A sonar image denoising and bridge foundation scour pit feature reconstruction method based on frequency band decomposition

Abstract

The application discloses a sonar image denoising method based on frequency band decomposition, comprising the following steps: S1: obtaining a sonar image; S2: decomposing the sonar image into a low-frequency image and high-frequency sub-band images through NSCT domain decomposition; S3: denoising the low-frequency image to obtain a low-frequency filtering result; S4: performing joint denoising based on edge detection on each high-frequency sub-band image to obtain a high-frequency denoising result; and S5: obtaining an NSCT reconstruction image according to the low-frequency filtering result and the loss high-frequency denoising result; the application can effectively denoise the sonar image, thereby helping to improve the efficiency and accuracy of feature extraction and reconstruction.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for denoising sonar images and reconstructing the features of scour pits in bridge foundations based on frequency band decomposition. Background Technology

[0002] As a crucial component of transportation infrastructure, the safety of bridges is paramount. Among the various factors threatening bridge safety, the problem of bridge foundation scour pits cannot be ignored. Bridge foundations, constantly exposed to water in environments such as rivers and lakes, are continuously eroded by water flow, causing the surrounding soil to gradually disappear and form scour pits. This phenomenon not only significantly weakens the load-bearing capacity of the bridge foundation but, in severe cases, can even lead to catastrophic consequences such as pier tilting and collapse, posing a significant threat to the safe operation and service life of the bridge.

[0003] Traditional methods for monitoring bridge foundation scour have many drawbacks. Sonar technology, an electronic device that utilizes the underwater propagation characteristics of sound waves to perform underwater detection and communication tasks through electroacoustic conversion and information processing, offers a new solution for underwater bridge inspection. It can be divided into active and passive types. In the field of bridge inspection, sonar technology can emit sound waves and receive reflected signals to detect the morphology and damage of underwater structures. Single-beam sonar is mainly used to acquire vertical cross-sectional images of underwater structures, while multi-beam sonar can generate three-dimensional images of underwater structures, providing a more comprehensive inspection perspective.

[0004] However, there are still some challenges in using sonar technology to detect scour pits in bridge foundations. For example, the raw sonar data files need to undergo complex processing before they can be used for analysis; sonar images are easily affected by noise, which affects the detection accuracy; there are gaps in the stitched sonar strip images that need to be filled; and the accurate extraction of scour pit features and reconstruction of their three-dimensional models all require further optimization and improvement.

[0005] Therefore, how to quickly and accurately extract and reconstruct the features of scour pits using sonar technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method for denoising sonar images and reconstructing bridge foundation scour pit features based on frequency band decomposition, which can denoise sonar images, thereby helping to improve the efficiency and accuracy of feature extraction and reconstruction.

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

[0008] A sonar image denoising method based on frequency band decomposition includes the following steps:

[0009] S1: Acquire sonar images;

[0010] S2: The sonar image is decomposed into a low-frequency image and a high-frequency sub-band image by NSCT domain decomposition;

[0011] S3: Denoise the low-frequency image to obtain the low-frequency filtering result;

[0012] S4: Perform joint denoising based on edge detection on each of the high-frequency sub-band images to obtain the high-frequency denoising result;

[0013] S5: Based on the low-frequency filtering results and the loss of high-frequency denoising results, the NSCT reconstructed image is obtained.

[0014] Preferably, the denoising method for the low-frequency image in step S3 employs a three-dimensional block matching filter algorithm.

[0015] Preferably, when using the three-dimensional block matching filtering algorithm for denoising, a hybrid distance of Euclidean distance and structural similarity is used to measure the similarity between each image block and the reference block.

[0016] Preferably, the formula for calculating the mixing distance is:

[0017]

[0018] Where SSIM stands for structural similarity. For the mixing distance, μ x μ y Let σ represent the mean of the reference block P and the candidate block Q, respectively. x σ y Let σ represent the variances of the reference block P and the candidate block Q, respectively. xy Let c1 represent the covariance between the reference block P and the candidate block Q, where c1 = (k1L). 2 c2 = (k2L) 2 To avoid denominators being 0, k1 and k2 are assumed to be 0.01 and 0.03 respectively. L represents the range of image pixel values, and α is the adaptive weighting coefficient.

[0019] Preferably, the joint denoising based on edge detection in S4 includes the following steps:

[0020] S41: Nonmaximum suppression in any direction;

[0021] Pixels are discrete points. The gradient of a point in any direction is calculated by interpolation. If the gradient of a pixel is a local maximum along the positive or negative gradient direction, the pixel is retained; otherwise, it is suppressed.

[0022] Where F represents the current pixel, and α represents the direction of gradient calculation, then the gradient linear interpolation of pixels F1 and F2 in its positive and negative directions is as follows:

[0023]

[0024] In the formula, tanα represents the ratio of the gradient magnitudes of pixel F in the Y and X directions;

[0025] S42: Double threshold OTSU;

[0026] Define a three-class classification problem: Our goal is to find two thresholds that classify gradient magnitudes into three categories: background, weak edges, and strong edges.

[0027] S43: Maximize the inter-class variance in a three-class classification;

[0028] According to the OTSU criterion, the image is divided into three classes. The threshold that maximizes the inter-class variance is the optimal threshold. Let the overall mean be μ. T Images of the background, weak edge, and strong edge classes are denoted by C0, C1, and C2, respectively. Their class probabilities and means are calculated, and the expression for the inter-class variance is as follows:

[0029]

[0030] Where p(i) is the probability of gradient magnitude i;

[0031] Background class C0:

[0032]

[0033] Weak edge class C1:

[0034]

[0035] Strong edge class C2:

[0036]

[0037] Between-class variance:

[0038]

[0039] By traversing all possible combinations of t1 and t2, the inter-class variance is made... The largest value is the optimal threshold;

[0040] S44: High-frequency subband noise reduction processing;

[0041] The low-frequency filtering result is subjected to Canny operator edge detection. Then, the edge detection result is convolved with each high-frequency sub-band. This process removes all noise outside the edges and textures in the image, effectively removing noise in the high-frequency sub-band while retaining more edge information. The specific steps are as follows: Gaussian filtering is used for image preprocessing, the gradient and size of each pixel are pre-calculated using the Sobel operator, non-maximum suppression is used for edge thinning, an improved double threshold optimization algorithm is applied to obtain the optimal threshold, and a high and low double threshold detection algorithm is used to find the image edges.

[0042] A method for reconstructing the morphological features of bridge foundation scour pits based on sonar includes the following steps:

[0043] Obtain raw sonar data of the bridge foundation scour pit;

[0044] An initial sonar image is drawn based on the original sonar data;

[0045] The initial sonar image is processed using any one of the denoising methods described in claims 1-5;

[0046] Feature reconstruction is performed based on the denoised image.

[0047] Preferably, the feature reconstruction step includes: stitching together the denoised sonar strip image according to the multi-beam survey line, and then performing interpolation processing on the stitched image to fill in the missing values.

[0048] Preferably, the interpolation process includes:

[0049] Step 1: Select any element from the downsampled image matrix and interpolate the elements whose amplitude values ​​are equal to zero.

[0050] Step 2: Find element B in the reference image matrix that corresponds to the element to be interpolated in the downsampled image matrix;

[0051] Step 3: Centered on element B, set the window size to 3×3 elements and add a window to the reference image matrix;

[0052] Step 4: Calculate the amplitude value of the element to be interpolated after the first interpolation according to the following formula: f(A)=Σω*H(2i+k,2j+k), with the lower limit of the summation being k=-L and the upper limit being L; where f(A) represents the amplitude value of the element to be interpolated after the first interpolation, A represents the element to be interpolated, Σ represents the summation operation, L=win / 2 represents the constraint factor, win represents the window size, k represents an integer in the interval [-L,L], ω represents the weighting coefficient, ω=1 / num, num represents the total number of non-zero elements in the window, * represents the multiplication operation, H represents the reference image matrix, i and j represent the x and y coordinates of the corresponding coordinate positions of the elements in the reference image matrix H in the downsampled image matrix, i=1,2,...,M / 2, j=1,2,...,N / 2, M and N represent the number of rows and columns of the two-dimensional image matrix y, respectively;

[0053] Step 5: Determine whether all non-zero elements in the downsampled image matrix have been selected. If so, the first stage of interpolation of the downsampled image matrix is ​​completed; otherwise, proceed to step 1.

[0054] Preferably, after the interpolation process, the Canny operator is used to perform edge detection on the scour pit to obtain the edge contour of the scour pit.

[0055] Preferably, the feature reconstruction step includes:

[0056] Utilizing the principle of grayscale mapping, the imaging principle of sonar images is layered coloring. Different grayscale values ​​represent different echo intensities, and the echo intensity reflects the distance from the sonar device to the bottom of the water. Therefore, in a sonar image, if the depth of a certain point in a scour pit is known, and its corresponding grayscale value is used, a linear relationship between the two can be obtained, thereby determining the depth of the entire scour pit.

[0057] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for sonar image denoising and bridge foundation scour pit feature reconstruction based on frequency band decomposition. The sonar image is divided into low-frequency and high-frequency sub-bands by NSCT domain decomposition. The low-frequency sub-band adopts a three-dimensional block matching filtering algorithm that integrates Euclidean distance and structural similarity (SSIM). The high-frequency sub-band is jointly denoised by improving Canny edge detection (including arbitrary direction non-maximum suppression and double-threshold OTSU three-class classification optimization), which can effectively suppress speckle noise and Gaussian noise while preserving edge details. After denoising, the holes are filled by multi-beam survey line stitching and weighted neighborhood interpolation. Combined with the gray value mapping principle, the three-dimensional depth reconstruction of the scour pit is realized, which can accurately identify the location of the scour pit and obtain depth data, thereby improving the accuracy and completeness of sonar image detection. Attached Figure Description

[0058] 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.

[0059] Figure 1 This is a schematic diagram of a sonar image denoising method based on frequency band decomposition provided in an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the improved BM3D algorithm provided in an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of the gradient direction of a pixel in an embodiment of the present invention.

[0062] Figure 4 This is a schematic diagram of the sonar raw data file parsing process in an embodiment of the present invention. Detailed Implementation

[0063] 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.

[0064] Example 1

[0065] like Figure 1 This invention discloses a sonar image denoising method based on frequency band decomposition, comprising the following steps:

[0066] S1: Acquire sonar images;

[0067] S2: The sonar image is decomposed into a low-frequency image and a high-frequency sub-band image by NSCT domain decomposition;

[0068] S3: Denoise the low-frequency image to obtain the low-frequency filtering result;

[0069] S4: Perform joint denoising based on edge detection on each of the high-frequency sub-band images to obtain the high-frequency denoising result;

[0070] S5: Based on the low-frequency filtering results and the loss of high-frequency denoising results, the NSCT reconstructed image is obtained.

[0071] To further implement the above technical solution, the denoising method for the low-frequency image in S3 adopts a three-dimensional block matching filter algorithm.

[0072] Furthermore, when using the aforementioned 3D block matching filtering algorithm for denoising, a hybrid distance measure combining Euclidean distance and structural similarity is employed to measure the similarity between each image block and the reference block.

[0073] In this embodiment, the multi-directionality of SCT can better capture edges and textures in sonar images, while BM3D can effectively remove noise while preserving structure during low-frequency denoising, which is more effective than traditional wavelet thresholding. Non-subsampled contour wave transform (NSCT) captures different frequency features through multi-scale decomposition and separates edge directions through multi-directional decomposition, accurately separating structural information and detailed noise in sonar images. It avoids the directional limitations of traditional wavelet transform and is more suitable for the complex geometric structure of sonar images. The non-subsampling characteristics of NSCT avoid the pseudo-Gibbs effect. Combined with the sparse representation of BM3D, it effectively reduces ringing artifacts in the reconstructed image and improves the visual effect.

[0074] After NSCT decomposition, noisy images retain the main energy information in the low-frequency subband, but also contain low-frequency noise mixed with the main structure (such as echo scattering noise and sensor slowly varying noise, which is usually considered Gaussian white noise). However, speckle noise (high-frequency multiplicative noise) is relatively weak in this frequency band. The BM3D algorithm has the advantages of preserving image details and effectively removing Gaussian noise; however, it can cause ringing at the image edges when processing high-contrast images. BM3D utilizes block matching and collaborative filtering to group similar image blocks and perform joint filtering. While removing Gaussian noise, it preserves key features such as texture and contour of the sonar image, which is superior to traditional low-frequency mean filtering or soft thresholding methods. Speckle noise exhibits high-frequency characteristics in the frequency domain. Due to the unique nature of its block matching grouping, the BM3D algorithm's ability to denoise such edge details is significantly reduced. Therefore, joint denoising using Canny operator edge detection for each high-frequency sub-band image can remove noise while effectively preventing false negatives of edge features. This paper's algorithm integrates the advantages of the frequency and spatial domains. NSCT provides multi-resolution analysis in the frequency domain, while BM3D leverages non-local similarity in the spatial domain, forming a complementary approach that balances global structural stability and local detail restoration. Furthermore, the BM3D algorithm employs a hybrid distance method with improved similarity metrics for block matching grouping.

[0075] Specifically, such as Figure 2The block matching grouping process of the BM3D algorithm aims to find the image block most similar to the reference block. The parameter for evaluating the similarity between blocks is distance. The original algorithm chose Euclidean distance, but its shortcoming is that it only uses the statistical characteristics of the image when making distance judgments for block matching, ignoring the structural information of the image. Since sonar images contain many structurally similar regions, a hybrid distance combining Euclidean distance and structural similarity (SSIM) is adopted, utilizing both pixel-level differences and structural similarity.

[0076]

[0077] Where μ x μ y Let σ represent the mean of the reference block P and the candidate block Q, respectively. x σ y Let σ represent the variances of the reference block P and the candidate block Q, respectively. xy Let c1 represent the covariance between the reference block P and the candidate block Q, where c1 = (k1L). 2 c2 = (k2L) 2 This represents two constants to avoid a denominator of 0. k1 and k2 are defaulted to 0.01 and 0.03 respectively. L represents the range of image pixel values.

[0078] The structural similarity coefficient (SSIM) ranges from 0 to 1. A higher SSIM value indicates a higher structural similarity between the two images. By introducing an adaptive weight coefficient α, and combining Euclidean distance with the structural similarity coefficient, an improved similarity measurement method is obtained:

[0079]

[0080] Based on the above improvements, the complete denoising process for low-frequency images includes:

[0081] 1. Block matching grouping.

[0082] Since the main noise in noisy images is speckle noise, which is multiplicative noise, a logarithmic transformation must be performed on the image before block matching. Let P represent the current reference image block. Within a neighborhood of size n×n centered at P, search for image blocks similar to the current reference block P in the original noisy image; these are denoted as candidate blocks Q. The set of similar image blocks is simply defined as:

[0083] Where τ match This is the maximum error threshold for determining two similar blocks, and it is generally set based on experience. Q It is a point within the neighborhood. It is x Q Candidate blocks centered on the target. This is a reference image patch. 'd' represents the distance between the two image patches; common methods for representing distance include Euclidean distance and normalized distance. The original algorithm uses Euclidean distance.

[0084]

[0085] Where Υ′ represents the hard thresholding operation, and the corresponding threshold is λ. 2D σ is typically two standard deviations, Γ 2D For the corresponding separable two-dimensional orthogonal transformation, N hard This represents the number of pixels in the matching block. In this algorithm, a mixed distance is used to match similar blocks.

[0086]

[0087] After obtaining the set of similar blocks, we stack all the similar blocks together to get a shape of... 3D array To speed up processing, only [the following information is retained] in the three-dimensional array. The k image blocks that are closest to the reference block. The image blocks in the image will be based on their relationship with The distances are sorted to easily select the best k similar image patches.

[0088] 2. Collaborative hard threshold filtering.

[0089] After obtaining the three-dimensional array corresponding to the reference block, three-dimensional cooperative transformation and filtering can be performed, the mathematical expression of which is:

[0090] Where Γ 3D This is the corresponding three-dimensional orthogonal transformation, and it is usually a separable transformation because the result of the two-dimensional transformation already exists when calculating the distance, and it can be reused. Therefore, Γ 3D Represented as the first two dimensions Γ 2D and the 3rd dimension Γ 1D The combination of these elements is typically used. Generally, the first two dimensions are transformed using Fourier transform-based DCT or DST transforms, while the third dimension is transformed using wavelet transform-based Haar transforms. Υ represents another hard thresholding operation, with a corresponding threshold of λ. 3D σ is typically 2 to 3 standard deviations. Includes block This represents the result obtained after collaborative filtering, with the corresponding reference block coordinates being x. P .

[0091] 3. Aggregation.

[0092] The weights are assigned based on the number of non-zero coefficients remaining after collaborative filtering. Assuming all pixels within the 3D combination are independent, let the number of coefficients remaining after the hard thresholding operation be... Then there is

[0093] Where σ 2 Represents the noise variance, representing The number of elements after collaborative filtering of the 3D matrix. Finally, by weighted summation of all reference blocks and corresponding similar blocks to a given pixel, its basic estimate can be obtained, i.e.:

[0094]

[0095] We assume that each block has been padded with zeros to match the original. Figure 1 The size of the sample, and through To determine whether a pixel x is in block x Q The above shows that a pixel can belong to different reference blocks, or to different similar blocks of the same reference block, and the same block can also be a similar block of different reference blocks. Through this overcomplete collaborative filtering method, we can remove noise energy as much as possible while preserving image details well, and maintain low computational complexity.

[0096] 4. Final estimate.

[0097] The steps in the final estimation stage are basically the same as those in the basic estimation stage, and will not be repeated here. The slight difference lies in the collaborative filtering stage. The basic estimation uses hard thresholding, while the final estimation uses Wiener filtering, based on the output of the basic estimation. To design a Wiener filter for three-dimensional transformations. The Wiener filter used is...

[0098]

[0099] The mathematical expression for synergistic Wiener filtration is:

[0100]

[0101] Since most of the noise is removed after the basic estimation stage, the output of the basic estimation stage is used as the input image for the final estimation in order to achieve better denoising effect and improve the accuracy of block matching.

[0102] 5. Experimental verification of the algorithm was conducted. Two sonar images containing a crashed aircraft and a sunken ship were selected as experimental objects. Speckle noise with a variance of 0.1 was added. The experimental results show that the BM3D algorithm with improved similarity measure has improved the denoising ability of the image, and the peak signal-to-noise ratio has increased by 2.1 and 2.8 respectively, verifying the effectiveness of the algorithm.

[0103] To further implement the above technical solution, the joint denoising based on edge detection in S4 includes the following steps:

[0104] S41: Nonmaximum suppression in any direction;

[0105] Pixels are discrete points. The gradient of a point in any direction is calculated by interpolation. If the gradient of a pixel is a local maximum along the positive or negative gradient direction, the pixel is retained; otherwise, it is suppressed.

[0106] like Figure 3 In the diagram, F represents the current pixel, and α represents the direction of gradient calculation. The linear interpolation of the gradients at pixels F1 and F2 in the positive and negative directions is as follows:

[0107]

[0108]

[0109] In the formula, tanα represents the ratio of the gradient magnitudes of pixel F in the Y and X directions;

[0110] S42: Double threshold OTSU;

[0111] Define a three-class classification problem: Our goal is to find two thresholds that classify gradient magnitudes into three categories: background, weak edges, and strong edges.

[0112] S43: Maximize the inter-class variance in a three-class classification;

[0113] According to the OTSU criterion, the image is divided into three classes. The threshold that maximizes the inter-class variance is the optimal threshold. Let the overall mean be μ. T Images of the background, weak edge, and strong edge classes are denoted by C0, C1, and C2, respectively. Their class probabilities and means are calculated, and the expression for the inter-class variance is as follows:

[0114]

[0115] Where p(i) is the probability of gradient magnitude i;

[0116] Background class C0:

[0117]

[0118] Weak edge class C1:

[0119]

[0120] Strong edge class C2:

[0121]

[0122] Between-class variance:

[0123]

[0124] By traversing all possible combinations of t1 and t2, the inter-class variance is made... The largest value is the optimal threshold;

[0125] S44: High-frequency subband noise reduction processing;

[0126] The low-frequency filtering result is subjected to Canny operator edge detection. Then, the edge detection result is convolved with each high-frequency sub-band. This process removes all noise outside the edges and textures in the image, effectively removing noise in the high-frequency sub-band while retaining more edge information. The specific steps are as follows: Gaussian filtering is used for image preprocessing, the gradient and size of each pixel are pre-calculated using the Sobel operator, non-maximum suppression is used for edge thinning, an improved double threshold optimization algorithm is applied to obtain the optimal threshold, and a high and low double threshold detection algorithm is used to find the image edges.

[0127] Example 2

[0128] Based on the same inventive concept, this invention discloses a method for reconstructing the morphological features of bridge foundation scour pits based on sonar, comprising the following steps:

[0129] Obtain raw sonar data of the bridge foundation scour pit;

[0130] An initial sonar image is drawn based on the original sonar data;

[0131] The initial sonar image is processed using the denoising method described in Example 1;

[0132] Feature reconstruction is performed based on the denoised image.

[0133] In this embodiment, sonar data of the underwater riverbed topography near the bridge foundation is collected. After obtaining the raw data, it is parsed to obtain the raw sonar image, and noise reduction is performed using a three-dimensional block matching filter algorithm based on the NSCT domain. The process of parsing the raw sonar data file is as follows: Figure 4 As shown.

[0134] To verify the denoising performance of the proposed algorithm, an existing side-scan sonar image dataset was selected. The experimental environment consisted of MATLAB R2019b, a computer with 16GB of memory, and an Intel Core i7-9750H processor with a clock speed of 2.60GHz. For ease of comparison and analysis, the results were compared with those obtained using wavelet thresholding, the BM3D algorithm, and the NSCT transform. Wavelet thresholding was performed using two-dimensional wavelets for decomposition and reconstruction. The parameters for the BM3D and NSCT algorithms were the same as those in this paper. Three real sonar images were selected, representing three different scenes: shipwreck, crashed aircraft, and underwater terrain. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the images after denoising using different methods were compared, as shown in Tables 1 and 2.

[0135] Table 1 Peak signal-to-noise ratio values ​​of denoised images using several algorithms

[0136]

[0137]

[0138] Table 6-2 Image structural similarity values ​​of several algorithms for denoising

[0139]

[0140] It can be seen that the NSCT transform can achieve optimal sparse distribution approximation of the image. After decomposing the noisy image, a low-frequency image and multiple high-frequency sub-band images are obtained. For the Gaussian noise that may remain in the low-frequency image, an improved BM3D algorithm is used for filtering. BM3D leverages its advantage of spatial non-local similarity to form a complementary effect, balancing global structural stability and local detail restoration. Simultaneously, utilizing the special structural characteristics of sonar images, the block matching distance of the similarity metric is improved, significantly enhancing the grouping effect. Furthermore, joint denoising based on Canny operator edge detection is employed, fully utilizing the results of low-frequency filtering and integrating the correlation between different scales after NSCT decomposition to jointly denoise the high-frequency sub-band images. This results in better preservation of edge and texture information and removal of most of the noise in the high frequencies. Experimental results show that this algorithm, while removing noise, preserves edge details as much as possible, achieving both overall and local enhancement, improving the PSNR value of the restored image. The algorithm's subjective visual effect and objective evaluation criteria are both better than the comparison algorithms.

[0141] The noise-reduced sonar strip images are stitched together according to the multi-beam survey lines, and then interpolation is performed on the stitched images to fill in the missing values.

[0142] Image interpolation processing includes the following steps:

[0143] Step 1: Select any element from the downsampled image matrix and interpolate the elements whose amplitude values ​​are equal to zero.

[0144] Step 2: Find element B in the reference image matrix that corresponds to the element to be interpolated in the downsampled image matrix;

[0145] Step 3: Centered on element B, set the window size to 3×3 elements and add a window to the reference image matrix;

[0146] Step 4: Calculate the amplitude value of the element to be interpolated after the first interpolation according to the following formula: f(A)=Σω*H(2i+k,2j+k), with the lower limit of the summation being k=-L and the upper limit being L. Where f(A) represents the amplitude value of the element to be interpolated after the first interpolation, A represents the element to be interpolated, Σ represents the summation operation, L=win / 2 represents the constraint factor, win represents the window size, k represents an integer in the interval [-L,L], ω represents the weighting coefficient, ω=1 / num, num represents the total number of non-zero elements in the window, * represents the multiplication operation, H represents the reference image matrix, i and j represent the x and y coordinates of the corresponding coordinate positions of the elements in the reference image matrix H in the downsampled image matrix, i=1,2,...,M / 2, j=1,2,...,N / 2, M and N represent the number of rows and columns of the two-dimensional image matrix y, respectively;

[0147] Step 5: Determine whether all non-zero elements in the downsampled image matrix have been selected. If so, the first stage of interpolation of the downsampled image matrix is ​​completed; otherwise, proceed to step 1.

[0148] A method for denoising sonar images and reconstructing bridge foundation scour pit features based on frequency band decomposition is characterized in that, after interpolation, the Canny operator is used to perform edge detection on the scour pit to obtain the edge contour of the scour pit.

[0149] A method for denoising sonar images and reconstructing bridge foundation scour pit features based on frequency band decomposition is characterized by utilizing the principle of gray value mapping. The imaging principle of sonar images is layered coloring, where different gray values ​​represent different echo intensities. The echo intensity reflects the distance from the sonar device to the bottom of the water. Therefore, in a sonar image, if the depth of a certain point in the scour pit is known, and its corresponding gray value is used, a linear relationship between the two can be obtained, thereby determining the depth of the entire scour pit.

[0150] 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.

[0151] 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. A sonar image denoising method based on frequency band decomposition, characterized in that, Includes the following steps: S1: Acquire sonar images; S2: The sonar image is decomposed into a low-frequency image and a high-frequency sub-band image by NSCT domain decomposition; S3: Denoise the low-frequency image to obtain a low-frequency filtering result; the denoising method for the low-frequency image adopts a three-dimensional block matching filter algorithm. When using the three-dimensional block matching filter algorithm for denoising, a hybrid distance of Euclidean distance and structural similarity is used to measure the similarity between each image block and the reference block; the formula for calculating the hybrid distance is: ； ； Where SSIM stands for structural similarity. For mixed distance, , Let P and Q represent the mean values ​​of the reference block and the candidate block, respectively. , Let P and Q represent the variances of the reference block and the candidate block, respectively. This represents the covariance between the reference block P and the candidate block Q. , To represent two constants, avoid having a denominator of 0. and The default values ​​are 0.01 and 0.03 respectively, where L represents the range of image pixel values. These are adaptive weighting coefficients; S4: Perform joint denoising based on edge detection on each of the high-frequency sub-band images to obtain high-frequency denoising results; the steps include: S41: High-frequency subband noise reduction processing; The low-frequency filtering result is subjected to Canny operator edge detection. Then, the edge detection result is convolved with each high-frequency sub-band. This process removes all noise outside the edges and textures in the image, effectively removing noise in the high-frequency sub-band while retaining more edge information. The specific steps are as follows: Gaussian filtering is used for image preprocessing. The gradient and size of each pixel are pre-calculated using the Sobel operator. Non-maximum suppression is used for edge thinning. An improved double threshold optimization algorithm is applied to obtain the optimal threshold. The high and low double threshold detection algorithm is used to find the image edges. S42: Nonmaximum suppression in any direction; Pixels are discrete points. The gradient of a point in any direction is calculated by interpolation. If the gradient of a pixel is a local maximum along the positive or negative gradient direction, the pixel is retained; otherwise, it is suppressed. Where F represents the current pixel, α represents the direction of gradient calculation, and the pixels in its positive and negative directions are... and Gradient linear interpolation: ； ； In the formula This represents the ratio of the gradient magnitudes of pixel F in the Y and X directions. S43: Double threshold OTSU; Define a three-class classification problem: Our goal is to find two thresholds that classify gradient magnitudes into three categories: background, weak edges, and strong edges. S44: Maximize the inter-class variance in a three-class classification; According to the OTSU criterion, images are divided into three classes. The threshold that maximizes the inter-class variance is the optimal threshold. Let the overall mean be... Images of background, weak edge, and strong edge classes are respectively... , and Let's represent the classes, calculate their probabilities and means respectively, and the expression for calculating the between-class variance is: ； in It is the probability of the gradient magnitude i; Background Class : ； weak edge class : ； Strong edge class : ； Between-class variance: ； By iterating through all possible... and Combinations make the inter-class variance The largest value is the optimal threshold; S5: Based on the low-frequency filtering results and the loss of high-frequency denoising results, the NSCT reconstructed image is obtained.

2. A method for reconstructing the morphological features of bridge foundation scour pits based on sonar, characterized in that, Includes the following steps: Obtain raw sonar data of the bridge foundation scour pit; An initial sonar image is drawn based on the original sonar data; The initial sonar image is processed using the denoising method described in claim 1; Feature reconstruction based on the denoised image includes the following steps: stitching the denoised sonar strip image according to the multi-beam survey line, then performing interpolation on the stitched image to fill in the missing values; after the interpolation, using the Canny operator to perform edge detection on the scour pit to obtain the edge contour of the scour pit. Utilizing the principle of grayscale mapping, the imaging principle of sonar images is layered coloring. Different grayscale values ​​represent different echo intensities, and the echo intensity reflects the distance from the sonar device to the bottom of the water. Therefore, in a sonar image, if the depth of a certain point in a scour pit is known, and its corresponding grayscale value is used, a linear relationship between the two can be obtained, thereby determining the depth of the entire scour pit.

3. The method for reconstructing the morphological features of bridge foundation scour pits based on sonar according to claim 2, characterized in that, The interpolation process includes the following steps: Step 1: Select any element from the downsampled image matrix and interpolate the elements whose amplitude values ​​are equal to zero. Step 2: Find element B in the reference image matrix that corresponds to the element to be interpolated in the downsampled image matrix; Step 3: Centered on element B, set the window size to 3×3 elements and add a window to the reference image matrix; Step 4: Calculate the amplitude value of the element to be interpolated after the first interpolation according to the following formula: f(A)=Σω*H(2i+k,2j+k), with the lower limit of the summation being k=-L and the upper limit being L; where f(A) represents the amplitude value of the element to be interpolated after the first interpolation, A represents the element to be interpolated, Σ represents the summation operation, L=win / 2 represents the constraint factor, win represents the window size, k represents an integer in the interval [-L,L], ω represents the weighting coefficient, ω=1 / num, num represents the total number of non-zero elements in the window, * represents the multiplication operation, H represents the reference image matrix, i and j represent the x and y coordinates of the corresponding coordinate positions of the elements in the reference image matrix H in the downsampled image matrix, i=1,2,...,M / 2, j=1,2,...,N / 2, M and N represent the number of rows and columns of the two-dimensional image matrix y, respectively; Step 5: Determine whether all non-zero elements in the downsampled image matrix have been selected. If so, the first stage of interpolation of the downsampled image matrix is ​​completed; otherwise, proceed to step 1.

Images