Infrared polarization noise suppression and edge enhancement method based on gradient vector analysis

Abstract

The invention belongs to the technical field of image processing, and discloses an infrared polarization noise suppression and edge enhancement method based on gradient vector analysis, which comprises the following steps of: according to spatial distribution similarity in a polarized image, performing preliminary highlight noise suppression on a polarization degree image based on a double-offset spatial neighborhood gradient weighting method; bidirectional feature coherence suppression is provided by using the difference of polarization characteristics of a small target in a sky background to obtain a significant target; the problem of excessive suppression of target edge information in the background suppression process is prevented, and target edge information enhancement is performed based on a gradient vector field connectivity statistical analysis method. According to the infrared polarization noise suppression and edge enhancement method based on gradient vector analysis, background highlight point noise of a polarization degree image is preliminarily suppressed based on a double-offset space neighborhood gradient weighting method; a saliency target is effectively obtained based on bidirectional feature coherence suppression; the shape information of the target is enhanced based on a gradient vector field connectivity statistical analysis method.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to an infrared polarization noise suppression and edge enhancement method based on gradient vector analysis. Background Art

[0002] Since infrared intensity images are easily disturbed by external heat sources, the image quality is affected. And when the temperature difference between the target and the background is not obvious, the intensity contrast between the target and the background is low, making the target difficult to detect. In addition, with the development of infrared camouflage and jamming technology, the difficulty of infrared detection has further increased. With the development of polarization detection technology, polarization imaging detection technology has become an important detection method. Polarization describes the regularity of the vibration direction of light waves. When all objects with temperatures above absolute zero emit electromagnetic radiation, the asymmetry of electromagnetic radiation will produce infrared polarization characteristics. Since the infrared polarization characteristics are not affected by external light changes and are only related to their own physical properties. Therefore, infrared polarization images can better distinguish between targets and backgrounds by capturing the polarization information of light. The research on infrared polarization images has good application prospects in military target detection, environmental monitoring, industrial detection and other fields. However, the basic parameters in polarization are calculated from the measured intensity by nonlinear operators, which is easy to amplify the noise in the intensity measurement, which means that the polarization information and the target are easily submerged in the noise, affecting the performance of polarization imaging. Especially for targets with shapes, the use of infrared polarization information to suppress background noise is easy to destroy the edge information of the target, resulting in a decrease in the detection quality of small targets. Therefore, studying how to suppress background noise of shaped targets to improve image quality is an urgent problem to be solved.

[0003] In order to enhance the edge information of small targets while suppressing background noise, researchers have proposed many related algorithms, which already include denoising steps in the process of infrared polarization image enhancement. Because noise and image structure (including edges and details) belong to the same frequency component, the enhancement and denoising algorithms can be divided into two categories: traditional enhancement and denoising and deep learning network enhancement and denoising. In traditional enhancement and denoising, most of them are based on the frequency domain decomposition and spatial domain decomposition of the input image: the frequency domain is decomposed into low-frequency and high-frequency images; the spatial domain is decomposed into detail layer (top layer) and base layer (other layers). After that, different strategies are used to process the decomposed image to achieve the purpose of suppressing noise and enhancing edge information. Since noise also belongs to high-frequency information, in order to maintain the structure of the image while denoising, the neighborhood weighting method is used to design the contribution weight of adjacent pixels to the central pixel. In deep learning network enhancement, the input image is deeply learned with relevant features according to the convolutional neural network architecture, and then enhanced and denoised. Common network architectures include those based on auto-encoding, Resnet, and generative adversarial networks GAN.

[0004] Although the above algorithms have certain effects in denoising and enhancing, the noise of small targets in infrared polarization images is still a challenge. This is because the application of the above algorithms is based on the premise that the target can be seen in the infrared intensity map, but when there is a low contrast between the target and the background, the intensity information is not applicable. In addition, although the target is visible in the polarization image, the target is small and easily submerged by noise, which further increases the difficulty of denoising. Summary of the invention

[0005] The purpose of the present invention is to provide an infrared polarization noise suppression and edge enhancement method based on gradient vector analysis. When the target and background have low contrast and the target information is submerged in the background noise, the difference in infrared polarization characteristics between the target and the background in the gradient vector characteristics is used to suppress noise and enhance the edge of the small target, thereby distinguishing the target from the background.

[0006] To achieve the above object, the present invention provides an infrared polarization noise suppression and edge enhancement method based on gradient vector analysis, comprising the following steps:

[0007] Step S1, first, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighting method, the polarization degree image is preliminarily suppressed to suppress the highlight point noise;

[0008] Step S2: Secondly, using the difference in polarization characteristics of small targets in the sky background, a bidirectional feature coherence suppression is proposed to obtain a salient target;

[0009] Step S3: Finally, in order to prevent the problem of excessive suppression of target edge information during background suppression, the target edge information is enhanced based on the connected statistical analysis method of the gradient vector field.

[0010] Preferably, in step S1, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighting method, the polarization degree image is preliminarily suppressed for highlight point noise, and the specific process is as follows:

[0011] Step S11: In a polarization imaging system, a Stokes vector is used to represent the polarization state of light. The general definition of the Stokes vector is as follows:

[0012] S= ( I,Q,U,V ) T ;

[0013] Among them, S represents the Stokes vector matrix; I represents the total radiation intensity of light; Q represents the difference in linear polarization intensity of light in the horizontal and vertical directions; U represents the difference in linear polarization intensity of light in the +45° and -45° directions; V represents the circular polarization characteristic of light, that is, the intensity difference between left-handed circular polarization and right-handed circular polarization; T represents matrix transposition;

[0014] In the natural environment, the V component of circularly polarized light is very small and can be ignored, so the Stokes vector degenerates into a third-order matrix as shown below:

[0015] S= ( I,Q,U ) T ;

[0016] Step S12: Select 0°, 60° and 120° to obtain objects I0, I 60 and I 120 The polarization intensity of the light is expressed by the Stokes parameter, where the Stokes vector is as follows:

[0017]

[0018] Step S13, in the Stokes parameters, Q and U describe the polarization information of the target surface in different directions. Since they have similarity in spatial distribution, spatial neighborhood gradient weighting is performed on two different polarization images;

[0019] Step S14: According to the obtained spatial neighborhood gradients of the polarized images Q and U, the Fimage image is obtained by linear weighted fusion, as shown below:

[0020]

[0021] Among them, Fimage represents the weighted fusion result of spatial neighborhood gradient; G Q_t represents the weighted gradient of image Q; G U_t Represents the weighted gradient of image U;

[0022] Step S15: Use the Fimage image to replace the original polarized image, perform preliminary suppression on the polarization degree image, and obtain the FDoP image, as shown below:

[0023]

[0024] Among them, FDoP represents the processing result of the double-biased spatial neighborhood gradient weighted method.

[0025] Preferably, in step S12, the polarization degree calculation formula is as follows:

[0026]

[0027] Among them, I pol Represents the intensity of the fully polarized component; I total Represents the total intensity in the light vector.

[0028] Preferably, in step S13, the calculation formula of the spatial neighborhood gradient weighting is as follows:

[0029] G t =G ν +G h ;

[0030] G h =∣X(x,y)-X(x,y-1)∣+∣X(x,y)-X(x,y+1)∣;

[0031] G ν =∣X(x,y)-X(x-1,y)∣+∣X(x,y)-X(x+1,y)∣;

[0032]

[0033] in, ( x,y ) is the coordinate of the pixel in the image; G in the image t It's pixels ( x,y ) The weighted gradient of G consists of two directions; h is the gradient in the horizontal direction; G v is the gradient in the vertical direction; | W | is the number of pixels in the neighborhood window; g ( x 1) is the Gaussian weight function; f ( x+x1 and f(x) are the grayscale values ​​of the pixels in the neighborhood and the central pixel respectively.

[0034] Preferably, in step S2, by utilizing the difference in polarization characteristics of the small target in the sky background, a bidirectional feature coherence suppression is proposed to obtain a significant target, and the specific process is as follows:

[0035] Step S21, first, remove the average direction of the column from the Fimage image; if the obtained column direction value is less than 0, it is represented as the background, and its grayscale value is set to 0; otherwise, the original grayscale value is maintained;

[0036] Step S22, remove the average direction of the row from the Fimage image again; if the obtained row direction value is less than 0, it is represented as the background, and its gray value is set to 0; otherwise, the original gray value is maintained;

[0037] Step S23, obtaining a CR_image image according to the operation of the bidirectional average direction of step S21 and step S22;

[0038] The calculation formula of bidirectional polarization coherence suppression is as follows:

[0039]

[0040] Among them, C_Fimage represents the average direction result of the original image without its columns; CR_image represents the result of bidirectional feature coherence suppression; Fimage i,j represents the weighted fusion result of the spatial neighborhood gradient at position (i, j); N represents the total number of column pixels; k represents the number of pixels; M represents the total number of row pixels; C_Fimage k,j Indicates at location ( k,j ) The average orientation result of the original image is removed from its columns.

[0041] Preferably, in step S3, the target edge information is enhanced based on the connected statistical analysis method of the gradient vector field, and the specific process is as follows:

[0042] Step S31, calculating the gradient vector field, and analyzing the characteristics of small targets in the gradient vector field;

[0043] Step S32: estimating the gradient in a local window based on all pixels in the area to obtain the horizontal gradient and vertical gradient of the image;

[0044] Step S33, using the least squares method to fit the grayscale value in the local window, fitting to obtain the light intensity change at a point in any direction, and obtaining the magnitude and direction of the gradient;

[0045] Step S34: Analyze the significant areas of gradient changes in the image by calculating the gradient vector field and using connected components.

[0046] Preferably, in step S31, each point in the gradient vector field represents the gradient of the corresponding point in the original image, and the gradient calculation formula is as follows:

[0047]

[0048] in, Represents image I at pixel point ( i,j ) The gradient vector at ; and Respectively expressed in pixels ( i,j ) The horizontal and vertical gradients of the image at a point.

[0049] Preferably, in step S32, the gradient in a local window is estimated based on all pixels in the area, and the specific process is as follows:

[0050] In a local window, a quadratic polynomial is used to represent the grayscale value of the image as follows:

[0051] I ( x,y ) ≈a0+a1x+a2y+a3x 2 +a4xy+a5y 2 ;

[0052] Among them, the horizontal gradient of the image at point (x, y) and vertical gradient As shown below:

[0053]

[0054] Among them, a1, a2, a3, a4, and a5 represent the estimated coefficients in the polynomial.

[0055] Preferably, in step S33, the grayscale value in the local window is fitted using the least squares method to obtain the light intensity change at a point in any direction, wherein the magnitude and direction of the gradient are as follows:

[0056]

[0057] Among them, M ( x,y ) represents the modulus of the gradient; D ( x,y ) Indicates the direction of the gradient.

[0058] Preferably, in step S34, by calculating the gradient vector field, the connected components are used to analyze the significant area of ​​gradient change in the image, and the specific process is as follows:

[0059] Step S341: Using the binary image obtained by thresholding, the pixels with gradient values ​​greater than the threshold are marked as significant change areas, and the other pixels are marked as background; wherein the binary image formula is as follows:

[0060]

[0061] Among them, Binary Map ( x,y ) Represents a binary image;

[0062] Step S342: Count the connected components of the binary image by marking them; the connected component marking is to calculate the area and centroid position of each connected component, and the area of ​​each connected component is the number of pixels it contains; the calculation formula of the area and centroid position is as follows:

[0063] Area ( k ) =∑ x,y 1 ( label( x,y ) =k ) ;

[0064]

[0065] Among them, Area ( k ) Indicates the area of ​​the connected component; label ( x,y ) Represents a label at position (x, y); Centroid ( k ) represents the centroid position of the connected component; 1(·) is an indicator function, which is 1 when the condition is true and 0 otherwise; k represents the proxy coefficient;

[0066] Step S343: Count the number of connected areas through connected component marking; the expression for the number of connected components is as follows:

[0067]

[0068] Among them, Num_D represents the number of connected components; num labels represents the number of connected component labels.

[0069] Therefore, the present invention adopts the above infrared polarization noise suppression and edge enhancement method based on gradient vector analysis, and the beneficial effects are as follows:

[0070] (1) The present invention proposes a double-biased spatial neighborhood gradient weighting method to preliminarily suppress background highlight noise in the polarization degree image;

[0071] (2) The present invention proposes bidirectional feature coherence suppression to effectively obtain salient targets;

[0072] (3) The present invention proposes a connected statistical analysis method based on the gradient vector field to enhance the shape information of the target.

[0073] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of the infrared polarization noise suppression and edge enhancement method based on gradient vector analysis of the present invention;

[0075] Figure 2 It is the overall framework diagram of the infrared polarization noise suppression and edge enhancement method based on gradient vector analysis of the present invention;

[0076] Figure 3 It is a bidirectional feature coherence suppression framework diagram of the present invention;

[0077] Figure 4 is a visualization diagram of the statistics of connected components of the present invention;

[0078] Figure 5 It is the final result diagram after the processing of the present invention. DETAILED DESCRIPTION

[0079] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0080] like Figure 1 and Figure 2 As shown, the infrared polarization noise suppression and edge enhancement method based on gradient vector analysis includes the following steps:

[0081] Step S1, first, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighting method, the polarization degree image is preliminarily suppressed to suppress the highlight point noise;

[0082] Step S2: Secondly, using the difference in polarization characteristics of small targets in the sky background, a bidirectional feature coherence suppression is proposed to obtain a salient target;

[0083] Step S3: Finally, in order to prevent the problem of excessive suppression of target edge information during background suppression, the target edge information is enhanced based on the connected statistical analysis method of the gradient vector field.

[0084] Example

[0085] Step S1: First, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighted method, the polarization degree image is preliminarily suppressed for highlight point noise.

[0086] Step S11: In a polarization imaging system, a Stokes vector is usually used to represent the polarization state of light. The general definition of the Stokes vector is as follows:

[0087] S=(I,Q,U,V) T ;

[0088] Among them, S represents the Stokes vector matrix; I represents the total radiation intensity of light; Q represents the difference in linear polarization intensity of light in the horizontal and vertical directions; U represents the difference in linear polarization intensity of light in the +45° and -45° directions; V represents the circular polarization characteristic of light, that is, the intensity difference between left-handed circular polarization and right-handed circular polarization; T represents matrix transpose.

[0089] In the natural environment, the V component of circularly polarized light is very small and can be ignored, so the Stokes vector degenerates into a third-order matrix as shown below:

[0090] S=(I,Q,U) T ;

[0091] Step S12: Select 0°, 60° and 120° to obtain objects I0, I 60 and I 120 The polarization intensity, the Stokes vector, is given by:

[0092]

[0093] The Stokes parameter is used to represent the polarization degree of light. The physical meaning of the polarization degree is the ratio of the intensity of the completely polarized component in the light vector to the total intensity in the light vector. The calculation formula for the polarization degree is as follows:

[0094]

[0095] Among them, I pol Represents the intensity of the fully polarized component; I total Represents the total intensity in the light vector.

[0096] Step S13: In the Stokes parameters, Q and U describe the polarization information of the target surface in different directions. Due to their similarity in spatial distribution, spatial neighborhood gradient weighting (BNGW) is performed on two different polarization images. The calculation formula of spatial neighborhood gradient weighting is as follows:

[0097] G t =G ν +G h ;

[0098] G h =|X ( x,y ) -X ( x,y-1 ) ∣+∣X ( x,y ) -X ( x,y+1 ) ∣;

[0099] G ν =|X ( x,y ) -X ( x-1,y ) ∣+∣X ( x,y ) -X ( x+1,y ) ∣;

[0100]

[0101] in, ( x,y ) is the coordinate of the pixel in the image; G in the image tIt's pixels ( x,y ) The weighted gradient of G consists of two directions; h is the gradient in the horizontal direction; G v is the gradient in the vertical direction; | W | is the number of pixels in the neighborhood window; g ( x 1) is the Gaussian weight function; f ( x+x1 and f(x) are the grayscale values ​​of the pixels in the neighborhood and the central pixel respectively.

[0102] Step S14, respectively obtain the spatial neighborhood gradients of the polarized images Q and U through the above formula, and fuse them through linear weighting to obtain the Fimage image, as shown below:

[0103]

[0104] Among them, Fimage represents the weighted fusion result of spatial neighborhood gradient; G Q_t represents the weighted gradient of image Q; G U_t Represents the weighted gradient of image U.

[0105] Step S15: Use the Fimage image to replace the original polarized image, perform preliminary suppression on the polarization degree image, and obtain the FDoP image, as shown below:

[0106]

[0107] Among them, FDoP represents the processing result of the double-biased spatial neighborhood gradient weighted method.

[0108] Step S2: Secondly, by utilizing the difference in polarization characteristics of small targets in the sky background, bidirectional feature coherence suppression is proposed to obtain salient targets.

[0109] Since the target and the background have different polarization characteristics, the background has a low polarization characteristic for the target, and the target has a high polarization characteristic, and the Fimage obtained by weighting the dual-bias spatial gradient has polarization information in different directions.

[0110] Therefore, the present invention proposes bidirectional feature coherence suppression to effectively obtain salient targets. The specific process is as follows:

[0111] Step S21: First, since the target occupies only a small part of the image compared to the background, the Fimage image is removed from the average direction of its columns. If the value of the column direction is less than 0, it is represented as the background and its grayscale value is set to 0; otherwise, the original grayscale value is maintained.

[0112] Step S22, after that, the average direction of the row is removed from the Fimage image again. If the obtained row direction value is less than 0, it is represented as background, and its grayscale value is set to 0; otherwise, the original grayscale value is maintained.

[0113] Step S23. Finally, according to the operation of the bidirectional average direction of step S21 and step S22, the CR_image image is obtained.

[0114] Among them, the framework diagram of bidirectional polarization coherence suppression is as follows: Figure 3 As shown, the calculation formula is as follows:

[0115]

[0116] Among them, C_Fimage represents the average direction result of the original image without its columns; CR_image represents the result of bidirectional feature coherence suppression; Fimage i,j represents the weighted fusion result of the spatial neighborhood gradient at position (i, j); N represents the total number of column pixels; k represents the number of pixels; M represents the total number of row pixels; C_Fimage k,j Indicates at location ( k,j ) The average orientation result of the original image is removed from its columns.

[0117] Step S3: Since the shape of the target is destroyed after bidirectional polarization coherence suppression and some noise is not completely removed, the present invention proposes a connected statistical analysis method based on the gradient vector field to enhance the target edge information and maintain the shape of the target while performing denoising.

[0118] Step S31, calculating the gradient vector field, and analyzing the characteristics of small targets in the gradient vector field.

[0119] In the gradient vector field, each point represents the gradient of the corresponding point in the original image. The gradient calculation formula is as follows:

[0120]

[0121] in, Represents image I at pixel point ( i,j ) The gradient vector at ; and Respectively expressed in pixels ( i,j ) At the point, the horizontal and vertical gradients of the image.

[0122] Step S32: Generally, the method for calculating the gradient of an image is to use the information of two adjacent points for calculation. However, in order to consider the correlation of the grayscale values ​​of other pixels, it is proposed to estimate the gradient in the area based on all the pixels in a local window.

[0123] Assume that within a local window (usually a w×w window), the grayscale value of the image is approximated by a quadratic polynomial:

[0124] I ( x,y ) ≈a0+a1x+a2y+a3x 2 +a4xy+a5y 2 ;

[0125] Under this model, the horizontal gradient of the image at point (x, y) is and vertical gradient As shown below:

[0126]

[0127] Among them, a1, a2, a3, a4, and a5 represent the estimated coefficients in the polynomial.

[0128] Step S33: In order to estimate the coefficient, the gray value in the local window is fitted using the least square method. The light intensity change at a point in any direction can be fitted using the above formula.

[0129] like Figure 4 As shown in the figure, two small target shapes of gradient vector blocks are given. Each point in the block is represented by an arrow, which consists of a direction and a length. The direction of the arrow represents the gradient direction of this point, and the length of the arrow represents the magnitude of the gradient. The magnitude of the gradient M ( x,y ) and direction D ( x,y ) , as shown below:

[0130]

[0131] Step S34: After calculating the gradient vector field, use connected components to analyze the significant areas of gradient change in the image.

[0132] Step S341: First, using the binary image obtained by thresholding, the pixels with gradient values ​​greater than a certain threshold (threshold T=10) are marked as significant change areas, and the other pixels are marked as background. The binary image formula is as follows:

[0133]

[0134] Among them, Binary Map( x,y ) Represents a binary image.

[0135] Step S342: Count the number of connected components by marking the binary image. Connected component marking is to calculate the area and centroid position of each connected component, and the area of ​​each connected component is the number of pixels it contains. The calculation formula of the area and centroid position is as follows:

[0136] Area ( k ) =∑ x,y 1 ( label ( x,y ) =k ) ;

[0137]

[0138] Among them, Area ( k ) Indicates the area of ​​the connected component; label ( x,y ) Represents a label at position (x, y); Centroid ( k ) represents the centroid position of the connected component; 1(·) is an indicator function, which is 1 when the condition is true and 0 otherwise; k represents the proxy coefficient.

[0139] Step S343, since the number of connected components represents the number of independent regions with the same pixel value in the image. Each pixel in these regions is directly or indirectly connected to other pixels in the region. The number of connected regions can be counted by connecting component marking. The expression of the number of connected components is as follows:

[0140]

[0141] Among them, Num_D represents the number of connected components; num labels represents the number of connected component labels.

[0142] According to the analysis of the gradient vector field, different image blocks have different gradient characteristics. Figure 4 As shown in the figure, the arrows in the gradient blocks of the shapeless small target all point to the same position, and the length of the arrows is very long, which means that the modulus vector of the gradient is large and the gradient blocks are concentrated. There are multiple blocks with arrows pointing to the same position in the gradient blocks of the shaped small target, and the gradient blocks are more dispersed. According to the uniqueness of the small target in the gradient vector field, the distribution of the gradient blocks is statistically analyzed using the connected statistical analysis method to determine the different shapes of small targets using different methods to maintain edge information. Figure 4As shown, the third column is a visualization of the connected component statistics.

[0143] Therefore, the present invention adopts the above-mentioned infrared polarization noise suppression and edge enhancement method based on gradient vector analysis. By utilizing the polarization image information in different directions and according to the concept of spatial distribution similarity, the obtained weighted image replaces the original polarization image, so as to achieve the purpose of preliminarily suppressing the highlight point noise of the polarization image; since the small target occupies a small number of pixels in the sky background, the gradient mean in the horizontal and vertical directions is utilized to obtain the significant target based on bidirectional feature coherence suppression; for small targets in different situations, a connected statistical analysis method based on the gradient vector field is proposed to suppress the background while maintaining the target edge information; if there is a small target with a shape, the edge information of the target is easily destroyed when the background is suppressed, so the double-biased weighted image is used as a guide image to filter the image after the preliminary suppression, which can effectively protect the edge information of the target. The final processing result is as follows: Figure 5 shown.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. Infrared polarization noise suppression and edge enhancement method based on gradient vector analysis, characterized in that: The following steps are involved: Step S1, first, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighting method, the polarization degree image is preliminarily suppressed to suppress the highlight point noise; Step S2: Secondly, using the difference in polarization characteristics of small targets in the sky background, a bidirectional feature coherence suppression is proposed to obtain a salient target; Step S3: Finally, in order to prevent the problem of excessive suppression of target edge information during background suppression, the target edge information is enhanced based on the connected statistical analysis method of the gradient vector field.

2. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 1, characterized in that: In step S1, according to the spatial distribution similarity in the polarized image, based on the double-biased spatial neighborhood gradient weighting method, the polarization degree image is preliminarily suppressed for highlight point noise. The specific process is as follows: Step S11: In a polarization imaging system, a Stokes vector is used to represent the polarization state of light. The general definition of the Stokes vector is as follows: S=(I,Q,U,V) T ; Among them, S represents the Stokes vector matrix; I represents the total radiation intensity of light; Q represents the difference in linear polarization intensity of light in the horizontal and vertical directions; U represents the difference in linear polarization intensity of light in the +45° and -45° directions; V represents the circular polarization characteristic of light, that is, the intensity difference between left-handed circular polarization and right-handed circular polarization; T represents matrix transposition; In the natural environment, the V component of circularly polarized light is very small and can be ignored, so the Stokes vector degenerates into a third-order matrix as shown below: S= ( I,Q,U ) T ; Step S12: Select 0°, 60° and 120° to obtain objects I0, I 60 and I 120 The polarization intensity of the light is expressed by the Stokes parameter, where the Stokes vector is as follows: Step S13, in the Stokes parameters, Q and U describe the polarization information of the target surface in different directions. Since they have similarity in spatial distribution, spatial neighborhood gradient weighting is performed on two different polarization images; Step S14: According to the obtained spatial neighborhood gradients of the polarized images Q and U, the Fimage image is obtained by linear weighted fusion, as shown below: Among them, Fimage represents the weighted fusion result of spatial neighborhood gradient; G Q_t represents the weighted gradient of image Q; G U_t Represents the weighted gradient of image U; Step S15: Use the Fimage image to replace the original polarized image, perform preliminary suppression on the polarization degree image, and obtain the FDoP image, as shown below: Among them, FDoP represents the processing result of the double-biased spatial neighborhood gradient weighted method.

3. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 2 is characterized in that: In step S12, the polarization degree calculation formula is as follows: Among them, I pol Represents the intensity of the fully polarized component; I total Represents the total intensity in the light vector.

4. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 2, characterized in that: In step S13, the calculation formula of the spatial neighborhood gradient weight is as follows: G t =G ν +G h ; G h =∣X ( x,y ) -X ( x,y-1 ) ∣+∣X ( x,y ) -X ( x,y+1 ) ∣; G ν =∣X ( x,y ) -X ( x-1,y ) ∣+∣X ( x,y ) -X ( x+1,y ) ∣; in, ( x,y ) is the coordinate of the pixel in the image; G in the image t It's pixels ( x,y ) The weighted gradient of G consists of two directions; h is the gradient in the horizontal direction; G v is the gradient in the vertical direction; | W | is the number of pixels in the neighborhood window; g ( x 1) is the Gaussian weight function; f ( x+x1 and f(x) are the grayscale values ​​of the pixels in the neighborhood and the central pixel respectively.

5. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 1, characterized in that: In step S2, the difference in polarization characteristics of small targets in the sky background is used to propose bidirectional feature coherence suppression to obtain salient targets. The specific process is as follows: Step S21, first, remove the average direction of the column from the Fimage image; if the obtained column direction value is less than 0, it is represented as the background, and its grayscale value is set to 0; otherwise, the original grayscale value is maintained; Step S22, remove the average direction of the row from the Fimage image again; if the obtained row direction value is less than 0, it is represented as the background, and its gray value is set to 0; otherwise, the original gray value is maintained; Step S23, obtaining a CR_image image according to the operation of the bidirectional average direction of step S21 and step S22; The calculation formula of bidirectional polarization coherence suppression is as follows: Among them, C_Fimage represents the average direction result of the original image without its columns; CR_image represents the result of bidirectional feature coherence suppression; Fimage i,j represents the weighted fusion result of the spatial neighborhood gradient at position (i, j); N represents the total number of column pixels; k represents the number of pixels; M represents the total number of row pixels; C_Fimage k,j Indicates at location ( k,j ) The average orientation result of the original image is removed from its columns.

6. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 1, characterized in that: In step S3, the target edge information is enhanced based on the connected statistical analysis method of the gradient vector field. The specific process is as follows: Step S31, calculating the gradient vector field, and analyzing the characteristics of small targets in the gradient vector field; Step S32: estimating the gradient in a local window based on all pixels in the area to obtain the horizontal gradient and vertical gradient of the image; Step S33, using the least squares method to fit the grayscale value in the local window, fitting to obtain the light intensity change at a point in any direction, and obtaining the magnitude and direction of the gradient; Step S34: Analyze the significant areas of gradient changes in the image by calculating the gradient vector field and using connected components.

7. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 6, characterized in that: In step S31, each point in the gradient vector field represents the gradient of the corresponding point in the original image, and the gradient calculation formula is as follows: in, Represents image I at pixel point ( i,j ) The gradient vector at ; and Respectively expressed in pixels ( i,j ) The horizontal and vertical gradients of the image at the point.

8. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 6, characterized in that: In step S32, the gradient in a local window is estimated based on all the pixels in the area. The specific process is as follows: In a local window, a quadratic polynomial is used to represent the grayscale value of the image as follows: I ( x,y ) ≈a0+a1x+a2y+a3x 2 +a4xy+a5y 2 ; Among them, the horizontal gradient of the image at point (x, y) and vertical gradient As shown below: Among them, a1, a2, a3, a4, and a5 represent the estimated coefficients in the polynomial.

9. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 6, characterized in that: In step S33, the grayscale value in the local window is fitted using the least squares method to obtain the light intensity change at a point in any direction, where the magnitude and direction of the gradient are as follows: Among them, M ( x,y ) represents the modulus of the gradient; D ( x,y ) Indicates the direction of the gradient.

10. The infrared polarization noise suppression and edge enhancement method based on gradient vector analysis according to claim 6, characterized in that: In step S34, by calculating the gradient vector field, the connected components are used to analyze the significant areas of gradient changes in the image. The specific process is as follows: Step S341: Using the binary image obtained by thresholding, the pixels with gradient values ​​greater than the threshold are marked as significant change areas, and the other pixels are marked as background; wherein the binary image formula is as follows: Among them, Binary Map ( x,y ) Represents a binary image; Step S342: Count the connected components of the binary image by marking them; the connected component marking is to calculate the area and centroid position of each connected component, and the area of ​​each connected component is the number of pixels it contains; the calculation formula of the area and centroid position is as follows: Area ( k ) =∑ x,y 1 ( label ( x,y ) =k ) ; Among them, Area ( k ) Indicates the area of ​​the connected component; label ( x,y ) Represents a label at position (x, y); Centroid ( k ) represents the centroid position of the connected component; 1(·) is an indicator function, which is 1 when the condition is true and 0 otherwise; k represents the proxy coefficient; Step S343: Count the number of connected areas through connected component marking; the expression for the number of connected components is as follows: Among them, Num_D represents the number of connected components; numlabels represents the number of connected component labels.

Images