Fan blade aerial image deblurring method fusing auxiliary image prior

Abstract

The invention discloses a fan blade aerial image deblurring method fused with auxiliary image prior, which comprises the following steps of: establishing an image degradation model, estimating a blurring kernel and a potential image in the image degradation model by optimizing a target function, and after the blurring kernel is obtained, performing image deblurring on the image degradation model; non-blind deconvolution is applied to the input blurred image to generate a final clear image. According to the method, the auxiliary image is introduced in the process of estimating the blurring kernel, the structural consistency and defect identification degree of the recovered image are remarkably improved, a breakthrough from natural image priori to task specific priori is achieved, after the blurring kernel estimation is completed, standard deconvolution is not directly used, and the algorithm is simple and convenient to operate. Compared with the prior art, the non-blind deconvolution energy model fusing multiple image structure priori is innovatively proposed, the common problems of artifacts, edge blurring and the like in a traditional deconvolution image are greatly solved, high-fidelity image reconstruction is achieved, and the accuracy and robustness of a subsequent defect detection system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence, computer vision, image processing and wind power equipment operation and maintenance, and in particular to a deblurring method for aerial images of wind turbine blades. Background Art

[0002] As a key component of wind turbine generators, the operating condition of wind turbine blades directly impacts power generation efficiency and equipment safety. However, wind turbine blades are constantly exposed to complex and changing environments, such as strong winds, dust, and salt spray, making them susceptible to defects such as cracks, corrosion, and wear. If not promptly detected and repaired, these defects can further expand, leading to blade breakage or even unit failure, resulting in severe economic losses and safety risks. Therefore, efficient and accurate wind turbine blade defect detection technology has become a key research area in wind power operation and maintenance.

[0003] Detecting surface defects in wind turbine blades is a critical component of wind turbine operation and maintenance. Unmanned aerial vehicle (UAV)-based inspection methods have become a mainstream approach due to their flexibility, high inspection efficiency, and low cost. However, in practical applications, this technology faces significant image quality challenges. First, wind turbine blades typically rotate continuously at 10-20 rpm, causing significant surface movement during the camera exposure time, resulting in motion blur in the captured images. Second, UAVs are susceptible to interference from atmospheric turbulence and mechanical vibration during inspections, making it difficult for them to remain absolutely stationary. This platform vibration further exacerbates image blur. Furthermore, complex environmental conditions such as sudden changes in lighting, haze, rain, and snow can significantly reduce the image signal-to-noise ratio, exacerbating the blurring effect. These image degradation issues can lead to a significant loss of texture features for defects such as microcracks and corrosion on the blade surface. This not only reduces the detection accuracy of computer vision algorithms but can also lead to serious consequences such as missed or false detections. Therefore, it is of great engineering value to develop an efficient image restoration method to address the problem of wind blade defect image degradation caused by motion blur. This is of great significance for improving the accuracy and robustness of defect detection algorithms and ensuring the safe and stable operation of wind turbines.

[0004] In single image deblurring, image degradation is usually modeled as:

[0005] b=k*l+n

[0006] Where b is the observed blurred image, k is the blur kernel, * is the convolution operator, l is the potential clear image, and n is the image noise. The goal of blind deconvolution is to recover k and l from a single input b to achieve image deblurring.

[0007] This modeling approach generally performs well in natural image deblurring tasks, but its effectiveness degrades significantly when faced with images with specific characteristics, such as those used in drone-generated wind turbine blade defect detection. This is primarily due to the insufficient constraints of the natural image statistical prior in such scenarios, making it difficult to effectively capture the texture characteristics of blade defect images, which lack internal color gradients and have sharp edges. In particular, in the presence of a large blur kernel, the natural image model struggles to accurately recover the details of the blade defect. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a deblurring method for aerial images of wind blades that integrates auxiliary image priors to solve the blurring problem of aerial images of wind blades and improve the recognizability of defective areas in blurred images.

[0009] The deblurring method for aerial images of wind turbine blades by integrating auxiliary image priors of the present invention comprises the following steps:

[0010] 1) Establishing image degradation model:

[0011] b=k*l+n (1)

[0012] Where b is the blurred image of the wind turbine blade, k is the blur kernel, * is the convolution operator, l is the potential clear image, and n is the image noise;

[0013] 2) Estimate a pair of solutions of k and l in the image degradation model by optimizing the following objective function:

[0014] argmin l,k,a ‖bk*l‖ 2 +ρ l (l)+ρ k (k)+ρ a (a)+β‖la‖ 2 (2)

[0015] Among them, ‖bk*l‖ 2 is the data fitting term, ρ l (l) is the prior term for the potential clear image, ρ k (k) is the prior term for the blur kernel, a is the auxiliary image, ρ a (a) is the cost function, ‖la‖ 2 is a penalty term that measures the similarity between the latent image l and the auxiliary image a, and β is the weight of the penalty term;

[0016] where ρ l (l) Adjustment is made by:

[0017]

[0018] Where i is the pixel index in the image, and the sum is performed for each pixel in the entire image; Represents the horizontal gradient of image l at pixel i, Represents the vertical gradient of image l at pixel i, and α is used to impose constraints on the sparsity of image gradients;

[0019] Among them, the cost function ρ a (a) is defined as follows:

[0020]

[0021] Among them, a i is the pixel value of the auxiliary image a at the position of the i-th pixel, is the ideal image a P The pixel value at the position of the i-th pixel, i is the pixel index, d MAX is a constant whose value is greater than the maximum pixel value of l; the meaning of formula (4) is ρ a (a) will be compared with the ideal image a that meets the texture characteristics of the leaf defect P Penalize inconsistent pixels;

[0022] where ρ k (k) is defined as follows:

[0023]

[0024] Among them, λ k is the regularization coefficient, which controls the weight of the regularization term. The larger the value, the stronger the smoothness constraint on k; represents the horizontal gradient of k at the i-th pixel, Represents the gradient of k in the vertical direction at the i-th pixel;

[0025] The steps to estimate the clear image l and blur kernel k are as follows:

[0026] 1) Parameter initialization: The blur kernel k is initialized to unit impact, the potential clear image l and the auxiliary image a are initialized to the blurred image b; β is initialized to β0 = 0.01; set the parameter δ = 1.5, β MAX =5;

[0027] 2) Calculate l and a alternately using the following formula:

[0028] arg min l,a ||bk*l|| 2 +ρ l (l)+ρ a (a)+β||la|| 2 (6)

[0029] In formula (6), one of the variables l and a is fixed to calculate the other variable, and the fixed value comes from the result of the previous iteration;

[0030] ①Calculate l:

[0031] Given a blurred image b, blur kernel k and auxiliary image a, solve:

[0032] arg min l ||bk*l|| 2 +ρ l (l)+β||la|| 2 (7)

[0033] in Used to regularize the gradient of l; Formula (7) is solved by fast Fourier transform:

[0034]

[0035] in: and are the Fourier transform and its complex conjugate, is the Fourier transform of the unit impulse function; is the gradient operator; let λ l = 0.1 and use it as an initial estimate of l;

[0036] ②Calculate a:

[0037] First, minimize the L0 gradient of l calculated in step ①;

[0038] Then use the image l obtained by L0 gradient minimization to construct the ideal image a P : Use Gabor to segment the image l, dividing the image l into defect areas and normal background areas, uniformly recolor the pixels in each defect area and set them to the maximum value of the area; the pixels in the non-defective area are assigned zero; thus constructing the ideal image a P ;

[0039] Then calculate a using the following formula:

[0040] arg min a ρ a (a)+β||la|| 2 (9)

[0041] Since ρ a (a) is defined pixel by pixel, and formula (9) is solved independently for each pixel, and the final solution is a i * for:

[0042]

[0043] Where l i is the value of the latent image l at pixel position i in the current iteration. In the early iteration of formula (9), β is small and the second term The penalty is lighter, tending to make a strongly consistent with the prior; as the iteration proceeds, β increases, the term becomes more significant, and a is forced to be close to l;

[0044] 3) Estimate blur kernel k:

[0045] After introducing a, the blur kernel k is iteratively updated by minimizing the following formula:

[0046]

[0047] in Represents the set of image derivative operators involved in the calculation, which is used to capture structural information such as edges and textures of the image. Represents the first-order partial derivative of the image along the horizontal x-axis, Represents the first-order partial derivative of the image along the vertical y-axis, Represents the second-order partial derivative of the image along the horizontal x-axis, represents the mixed partial derivatives of the image in the x and y directions, Represents the second-order partial derivative of the image along the vertical y-axis; Represents the weight coefficient of each derivative direction, which is used to adjust the contribution of different derivative terms to the overall loss function; is a data term that measures the residual between the blurred image b and the auxiliary image a convolved with blur kernel k in different derivative directions; γ||k|| 2 is the Tikhonov regularization term, which is used to prevent the blur kernel from overfitting and encourage the blur kernel to remain small and smooth. γ is the regularization weight, which controls the degree of penalty on the blur kernel size.

[0048] 4) After each iteration, multiply the weight β by δ. If the product of δ and β is greater than or equal to 5, set β = 5 and repeat steps 2) and 3) until l and a converge.

[0049] 3) After the blur kernel k is obtained by iterative calculation in step 2), non-blind deconvolution is applied to the input blurred image b to generate the final clear image l. The overall energy function of non-blind deconvolution is defined as follows:

[0050] E(l)=||bk*l|| 2 +ρ T (l) (12)

[0051] in:

[0052]

[0053] Where T is a P The pixel set of the leaf defect area detected in ; represents the gradient of image l at pixel i, indicating the local intensity change; The gradient value in the defective area of the blade is encouraged to be zero; The term makes the pixel value of the defective area of the blade consistent with the ideal image a P consistent; The background area is encouraged to have sparse gradients, and α is set to 0.8; λ1, λ2, and λ2 are the weights of each item respectively.

[0054] Beneficial effects of the present invention:

[0055] 1. Most traditional image deblurring methods rely on natural image statistical priors, making it difficult to accurately restore clear details of defective areas of wind blades. To overcome this limitation, the present invention proposes the introduction of an "auxiliary image prior" a, which is embedded into the optimization framework through a semi-quadratic penalty method to explicitly model the structural characteristics of wind blade images. The auxiliary image a is generated from the latent clear image l and is continuously refined during iteration. Its constraint effect gradually increases with the β factor, guiding the image restoration process to more accurately meet the physical characteristics of defective areas such as cracks and corrosion, which require "clear boundaries and small internal gradients." This method significantly improves the structural consistency and defect recognition of the restored image, achieving a breakthrough from natural image priors to task-specific priors.

[0056] 2. Aiming at the image characteristics of "sparse gradients and significant boundaries" in the defective area of wind turbine blades, this paper constructs an optimization mechanism based on L0 gradient minimization and regional structure recoloring. L0 regularization is used to retain the significant gradient at the defect edge, suppressing the ringing artifacts and image noise caused by blurring; at the same time, a Gabor filter is used to perform texture segmentation on image l, and the pixels in each defect area are forced to tend to a uniform grayscale to achieve internal color consistency, thus forming an ideal image a. P The above processing not only improves the image quality of the defect area, but also provides a more accurate image structure prior for subsequent blur kernel estimation and final deconvolution, thereby achieving more accurate blur removal and defect reconstruction.

[0057] 3. After the blur kernel estimation is completed, the present invention does not directly use the standard deconvolution, but innovatively proposes a non-blind deconvolution energy model that integrates multiple image structure priors. This model sets the internal gradient of the defect area to zero and the pixel value to the ideal image a. P The three priors of close proximity and background area gradient sparseness are jointly embedded in the optimization objective function to form a constraint mechanism for the characteristics of the fan blade defect image. In the iterative process, not only the defect area T is dynamically updated, but also the ideal image a is adjusted in real time.P , to maintain the authenticity and consistency of structural information. This method greatly improves the common artifacts and edge blurring problems in traditional deconvolution images, achieves high-fidelity image reconstruction, and enhances the accuracy and robustness of subsequent defect detection systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The overall process diagram of wind blade image deblurring. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings and examples.

[0060] The deblurring method for aerial images of wind turbine blades by integrating auxiliary image priors in this embodiment includes the following steps:

[0061] 1) Establishing image degradation model:

[0062] b=k*l+n (1)

[0063] Where b is the blurred image of the wind blade, k is the blur kernel, * is the convolution operator, l is the latent sharp image, and n is the image noise.

[0064] Texture features of the leaf defect image that need to be considered in the deblurring process:

[0065] Feature 1: Defects on wind turbine blades, such as cracks and holes, usually have high contrast with the adjacent background areas.

[0066] Feature 2: The overall color of the leaf defect is close to uniform, and its internal gradient value should be close to zero.

[0067] Feature 3: Although the background of many aerial wind turbine blade images is the color of the wind turbine, due to the varying shooting angles, backgrounds such as clouds and the ground also occupy a certain proportion of the image. Therefore, it is necessary to assume that the background gradient obeys natural image statistics and passes the sparse constraint.

[0068] 2) In order to fuse these features in the deblurring process, a pair of solutions of k and l in the image degradation model is estimated by optimizing the objective function as follows:

[0069] argmin l,k,a ‖bk*l‖ 2 +ρ l (l)+ρ k (k)+ρ a (a)+β‖la‖ 2 (2)

[0070] Among them, ‖bk*l‖ 2 is the data fitting term, ρ l(l) is the prior term for the potential clear image, ρ k (k) is the prior term for the blur kernel, a is the auxiliary image, ρ a (a) is the cost function, ‖la‖ 2 is a penalty term that measures the similarity between the latent image l and the auxiliary image a, and β is the weight of the penalty term.

[0071] Since the blade defect features act as image priors, ρ needs to be adjusted in some way l (1). In this embodiment, p l (l) Adjustment is made by:

[0072]

[0073] Where i is the pixel index in the image, and the sum is performed for each pixel in the entire image; Represents the horizontal gradient of image l at pixel i, Represents the vertical gradient of image l at pixel i, and α is used to impose constraints on the sparsity of image gradients.

[0074] Since the three potential blade defect image features analyzed above require a detailed separation of the defect and background areas, and this spatial segmentation is difficult to define using a simple gradient, this embodiment introduces an auxiliary image a in formula (2) to enforce the specific features of the blade defect on the latent image l. When β is large enough, formula (2) will approach argmin l,k ‖bk*l‖ 2 +ρ l (l)+ρ k (k), and ρ l (l)≈ρ a (a). In this form, the domain-specific prior ρ a (a) is explicitly embedded into the optimization framework, while the parameter β is automatically adjusted during the iteration process to control when and to what extent ρ a The strong constraints of (a) are reflected in the latent image l.

[0075] ρ a (a) is used to measure whether the auxiliary image a satisfies the image texture characteristics given by the above analysis. Cost function ρ a (a) is defined as follows:

[0076]

[0077] Among them, a i is the pixel value of the auxiliary image a at the position of the i-th pixel, is the ideal image aP The pixel value at the position of the i-th pixel, i is the pixel index, d MAX is a constant whose value is greater than the maximum pixel value of l; the meaning of formula (4) is ρ a (a) will be compared with the ideal image a that meets the texture characteristics of the leaf defect P Inconsistent pixels are penalized.

[0078] where ρ k (k) is defined as follows:

[0079]

[0080] Among them, λ k Regularization coefficient, which controls the weight of the regularization term. The larger the value, the stronger the smoothness constraint on k; represents the horizontal gradient of k at the i-th pixel, represents the gradient of k in the vertical direction at the i-th pixel. The main function of formula (5) is to constrain the smoothness of the blur kernel k.

[0081] The steps to estimate the clear image l and blur kernel k are as follows:

[0082] 1) Parameter initialization: The blur kernel k is initialized to unit impact, the potential clear image l and the auxiliary image a are initialized to the blurred image b; β is initialized to β0 = 0.01; set the parameter δ = 1.5, β MAX =5.

[0083] 2) Calculate l and a alternately using the following formula:

[0084]

[0085] In formula (6), one of the variables l and a is fixed to calculate the other variable, and the fixed value comes from the result of the previous iteration.

[0086] ①Calculate l:

[0087] Given a blurred image b, blur kernel k and auxiliary image a, solve:

[0088]

[0089] in Used to regularize the gradient of l; Formula (7) is solved by fast Fourier transform:

[0090]

[0091] in: and are the Fourier transform and its complex conjugate, is the Fourier transform of the unit impulse function; It is a gradient operator that reflects the degree of change of the pixel value of the leaf image in space and has a higher response in the defect edge area; is the Fourier transform of the gradient operator, which is used to perform regularized calculation of image gradient in the frequency domain; let λ l = 0.1 and use it as an initial estimate of l.

[0092] ②Calculate a:

[0093] First, perform L0 gradient minimization on l calculated in step ①. The L0 gradient minimization method can effectively remove the noise caused by the simple deconvolution in formula (8) and preserve small edges to generate the initial auxiliary image a, which also meets the requirements of features 1 and 2 mentioned above. The boundaries of blade defects usually have large gradients and are therefore well preserved in the L0 gradient minimization. Noise and ringing artifacts have small gradients and are therefore suppressed within the defect area.

[0094] Then calculate a using the following formula:

[0095] arg min a ρ a (a)+β||la|| 2 (9)

[0096] Then use the image l obtained by L0 gradient minimization to construct the ideal image a P Since the ideal image a in formula (4) P is unknown, and it needs to satisfy the structural characteristic constraints of the fan blade defects, so it is necessary to automatically extract structural information from the current estimated image l to construct a P This embodiment uses Gabor filters to perform texture segmentation on l. By applying multi-scale and multi-directional Gabor filter responses to the image, local texture changes are captured and the image is divided into defect areas and normal background areas based on the response characteristics. The pixels in each defect area are uniformly recolored and set to the maximum value of the area. The pixels in non-defective areas are assigned zero values, thereby constructing an ideal image a. P The ideal image a is constructed P The uniform texture response and clear boundaries in the defect area can effectively reflect the structural prior characteristics in the blade image and provide powerful prior guidance for the blur kernel estimation and image restoration process.

[0097] Since ρ a (a) is defined pixel by pixel, and formula (9) is solved independently for each pixel, and the final solution is a i * for:

[0098]

[0099] Where l i is the value of the latent image l at pixel position i in the current iteration. In the early iteration of formula (10), β is small and the second term The penalty is relatively light, tending to make a strongly consistent with the prior. As the iterations proceed, β increases, the term becomes more significant, and a is forced to approach l. Since l is updated based on a in each iteration, eventually l will gradually conform to the characteristics of the blade defect image.

[0100] Pixels inside the leaf defect ultimately have the same intensity, thus satisfying Feature 2. By minimizing the L0 gradient and recoloring the leaf defect region, significant gradients at the defect boundary are preserved, thus satisfying Feature 1. However, a sparse prior (Feature 3) is not introduced for the background in this step because the current goal is to recover the sharp edges of the defect boundary, which are important for kernel estimation. The sparse prior will be added in the final deconvolution step.

[0101] 3) Estimate blur kernel k:

[0102] After introducing a, the blur kernel k is iteratively updated by minimizing the following formula:

[0103]

[0104] in Represents the set of image derivative operators involved in the calculation, which is used to capture structural information such as edges and textures of the image. Represents the first-order partial derivative of the image along the horizontal x-axis, Represents the first-order partial derivative of the image along the vertical y-axis, Represents the second-order partial derivative of the image along the horizontal x-axis, represents the mixed partial derivatives of the image in the x and y directions, Represents the second-order partial derivative of the image along the vertical y-axis; Represents the weight coefficient of each derivative direction, which is used to adjust the contribution of different derivative terms to the overall loss function; is a data term that measures the residual between the blurred image b and the auxiliary image a convolved with blur kernel k in different derivative directions; γ||k|| 2 is the Tikhonov regularization term, which is used to prevent the blur kernel from overfitting and encourage the blur kernel to remain small and smooth. γ is the regularization weight, which controls the degree of penalty on the blur kernel size to prevent overfitting.

[0105] In formula (11), which updates the blur kernel k, the image edge information of the auxiliary image a is used instead of the image l. Although l and a should eventually converge to the same image according to the alternating optimization process described in formula (7), l calculated using the fast Fourier transform (FFT) in formula (8) is often accompanied by ringing artifacts and noise. These artifacts have been effectively suppressed by optimizing a in step 2. Therefore, using a as the potential image for blur kernel estimation in formula (11) can significantly improve the estimation accuracy and deblurring effect.

[0106] 4) After each round of iteration, multiply the weight β by δ. If the product of δ and β is greater than or equal to 5, set β = 5, and then repeat steps 2) and 3) until l and a converge.

[0107] 3) Since the blur kernel estimation stage primarily optimizes the restoration of the blade defect edge region, the resulting latent image is typically suboptimal in visual quality. To obtain a high-quality output image, after iteratively calculating the blur kernel k in step 2), non-blind deconvolution is applied to the input blurred image b to generate the final clear image l. The overall energy function of non-blind deconvolution is defined as follows:

[0108] E(l)=||bk*l|| 2 +ρ T (l) (12)

[0109] in:

[0110]

[0111] Where T is a P The pixel set of the leaf defect area detected in ; represents the gradient of image l at pixel i, indicating the local intensity change; The gradient value in the defective area of the blade is encouraged to be zero; The term makes the pixel value of the defective area of the blade consistent with the ideal image a P consistent; The background area is encouraged to have sparse gradients, and α is set to 0.8; λ1, λ2, and λ2 are the weights of each item, which respectively control the gradient constraint strength inside the blade defect area, the distance between the blade defect area and the ideal image a. P The pixel difference and the gradient sparsity intensity of the background area.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for deblurring aerial images of wind turbine blades by integrating auxiliary image priors, characterized by: The following steps are involved: 1) Establishing image degradation model: b=k*l+n (1) Where b is the blurred image of the wind turbine blade, k is the blur kernel, * is the convolution operator, l is the potential clear image, and n is the image noise; 2) Estimate a pair of solutions of k and l in the image degradation model by optimizing the following objective function: argmin l,k,a ‖bk*l‖ 2 +r l (l)+p k (k)+p a (a)+β‖la‖ 2 (2) Among them, ‖bk*l‖ 2 is the data fitting term, ρ l (l) is the prior term for the potential clear image, ρ k (k) is the prior term for the blur kernel, a is the auxiliary image, ρ a (a) is the cost function, ‖la‖ 2 is a penalty term that measures the similarity between the latent image l and the auxiliary image a, and β is the weight of the penalty term; where ρ l (l) Adjustment is made by: Where i is the pixel index in the image, and the sum is performed for each pixel in the entire image; Represents the horizontal gradient of image l at pixel i, Represents the vertical gradient of image l at pixel i, and α is used to impose constraints on the sparsity of image gradients; Among them, the cost function ρ a (a) is defined as follows: Among them, a i is the pixel value of the auxiliary image a at the position of the i-th pixel, is the ideal image a P The pixel value at the position of the i-th pixel, i is the pixel index, d MAX is a constant whose value is greater than the maximum pixel value of l; the meaning of formula (4) is ρ a (a) will be compared with the ideal image a that meets the texture characteristics of the leaf defect P Penalize inconsistent pixels; where ρ k (k) is defined as follows: Among them, λ k is the regularization coefficient, which controls the weight of the regularization term. The larger the value, the stronger the smoothness constraint on k; represents the horizontal gradient of k at the i-th pixel, Represents the gradient of k in the vertical direction at the i-th pixel; The steps to estimate the clear image l and blur kernel k are as follows: 1) Parameter initialization: The blur kernel k is initialized to unit impact, the potential clear image l and the auxiliary image a are initialized to the blurred image b; β is initialized to β0 = 0.01; set the parameter δ = 1.5, β MAX =5; 2) Calculate l and a alternately using the following formula: argmin l,a ||bk*l|| 2 +r l (l)+p a (a)+β||la|| 2 (6) In formula (6), one of the variables l and a is fixed to calculate the other variable, and the fixed value comes from the result of the previous iteration; ①Calculate l: Given a blurred image b, blur kernel k and auxiliary image a, solve: argmin l ||bk*l|| 2 +r l (l)+β||la|| 2 (7) where ρ l (l) = λ l ‖▽l‖ 2 Used to regularize the gradient of l; Formula (7) is solved by fast Fourier transform: in: and are the Fourier transform and its complex conjugate, is the Fourier transform of the unit impulse function; is the gradient operator; let λ l = 0.1 and use it as an initial estimate of l; ②Calculate a: First, minimize the L0 gradient of l calculated in step ①; Then use the image l obtained by L0 gradient minimization to construct the ideal image a P : Use Gabor to segment the image l, dividing the image l into defect areas and normal background areas, uniformly recolor the pixels in each defect area and set them to the maximum value of the area; the pixels in the non-defective area are assigned zero; thus constructing the ideal image a P ; Then calculate a using the following formula: argmin a r a (a)+β||la|| 2 (9) Since ρ a (a) is defined pixel by pixel, and formula (9) is solved independently for each pixel. The final solution is for: Where l i is the value of the latent image l at pixel position i in the current iteration. In the early iteration of formula (9), β is small and the second term The penalty is lighter, tending to make a strongly consistent with the prior; as the iteration proceeds, β increases, the term becomes more significant, and a is forced to be close to l; 3) Estimate blur kernel k: After introducing a, the blur kernel k is iteratively updated by minimizing the following formula: in Represents the set of image derivative operators involved in the calculation, which is used to capture structural information such as edges and textures of the image. Represents the first-order partial derivative of the image along the horizontal x-axis, Represents the first-order partial derivative of the image along the vertical y-axis, Represents the second-order partial derivative of the image along the horizontal x-axis, represents the mixed partial derivatives of the image in the x and y directions, Represents the second-order partial derivative of the image along the vertical y-axis; Represents the weight coefficient of each derivative direction, which is used to adjust the contribution of different derivative terms to the overall loss function; is a data term that measures the residual between the blurred image b and the auxiliary image a convolved with blur kernel k in different derivative directions; γ||k|| 2 is the Tikhonov regularization term, which is used to prevent the blur kernel from overfitting and encourage the blur kernel to remain small and smooth. γ is the regularization weight, which controls the degree of penalty on the blur kernel size. 4) After each iteration, multiply the weight β by δ. If the product of δ and β is greater than or equal to 5, set β = 5 and repeat steps 2) and 3) until l and a converge. 3) After the blur kernel k is obtained by iterative calculation in step 2), non-blind deconvolution is applied to the input blurred image b to generate the final clear image l. The overall energy function of non-blind deconvolution is defined as follows: E(l)=||bk*l|| 2 +ρ T (l) (12) in: Where T is a P The pixel set of the leaf defect area detected in ; represents the gradient of image l at pixel i, indicating the local intensity change; The gradient value in the defective area of the blade is encouraged to be zero; The term makes the pixel value of the defective area of the blade consistent with the ideal image a P consistent; The background area is encouraged to have sparse gradients, and α is set to 0.8; λ1, λ2, and λ2 are the weights of each item respectively.

Images