LRA-SVD adaptive detail preserving denoising algorithm based on improved noise estimation

Abstract

The application discloses an LRA-SVD adaptive detail-preserving noise reduction algorithm based on improved noise estimation, and comprises the following steps: performing noise estimation on a subsurface defect image; performing noise reduction on an original image by using a noise estimation value and an LRA-SVD algorithm; obtaining an initial method denoising image by subtracting the original image from a preliminary noise reduction algorithm; calculating a correlation coefficient between the denoising image and a method noise, and if the correlation coefficient is greater than a threshold value, performing an iterative regularization noise reduction process: multiplying the method noise by a certain coefficient and adding the method noise to the denoising image as a new noisy image; performing noise estimation on the new noisy image by using an improved noise estimation method, performing noise reduction on the new noisy image by using a noise estimation value and the LRA-SVD algorithm; subtracting the original image from the new denoising image; calculating a correlation coefficient between the new denoising image and a new method noise, and if the correlation coefficient is less than the threshold value, stopping iteration to obtain a denoised subsurface defect image; otherwise, performing a new round of regularization noise reduction.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an LRA-SVD adaptive detail-preserving denoising algorithm based on improved noise estimation. Background Technology

[0002] Conventional filtering-based noise reduction methods can lead to the loss of high-frequency information of subsurface defects while eliminating noise. One of the problems that needs to be solved in dark field confocal image processing is how to remove as much noise as possible while preserving the detailed information of subsurface defects.

[0003] In recent years, the nonlocal self-similarity prior of images has been widely used in denoising algorithms. Low-rank denoising algorithms based on nonlocality theory and low-rank theory group similar blocks in an image into image block matrix; then, singular value decomposition is performed on each combined matrix, retaining large singular values ​​to approximate the matrix with low rank, thus achieving good denoising results. However, how to estimate the number of singular values ​​to retain or how to estimate the rank of the noise-free matrix are problems that need to be solved by such methods. The LRA-SVD algorithm proposed by Guo et al. effectively solves the problem of reasonable estimation of matrix rank. This method utilizes the optimal energy compression property in the SVD domain and retains large singular values ​​through hard thresholding to achieve low-rank estimation of the matrix. However, this method relies on noise priors and requires high accuracy in noise estimation. Noise bias can lead to inaccurate rank determination. Too low a rank can cause loss of detail, while too high a rank can leave noise residue. Furthermore, this method ignores the fact that noise also exists in larger singular values. It uses a hard thresholding method to discard all small singular values, resulting in some noise information distributed in the large singular value vector space remaining. At the same time, it causes the detailed information distributed in the smaller singular value vector space to be discarded, thus causing the denoised image to lose detailed information.

[0004] Furthermore, existing algorithms use the same noise estimation method as the original image during iterative regularization denoising, which leads to inaccurate noise estimation for the newly combined noisy image, thereby reducing the accuracy of the iterative denoising results and increasing the time complexity of the algorithm.

[0005] Therefore, how to provide an adaptive detail-preserving denoising algorithm for subsurface defect images that can overcome the above-mentioned defects is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, this invention provides an adaptive detail-preserving denoising algorithm for LRA-SVD based on improved noise estimation. This algorithm overcomes the problems of traditional LRA-SVD algorithms, which rely on noise priors, require high accuracy in noise estimation, and suffer from inaccurate rank determination due to noise bias. Too low a rank results in detail loss, while too high a rank leads to residual noise. Furthermore, it overcomes the problem that LRA-SVD ignores the existence of noise in larger singular values, using a hard thresholding method to discard all small singular values. This results in residual noise information distributed in the large singular value vector space, while simultaneously discarding detail information distributed in the smaller singular value vector space, leading to loss of detail information in the denoised image. This invention reduces the high-precision requirements of the noise estimation algorithm, achieving adaptive denoising and reducing residual noise while preserving image detail to the greatest extent. By improving the noise estimation algorithm in the iterative denoising process, computational efficiency is significantly improved, and computational overhead is greatly reduced. This invention is used for denoising dark-field confocal subsurface defect images.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] An adaptive detail-preserving denoising algorithm based on improved noise estimation using LRA-SVD includes:

[0009] S1: Obtain the original subsurface defect image and perform noise estimation on the original subsurface defect image using a noise estimation algorithm;

[0010] S2: Denoise the original subsurface defect image by using noise estimation and the LRA-SVD algorithm to obtain the initial denoised image;

[0011] S3: Obtain the initial method noise image by subtracting the original subsurface defect image from the initial denoised image;

[0012] S4: Calculate the correlation coefficient between the initial denoised image and the initial method noise image, and determine whether the correlation coefficient is greater than the preset threshold. If it is greater, proceed with the S5 iterative regularization denoising process. If it is not greater, obtain the denoised subsurface defect image.

[0013] S5: Multiply the initial method noise image by a scaling factor and add it to the initial denoised image to obtain a new noisy image; use the noise estimation model to estimate the noise in the new noisy image, and use the noise estimate and LRA-SVD algorithm to denoise the new noisy image to obtain a new denoised image; calculate the difference between the new noisy image and the new denoised image to obtain a new method noise image; calculate the correlation coefficient between the new denoised image and the new method noise image, and determine whether the correlation coefficient is greater than a preset threshold. If it is greater, perform a new round of S5 iteration regularization denoising process; if it is not greater, obtain the denoised subsurface defect image.

[0014] The noise estimation model is

[0015] in, This is the noise estimate for the (k+1)th iteration. This represents the noise estimate obtained in the (k+1)th iteration using a noise estimation method based on image patch clustering. c is a noise estimate of the original noisy image obtained using a noise estimation method based on image patch clustering, where c is a constant.

[0016] Preferably, the correlation coefficient (corrcoef) is calculated using the following formula:

[0017]

[0018]

[0019] In the formula, Y represents the original subsurface defect image. This represents the initial denoised image.

[0020] Preferably, the new method for acquiring noisy images is as follows:

[0021]

[0022] In the formula, δ is the proportionality coefficient, set to a value between 0 and 1, Y represents the original subsurface defect image, and k represents the number of iterations. This represents the initial denoised image.

[0023] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an LRA-SVD adaptive detail-preserving denoising algorithm based on improved noise estimation. This algorithm first uses a noise estimation algorithm to estimate the noise in the image to be denoised, reducing noise residue or detail loss caused by blindly selecting noise variance, as well as the high computational cost due to inaccurate denoising results. It calculates the method noise and multiplies it by a certain coefficient, adding it to the denoised image to form a new noisy image. By iteratively regularizing the denoising of the new noisy image, noise information remaining in the large singular value vector space can be further removed, and detail information in the method noise can be restored to the denoised image, avoiding image detail loss and thus improving the algorithm's detail preservation capability. By using an improved noise estimation algorithm to estimate the noise in the new noisy image during the iterative regularization denoising step, the denoising result of each iteration is more accurate, thereby reducing the number of iterations and improving computational efficiency. By using the correlation coefficient between the denoised image and the method noise as the iteration stopping criterion, the denoising effect decreases due to a fixed number of iterations, ensuring that different noisy images can obtain the best denoising effect after different numbers of iterations. Based on the above characteristics, this algorithm reduces the high accuracy requirements of noise estimation algorithms. While achieving adaptive noise reduction to reduce noise residue, it can preserve the detailed information of the image to the greatest extent, greatly improve the computational efficiency, and significantly reduce the computational overhead. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 The attached figure is a flowchart of an LRA-SVD adaptive detail-preserving denoising algorithm based on improved noise estimation provided by the present invention.

[0026] Figure 2 , Figure 3 This paper compares the denoising results obtained by the adaptive detail-preserving denoising algorithm that uses the same noise estimation LRA-SVD as the original image during iterative regularization denoising and the algorithm of this invention.

[0027] Figure 4 , 5 A comparison is made between the noise images obtained by the adaptive detail-preserving denoising algorithm that uses the same noise estimation LRA-SVD as the original image during the iterative regularization denoising process and the algorithm of this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] This invention discloses an LRA-SVD adaptive detail-preserving denoising algorithm based on improved noise estimation, such as... Figure 1 As shown, it includes:

[0030] S1: Obtain subsurface defect images and perform noise estimation on the original subsurface defect images using a noise estimation algorithm based on image patch clustering;

[0031] S2: Denoise the original subsurface defect image using noise estimation and the LRA-SVD algorithm to obtain the initial denoised image.

[0032] S3: Compare the original subsurface defect image Y with the initial denoised image Make a mistake Obtain the initial method noise image MN;

[0033] S4: Calculate the correlation coefficient between the initial denoised image and the initial method noise image, and determine whether the correlation coefficient is greater than the preset threshold. If it is greater, proceed with the S5 iterative regularization denoising process. If it is not greater, obtain the denoised subsurface defect image.

[0034] S5: Iterative Regularized Denoising Process: Multiply the initial method noise image by a scaling factor and add it to the initial denoised image to obtain a new noisy image; use a noise estimation model to estimate the noise in the new noisy image, and use the noise estimate and the LRA-SVD algorithm to denoise the new noisy image to obtain a new denoised image; calculate the difference between the new noisy image and the new denoised image to obtain a new method noise image; calculate the correlation coefficient between the new denoised image and the new method noise image, and determine whether the correlation coefficient is greater than a preset threshold. If it is not greater, obtain the denoised subsurface defect image; if it is greater, perform a new round of S5 iterative regularized denoising process, that is, multiply the new method noise image by a scaling factor and add it to the new noisy image to obtain a combined new noisy image, and use a noise estimation model to estimate the noise in the combined new noisy image, etc.

[0035] The noise estimation model is

[0036] in, This is the noise estimate for the (k+1)th iteration. This represents the noise estimate obtained in the (k+1)th iteration using a noise estimation method based on image patch clustering. c is a noise estimate of the original noisy image obtained using a noise estimation method based on image patch clustering, where c is a constant.

[0037] The second term of the formula uses method noise to compensate for noise estimation errors. As the number of iterations increases, the estimated image becomes closer to the noise-free real image, and the estimated noise also becomes closer to the real noise. Therefore, the second term gradually approaches zero. The noise estimation obtained by the new noise model is closer to the real noise, effectively reducing noise residue, thereby reducing the number of iterations and the time complexity of the algorithm.

[0038] This invention uses the initial LRA-SVD denoised image and the initial method noise as the input image for the first iteration. In each iterative regularization denoising step, a portion of the method noise is added to the input image for the next iteration to improve the detail preservation performance of the denoising algorithm. Since different noisy images converge at different speeds during the iteration process, a stopping criterion is designed to avoid a decrease in denoising effect caused by a fixed number of iterations. This ensures that different noisy images can achieve the best denoising effect after different numbers of iterations.

[0039] In this embodiment, S1 uses a noise estimation algorithm based on image patch clustering to estimate the noise in the original noisy image, avoiding blindly inputting the noise variance, which would cause noise residue or loss of details in the initial denoised image and thus affect the final denoising effect.

[0040] In this embodiment, the correlation coefficient (corrcoef) is calculated using the following formula:

[0041]

[0042]

[0043] In the formula, Y represents the original subsurface defect image. This represents the initial denoised image.

[0044] In this embodiment, the method for acquiring the new noisy image is as follows:

[0045]

[0046] In the formula, δ is the proportionality coefficient, set to a value between 0 and 1, Y represents the original subsurface defect image, and k represents the number of iterations. This represents the initial denoised image.

[0047] To verify the superiority of the method of this invention, a comparison was made between the denoising results obtained by the traditional method of using the same noise estimation method as the original image during iterative regularization denoising and the algorithm of this invention. Figure 2 and Figure 3 As shown, a comparison is made between the noise estimation method using the same noise estimation method as the original image in the traditional iterative regularization noise reduction process and the noise image obtained by the method of this invention. Figure 4 and Figure 5 As shown.

[0048] Table 1 compares the PSNR, SSIM, computation time, and number of iterations of the LRA-SVD algorithm (which uses the same noise estimation as the original image) and the noise reduction results of this invention during the iterative regularization denoising process.

[0049] Table 1

[0050]

[0051] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An LRA-SVD adaptive detail preserving noise reduction algorithm based on improved noise estimation, characterized in that, Comprise: S1: obtain the original subsurface defect image, and estimate the noise of the original subsurface defect image through the noise estimation algorithm based on image block clustering; S2: denoise the original subsurface defect image through the noise estimation value and LRA-SVD algorithm, and obtain the initial denoising image; S3: obtain the initial method noise image by subtracting the initial denoising image from the original subsurface defect image; S4: calculate the correlation coefficient between the initial denoising image and the initial method noise image, and judge whether the correlation coefficient is greater than the preset threshold value, if greater, proceed to S5 iterative regularization denoising process, if not greater, obtain the denoised subsurface defect image; S5: multiply the initial method noise image by the proportionality coefficient and add it to the initial denoising image to obtain a new noisy image; estimate the noise of the new noisy image using the noise estimation model, denoise the new noisy image using the noise estimation value and LRA-SVD algorithm, and obtain a new denoising image; obtain a new method noise image by subtracting the new denoising image from the new noisy image; calculate the correlation coefficient between the new denoising image and the new method noise image, and judge whether the correlation coefficient is greater than the preset threshold value, if greater, proceed to a new round of S5 iterative regularization denoising process, if not greater, obtain the denoised subsurface defect image; The noise estimation model is wherein, is the noise estimate for the k+1 iteration, is the noise estimate for the k+1 iteration obtained using the image block clustering based noise estimation method, is the noise estimate for the original noisy image obtained using the image block clustering based noise estimation method, c is a constant.

2. The LRA-SVD adaptive detail preserving noise reduction algorithm based on improved noise estimation according to claim 1, characterized in that, The correlation coefficient corrcoef calculation formula is: In the formula, Y represents the original subsurface defect image, represents the initial denoised image.

3. The LRA-SVD adaptive detail preserving noise reduction algorithm based on improved noise estimation according to claim 1, characterized in that, The new noisy image acquisition method is: where δ is a proportionality coefficient set to a value between 0 and 1, Y represents the original subsurface defect image, k represents the number of iterations, denotes the initial denoised image.

Images