A display screen weak cross defect detection method based on frequency domain filtering of significant color channels

Abstract

The invention discloses a method for detecting weak cross-defects of a display screen by frequency domain filtering of significant color channels. Specifically, a relative entropy maximum criterion is adopted to select significant color channels in RGBY, and an image color quaternion matrix is ​​formed with the Lab color space. The image color quaternion matrix is ​​converted into a logarithmic spectrum of frequency domain significance by hypercomplex Fourier transform. Then, Gaussian convolution, threshold screening, bandpass filtering and hypercomplex Fourier inverse transform are adopted to obtain a cross-defect saliency map. Finally, the mean and standard deviation of the defect saliency map are calculated and combined with the structural information of the image to perform weak cross-defect detection by binary segmentation of the defect saliency map. The problem that weak cross-defects of a display screen cannot be detected is solved. A method for detecting weak cross-defects of a display screen with high accuracy and fast processing speed is provided. The cross-defect detection of a display screen with high accuracy under weak contrast is effectively realized. The method can be applied to weak cross-defect detection of a display screen.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a display screen weak cross defect detection method using significant color channel frequency domain filtering. Background Art

[0002] Display quality inspection plays a key role in the display production process. Existing display quality inspection still relies on manual inspection, which is affected by the subjective feelings of the inspectors. Manual inspection has the problems of low efficiency and unstable accuracy. Especially on small wearable devices, it is difficult for the human eye to detect defects when they are close to the background. Therefore, the use of machine vision and digital image processing technology to realize display defect detection has become an urgent problem to be solved.

[0003] The cross test screen of the display is a specific display screen in the process of display quality inspection, which can test the cross display capability of the display. The cross screen that cannot be displayed normally by a defective display is called a cross defect. There are two main difficulties in detecting cross defects: 1. The defect area is large, close to half of the display screen, and the defect area is displayed as a continuous color linear defect feature with an irregular color; 2. The contrast between the defect area and the background is low, and the low-contrast defect area is difficult to detect.

[0004] With the rapid development of display technology in the past decade, the methods for detecting weak contrast defects in display screens mainly use background reconstruction defect detection methods such as frequency domain-based reconstruction and fitting-based reconstruction. However, the frequency domain-based reconstruction method cannot effectively obtain the defect-free background information of cross defects in the frequency domain, which makes it impossible to separate defects from the background; the fitting-based reconstruction method requires a uniform background image, but the fitting method cannot be used in the cross test image due to the special shape of the cross test image. Summary of the invention

[0005] The present invention proposes a display screen weak cross defect detection method using frequency domain filtering of significant color channels, which realizes efficient and accurate detection of weak cross defects. The specific steps are as follows:

[0006] Step S1, edge gradient suppression of the original image,

[0007] Calculate the grayscale mean map of the input image, then calculate the gradient map of the image, and use the obtained gradient map to suppress the edge of the original image;

[0008] Step S2, adaptive selection of color features,

[0009] For the input display screen image, the RGBY color space and Lab color space are calculated respectively, and the relative entropy is used to adaptively select the significant color channel in the RGBY color space;

[0010] Step S3, constructing the image quaternion matrix and the hypercomplex Fourier transform,

[0011] The color information of the four channels extracted is used to construct the image color quaternion matrix, and then the hypercomplex Fourier transform is used to calculate the logarithmic magnitude spectrum of the input features;

[0012] Step S4, Gaussian adaptive smoothing parameter calculation,

[0013] Use the Gaussian template and the amplitude spectrum for convolution operation, first fix the standard deviation value of the Gaussian template, change the size of the Gaussian template and perform convolution operation with the amplitude spectrum, use one-dimensional entropy to calculate the information content of the image, retain the template size with minimum entropy, then fix the template size and change the standard deviation value of the Gaussian window, retain the template standard deviation value with minimum entropy, and obtain the optimal amplitude spectrum parameters of the defect;

[0014] Step S5, threshold screening and bandpass filtering of the amplitude spectrum,

[0015] Calculate the mean and standard deviation of the original amplitude spectrum, use a threshold to filter the smoothed amplitude values, retain the amplitude spectrum values ​​greater than the threshold, and then use a bandpass filter to select the frequency to obtain the amplitude spectrum of the defect;

[0016] Step S6, hypercomplex inverse Fourier transform and adaptive threshold segmentation,

[0017] Perform inverse hypercomplex Fourier transform on the obtained defect amplitude spectrum to obtain the defect image, extract the mean and standard deviation of the defect image, and perform binary segmentation on the image;

[0018] Step S7, adaptive selection of defective areas,

[0019] In order to achieve accurate extraction of defective areas, the segmented binary image is constrained using structural information, thereby achieving accurate segmentation of defective areas.

[0020] Beneficial effects of the present invention: The method of the present invention specifically adopts the maximum relative entropy criterion to select the significant color channel in RGBY, and forms an image color quaternion matrix with the Lab color space. The image color quaternion matrix is ​​converted into a logarithmic spectrum of frequency domain significance through hypercomplex Fourier transform, and then Gaussian convolution, threshold screening, bandpass filtering and hypercomplex Fourier inverse transform are used to obtain a cross defect saliency map. Finally, the mean and standard deviation of the defect saliency map are calculated and combined with the structural information of the image to perform defect saliency. Figure 2 The weak cross defect detection based on value segmentation solves the problem that weak cross defects in display screens cannot be detected, and provides a weak cross defect detection method with high accuracy and fast processing speed in display screen images. It effectively realizes the cross defect detection of display screens with high accuracy under weak contrast, and can be applied to weak cross defect detection of display screens. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flow chart of a method for detecting weak cross-defects of display screens by frequency domain filtering of significant color channels.

[0022] Figure 2 The figure shows the sample of the original image and the edge gradient suppressed mask image of the present invention.

[0023] Figure 3 The following is a sample display diagram of four feature channel images represented by the quaternion of the present invention.

[0024] Figure 4 This is a diagram showing an example of processing of the hypercomplex amplitude spectrum mode of the present invention.

[0025] Figure 5 This is a sample display diagram of defect image segmentation and region selection according to the present invention.

[0026] Figure 6 A sample display diagram of coordinate selection for the defect image area of ​​the present invention.

[0027] Figure 7 This is a sample display diagram of the test results of the present invention. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention is described below in conjunction with the accompanying drawings:

[0029] See also Figure 1 In the embodiment described below, a method for detecting weak cross-defects of a display screen by frequency domain filtering of a significant color channel has the following steps:

[0030] Step S1, edge gradient suppression of the original image,

[0031] S11: Grayscale mean image I calculation,

[0032] The calculation is as follows:

[0033]

[0034] in Figure 2 (a) is the original image, r, g, b represent the red, green, and blue channels of the original image.

[0035] S12: Calculate the gradient map for the grayscale mean image I,

[0036] The grayscale mean image I calculates the horizontal and vertical gradients, then modulos the gradients in the two directions, and finally normalizes the gradient image. The normalized gradient image is subtracted from 1 to obtain the defect edge image G. The result is as follows: Figure 2 (b) Defect edge image.

[0037] S13: Original image edge suppression,

[0038] The defect edge image G is calculated with the original image to obtain the input image with suppressed edges. The calculation formula is as follows:

[0039]

[0040] Among them, r, g, and b represent the red, green, and blue channels of the input image. i , g i 、b i Represents the red, green, and blue channels of the input image with suppressed edges.

[0041] Step S2, adaptive selection of color features,

[0042] S21: RGBY color space calculation,

[0043] The color space calculation is performed on the input image. The calculation formula is as follows:

[0044]

[0045]

[0046] Among them, r i , g i 、b i Represents the red, green, and blue channels of the input image with suppressed edges, RG is the red-green opposing color channel, GR is the green-red opposing color channel, BY is the blue-yellow opposing color channel, and YB is the yellow-blue opposing color channel.

[0047] S22: Lab color space calculation,

[0048] The edge suppression image obtained in step S13 is subjected to Lab color space calculation to obtain an L brightness channel, an a red-green color channel, and a b1 yellow-blue color channel, and the a channel and the b1 channel are retained.

[0049] S23: Selection of significant color channels in RGBY color space,

[0050] Input RG and I, GR and I and BY and I, YB and I respectively;

[0051] Relative entropy maximum criterion: According to the value of relative entropy, when H kl (RG,I)>H kl When (GR, I), the color feature retains RG, otherwise it retains GR, and the BY feature is processed in the same way. The selection method is as follows:

[0052]

[0053] Among them, f1 represents the significant color channel in RG and GR, f2 represents the significant color channel in BY and YB, and H kl Represents the relative entropy calculation, and the calculation formula is as follows:

[0054]

[0055] Among them, P(x) represents the grayscale probability distribution function of input P, ​​and Q(x) represents the grayscale probability distribution function of input Q.

[0056] Step S3, constructing the image quaternion matrix and the hypercomplex Fourier transform,

[0057] S31: Image color quaternion matrix representation,

[0058] Quaternion matrix f h (n,m) is expressed as follows:

[0059] f h (n,m)=f1+f2i+f3j+f4k

[0060] Among them, i 2 =-1,j 2 =-1, k 2 = -1, ijk = -1, n, m represent image space coordinates, f1, f2 are dynamically selected by step S23, f3 represents the a channel of the Lab color space, and f4 represents the b1 channel of the Lab space. The result is as follows Figure 3 As shown, among which Figure 3 (a) is the f1 channel feature map result, Figure 3 (b) is the f2 channel feature map result, Figure 3 (c) is the result of the f3 channel feature map. Figure 3 (d) is the f4 channel feature map result.

[0061] S32: Hypercomplex Fourier Transform,

[0062] The hypercomplex Fourier transform is used to calculate the amplitude spectrum of the hypercomplex features represented by the quaternion. The hypercomplex Fourier transform formula is as follows:

[0063]

[0064] Among them, F H [u,v] represents the hypercomplex frequency domain space, [u,v] represents the frequency coordinate, M,N are the length and width of the input image, μ is the quaternion unit, and μ 2 = -1;

[0065] The inverse transform formula of the hypercomplex Fourier transform is as follows:

[0066]

[0067] F H The formula for calculating the [u,v] amplitude spectrum is:

[0068] Am(u,v)=|F H [u,v]|

[0069] F H The calculation formula of [u,v] phase spectrum is:

[0070]

[0071] The calculation formula of characteristic axis spectrum is:

[0072]

[0073] Where Am(u,v) represents the amplitude spectrum of the hypercomplex Fourier transform, |·| represents the modulus of each element in the hypercomplex matrix, P(u,v) represents the phase spectrum of the hypercomplex Fourier transform, and χ(u,v) represents the characteristic axis spectrum of the hypercomplex Fourier transform.

[0074] S33: Logarithmic magnitude spectrum transform,

[0075] In order to effectively represent the amplitude spectrum information, logarithmic calculation is used to perform logarithmic transformation on the amplitude spectrum. The calculation formula is as follows:

[0076] A(u,v)=log(Am(u,v)+1)

[0077] Where A(u,v) represents the logarithmic spectrum of the amplitude spectrum.

[0078] Step S4: Gaussian adaptive smoothing filtering.

[0079] In order to effectively extract the defect part, it is necessary to perform convolution operation on the amplitude spectrum A(u,v) using a Gaussian filter template. Therefore, the specific steps of the Gaussian template parameter determination process are as follows:

[0080] S41: define the initial Gaussian template,

[0081] The calculation formula of Gaussian template g(u,v) is as follows:

[0082]

[0083] Among them, the parameters that need to be determined are k and σ, k represents the Gaussian template size parameter, the template size is (2k+1,2k+1), and σ is the standard deviation of the Gaussian template.

[0084] S42: Initialize Gaussian template,

[0085] The σ value is fixed to 0.04×min{M,N}, and the Gaussian template size parameter k is set to the range:

[0086] k=min{M,N}×0.01×n1, (n1=1,2,3,…,8)

[0087] Therefore, a k-based Gaussian template scale space is established.

[0088] S43: Gaussian convolution operation, using the template established above to perform Gaussian convolution operation on the amplitude spectrum, the calculation formula is as follows:

[0089] A(u,v) k ′ =g(u,v)*A(u,v)

[0090] Among them, A(u,v) k ′ is the amplitude spectrum after Gaussian template convolution when the σ parameter is fixed, and * represents the convolution operation.

[0091] S44: Amplitude spectrum entropy calculation,

[0092] Use one-dimensional entropy to calculate the filtered amplitude spectrum and determine the optimal k parameter:

[0093] H = -∑(p×log(p))

[0094] Among them, H represents the information entropy of the amplitude spectrum, p represents the statistical value of the grayscale histogram of the amplitude spectrum, and the entropy value of the amplitude spectrum after filtering is calculated respectively, and the k parameter at the minimum entropy is saved. The calculation method is as follows:

[0095] k=min(H(A(u,v) k ′ ))

[0096] Therefore, the optimal k parameter of Gaussian template size is obtained by calculation.

[0097] S45: Determination of σ parameter,

[0098] Establish the value range of the σ parameter, the range is as follows:

[0099] σ=3 n1-2 ,(n1=1,2,3,…,8)

[0100] Thus, the Gaussian template σ parameter space is established. The Gaussian template is established using the σ parameter and the convolution operation is performed. The calculation formula is as follows:

[0101] A(u,v) σ ′ =g(u,v)*A(u,v)

[0102] Among them, A(u,v) σ ′ is the amplitude spectrum after Gaussian template convolution when the k parameter is fixed.

[0103] Use one-dimensional entropy to calculate the amplitude spectrum after filtering and calculate the optimal σ parameter, where the σ parameter with minimum entropy is saved:

[0104] σ=min(H(A(u,v) σ ′ ))

[0105] Therefore, the optimal Gaussian template filter parameters (k, σ) are determined.

[0106] S46: Optimal smoothed amplitude spectrum calculation,

[0107] The original amplitude spectrum is convolved with the obtained optimal Gaussian template parameters to obtain the optimal smoothed amplitude spectrum. The result is as follows: Figure 4 (a) Original amplitude spectrum, Figure 4 (b) The optimal smoothed amplitude spectrum is shown.

[0108] Step S5, threshold screening and bandpass filtering of the amplitude spectrum,

[0109] In order to reduce the influence of speckle noise on defect image extraction, it is necessary to suppress the speckle noise and enhance the defect area.

[0110] S51: Adaptive Speckle Noise Suppression,

[0111] Calculate the mean and standard deviation of the original amplitude spectrum using the following formula:

[0112] th=μs+Kδs

[0113]

[0114]

[0115] Among them, μs represents the mean of the original amplitude spectrum, δs represents the standard deviation of the original amplitude spectrum, and K represents the parameter of the standard deviation. According to the actual image selection, the best smoothed amplitude spectrum is threshold-screened, and the amplitude values ​​greater than the threshold in the amplitude spectrum are retained to obtain the amplitude spectrum after threshold screening. The screening conditions are as follows:

[0116]

[0117] Among them, F(w) is the amplitude value of the optimal smoothing amplitude, th is the threshold, and the result is as follows Figure 4 (c) Threshold-screened amplitude spectrum.

[0118] S52: Threshold screening amplitude spectrum bandpass filtering,

[0119] The selection of the passband requires comparing the threshold amplitude spectrum without defects with the amplitude spectrum of the threshold screening with defects. Therefore, a suitable bandpass filter is designed based on this, and the amplitude spectrum of the threshold screening is bandpass filtered:

[0120]

[0121] Among them, H(u,v) is a bandpass filter, D0 represents the frequency point selected by the bandpass filter, and the passband range is (D0, -D0), which is determined by the actual situation. After using the bandpass filter, the defect amplitude spectrum information is obtained. The result is as follows Figure 4 (d) Bandpass filtered amplitude spectrum.

[0122] Step S6, hypercomplex inverse Fourier transform and adaptive threshold segmentation,

[0123] S61: Defect saliency map calculation,

[0124] Perform inverse hypercomplex Fourier transform on the defect amplitude spectrum:

[0125]

[0126] Among them, A(u,v)" represents the amplitude spectrum retained after bandpass filtering, and S(n,m) represents the defect saliency map obtained. The results are as follows Figure 5 (a) Hypercomplex inverse Fourier transform saliency map.

[0127] S62: Statistical feature extraction of defect saliency map,

[0128] The mean and standard deviation of the obtained S(n,m) defect saliency map are calculated using the following formula:

[0129] ths=μ m +Kδ m

[0130]

[0131]

[0132] Among them, ths is the binary segmentation threshold of the defect image, μ m is the mean of the defect saliency map, δ m is the standard deviation of the defect saliency map. Use the ths value to binarize the defect image to obtain a binary defect image, such as Figure 5 (b) The adaptive threshold segmentation result is shown in the figure.

[0133] Step S7, adaptive selection of defective areas, the implementation steps are as follows:

[0134] S71: Extraction of structural information of the original image,

[0135] The defect edge image G in step S1 is segmented using a threshold to obtain the contour information of the target, and the contour analysis function is used to obtain the coordinates of the four corner points of the internal contour. Among them, n and m are image space coordinates, l represents left, r represents right, up represents up, and dn represents down. The result is as follows Figure 6 (a) Coordinate diagram of the inner contour corner points.

[0136] S72: Regional statistics of defect binary images,

[0137] In order to effectively calculate the edge coordinates of the defect area, it is necessary to refer to the corner points of the internal contour and the corner points of the actual segmentation area. The calculation method is as follows:

[0138]

[0139] Among them, n cen represents the n coordinate of the center point of the defect binary image, Indicates the minimum n coordinates on the left side of the defect binary image that meet the conditions, Indicates the maximum n coordinates on the left side of the defect binary image that meet the conditions, Indicates the minimum n coordinates on the right side of the defect binary image that meet the conditions, It represents the maximum n coordinates on the right side of the defect binary image that meet the conditions. The result is as follows Figure 6 (b) Defect area segmentation coordinate diagram.

[0140] S73: Defect area segmentation,

[0141] The defect area segmentation is updated based on the inner contour coordinates of the defect and the new coordinates screened. The defect is divided into two areas, where the endpoints of the left area are The right region ends at The new binary image is reconstructed with these four endpoints as the edges to obtain the defect area segmentation image. The result is as follows: Figure 7 shown.

[0142] The method of the present invention specifically adopts the maximum relative entropy criterion to select the significant color channel in RGBY, and forms an image color quaternion matrix with the Lab color space. The image color quaternion matrix is ​​converted into a logarithmic spectrum of frequency domain significance through hypercomplex Fourier transform, and then Gaussian convolution, threshold screening, bandpass filtering and hypercomplex Fourier inverse transform are used to obtain a cross defect saliency map. Finally, the mean and standard deviation of the defect saliency map are calculated and combined with the structural information of the image to perform defect saliency. Figure 2The weak cross defect detection based on value segmentation solves the problem that weak cross defects in display screens cannot be detected, and provides a weak cross defect detection method with high accuracy and fast processing speed in display screen images. It effectively realizes the cross defect detection of display screens with high accuracy under weak contrast, and can be applied to weak cross defect detection of display screens.

Claims

1. A method for detecting weak cross-defects of display screens by frequency domain filtering of significant color channels, comprising the following steps: Step S1, edge gradient suppression of the original image, Calculate the grayscale mean map of the input image, then calculate the gradient map of the image, and use the obtained gradient map to suppress the edge of the original image; Step S2, adaptive selection of color features, For the input display screen image, the RGBY color space and Lab color space are calculated respectively, and the relative entropy is used to adaptively select the significant color channel in the RGBY color space; Step S3, constructing the image quaternion matrix and the hypercomplex Fourier transform, The color information of the four channels extracted is used to construct the image color quaternion matrix, and then the hypercomplex Fourier transform is used to calculate the logarithmic magnitude spectrum of the input features; Step S4, Gaussian adaptive smoothing parameter calculation, Use the Gaussian template and the amplitude spectrum for convolution operation, first fix the standard deviation value of the Gaussian template, change the size of the Gaussian template and perform convolution operation with the amplitude spectrum, use one-dimensional entropy to calculate the information content of the image, retain the template size with minimum entropy, then fix the template size and change the standard deviation value of the Gaussian window, retain the template standard deviation value with minimum entropy, and obtain the optimal amplitude spectrum parameters of the defect; Step S5, threshold screening and bandpass filtering of the amplitude spectrum, Calculate the mean and standard deviation of the original amplitude spectrum, use a threshold to filter the smoothed amplitude values, retain the amplitude spectrum values ​​greater than the threshold, and then use a bandpass filter to select the frequency to obtain the amplitude spectrum of the defect; Step S6, hypercomplex inverse Fourier transform and adaptive threshold segmentation, Perform inverse hypercomplex Fourier transform on the obtained defect amplitude spectrum to obtain the defect image, extract the mean and standard deviation of the defect image, and perform binary segmentation on the image; Step S7, adaptive selection of defective areas, Use structural information to constrain the segmented binary image, so as to achieve accurate segmentation of the defect area; The specific steps of step S1 are as follows: S11: Grayscale mean image I calculation, The calculation is as follows: Among them, r, g, and b represent the red, green, and blue channels of the original image; S12: Calculate the gradient map for the grayscale mean image I, The grayscale mean image I calculates the gradients in the horizontal and vertical directions, then modulos the gradients in the two directions, and finally normalizes the gradient image. The normalized gradient image is subtracted from 1 to obtain the defect edge image G; S13: Original image edge suppression, The defect edge image G is calculated with the original image to obtain the input image with suppressed edges. The calculation formula is as follows: Among them, r i , g i 、b i represents the red, green, and blue channels of the input image with suppressed edges; The specific steps of step S2 are as follows: S21: RGBY color space calculation, The color space calculation is performed on the input image. The calculation formula is as follows: Among them, RG is the red-green opposing color channel, GR is the green-red opposing color channel, BY is the blue-yellow opposing color channel, and YB is the yellow-blue opposing color channel; S22: Lab color space calculation, Perform Lab color space calculation on the edge suppression image obtained in step S13 to obtain L brightness channel, a red and green color channel, b1 yellow and blue color channel, and retain a channel and b1 channel; S23: Selection of significant color channels in RGBY color space, The selection method is as follows: Among them, f1 represents the significant color channel in RG and GR, f2 represents the significant color channel in BY and YB, and H kl Represents the relative entropy calculation, and the calculation formula is as follows: Where P(x) represents the grayscale probability distribution function of input P, ​​and Q(x) represents the grayscale probability distribution function of input Q; The specific steps of step S3 are as follows: S31: Image color quaternion matrix representation, Quaternion matrix f h (n,m) is expressed as follows: f h (n,m)=f1+f2i+f3j+f4k Among them, i 2 =-1,j 2 =-1, k 2 = -1, ijk = -1, n, m represent image space coordinates, f1, f2 are dynamically selected by step S23, f3 represents the a channel of the Lab color space, and f4 represents the b1 channel of the Lab space; S32: Hypercomplex Fourier Transform, The hypercomplex Fourier transform formula is as follows: Among them, F H [u,v] represents the hypercomplex frequency domain space, [u,v] represents the frequency coordinate, M,N are the length and width of the input image, μ is the quaternion unit, and μ 2 = -1; The inverse transform formula of the hypercomplex Fourier transform is as follows: F H The formula for calculating the [u,v] amplitude spectrum is: Am(u,v)=|F H [u,v]| F H The calculation formula of [u,v] phase spectrum is: The calculation formula of characteristic axis spectrum is: Where Am(u,v) represents the amplitude spectrum of the hypercomplex Fourier transform, |·| represents the modulus of each element in the hypercomplex matrix, P(u,v) represents the phase spectrum of the hypercomplex Fourier transform, and χ(u,v) represents the characteristic axis spectrum of the hypercomplex Fourier transform; S33: Logarithmic magnitude spectrum transform, Use logarithmic calculation to logarithmically transform the amplitude spectrum. The calculation formula is as follows: A(u,v)=log(Am(u,v)+1) Where A(u,v) represents the logarithmic spectrum of the amplitude spectrum; The specific steps of step S4 are as follows: S41: define the initial Gaussian template, The calculation formula of Gaussian template g(u,v) is as follows: Among them, the parameters that need to be determined are k and σ, k represents the Gaussian template size parameter, the template size is (2k+1,2k+1), and σ is the standard deviation of the Gaussian template; S42: Initialize Gaussian template, The σ value is fixed to 0.04×min{M,N}, and the Gaussian template size parameter k is set to the range: k=min{M,N}×0.01×n1,n1=1,2,3,…,8 Therefore, a k-based Gaussian template scale space is established; S43: Gaussian convolution operation, The calculation formula is as follows: A(u,v) k ′ =g(u,v)*A(u,v) Among them, A(u,v) k ′ is the amplitude spectrum after Gaussian template convolution when the σ parameter is fixed, * indicates the convolution operation; S44: Amplitude spectrum entropy calculation, Use one-dimensional entropy to calculate the filtered amplitude spectrum and determine the optimal k parameter: H = -∑(p×log(p)) Among them, H represents the information entropy of the amplitude spectrum, p represents the statistical value of the grayscale histogram of the amplitude spectrum, and the entropy value of the amplitude spectrum after filtering is calculated respectively, and the k parameter at the minimum entropy is saved. The calculation method is as follows: k=min(H(A(u,v) k ′ )) The optimal k parameter of Gaussian template size was obtained by calculation; S45: Determination of σ parameter, Establish the value range of the σ parameter, the range is as follows: σ=3 n1-2 ,n1=1,2,3,…,8 Establish the Gaussian template σ parameter space, use the σ parameter to establish the Gaussian template, and perform convolution operation. The calculation formula is as follows: A(u,v) σ ′ =g(u,v)*A(u,v) Among them, A(u,v) σ ′ is the amplitude spectrum after Gaussian template convolution when the k parameter is fixed; Use one-dimensional entropy to calculate the amplitude spectrum after filtering and calculate the optimal σ parameter, where the σ parameter with minimum entropy is saved: σ=min(H(A(u,v) σ ′ )) Determine the optimal Gaussian template filter parameters (k, σ); S46: Optimal smoothed amplitude spectrum calculation, The original amplitude spectrum is convolved with the obtained optimal Gaussian template parameters to obtain the optimal smoothed amplitude spectrum; The specific steps of step S5 are as follows: S51: Adaptive Speckle Noise Suppression, Calculate the mean and standard deviation of the original amplitude spectrum using the following formula: th=μs+Kδs Among them, μs represents the mean of the original amplitude spectrum, δs represents the standard deviation of the original amplitude spectrum, and K represents the parameter of the standard deviation. According to the actual image selection, the best smoothed amplitude spectrum is threshold-screened, and the amplitude values ​​greater than the threshold in the amplitude spectrum are retained to obtain the amplitude spectrum after threshold screening. The screening conditions are as follows: Among them, F(w) is the amplitude value of the optimal smoothing amplitude, and th is the threshold; S52: Threshold screening amplitude spectrum bandpass filtering, Bandpass filter the thresholded amplitude spectrum: Among them, H(u,v) is a bandpass filter, D0 represents the frequency point selected by the bandpass filter, the passband range is (D0, -D0), and the defect amplitude spectrum information is obtained after using the bandpass filter; The specific steps of step S6 are as follows: S61: Defect saliency map calculation, Perform inverse hypercomplex Fourier transform on the defect amplitude spectrum: Where A(u,v)" represents the amplitude spectrum retained after bandpass filtering, and S(n,m) represents the defect saliency map obtained; S62: Statistical feature extraction of defect saliency map, Calculate the mean and standard deviation of the obtained S(n,m) defect saliency map: ths=μ m +Kδ m Among them, ths is the binary segmentation threshold of the defect image, μ m is the mean of the defect saliency map, δ m is the standard deviation of the defect saliency map; the defect image is binarized using the ths value to obtain a binary defect image; The specific steps of step S7 are as follows: S71: Extraction of structural information of the original image, The defect edge image G in step S1 is segmented using a threshold to obtain the contour information of the target and the coordinates of the four corner points of the internal contour. Among them, n and m are image space coordinates, l represents left, r represents right, up represents up, and dn represents down; S72: Regional statistics of defect binary images, Calculate the edge coordinates of the defect area, refer to the corner points of the internal contour, and the corner points of the actual segmented area: Among them, n cen represents the n coordinate of the center point of the defect binary image, Indicates the minimum n coordinates on the left side of the defect binary image that meet the conditions, Indicates the maximum n coordinates on the left side of the defect binary image that meet the conditions, Indicates the minimum n coordinates on the right side of the defect binary image that meet the conditions, Indicates the maximum n coordinates of the right end that meet the conditions in the defect binary image; S73: Defect area segmentation, The defect area segmentation is updated based on the inner contour coordinates of the defect and the new filtered coordinates; the defect is divided into two areas, where the endpoint of the left area is The right region ends at A new binary image is reconstructed with these four endpoints as edges to obtain the defect area segmentation image.

Images