Polarization image fusion method based on multi-scale transformation and image enhancement

Through the polarized image fusion method of multi-scale transformation and image enhancement, the problem of difficult information capture caused by light reflection and scattering in the detection of reflective metal surface defects is solved, and high-precision defect detection is achieved, which improves detection efficiency and accuracy.

CN118982467AActive Publication Date: 2024-11-19SOUTHEAST UNIV

Patent Information

Application Number
CN202411007139.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-19
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

In the detection of reflective metal surface defects, due to light reflection and scattering, defect information is difficult to capture and process, and the detection accuracy is reduced.

Method used

Using a polarized image fusion method based on multi-scale transformation and image enhancement, images of different polarization angles are collected through the polarization camera, and fast-guided filtering, CLAHE enhancement, Laplace pyramid decomposition, bilateral filtering and Laplace sharpening processing are carried out. Combining the brightness weight pyramid for weighting, the final fusion image is generated.

Benefits of technology

It significantly improves the accuracy and efficiency of defect detection on reflective metal surfaces, overcomes the problem of low image clarity under complex lighting conditions, ensures that the shape and position of the defect are clearly distinguishable, and improves the accuracy of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982467A_ABST
    Figure CN118982467A_ABST
Patent Text Reader

Abstract

The invention discloses a polarization image fusion method based on multi-scale transformation and image enhancement, and the method comprises the steps: firstly obtaining polarization images at four different polarization angles through a polarization camera, and calculating and synthesizing the required polarization degree, polarization angle and visible light image; performing filtering enhancement on the image by using fast steering filtering and enhanced CLAHE; decomposing the polarization degree, the polarization angle and the visible light image into a Laplacian image pyramid; for high-frequency layers of the three Laplacian pyramid images, bilateral filtering and Laplacian sharpening processing are used for enhancing high-frequency information; calculating the brightness of the original image, generating a brightness weight pyramid, and performing weighted fusion on each pyramid hierarchy; and reconstructing the fused Laplacian pyramid to obtain a final fused image. According to the invention, defect feature information of different dimensions of the reflective metal surface can be captured and organically fused together, the method has very important practical significance for subsequent defect detection, and the efficiency and accuracy of reflective metal surface defect detection are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a polarization image fusion method based on multi-scale transformation and image enhancement. Background Art

[0002] Traditional metal surface defect detection methods, such as visual inspection and ultrasonic inspection, can work in some cases, but they often face challenges in the detection of reflective metal surfaces. The strong reflection and scattering of reflective metal surfaces make it difficult to directly capture defect information, resulting in reduced detection accuracy. Polarized light can be used to solve the problem that it is difficult to capture the characteristic information of reflective metal surface defects.

[0003] Image acquisition with the help of polarization can suppress reflections to a certain extent. At the same time, if polarization images at different angles are collected, polarization degree images and polarization angle images can be calculated. Compared with images collected by ordinary cameras, they contain metal surface defect information of different dimensions. However, this information is scattered and may exist in different images, so they need to be integrated.

[0004] In the field of image processing, image fusion technology is widely used in various scenarios because it can integrate image information from different sensors or different imaging methods. In the detection of reflective metal surface defects, the accuracy and reliability of defect detection can be significantly improved by fusing image information of different scales and angles. Polarization image refers to the proportion of polarized light to the total light intensity, which reflects the degree of polarization of light. Polarization image is less affected by uneven illumination and dark illumination, and can effectively improve the signal-to-noise ratio and contrast of detection images in complex lighting environments or poor lighting conditions. Polarization angle image describes the polarization direction of light and can reflect the surface reflection and scattering characteristics of the target object. Visible light image provides the basic shape and color information of the target object, which is crucial for the preliminary identification and rapid positioning of defects. To this end, a polarization image fusion method based on multi-scale transformation and image enhancement is designed. Summary of the invention

[0005] The technical problem to be solved by the present invention is: to design a polarization image fusion method based on multi-scale transformation and image enhancement to solve the problem that when detecting defects on reflective metal surfaces, defect information is difficult to capture and process due to light reflection and scattering.

[0006] To achieve the above object, the present invention provides a technical solution: a polarization image fusion method based on multi-scale transformation and image enhancement, comprising the following steps:

[0007] Step 1: The polarization camera collects polarization images at four different polarization angles, and the polarization degree, polarization angle and visible light image are obtained through calculation;

[0008] Step 2: Perform fast guided filtering and CLAHE enhancement on the polarization degree, polarization angle and visible light images to improve image quality and contrast;

[0009] Step 3: Decompose the three images (polarization degree, polarization angle, and visible light) into a Laplacian image pyramid;

[0010] Step 4: Perform bilateral filtering on the high-frequency layer image in the Laplacian pyramid to smooth and preserve edge information, and apply Laplacian sharpening to enhance image details;

[0011] Step 5: Calculate the brightness of the original image and generate a brightness weight pyramid. According to the brightness weight pyramid, perform weighted fusion on each pyramid level;

[0012] Step 6: Reconstruct the fused Laplacian pyramid to obtain the final fused image.

[0013] Furthermore, in step 1, a polarization camera is used to collect polarization images with polarization angles of 0°, 45°, 90°, and 135°, which are expressed by the Stokes formula:

[0014]

[0015] Where S0 is the total incident light intensity of the system, expressed as the sum of the linear polarized light intensities passing through the 0° and 90° directions; S1 is the difference between the linear polarized light intensities at 0° and 90°; S2 is the difference between the linear polarized light intensities at 45° and 135°; S3 is the difference between the right-handed and left-handed circularly polarized light intensities; (x, y) represents the pixel coordinates.

[0016] Calculate the degree of polarization image (Dop) and the angle of polarization image (Aop):

[0017]

[0018] Furthermore, in step 2, fast guided filtering and CLAHE enhancement are performed on the polarization degree, polarization angle and visible light image. The local linear filter of the guided filtering is:

[0019]

[0020] Where q is the value of the output pixel, I is the value of the input image, i and k are pixel indices, and a and b are the coefficients of the linear function when the window is centered at k.

[0021] Find the coefficient a of the linear function k and b k , is to minimize the gap between the output value q of the fitting function and the true value p, which is converted into an optimization problem, that is, to minimize the following formula:

[0022]

[0023] Through least squares, we can get:

[0024]

[0025] Among them, μ k is the image I in window ω k The average value in is the image p to be filtered in the window ω k The mean value in I am in the window ω k The variance in |ω| is the window ω k The quality of the pixels in , ε is the regularization parameter.

[0026] To find the output value of a specific point, you need to average all the linear function values ​​that include this point, as follows:

[0027]

[0028] In the formula, ω k are all windows containing pixel i, and k is its center position.

[0029] Fast guided filtering is based on the above guided filtering, which downsamples the image to reduce the filtering points and calculates the coefficient a. k and b k After taking the mean of the image, upsampling is performed to restore the original image size.

[0030] The core process of the enhanced CLAHE algorithm is as follows:

[0031] (1) Usually, the input image is divided into 8*8 grids for calculation. The enhanced CLAHE first calculates the mean square error of the image brightness. The smaller the mean square error of the image brightness, the more consistent the brightness and darkness of the image is. A larger number of grids, i.e. 8*8, is used; otherwise, a smaller number of grids, i.e. 4*4, is used.

[0032] (2) Calculate the sub-block histogram.

[0033] (3) The histogram is cropped and redistributed, and then the cropped histogram is equalized to obtain a mapping table for each small block.

[0034] (4) Smoothing the mapping table. Perform Gaussian smoothing on the mapping table in one dimension. For areas where the image changes slowly, this smoothing can reduce the color block feeling caused by enhancement to a certain extent.

[0035] (5) Perform bilinear interpolation on each small block to obtain the final enhancement effect.

[0036] Furthermore, in step 3, the three images (polarization degree, polarization angle, and visible light) are decomposed into a Laplacian image pyramid. The specific process is as follows:

[0037] The polarization degree image, polarization angle image and visible light image are used as the bottom of the tower G0, and then Gaussian low-pass filtering and downsampling are performed to construct the Gaussian pyramid of the image. The l-th layer image G l for:

[0038]

[0039] Where, 1≤l≤N, 0<i<C l , 0<j<R l ; N is the total number of layers, C l and R l G l The number of rows and columns; ω(n,m) is the Gaussian kernel function.

[0040] Each layer image G of the Gaussian pyramid l Subtract the image after upsampling and Gaussian convolution of the previous layer The obtained series of difference images is the Laplacian pyramid LP l ,Right now:

[0041]

[0042] Where ω(n,m) is the Gaussian kernel function. When (i+m) and (j+n) are even numbers, G' l+1 Equal to G l+1 , in other cases, G' l+1 is 0.

[0043] Decompose the polarization degree image, polarization angle image and visible light image, and the resulting Laplace pyramid is denoted as: Dop_LP l 、Aop_LP l and I_LP l , where 0≤l≤N, N is the number of decomposition levels.

[0044] Furthermore, in step 4, bilateral filtering is performed on the high-frequency layer image in the Laplacian pyramid to smooth and preserve edge information, and Laplacian sharpening is applied to enhance image details.

[0045] The working principle of bilateral filtering is as follows:

[0046]

[0047] Where i and j are the coordinates of the current center pixel, g(i,j) is the output after filtering, k and l are the coordinates of the neighboring pixels, f(k,l) is the neighboring pixel value, w(i,j,k,l) ​​is the weighting coefficient, σd is the standard deviation of the spatial position Gaussian function, σ r is the standard deviation of the Gaussian function of pixel similarity.

[0048] The Laplace sharpening formula is:

[0049]

[0050] In the formula, g(i,j) is the image after sharpening, f(i,j) is the image before sharpening, and the convolution kernel H(r,s) is

[0051]

[0052] Furthermore, in step 5, the brightness of the original image is calculated to generate a brightness weight pyramid. According to the brightness weight pyramid, each pyramid level is weightedly fused.

[0053] Calculate the brightness of an image:

[0054]

[0055] Among them, E(i,j) is the brightness image after calculation, I(i,j) is the value of the original image pixel after normalization to 0-1, and σ is the variance.

[0056] The polarization degree, polarization angle and brightness image corresponding to the visible light image are decomposed into a Gaussian pyramid as the fusion weight of the original image Laplacian pyramid. The calculated brightness weight pyramid is denoted as Dop_G l 、Aop_G l and I_G l , where 0≤l≤N, N is the number of decomposition levels.

[0057] The calculation formula of each layer's fused image F(i,j) is:

[0058]

[0059] Each layer of the pyramid is fused in this way to form the Laplacian pyramid LP of the fused image. l F , where 0≤l≤N, N is the number of decomposition levels.

[0060] Furthermore, in step six, the fused Laplacian pyramid is reconstructed to obtain the final fused image.

[0061] After fusing the three images using the above method, a Laplacian pyramid of the fused image is formed. Reconstructing the Gaussian pyramid from the Laplacian pyramid The resulting underlying image That is the final fused image, expressed as:

[0062]

[0063] Where ω(n,m) is the Gaussian kernel function, is the upsampled image of the Gaussian pyramid image. When (i+m) and (j+n) are even numbers, equal In other cases, is 0.

[0064] Compared with the existing methods, the beneficial effects of the present invention are as follows: a polarization image fusion method based on multi-scale transformation and image enhancement is proposed. Compared with the traditional method, it can capture the defect feature information of reflective metal from different dimensions, and overcome the problem that the traditional method is difficult to obtain clear images under complex lighting conditions. Before and during the multi-scale transformation, image enhancement is performed multiple times to enhance the edge and detail information in the image, especially the defect part, so that the shape and position of the defect are clearer and more identifiable, greatly improving the accuracy of defect detection. The brightness weighted fusion strategy adopted in the fusion stage dynamically adjusts the fusion weights of each image layer according to the fusion source image, ensuring that the key defect objects with high brightness are highlighted in the fusion process, while retaining the background and detail information of the low brightness area. It has very important practical significance for subsequent defect detection, and can effectively improve the efficiency and accuracy of reflective metal surface defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is the overall flow chart of the present invention;

[0066] Figure 2 (a), (b), and (c) are respectively the polarization degree, polarization angle, and visible light image calculated by the present invention;

[0067] Figure 3 (a), (b), and (c) are the polarization degree, polarization angle, and visible light image after the fast guided filtering of the present invention;

[0068] Figure 4 (a), (b), and (c) are respectively the polarization degree, polarization angle, and visible light image after enhancement according to the present invention;

[0069] Figure 5 (a), (b), and (c) are respectively the polarization degree, polarization angle, and brightness images corresponding to the visible light images generated by the present invention;

[0070] Figure 6 This is the fused image obtained by the present invention. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention.

[0072] Example:

[0073] See also Figure 1 The polarization image fusion method based on multi-scale transformation and image enhancement of this embodiment includes the following steps:

[0074] Step 1: Based on the polarization images of four different polarization angles collected by the polarization camera, synthesize the polarization degree, polarization angle, and visible light image. The polarization camera collects polarization images with polarization angles of 0°, 45°, 90°, and 135°, and calculates the relevant parameters using the Stokes formula:

[0075]

[0076] Where S0 is the total incident light intensity of the system, expressed as the sum of the linear polarized light intensities passing through the 0° and 90° directions; S1 is the difference between the linear polarized light intensities at 0° and 90°; S2 is the difference between the linear polarized light intensities at 45° and 135°; S3 is the difference between the right-handed and left-handed circularly polarized light intensities; (x, y) represents the pixel coordinates.

[0077] Calculate the degree of polarization image (Dop) and the angle of polarization image (Aop):

[0078]

[0079] The polarization degree, polarization angle, and visible light image of this embodiment refer to Figure 2 .

[0080] Step 2: Use fast guided filtering and enhanced CLAHE to filter and enhance the polarization degree, polarization angle, and visible light image. The local linear filter of the guided filtering is:

[0081]

[0082] Where q is the value of the output pixel, I is the value of the input image, i and k are pixel indices, and a and b are the coefficients of the linear function when the window is centered at k.

[0083] Find the coefficient a of the linear function k and b k , is to minimize the gap between the output value q of the fitting function and the true value p, which is converted into an optimization problem, that is, to minimize the following formula:

[0084]

[0085] Through least squares, we can get:

[0086]

[0087]

[0088] Among them, μ k is the image I in window ω k The average value in is the image p to be filtered in the window ω k The mean value in I am in the window ω k The variance in |ω| is the window ω k The quality of the pixels in , ε is the regularization parameter.

[0089] To find the output value of a specific point, you need to average all the linear function values ​​that include this point, as follows:

[0090]

[0091] In the formula, ω k are all windows containing pixel i, and k is its center position.

[0092] Fast guided filtering is based on the above guided filtering, which downsamples the image to reduce the filtering points and calculates the coefficient a. k and b k After the average of , upsampling is performed to restore the original size of the image. In this embodiment, the downsampling factor is 4 and the window size is 5×5.

[0093] In this embodiment, the polarization degree, polarization angle, and visible light image after fast guidance filtering refer to Figure 3 .

[0094] The core process of the enhanced CLAHE algorithm is as follows:

[0095] (1) Usually, the input image is divided into 8*8 grids for calculation. The enhanced CLAHE first calculates the mean square error of the image brightness. The smaller the mean square error of the image brightness, the more consistent the brightness and darkness of the image is. A larger number of grids, i.e. 8*8, is used; otherwise, a smaller number of grids, i.e. 4*4, is used.

[0096] (2) Calculate the sub-block histogram.

[0097] (3) The histogram is cropped and redistributed, and then the cropped histogram is equalized to obtain a mapping table for each small block.

[0098] (4) Smoothing the mapping table. Perform Gaussian smoothing on the mapping table in one dimension. For areas where the image changes slowly, this smoothing can reduce the color block feeling caused by enhancement to a certain extent.

[0099] (5) Perform bilinear interpolation on each small block to obtain the final enhancement effect.

[0100] The polarization degree, polarization angle, and visible light image after enhancement in this embodiment refer to Figure 4 .

[0101] Step 3: Decompose the polarization degree, polarization angle, and visible light image into a Laplacian image pyramid, with the three original images as the base G0, and then perform Gaussian low-pass filtering and downsampling to construct a Gaussian pyramid of the image. The lth layer image G l for:

[0102]

[0103] Where, 1≤l≤N, 0<i<C l , 0<j<R l ; N is the total number of layers, C l and R l G l The number of rows and columns of the image; ω(n,m) is the Gaussian kernel function. In this embodiment, log(min(rows,cols)) / log(2) is taken, where rows is the number of rows of image pixels, cols is the number of columns of image pixels, the number of rows rows=1024, the number of columns cols=1224, and N=10 is calculated.

[0104] Each layer image G of the Gaussian pyramid l Subtract the image after upsampling and Gaussian convolution of the previous layer The obtained series of difference images is the Laplacian pyramid LP l ,Right now:

[0105]

[0106] Where ω(n,m) is the Gaussian kernel function. When (i+m) and (j+n) are even numbers, G' l+1 Equal to G l+1 , in other cases, G' l+1 is 0.

[0107] Decompose the polarization degree image, polarization angle image and visible light image, and the resulting Laplace pyramid is denoted as: Dop_LP l 、Aop_LP l and I_LP l , where 0≤l≤N, N is the number of decomposition levels. In this embodiment, log(min(rows,cols)) / log(2) is taken, where rows is the number of rows of image pixels, cols is the number of columns of image pixels, the number of rows rows=1024, the number of columns cols=1224, and N=10 is calculated.

[0108] In step 4, bilateral filtering is performed on the high-frequency layer image in the Laplacian pyramid to smooth and preserve edge information, and Laplacian sharpening is applied to enhance image details.

[0109] The formula for bilateral filtering is as follows:

[0110]

[0111] Where i and j are the coordinates of the current center pixel, g(i,j) is the output after filtering, k and l are the coordinates of the neighboring pixels, f(k,l) is the neighboring pixel value, w(i,j,k,l) ​​is the weighting coefficient, σ d is the standard deviation of the spatial position Gaussian function, σ r is the standard deviation of the Gaussian function of pixel similarity.

[0112] The Laplace sharpening formula is:

[0113]

[0114] In the formula, g(i,j) is the image after sharpening, f(i,j) is the image before sharpening, and the convolution kernel H(r,s) is

[0115]

[0116] Step 5: Calculate the brightness of the original image and generate a brightness weight pyramid. According to the brightness weight pyramid, perform weighted fusion on each pyramid level.

[0117] Calculate the brightness of an image:

[0118]

[0119] Wherein, E(i, j) is the brightness image after calculation, I(i, j) is the value of the original image pixel after normalization to 0-1, and σ is the variance, which is 0.2 in this embodiment.

[0120] The polarization degree, polarization angle, and brightness image corresponding to the visible light image generated in this embodiment are shown in FIG. Figure 5 .

[0121] The polarization degree, polarization angle and brightness image corresponding to the visible light image are decomposed into a Gaussian pyramid as the fusion weight of the original image Laplacian pyramid. The calculated brightness weight pyramid is denoted as Dop_G l 、Aop_G l and I_G l , where 0≤l≤N, N is the number of decomposition levels.

[0122] The calculation formula of each layer's fused image F(i,j) is:

[0123]

[0124] Each layer of the pyramid is fused in this way to form the Laplacian pyramid LP of the fused image. l F , where 0≤l≤N, N is the number of decomposition layers. In this embodiment, log(min(rows,cols)) / log(2) is taken, where rows is the number of rows of image pixels, cols is the number of columns of image pixels, the number of rows rows=1024, the number of columns cols=1224, and N=10 is calculated.

[0125] Step six, reconstruct the fused Laplacian pyramid to obtain the final fused image.

[0126] After fusing the three images using the above method, a Laplacian pyramid of the fused image is formed. Reconstructing the Gaussian pyramid from the Laplacian pyramid The resulting underlying image That is the final fused image, expressed as:

[0127]

[0128]

[0129] Where ω(n,m) is the Gaussian kernel function, is the upsampled image of the Gaussian pyramid image. When (i+m) and (j+n) are even numbers, equal In other cases, is 0.

[0130] In this embodiment, the obtained fused image is referred to Figure 6 .

[0131] The above description of the present invention and its embodiments is not restrictive. The accompanying drawings show only one embodiment of the present invention, and the practical application is not limited thereto. If a person skilled in the art is inspired by it and, without departing from the purpose of the invention, designs a structure and an embodiment similar to the technical solution without creativity, they shall all fall within the protection scope of the present invention.

Claims

1. A polarization image fusion method based on multi-scale transformation and image enhancement, It is characterized in that The method comprises the following steps: Step 1: The polarization camera collects polarization images at four different polarization angles, and the polarization degree, polarization angle and visible light image are obtained through calculation; Step 2: Perform fast guided filtering and CLAHE enhancement on the polarization degree, polarization angle and visible light images to improve image quality and contrast; Step 3: Decompose the three images, namely polarization degree, polarization angle, and visible light, into a Laplacian image pyramid; Step 4: Perform bilateral filtering on the high-frequency layer image in the Laplacian pyramid to smooth and preserve edge information, and apply Laplacian sharpening to enhance image details; Step 5: Calculate the brightness of the original image, generate a brightness weight pyramid, and perform weighted fusion on each pyramid level based on the brightness weight pyramid; Step 6: Reconstruct the fused Laplacian pyramid to obtain the final fused image.

2. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that In step 1, the polarization camera collects polarization images at four different polarization angles, and the polarization degree, polarization angle and visible light image are obtained through calculation. The specific operation is as follows: Use a polarization camera to collect polarization images with polarization angles of 0°, 45°, 90°, and 135°, which are expressed by the Stokes equation: In the formula, S 0 is the total incident light intensity of the system, expressed as the sum of the linear polarized light intensities passing through the 0° and 90° directions; S 1 The difference between the linear polarized light intensities at 0° and 90°; S 2 The difference between linear polarization of light at 45° and 135°; S 3 is the difference between the intensity of right-handed and left-handed circularly polarized light; (x, y) represents the coordinates of the pixel point, Calculate the degree of polarization image (Dop) and the angle of polarization image (Aop):

3. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that The specific operations for preprocessing the polarization degree, polarization angle and visible light image in step 2 are as follows: The local linear filter for guided filtering is: Where q is the value of the output pixel, I is the value of the input image, i and k are pixel indices, and a and b are the coefficients of the linear function when the window center is at k. Find the coefficient a of the linear function k and b k , is to minimize the gap between the output value q of the fitting function and the true value p, which is converted into an optimization problem, that is, to minimize the following formula: Through least squares, we get: Among them, μ k is the image I in window ω k The average value in is the image p to be filtered in the window ω k The mean value in I am in the window ω k The variance in |ω| is the window ω k The quality of the pixels in , ε is the regularization parameter, To find the output value of a specific point, you need to average all the linear function values ​​that include this point, as follows: In the formula, ω k are all windows containing pixel i, k is its center position, Fast guided filtering is based on the above guided filtering, which downsamples the image to reduce the filtering points and calculates the coefficient a. k and b k After taking the mean of , up-sample to restore the original image size. The core process of the enhanced CLAHE algorithm is as follows: (1) Usually, the input image is divided into 8*8 grids for calculation. The enhanced CLAHE first calculates the mean square error of the image brightness. The smaller the mean square error of the image brightness, the more consistent the brightness and darkness changes of the image. A larger number of grids, i.e., 8*8, is used. On the contrary, a smaller number of grids, i.e., 4*4, is used. (2) Calculate the sub-block histogram, (3) The histogram is cropped and redistributed, and then the cropped histogram is equalized to obtain a mapping table for each small block. (4) Smoothing the mapping table: Gaussian smoothing is performed on the mapping table in one dimension. For areas where the image changes slowly, this smoothing can reduce the color block feeling caused by enhancement to a certain extent. (5) Perform bilinear interpolation on each small block to obtain the final enhancement effect.

4. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that In step 3, the three images (polarization degree, polarization angle, and visible light) are decomposed into a Laplacian image pyramid. The specific operations are as follows: The polarization degree image, polarization angle image and visible light image are used as the bottom G 0 , and then perform Gaussian low-pass filtering and downsampling to construct the Gaussian pyramid of the image. The l-th layer image G l for: Where, 1≤l≤N, 0<i<C l , 0<j<R l ; N is the total number of layers, C l and R l G l The number of rows and columns; ω(n,m) is the Gaussian kernel function, Each layer image G of the Gaussian pyramid l Subtract the image after upsampling and Gaussian convolution of the previous layer The obtained series of difference images is the Laplacian pyramid LP l ,Right now: Where ω(n,m) is the Gaussian kernel function. When (i+m) and (j+n) are even numbers, G l ' +1 Equal to G l+1 , in other cases, G l ' +1 is 0, Decompose the polarization degree image, polarization angle image and visible light image, and the resulting Laplace pyramid is denoted as: Dop_LP l 、Aop_LP l and I_LP l , where 0≤l≤N, N is the number of decomposition levels.

5. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that In step 4, bilateral filtering is performed on the high-frequency layer image in the Laplacian pyramid to smooth and retain edge information, and Laplacian sharpening is applied to enhance image details. The specific operations are as follows: The working principle of bilateral filtering is as follows: Where i and j are the coordinates of the current center pixel, g(i,j) is the output after filtering, k and l are the coordinates of the neighboring pixels, f(k,l) is the neighboring pixel value, w(i,j,k,l) ​​is the weighting coefficient, σ d is the standard deviation of the spatial position Gaussian function, σ r is the standard deviation of the Gaussian function of pixel similarity, The Laplace sharpening formula is: In the formula, g(i,j) is the image after sharpening, f(i,j) is the image before sharpening, and the convolution kernel H(r,s) is 6. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that In step 5, the brightness of the original image is calculated to generate a brightness weight pyramid. According to the brightness weight pyramid, the specific operation of weighted fusion of each pyramid level is as follows: Calculate the brightness of an image: Among them, E(i,j) is the brightness image after calculation, I(i,j) is the value of the original image pixel after normalization to 0-1, σ is the variance, The polarization degree, polarization angle and brightness image corresponding to the visible light image are decomposed into a Gaussian pyramid as the fusion weight of the Laplacian pyramid of the original image. The calculated brightness weight pyramid is recorded as Dop_G l 、Aop_G l and I_G l , where 0≤l≤N, N is the number of decomposition levels, The calculation formula of each layer's fused image F(i,j) is: Each layer of the pyramid is fused in this way to form the Laplacian pyramid LP of the fused image. l F , where 0≤l≤N, N is the number of decomposition levels.

7. The polarization image fusion method based on multi-scale transformation and image enhancement according to claim 1, It is characterized in that In step six, the fused Laplacian pyramid is reconstructed to obtain the final fused image. The specific operations are as follows: After fusing the three images using the above method, a Laplacian pyramid of the fused image is formed. Reconstructing the Gaussian pyramid from the Laplacian pyramid The resulting underlying image That is the final fused image, expressed as: Where ω(n,m) is the Gaussian kernel function, is the upsampled image of the Gaussian pyramid image. When (i+m) and (j+n) are even numbers, equal In other cases, is 0.

Citation Information

Patent Citations

  • Image contrast enhancement method and system, and terminal device

    CN110751603A

  • Electric power system oil leakage detection method and device, model training method and device and computer equipment

    CN116503347A

  • Image pyramid generation for image keypoint detection and descriptor generation

    US20220301110A1

Cited By

  • Method and system for detecting inclusion defects on surface of wiredrawing steel plate based on machine vision

    CN119313653A

  • Method and system for detecting inclusion defects on the surface of drawn steel plate based on machine vision

    CN119313653B

  • Liquid crystal display screen defect detection method

    CN119827522A

  • High-temperature plate blank image processing method and system based on artificial intelligence

    CN120236097A

  • Visual positioning method for mobile phone part assembly

    CN120912673A