the region, the intra-class difference Int(C) of the region is the maximum edge weight in region C.

2. The polarization image segmentation method according to claim 1, wherein: In the step S1, performing data enhancement refers to processing the S0 image using histogram equalization.

3. The polarization image segmentation method according to claim 1, wherein: In the step S1, the S1 image, the S2 image and the DoP image are sharpened using an exponential high-pass filtering method.

4. The polarization image segmentation method according to claim 3, wherein: The sharpening process is specifically as follows: high-pass filtering is performed on the Fourier spectra in the S1 image, the S2 image, and the DoP image using the following transfer function, where the transfer function H(u,v) is: Where D0 is the cutoff frequency from the origin of the frequency domain plane, and D(u,v) is n is the filter order.

5. A polarization image segmentation method according to any one of claims 1 to 4, characterized in that: In the step S3, when filtering is performed: a Gaussian filter function is used to filter the stacked image.

Images