A concrete flaw detection ultrasonic image speckle noise suppression method and system based on multi-modal signal processing

Abstract

The application relates to the technical field of concrete ultrasonic flaw detection image processing, and discloses a concrete flaw detection ultrasonic image speckle noise suppression method and system based on multi-modal signal processing, which is characterized in that the original ultrasonic signal is subjected to adaptive white noise injection and EMD decomposition, non-aliasing IMF components are obtained through modal alignment and averaging, and high-frequency noise dominant components and medium-low-frequency effective signal components are divided according to the kurtosis criterion; subsequently, the high-frequency components are subjected to VMD optimization, the frequency band is accurately divided by constructing a variational constraint model and iteratively solving; the improved adaptive wavelet threshold function and the hierarchical hybrid threshold strategy are adopted for the medium-low-frequency components, so that the noise is accurately filtered out and the effective signal is retained; finally, the de-noised components of various frequency bands are fused to generate a one-dimensional time domain signal, the residual noise points are removed through median filtering after being restored into a two-dimensional image, and the de-noised enhanced image is output. The application realizes accurate separation of the noise and the effective signal, and completely retains the fine defect features in the concrete.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology for ultrasonic flaw detection of concrete, specifically relating to a method and system for suppressing speckle noise in ultrasonic images of ultrasonic flaw detection of concrete based on multimodal signal processing, which is applied to the suppression of speckle noise and optimization of image quality in ultrasonic monitoring images. Background Technology

[0002] Concrete structures are the core load-bearing components of major engineering projects such as bridges, tunnels, and high-rise buildings. Their internal defects directly affect the safe service life of the engineering structure, thus requiring accurate identification of defects through non-destructive testing technology. Ultrasonic testing technology has become the mainstream technique for detecting internal defects in concrete due to its advantages of being non-destructive, portable, and low-cost. However, multiplicative speckle noise is commonly found in ultrasonic images, leading to blurred defect boundaries and reduced identification accuracy. This is the core technical challenge in the current field of ultrasonic testing for concrete.

[0003] To address the issue of speckle noise suppression in ultrasound images, existing technologies primarily employ the following approaches: First, the original ultrasound signal is decomposed into multiple intrinsic mode function (IMF) components using EMD (Enhanced Motion Discharge), and noise-dominant high-frequency IMF components are screened and removed to achieve noise suppression. Second, a fixed wavelet basis (such as db4) is used to perform multi-scale decomposition of the signal, and the wavelet coefficients are truncated using preset hard / soft threshold functions before reconstructing the signal to remove noise. Third, the ultrasound signal is converted to the frequency domain, and high-frequency / specific frequency band components corresponding to noise are filtered out using bandpass / lowpass filters, and then inversely transformed back to the time domain to obtain the denoised signal.

[0004] The above-mentioned methods have significant limitations in practical engineering flaw detection scenarios: single-mode decomposition is prone to signal mode mixing, making it difficult to distinguish between noise signals and effective defect signals; it cannot adapt to the nonlinear variation characteristics of ultrasonic signals, and its preservation effect on fine concrete structures of 0.3 mm and below is poor; it cannot be adapted to ultrasonic testing probes in the full frequency band of 5 to 15 MHz; and its adaptability in actual engineering testing scenarios is insufficient. These are the shortcomings of existing technologies.

[0005] In view of this, it is very necessary to provide a multimodal signal processing method that combines fully adaptive noise ensemble empirical mode decomposition (CEEMDAN), variational mode decomposition (VMD), and wavelet threshold denoising to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the deficiencies of the prior art by providing a method and system for suppressing speckle noise in ultrasonic images used for concrete flaw detection based on multimodal signal processing, thereby solving the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for suppressing speckle noise in ultrasonic images used for concrete flaw detection based on multimodal signal processing, specifically comprising the following steps: Step S1, the CEEMDAN preprocessing step, in which: For the original ultrasonic signal from concrete flaw detection, adaptive attenuated white noise adapted to the near-field noise characteristics of ultrasound is injected and EMD decomposition is performed. After mode alignment and averaging, 8th-order IMF components without mode mixing are obtained. Then, the components are classified by improving the kurtosis criterion, and the high-frequency noise-dominant components (IMF1-3) and the mid-to-low frequency effective signal components (IMF4-8) are identified, providing accurate component classification basis for subsequent layered differential denoising. Step S2, VMD optimization steps The high-frequency noise dominant component obtained in step S1 is decomposed in a secondary manner. By constructing a variational constraint optimization model and completing iterative solution, the frequency band is accurately divided. At the same time, the optimal parameters verified by engineering flaw detection data are combined to ensure the accuracy and engineering adaptability of the high-frequency component decomposition, and provide high-quality high-frequency processing components for subsequent multimodal signal reconstruction. Step S3, the wavelet joint denoising step, in which: Targeted denoising processing is performed on the mid-to-low frequency IMF components obtained from CEEMDAN preprocessing. By using an improved adaptive wavelet threshold function and a hierarchical hybrid thresholding strategy, the noise in the mid-to-low frequency region is accurately suppressed while the effective signal is completely preserved, providing high-quality mid-to-low frequency processed components for the final multimodal signal reconstruction. Step S4, the multimodal signal reconstruction and post-processing step, in which: The high-frequency denoised component integrated in step S2 is fused with the denoised mid-low frequency IMF component obtained in step S3 across the entire frequency band to generate a one-dimensional complete time-domain denoised signal. The signal is restored to a two-dimensional matrix of the original image dimension, and residual isolated noise points are eliminated by median filtering to output a denoised and enhanced concrete ultrasonic image that can be used for engineering flaw detection.

[0008] Step S1, the CEEMDAN preprocessing step, specifically includes the following steps: Step S101, the noise injection and decomposition step, in which: Based on the input raw ultrasound signal, the noise standard deviation of CEEMDAN preprocessing is calculated using the following formula:

[0009] in, Let L2 be the norm of the input signal; Noise decay time constant T represents the total duration of the input ultrasound signal; the total number of IMF components decomposed in the CEEMDAN preprocessing step is set to 8, and the maximum number of iterations is 100. Simultaneously, based on the characteristics of the original ultrasound signal, the attenuation coefficient of the adaptive noise is calculated using the following formula:

[0010] in, is the attenuation coefficient for adaptive noise, and 0.2 is the fundamental noise intensity coefficient. Let x be the standard deviation of the original signal x, and t be the time-domain sampling point of the signal.

[0011] Furthermore, noise is injected sequentially and EMD decomposition is performed. In each loop, attenuated white noise is superimposed on the original signal to obtain a noisy signal. Then, conventional EMD decomposition is performed on the noisy signal to obtain a set of temporary IMF components. This process is repeated 100 times to obtain 100 sets of IMF components corresponding to different noisy signals. The relevant calculation formula is as follows:

[0012]

[0013] in, The noisy ultrasonic signal in the nth iteration; A function to generate standard normally distributed white noise with the same dimension as the original signal x; It is the empirical mode decomposition function; This is the set of temporary IMF components obtained from the nth noisy signal decomposition. This is the attenuation coefficient for adaptive noise.

[0014] Furthermore, after completing 100 EMD decompositions, modal alignment and averaging of IMF components are performed on all temporary IMF components. This involves matching IMF components of the same level among the 100 components in descending frequency order. Then, the average value of the IMF components at the same level is calculated, ultimately yielding a set of stable, mode-free CEEMDAN decomposed IMF components. This operation effectively cancels out interference from multiple injections of white noise, preserving the true characteristics of the original signal. The specific formula is as follows:

[0015] in, This represents the final k-th order IMF component; Let k be the temporary IMF component of order k obtained from the nth iteration decomposition; k is the order index of the IMF component.

[0016] Furthermore, during the screening process of a single EMD decomposition, the continuous screening standard deviation (SD) of two adjacent screening results needs to be monitored in real time. When SD < 0.3, it is determined that the current IMF component has been completely decomposed, and the screening iteration of that component is immediately terminated, and the decomposition of the next IMF component is carried out until all components are extracted. This stopping condition can avoid signal feature distortion caused by over-screening.

[0017] Step S102, the noise-driven IMF identification step, in which: For each IMF component obtained from CEEMDAN decomposition, calculate its kurtosis coefficient using the following formula:

[0018] in, is the kurtosis coefficient of the i-th order IMF component; The function is used to calculate the mathematical expectation. For the i-th order IMF component; Let be the mean of the i-th order IMF component; Let be the standard deviation of the i-th order IMF component. Let be the fourth central moment of the i-th IMF component.

[0019] Furthermore, kurtosis is an indicator that measures the steepness of the data distribution and the tail characteristics. Noise signals typically have higher kurtosis values: steeper distribution and thicker tails.

[0020] Furthermore, a kurtosis threshold K>3.5 is set for screening noise-dominant components. IMF components with kurtosis values ​​exceeding this threshold are identified as noise-dominant IMF components. These components have a higher proportion of speckle noise and are sent to the VMD optimization layer for secondary precise decomposition. IMF components with kurtosis values ​​below the threshold are identified as mid-to-low frequency components containing more effective signal and will be processed by the wavelet joint denoising layer.

[0021] Furthermore, step S2, the VMD optimization step, specifically includes the following steps: Step S201, the steps for solving the variational problem, in which: A variational optimization objective model is established to achieve the separation of signals in different frequency bands by minimizing the bandwidth and sparsity constraints of each modal component. The specific expression is as follows:

[0022] in, Let be the set of VMD modal components to be solved. This is the set of center frequencies corresponding to each modal component; For a partial differential operator with respect to time t; It is the Fourier frequency shift factor; It is the square of the L2 norm, used to constrain the smoothness of the modal components; It serves as a penalty factor, used to control the compactness of modal components; for The frequency domain expression; It is the L1 norm, used to constrain the sparsity of frequency domain components.

[0023] Furthermore, before iteratively solving, the center frequencies of each VMD modal component are initialized to provide an initial reference for subsequent iterations. The calculation formula is as follows:

[0024] in, Let be the initial center frequency of the k-th modal component; For the index of the modal components; The sampling frequency of the ultrasound signal; represents the number of modes in the VMD decomposition.

[0025] Furthermore, based on the initial center frequency, the alternating direction multiplier method (ADMM) is used to iteratively solve the variational optimization model. During the iteration process, the time-domain expression and frequency-domain center frequency of each modal component are updated sequentially to gradually approach the optimal solution of the variational model and realize the frequency band separation of the high-frequency IMF components.

[0026] Furthermore, after each iteration, the convergence of the solution results needs to be checked. The iteration is terminated when the convergence condition is met to ensure the stability of the decomposition results. The convergence check formula is as follows:

[0027] in, This represents the frequency domain value of the k-th modal component in the (n+1)-th iteration. This represents the frequency domain value of the k-th modal component in the nth iteration. It is an L2 norm; The convergence threshold is set when the condition is met, indicating that the current modal component has been solved.

[0028] Step S202, the step of optimizing engineering flaw detection parameters, in which: Based on grid search verification using simulated flaw detection data of non-compact concrete, the optimal parameter combination for VMD decomposition is determined to ensure the adaptability and decomposition effect of the VMD optimization layer in engineering flaw detection scenarios.

[0029] Furthermore, step S3 and the wavelet joint denoising layer steps specifically include the following steps: Step S301, the step of improving the threshold function implementation, in which: Design an exponentially decaying adaptive threshold function, the specific expression of which is:

[0030] in, The input is a sequence of mid-to-low frequency IMF component signals, which is the original mid-to-low frequency signal data to be denoised. It is a symbolic function; This is the core logic for threshold processing; This is the threshold strength reference coefficient, used to control the initial magnitude of the threshold. Signal standard deviation; This is the adaptive decay threshold term.

[0031] Step S302, the step of executing the hybrid threshold strategy, in which: The core parameters of wavelet decomposition are set to ensure the adaptability of the decomposition to the signal characteristics; the sym8 wavelet basis is used for decomposition; the formula for calculating the number of decomposition levels L is:

[0032] Where N is the signal length of the low-to-medium frequency IMF component. () is the rounding function. Engineering verification shows that this number of layers can achieve a balance between decomposition accuracy and computational efficiency.

[0033] Furthermore, after performing wavelet decomposition on the mid- and low-frequency IMF components, a differentiated threshold strategy is adopted for the noise distribution characteristics of different frequency sub-bands: for high-frequency sub-bands, the SURE (Stein unbiased risk estimation) threshold is used; for low-frequency sub-bands, the Minimax threshold is used.

[0034] Furthermore, the wavelet coefficients after thresholding are reconstructed to recover the low-frequency signal in the time domain. The specific expression is as follows:

[0035] in, These are the wavelet coefficients after thresholding. This is the layer index for wavelet decomposition. It is a sym8 wavelet basis. Let i be the denoised signal of the i-th mid-to-low frequency IMF component, where i takes the value from 4 to 8 (corresponding to IMF4-IMF8 components).

[0036] Furthermore, step S4, the multimodal signal reconstruction and post-processing step, includes the following steps: Step S401, the high-frequency VMD component integration step, in which: The high-frequency modal components obtained from VMD decomposition are summed along the modality number dimension, and then redundant dimensions are eliminated. The specific expression is as follows:

[0037] in, This represents the m-th low-frequency IMF component after wavelet denoising (m is 4-8). Let be the number of modes in the VMD decomposition. The result of VMD decomposition is a three-dimensional matrix with dimension 1. (N is the signal length, and 3 corresponds to 3 high-frequency IMF components;) This indicates summation along the VMD modality number dimension; This is a dimension compression operator used to eliminate redundant dimensions of length 1 after summation, ultimately... Dimensions .

[0038] Step S402, the full-band signal fusion step, in which: After vertically concatenating the integrated high-frequency components with the low-frequency components from wavelet denoising, the summation is performed along the component dimensions to obtain the complete denoised signal. The specific expression is as follows:

[0039] in, For a one-dimensional complete time-domain denoised signal, ; The spliced ​​full-band component matrix is ​​expressed as follows:

[0040] in, This represents the m-th low-frequency IMF component after wavelet denoising. It is a one-dimensional complete time-domain denoised signal with a dimension of 1×N.

[0041] Step S403, post-processing step, in which: The reconstructed one-dimensional time-domain denoised signal is restored to a two-dimensional matrix with the same dimensions as the original input image. The specific expression is as follows:

[0042] in, This is the one-dimensional time-domain denoised signal after multimodal fusion; The original ultrasonic flaw detection image of concrete has two dimensions; reshape() is the dimension reshaping operator. This is the two-dimensional image matrix after dimension restoration.

[0043] Furthermore, median filtering is performed on the dimension-restored image to eliminate residual isolated noise points, as expressed by:

[0044] in, It is a two-dimensional median filter operator with a 3×3 window. It replaces the center pixel with the median of the pixels in the window to suppress residual speckle noise and preserve defect edges. This is for the final output denoised and enhanced ultrasound image.

[0045] Secondly, the present invention also provides a concrete flaw detection ultrasonic image speckle noise suppression system based on multimodal signal processing, including a CEEMDAN preprocessing module, a VMD optimization module, a wavelet joint denoising module, and a multimodal signal reconstruction and post-processing module; The CEEMDAN preprocessing module, in which: For the original ultrasonic flaw detection signal of concrete, firstly, adaptive attenuated white noise adapted to the near-field noise characteristics of ultrasound is injected 100 times and EMD decomposition is performed. After mode alignment and averaging, 8th-order IMF components without mode mixing are obtained. Then, the improved kurtosis criterion is used to complete component classification, and high-frequency noise dominant components (IMF1-3) and mid-to-low frequency effective signal components (IMF4-8) are identified, providing accurate component classification basis for subsequent layered differential denoising. The VMD optimization module, in which: The obtained high-frequency noise dominant components are decomposed in a secondary manner. By constructing a variational constraint optimization model and completing iterative solution, the frequency band is accurately divided. At the same time, the optimal parameters verified by engineering flaw detection data are combined to ensure the accuracy and engineering adaptability of high-frequency component decomposition, and provide high-quality high-frequency processing components for subsequent multimodal signal reconstruction. The wavelet joint denoising module, in which: Targeted denoising processing is performed on the mid-to-low frequency IMF components obtained from CEEMDAN preprocessing. By using an improved adaptive wavelet threshold function and a hierarchical hybrid thresholding strategy, the noise in the mid-to-low frequency region is accurately suppressed while the effective signal is completely preserved, providing high-quality mid-to-low frequency processed components for the final multimodal signal reconstruction. The multimodal signal reconstruction and post-processing module, in which: The integrated high-frequency denoised component is fused with the obtained denoised mid-to-low frequency IMF component across the entire frequency band to generate a one-dimensional complete time-domain denoised signal. The signal is then restored to a two-dimensional matrix in the original image dimension, and residual isolated noise points are eliminated by 3×3 window mid-range filtering. The final output is a denoised and enhanced concrete ultrasonic image that can be used for engineering flaw detection, achieving an engineering balance between noise suppression and defect edge preservation.

[0046] Furthermore, the system also includes an initialization setup module, in which: This invention has clear protocol and frequency compatibility requirements for ultrasonic acquisition hardware. It requires the use of color Doppler ultrasound testing equipment that supports the DICOM3.0 medical image transmission protocol. The operating frequency of the equipment probe needs to cover the 5-15MHz range. The corresponding frequency probe can be selected according to the actual concrete flaw detection scenario to ensure that the acquired ultrasonic images meet the signal input standards of subsequent algorithms.

[0047] Furthermore, through a custom parameter calculation model, the original ultrasound image and probe sampling frequency are received as input, and the baseline parameters for the three core steps of CEEMDAN preprocessing, VMD optimization, and wavelet joint denoising are calculated sequentially. Finally, a standardized parameter set is output, providing a unified parameter benchmark for subsequent algorithm modules. The specific execution steps are as follows: Input the original ultrasound image matrix and the probe sampling frequency to determine the one-dimensional unfolded length of the image pixels; the expression for the signal unfolded length is:

[0048] Where img is the original ultrasound image matrix. The length of the signal spread; Furthermore, the CEEMDAN noise standard deviation parameter is calculated to control the injection intensity of adaptive white noise. The calculation formula is as follows:

[0049] in, is the L2 norm of the one-dimensional pixel sequence of the original ultrasound image, and 0.2 is the engineering-verified fundamental noise intensity coefficient. This is the standard deviation parameter for CEEMDAN noise.

[0050] Furthermore, the VMD adaptive penalty factor is calculated to achieve dynamic adaptation of the penalty factor with the probe frequency. The calculation formula is as follows:

[0051] in, The sampling frequency of the ultrasonic probe. The reference frequency coefficient (corresponding to a 10MHz reference probe) is 2000, which is the optimal penalty factor reference value verified by 200 engineering flaw detection data. When the probe frequency is 10MHz, the penalty factor is the reference value of 2000. When the frequency is higher or lower than 10MHz, the penalty factor increases or decreases synchronously to ensure the accuracy of VMD frequency band division under different probes. This is the VMD adaptive penalty factor.

[0052] Furthermore, the number of wavelet decomposition layers is calculated to balance the accuracy and computational efficiency of wavelet decomposition. The calculation formula is as follows:

[0053] Where N is the signal length of the low-to-medium frequency IMF component. () is the binary logarithm of the signal length. This is the rounding function. The number of wavelet decomposition layers is given by this formula. This formula can adaptively adjust the number of decomposition layers according to the signal length. Engineering verification has shown that it can achieve the optimal balance between decomposition accuracy and computational efficiency.

[0054] The three parameters mentioned above—CEEMDAN noise standard deviation, VMD adaptive penalty factor, and wavelet decomposition level—are integrated to form a complete initialization parameter set, params, which is then output.

[0055] Furthermore, the system also includes a DICOM standard access and real-time video processing module, in which: The DICOM file parameter reading function calls the `dicom_interface` interface. The `dicominfo` function reads the header information of the input DICOM file, and the `dicomread` function extracts the ultrasound image data. Simultaneously, it automatically reads the probe parameters from the file header information and calculates the signal sampling frequency.

[0056] The probe frequency adaptive parameter adjustment adjusts the core parameters of CEEMDAN, VMD, and wavelet denoising based on the probe center frequency (5-15MHz range).

[0057] The standard enhanced image output will complete the denoising signal after multimodal reconstruction and median filtering post-processing, and encapsulate it into a DICOM standard format image; at the same time, it supports three-dimensional ultrasound volume reconstruction based on the image, realizing three-dimensional imaging with isotropic resolution of 0.4mm, and seamlessly connecting with AI diagnostic systems.

[0058] Furthermore, the `realtime_processing` function is called to access the target video source via the `VideoReader` interface, entering a frame-by-frame reading process to acquire the original image frames of the ultrasound video one by one. For single-frame denoising, the `ultrasound_denoising` function is called for each image frame, with parameters configured by default at a 15MHz probe frequency. This sequentially completes CEEMDAN preprocessing, VMD optimization, wavelet denoising, multimodal reconstruction, and post-processing to obtain the denoised image frame.

[0059] Real-time output integration with OpenCV can be achieved by building an integrated solution based on the OpenCV framework, defining the process_frame function, and integrating tools such as pywt to realize CEEMD-VMD parallel computing, outputting and displaying the processed image frames in real time, and completing the real-time noise reduction presentation of ultrasound video.

[0060] The beneficial effects of this invention are as follows: This invention proposes a method and system for suppressing speckle noise in ultrasonic images used for concrete flaw detection based on multimodal signal processing. This effectively improves the reliability and accuracy of ultrasonic image quality. It solves the problem of modal aliasing in traditional EMD by injecting adaptive attenuated white noise and averaging modal alignment to obtain aliased IMF components. Combined with kurtosis criteria, it accurately distinguishes high-frequency noise from effective mid- and low-frequency components, avoiding the accidental deletion of effective defect information. Through a variational constraint model and adaptive parameter system, it achieves precise frequency band division for high-frequency components, adaptable to 5-15MHz full-band probes, and effectively suppresses multi-band noise harmonics. An improved exponentially decaying wavelet threshold and layered mixing strategy can completely preserve the fine structure of concrete ≤0.3mm, significantly improving defect boundary sharpness. The DICOM interface and real-time video module can cover both static and dynamic flaw detection scenarios, and also support 3D reconstruction and integration with AI diagnostic systems, significantly enhancing its practical engineering value.

[0061] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0062] 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.

[0063] Figure 1 This is a flowchart of a method for suppressing speckle noise in ultrasonic images used in concrete flaw detection based on multimodal signal processing.

[0064] Figure 2 This is a schematic diagram of the principle framework of a concrete flaw detection ultrasonic image speckle noise suppression system based on multimodal signal processing.

[0065] Among them, 1-CEEMDAN preprocessing module, 2-VMD optimization module, 3-wavelet joint denoising module, 4-multimodal signal reconstruction and post-processing module, and 5-DICOM standard access and real-time video processing module. Detailed Implementation

[0066] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0067] Example 1: In a first aspect, the present invention provides a method for suppressing speckle noise in ultrasonic images used for concrete flaw detection based on multimodal signal processing, such as... Figure 1 As shown, the specific steps include: Step S0, the initialization setup step, in which: The ultrasound testing equipment supports the DICOM3.0 medical image transmission protocol and the working frequency of the probe needs to cover the 5-15MHz range. The original ultrasound image and the probe sampling frequency are received as inputs, and the CEEMDAN noise standard deviation parameter, VMD adaptive penalty factor and wavelet decomposition level are calculated in sequence. Finally, a standardized parameter set is output to provide a unified parameter benchmark for subsequent algorithm modules. Step S1, the CEEMDAN preprocessing step, in which: By injecting adaptive attenuated white noise, which is adapted to the near-field noise characteristics of ultrasound, into the original ultrasonic flaw detection signal of concrete 100 times and performing EMD decomposition, the 8th order IMF components without mode mixing are obtained after mode alignment and averaging. The components are classified by improving the kurtosis criterion (K>3.5), and the high-frequency noise dominant components (IMF1-3) and the mid-to-low frequency effective signal components (IMF4-8) are identified, providing a precise component classification basis for subsequent layered differential denoising. Step S2, VMD optimization steps The high-frequency noise dominant component obtained in step S1 is decomposed in a secondary manner. By constructing a variational constraint optimization model and completing iterative solution, the frequency band is accurately divided. At the same time, the optimal parameters verified by engineering flaw detection data are combined to ensure the accuracy and engineering adaptability of the high-frequency component decomposition, and provide high-quality high-frequency processing components for subsequent multimodal signal reconstruction. Step S3, the wavelet joint denoising step, in which: Targeted denoising processing is performed on the mid-to-low frequency IMF components obtained from CEEMDAN preprocessing. By using an improved adaptive wavelet threshold function and a hierarchical hybrid thresholding strategy, the noise in the mid-to-low frequency region is accurately suppressed while the effective signal is completely preserved, providing high-quality mid-to-low frequency processed components for the final multimodal signal reconstruction. Step S4, the multimodal signal reconstruction and post-processing step, in which: The high-frequency denoised component integrated in step S2 is fused with the denoised mid-to-low frequency IMF component obtained in step S3 across the entire frequency band to generate a one-dimensional complete time-domain denoised signal. The signal is then restored to a two-dimensional matrix in the original image dimension, and residual isolated noise points are eliminated by 3×3 window mid-range filtering. Finally, a denoised and enhanced concrete ultrasonic image that can be used for engineering flaw detection is output, achieving an engineering balance between noise suppression and defect edge preservation.

[0068] Step S5, DICOM standard access and real-time video processing steps, in which: The DICOM file parameter reading interface function reads the header information of the input DICOM file and extracts the ultrasound image data. Simultaneously, it automatically reads the probe parameters from the file header information, calculates the signal sampling frequency, and adjusts the core parameters of CEEMDAN, VMD, and wavelet denoising based on the read probe center frequency (5-15MHz range). The function then connects to the target video source and enters a frame loop reading process, acquiring the original image frames of the ultrasound video frame by frame.

[0069] The steps S0 and initialization settings specifically include: This method has clear protocol and frequency compatibility requirements for the hardware used in ultrasonic flaw detection of concrete. It employs color Doppler ultrasound equipment that supports the DICOM 3.0 medical image transmission protocol, with the probes operating in the 5-15MHz frequency range. Specifically, for detecting shallow concrete defects (depth ≤ 50mm), a 10-15MHz high-frequency probe is used to ensure the accuracy of signal acquisition for subtle shallow defects. For detecting deep concrete defects (depth > 50mm), a 5-8MHz low-frequency probe is used to enhance the penetration ability of ultrasound waves in the concrete medium, ensuring that the acquired ultrasonic images meet the input standards for subsequent multimodal signal processing algorithms, thus achieving full coverage of defect detection scenarios at different depths.

[0070] Furthermore, through a custom parameter calculation model, the original ultrasound image and probe sampling frequency are received as input, and the baseline parameters for the three core steps of CEEMDAN preprocessing, VMD optimization, and wavelet joint denoising are calculated sequentially. Finally, a standardized parameter set is output, providing a unified parameter benchmark for subsequent algorithm modules. The specific execution steps are as follows: Input the original ultrasound image matrix and the probe sampling frequency to determine the one-dimensional unfolded length of the image pixels; the expression for the signal unfolded length is:

[0071] Where img is the original ultrasound image matrix. The length of the signal spread; Furthermore, the CEEMDAN noise standard deviation parameter is calculated to control the injection intensity of adaptive white noise. The calculation formula is as follows:

[0072] in, is the L2 norm of the one-dimensional pixel sequence of the original ultrasound image, and 0.2 is the engineering-verified fundamental noise intensity coefficient. This is the standard deviation parameter for CEEMDAN noise.

[0073] Furthermore, the VMD adaptive penalty factor is calculated to achieve dynamic adaptation of the penalty factor with the probe frequency. The calculation formula is as follows:

[0074] in, The sampling frequency of the ultrasonic probe. The reference frequency coefficient (corresponding to a 10MHz reference probe) is 2000, which is the optimal penalty factor reference value verified by 200 engineering flaw detection data. When the probe frequency is 10MHz, the penalty factor is the reference value of 2000. When the frequency is higher or lower than 10MHz, the penalty factor increases or decreases synchronously to ensure the accuracy of VMD frequency band division under different probes. This is the VMD adaptive penalty factor.

[0075] Furthermore, the number of wavelet decomposition layers is calculated to balance the accuracy and computational efficiency of wavelet decomposition. The calculation formula is as follows:

[0076] Where N is the signal length of the low-to-medium frequency IMF component. () is the binary logarithm of the signal length. This is the rounding function. The number of wavelet decomposition layers is given by this formula. This formula can adaptively adjust the number of decomposition layers according to the signal length. Engineering verification has shown that it can achieve the optimal balance between decomposition accuracy and computational efficiency.

[0077] The three parameters mentioned above—CEEMDAN noise standard deviation, VMD adaptive penalty factor, and wavelet decomposition level—are integrated to form a complete initialization parameter set, params, which is then output.

[0078] Step S1, the CEEMDAN preprocessing step, specifically includes the following steps: Step S101, the noise injection and decomposition step, in which: By repeatedly injecting adaptively attenuated white noise, the mode aliasing problem in traditional EMD decomposition is solved, achieving signal stabilization decomposition. The specific execution steps are as follows: First, initialize the parameters: The total number of noise injections is set to 100. Simultaneously, based on the characteristics of the original ultrasonic signal x, the attenuation coefficient of the adaptive noise is calculated using the following formula:

[0079] in, is the attenuation coefficient for adaptive noise, and 0.2 is the fundamental noise intensity coefficient. Let x be the standard deviation of the original signal x, and t be the number of time-domain sampling points of the signal; The injected noise intensity decreases exponentially with the increase of signal sequence length, ensuring that the noise injection intensity is slightly higher at the front end (near field) and lower at the back end (far field), which is suitable for the strong near-field noise characteristics of ultrasonic signals.

[0080] Furthermore, noise is injected sequentially and EMD decomposition is performed. In each loop, attenuated white noise is superimposed on the original signal to obtain a noisy signal. Then, conventional EMD decomposition is performed on the noisy signal to obtain a set of temporary IMF components. This process is repeated 100 times to obtain 100 sets of IMF components corresponding to different noisy signals. The relevant calculation formula is as follows:

[0081]

[0082] in, The noisy ultrasonic signal in the nth iteration; A function to generate standard normally distributed white noise with the same dimension as the original signal x; It is the empirical mode decomposition function; This is the set of temporary IMF components obtained from the nth noisy signal decomposition. This is the attenuation coefficient for adaptive noise; Furthermore, after completing 100 EMD decompositions, modal alignment and averaging of IMF components are performed on all temporary IMF components. This involves matching IMF components of the same level among the 100 components in descending frequency order. Then, the average value of the IMF components at the same level is calculated, ultimately yielding a set of stable, mode-free CEEMDAN decomposed IMF components. This operation effectively cancels out interference from multiple injections of white noise, preserving the true characteristics of the original signal. The specific formula is as follows:

[0083] in, This represents the final k-th order IMF component; Let k be the temporary IMF component of order k obtained from the nth iteration decomposition; k is the order index of the IMF component.

[0084] Furthermore, during the screening process of a single EMD decomposition, the continuous screening standard deviation (SD) of two adjacent screening results needs to be monitored in real time. When SD < 0.3, it is determined that the current IMF component has been completely decomposed, and the screening iteration of that component is immediately terminated, and the decomposition of the next IMF component is carried out until all components are extracted. This stopping condition can avoid signal feature distortion caused by over-screening.

[0085] Step S102, the noise-driven IMF identification step, in which: After completing the CEEMDAN decomposition, the noise-dominated IMF component was identified using an improved kurtosis criterion, providing a basis for subsequent targeted processing. The specific steps are as follows: For each IMF component obtained from CEEMDAN decomposition, calculate its kurtosis coefficient using the following formula:

[0086] in, is the kurtosis coefficient of the i-th order IMF component; The function is used to calculate the mathematical expectation. For the i-th order IMF component; Let be the mean of the i-th order IMF component; Let be the standard deviation of the i-th order IMF component. Let be the fourth central moment of the i-th IMF component.

[0087] Furthermore, kurtosis is an indicator that measures the steepness of the data distribution and the tail characteristics. Noise signals typically have higher kurtosis values: steeper distribution and thicker tails.

[0088] Furthermore, a kurtosis threshold K>3.5 is set for screening noise-dominant components. IMF components with kurtosis values ​​exceeding this threshold are identified as noise-dominant IMF components. These components have a higher proportion of speckle noise and are sent to the VMD optimization layer for secondary precise decomposition. IMF components with kurtosis values ​​below the threshold are identified as mid-to-low frequency components containing more effective signal and will be processed by the wavelet joint denoising layer.

[0089] Furthermore, step S2, the VMD optimization step, specifically includes the following steps: Step S201, the steps for solving the variational problem, in which: By constructing a variational constraint optimization model and combining center frequency initialization with alternating direction multiplier method iteration, accurate frequency band decomposition of high-frequency IMF components is achieved. The specific execution steps are as follows: To achieve frequency band division of high-frequency IMF components, a variational optimization objective model is established. By minimizing the bandwidth and sparsity constraints of each modal component, the separation of signals in different frequency bands is achieved. The specific expression is as follows:

[0090] in, Let be the set of VMD modal components to be solved. This is the set of center frequencies corresponding to each modal component; For a partial differential operator with respect to time t; It is the Fourier frequency shift factor; It is the square of the L2 norm, used to constrain the smoothness of the modal components; It serves as a penalty factor, used to control the compactness of modal components; for The frequency domain expression; It is the L1 norm, used to constrain the sparsity of frequency domain components.

[0091] Furthermore, before iteratively solving, the center frequencies of each VMD modal component are initialized to provide an initial reference for subsequent iterations. The calculation formula is as follows:

[0092] in, Let be the initial center frequency of the k-th modal component; For the index of the modal components; The sampling frequency of the ultrasound signal; represents the number of modes in the VMD decomposition.

[0093] Furthermore, based on the initial center frequency, the alternating direction multiplier method (ADMM) is used to iteratively solve the variational optimization model. During the iteration process, the time-domain expression and frequency-domain center frequency of each modal component are updated sequentially to gradually approach the optimal solution of the variational model and realize the frequency band separation of the high-frequency IMF components.

[0094] Furthermore, after each iteration, the convergence of the solution results needs to be checked. The iteration is terminated when the convergence condition is met to ensure the stability of the decomposition results. The convergence check formula is as follows:

[0095] in, This represents the frequency domain value of the k-th modal component in the (n+1)-th iteration. This represents the frequency domain value of the k-th modal component in the nth iteration. It is an L2 norm; The convergence threshold is set when the condition is met, indicating that the current modal component has been solved.

[0096] Step S202, the step of optimizing engineering flaw detection parameters, in which: Based on grid search verification using simulated flaw detection data of non-compact concrete, the optimal parameter combination for VMD decomposition was determined to ensure the adaptability and decomposition effect of the VMD optimization layer in engineering flaw detection scenarios. The specific parameter statistics are shown in the table below: Table 1 Statistical Table of Optimal Combination Parameters

[0097] Among them, the number of modes K is used to control the number of modal components in VMD decomposition. The optimal value can achieve accurate multi-band splitting of high-frequency IMF components, avoiding band aliasing caused by insufficient mode number, or excessive splitting of effective signal caused by excessive mode number; penalty factor α The optimal value is used to constrain the frequency domain compactness of VMD modal components. It can preserve the effective defect features in the high-frequency signals of concrete flaw detection while ensuring the accuracy of frequency band division, and avoid distortion of the effective signals.

[0098] Furthermore, step S3 and the wavelet joint denoising layer steps specifically include the following steps: Step S301, the step of improving the threshold function implementation, in which: An adaptive wavelet threshold function is constructed to overcome the limitations of traditional hard / soft threshold functions, thereby reducing amplitude distortion and edge blurring of the effective signal while achieving noise filtering.

[0099] To accommodate the nonlinear characteristics of low-frequency components in ultrasound signals, an exponentially decaying adaptive threshold function is designed, with the following expression:

[0100] in, The input is a sequence of mid-to-low frequency IMF component signals, which is the original mid-to-low frequency signal data to be denoised. This is the sign function, used to preserve the polarity (positive / negative) of the original signal and avoid phase distortion of the signal after denoising; The core logic for threshold processing is to retain the portion of the signal that exceeds the adaptive threshold when the absolute value of the signal is greater than the threshold, and otherwise set it to 0 to filter out noise. This is the threshold strength reference coefficient, used to control the initial magnitude of the threshold. Signal standard deviation is used to match the fluctuation characteristics of the input signal; for As an adaptive attenuation threshold term, the larger the signal amplitude, the smaller the threshold. This ensures that small-amplitude noise is fully filtered out while avoiding excessive compression of large-amplitude effective signals, thus solving the signal distortion problem caused by the constant threshold in traditional threshold functions.

[0101] Step S302, the step of executing the hybrid threshold strategy, in which: Combining the sub-band characteristics of low- and mid-frequency IMF components after wavelet decomposition, a hierarchical hybrid thresholding strategy is adopted, along with a specified wavelet basis and decomposition level, to achieve differentiated denoising for different frequency sub-bands. The specific execution process is as follows: The core parameters of wavelet decomposition are set to ensure the compatibility of the decomposition with the signal characteristics; the sym8 wavelet basis is used for decomposition, which has good compact support and symmetry, and can match the waveform characteristics of low-frequency signals in concrete flaw detection ultrasound, reducing signal energy leakage during the decomposition process; the formula for calculating the number of decomposition layers L is:

[0102] Where N is the signal length of the low-to-medium frequency IMF component. () is the rounding function. Engineering verification shows that this number of layers can achieve a balance between decomposition accuracy and computational efficiency.

[0103] Furthermore, after performing wavelet decomposition on the mid- and low-frequency IMF components, a differentiated threshold strategy is adopted based on the noise distribution characteristics of different frequency sub-bands: For high-frequency sub-bands, the SURE (Stein unbiased risk estimation) threshold is used. This threshold can be adaptively estimated using statistical methods, making it suitable for high-frequency sub-bands with complex noise distributions. It can maximize noise filtering while minimizing the signal estimation risk. For low-frequency sub-bands, the Minimax threshold is used. This threshold is the optimal threshold with a minimum maximum value, which can ensure the integrity of the effective signal in the low-frequency sub-band and avoid the loss of fine structural features of concrete due to excessively high thresholds. It is also suitable for the characteristics of high effective signal ratio and low noise ratio in the low-frequency sub-bands.

[0104] Furthermore, the wavelet coefficients after thresholding are reconstructed to recover the low-frequency signal in the time domain. The specific expression is as follows:

[0105] in, These are the wavelet coefficients after thresholding. This is the layer index for wavelet decomposition. It is a sym8 wavelet basis. Let i be the denoised signal of the i-th mid-to-low frequency IMF component, where i takes the value from 4 to 8 (corresponding to IMF4-IMF8 components).

[0106] Furthermore, step S4, the multimodal signal reconstruction and post-processing step, includes the following steps: Step S401, the high-frequency VMD component integration step, in which: The high-frequency modal components obtained from VMD decomposition are summed along the modality number dimension, and then redundant dimensions are eliminated. The specific expression is as follows:

[0107] in, This represents the m-th low-frequency IMF component after wavelet denoising (m is 4-8). Let be the number of modes in the VMD decomposition. The result of VMD decomposition is a three-dimensional matrix with dimension 1. (N is the signal length, and 3 corresponds to 3 high-frequency IMF components;) This indicates summation along the VMD modality number dimension; This is a dimension compression operator used to eliminate redundant dimensions of length 1 after summation, ultimately... Dimensions .

[0108] Step S402, the full-band signal fusion step, in which: After vertically concatenating the integrated high-frequency components with the low-frequency components from wavelet denoising, the summation is performed along the component dimensions to obtain the complete denoised signal. The specific expression is as follows:

[0109] in, For a one-dimensional complete time-domain denoised signal, (3 high-frequency components, 5 mid-to-low-frequency components); The spliced ​​full-band component matrix is ​​expressed as follows:

[0110] in, This represents the m-th low-frequency IMF component after wavelet denoising (m is 4-8). It is a one-dimensional complete time-domain denoised signal with a dimension of 1×N.

[0111] Step S403, post-processing step, in which: The reconstructed one-dimensional time-domain denoised signal is restored to a two-dimensional matrix with the same dimensions as the original input image. The specific expression is as follows:

[0112] in, This is the one-dimensional time-domain denoised signal after multimodal fusion; Let H be the two-dimensional dimension of the original ultrasonic flaw detection image of concrete (H×W, where H is the image height and W is the image width); reshape() is the dimension reshaping operator. This is the two-dimensional image matrix after dimension restoration, with dimensions of H×W.

[0113] Furthermore, median filtering is performed on the dimension-restored image to eliminate residual isolated noise points, as expressed by:

[0114] in, It is a two-dimensional median filter operator with a 3×3 window. It replaces the center pixel with the median of the pixels in the window to suppress residual speckle noise and preserve defect edges. This is for the final output denoised and enhanced ultrasound image.

[0115] Furthermore, step S5, the DICOM standard access and real-time video processing steps, specifically also include: The DICOM file parameter reading function calls the `dicom_interface` interface. The `dicominfo` function reads the header information of the input DICOM file, and the `dicomread` function extracts the ultrasound image data. Simultaneously, it automatically reads the probe parameters from the file header information and calculates the signal sampling frequency.

[0116] The probe frequency adaptive parameter adjustment adjusts the core parameters of CEEMDAN, VMD, and wavelet denoising based on the probe center frequency (5-15MHz range).

[0117] The standard enhanced image output will complete the denoising signal after multimodal reconstruction and median filtering post-processing, and encapsulate it into a DICOM standard format image; at the same time, it supports three-dimensional ultrasound volume reconstruction based on the image, realizing three-dimensional imaging with isotropic resolution of 0.4mm, and seamlessly connecting with AI diagnostic systems.

[0118] Furthermore, the `realtime_processing` function is called to access the target video source via the `VideoReader` interface, entering a frame-by-frame reading process to acquire the original image frames of the ultrasound video one by one. For single-frame denoising, the `ultrasound_denoising` function is called for each image frame, with parameters configured by default at a 15MHz probe frequency. This sequentially completes CEEMDAN preprocessing, VMD optimization, wavelet denoising, multimodal reconstruction, and post-processing to obtain the denoised image frame.

[0119] Real-time output integration with OpenCV can be achieved by building an integrated solution based on the OpenCV framework, defining the process_frame function, and integrating tools such as pywt to realize CEEMD-VMD parallel computing, outputting and displaying the processed image frames in real time, and completing the real-time noise reduction presentation of ultrasound video.

[0120] Example 2: Secondly, the present invention also provides a concrete flaw detection ultrasonic image speckle noise suppression system based on multimodal signal processing, such as... Figure 2As shown, it includes an initialization setting module, a CEEMDAN preprocessing module 1, a VMD optimization module 2, a wavelet joint denoising module 3, a multimodal signal reconstruction and post-processing module 4, and a DICOM standard access and real-time video processing module 5. The initialization setting module, in which: The ultrasound testing equipment supports the DICOM3.0 medical image transmission protocol and the working frequency of the probe needs to cover the 5-15MHz range. The original ultrasound image and the probe sampling frequency are received as inputs, and the CEEMDAN noise standard deviation parameter, VMD adaptive penalty factor and wavelet decomposition level are calculated in sequence. Finally, a standardized parameter set is output to provide a unified parameter benchmark for subsequent algorithm modules. The CEEMDAN preprocessing module 1, in which: For the original ultrasonic flaw detection signal of concrete, adaptive attenuated white noise adapted to the near-field noise characteristics of ultrasound is injected 100 times and EMD decomposition is performed. After mode alignment and averaging, 8th order IMF components without mode mixing are obtained. Then, the improved kurtosis criterion (K>3.5) is used to complete component classification, and high-frequency noise dominant components (IMF1-3) and mid-to-low frequency effective signal components (IMF4-8) are identified, providing accurate component classification basis for subsequent layered differential denoising. The VMD optimization module 2, in which: The obtained high-frequency noise dominant components are decomposed in a secondary manner. By constructing a variational constraint optimization model and completing iterative solution, the frequency band is accurately divided. At the same time, the optimal parameters verified by engineering flaw detection data are combined to ensure the accuracy and engineering adaptability of high-frequency component decomposition, and provide high-quality high-frequency processing components for subsequent multimodal signal reconstruction. The wavelet joint denoising module 3, in which: Targeted denoising processing is performed on the effective mid- and low-frequency signal components obtained from CEEMDAN preprocessing. By using an improved adaptive wavelet threshold function and a hierarchical hybrid thresholding strategy, the noise in the mid- and low-frequency region is accurately suppressed while the effective signal is completely preserved, providing high-quality mid- and low-frequency processed components for the final multimodal signal reconstruction. The multimodal signal reconstruction and post-processing module 4, in which: The integrated high-frequency denoised component is fused with the obtained denoised mid-to-low frequency IMF component across the entire frequency band to generate a one-dimensional complete time-domain denoised signal. The signal is then restored to a two-dimensional matrix in the original image dimension, and residual isolated noise points are eliminated by 3×3 window mid-range filtering. The final output is a denoised and enhanced concrete ultrasonic image that can be used for engineering flaw detection, achieving an engineering balance between noise suppression and defect edge preservation.

[0121] The DICOM standard access and real-time video processing module 5, in which: The DICOM file parameter reading interface function reads the header information of the input DICOM file and extracts the ultrasound image data. Simultaneously, it automatically reads the probe parameters from the file header information, calculates the signal sampling frequency, and adjusts the core parameters of CEEMDAN, VMD, and wavelet denoising based on the read probe center frequency (5-15MHz range). The function then connects to the target video source and enters a frame loop reading process, acquiring the original image frames of the ultrasound video frame by frame.

[0122] The initialization setting module specifically includes: This invention has clear protocol and frequency compatibility requirements for ultrasound acquisition hardware. It requires the use of color Doppler ultrasound testing equipment that supports the DICOM3.0 medical image transmission protocol. The operating frequency of the equipment probe needs to cover the 5-15MHz range. The corresponding frequency probe can be selected according to the actual concrete flaw detection scenario (such as shallow / deep defect detection) (10-15MHz high-frequency probe for shallow detection and 5-8MHz low-frequency probe for deep detection) to ensure that the acquired ultrasound images meet the signal input standards of subsequent algorithms.

[0123] Furthermore, through a custom parameter calculation model, the original ultrasound image and probe sampling frequency are received as input, and the baseline parameters for the three core steps of CEEMDAN preprocessing, VMD optimization, and wavelet joint denoising are calculated sequentially. Finally, a standardized parameter set is output, providing a unified parameter benchmark for subsequent algorithm modules. The specific execution steps are as follows: Input the original ultrasound image matrix and the probe sampling frequency to determine the one-dimensional unfolded length of the image pixels; the expression for the signal unfolded length is:

[0124] Where img is the original ultrasound image matrix. The length of the signal spread; Furthermore, the CEEMDAN noise standard deviation parameter is calculated to control the injection intensity of adaptive white noise. The calculation formula is as follows:

[0125] in, is the L2 norm of the one-dimensional pixel sequence of the original ultrasound image, and 0.2 is the engineering-verified fundamental noise intensity coefficient. This is the standard deviation parameter for CEEMDAN noise.

[0126] Furthermore, the VMD adaptive penalty factor is calculated to achieve dynamic adaptation of the penalty factor with the probe frequency. The calculation formula is as follows:

[0127] in, The sampling frequency of the ultrasonic probe. The reference frequency coefficient (corresponding to a 10MHz reference probe) is 2000, which is the optimal penalty factor reference value verified by 200 engineering flaw detection data. When the probe frequency is 10MHz, the penalty factor is the reference value of 2000. When the frequency is higher or lower than 10MHz, the penalty factor increases or decreases synchronously to ensure the accuracy of VMD frequency band division under different probes. This is the VMD adaptive penalty factor.

[0128] Furthermore, the number of wavelet decomposition layers is calculated to balance the accuracy and computational efficiency of wavelet decomposition. The calculation formula is as follows:

[0129] Where N is the signal length of the low-to-medium frequency IMF component. () is the binary logarithm of the signal length. This is the rounding function. The number of wavelet decomposition layers is given by this formula. This formula can adaptively adjust the number of decomposition layers according to the signal length. Engineering verification has shown that it can achieve the optimal balance between decomposition accuracy and computational efficiency.

[0130] The three parameters mentioned above—CEEMDAN noise standard deviation, VMD adaptive penalty factor, and wavelet decomposition level—are integrated to form a complete initialization parameter set, params, which is then output.

[0131] Furthermore, the CEEMDAN preprocessing module 1 includes a noise injection and decomposition submodule and a noise-dominant IMF identification submodule; The noise injection and decomposition submodule, in which: By repeatedly injecting adaptively attenuated white noise, the mode aliasing problem in traditional EMD decomposition is solved, achieving signal stabilization decomposition. The specific execution steps are as follows: First, initialize the parameters: The total number of noise injections is set to 100. Simultaneously, based on the characteristics of the original ultrasonic signal x, the attenuation coefficient of the adaptive noise is calculated using the following formula:

[0132] in, is the attenuation coefficient for adaptive noise, and 0.2 is the fundamental noise intensity coefficient. Let x be the standard deviation of the original signal x, and t be the number of time-domain sampling points of the signal; The injected noise intensity decreases exponentially with the increase of signal sequence length, ensuring that the noise injection intensity is slightly higher at the front end (near field) and lower at the back end (far field), which is suitable for the strong near-field noise characteristics of ultrasonic signals.

[0133] Furthermore, noise is injected sequentially and EMD decomposition is performed. In each loop, attenuated white noise is superimposed on the original signal to obtain a noisy signal. Then, conventional EMD decomposition is performed on the noisy signal to obtain a set of temporary IMF components. This process is repeated 100 times to obtain 100 sets of IMF components corresponding to different noisy signals. The relevant calculation formula is as follows:

[0134]

[0135] in, The noisy ultrasonic signal in the nth iteration; A function to generate standard normally distributed white noise with the same dimension as the original signal x; It is the empirical mode decomposition function; This is the set of temporary IMF components obtained from the nth noisy signal decomposition. This is the attenuation coefficient for adaptive noise.

[0136] Furthermore, after completing 100 EMD decompositions, modal alignment and averaging of IMF components are performed on all temporary IMF components. This involves matching IMF components of the same level among the 100 components in descending frequency order. Then, the average value of the IMF components at the same level is calculated, ultimately yielding a set of stable, mode-free CEEMDAN decomposed IMF components. This operation effectively cancels out interference from multiple injections of white noise, preserving the true characteristics of the original signal. The specific formula is as follows:

[0137] in, This represents the final k-th order IMF component; Let k be the temporary IMF component of order k obtained from the nth iteration decomposition; k is the order index of the IMF component.

[0138] Furthermore, during the screening process of a single EMD decomposition, the continuous screening standard deviation (SD) of two adjacent screening results needs to be monitored in real time. When SD < 0.3, it is determined that the current IMF component has been completely decomposed, and the screening iteration of that component is immediately terminated, and the decomposition of the next IMF component is carried out until all components are extracted. This stopping condition can avoid signal feature distortion caused by over-screening.

[0139] The noise-dominated IMF identification submodule, in which: After completing the CEEMDAN decomposition, the noise-dominated IMF component was identified using an improved kurtosis criterion, providing a basis for subsequent targeted processing. The specific steps are as follows: For each IMF component obtained from CEEMDAN decomposition, calculate its kurtosis coefficient using the following formula:

[0140] in, is the kurtosis coefficient of the i-th order IMF component; The function is used to calculate the mathematical expectation. For the i-th order IMF component; Let be the mean of the i-th order IMF component; Let be the standard deviation of the i-th order IMF component. Let be the fourth central moment of the i-th IMF component.

[0141] Furthermore, kurtosis is an indicator that measures the steepness of the data distribution and the tail characteristics. Noise signals typically have higher kurtosis values: steeper distribution and thicker tails.

[0142] Furthermore, a kurtosis threshold K>3.5 is set for screening noise-dominant components. IMF components with kurtosis values ​​exceeding this threshold are identified as noise-dominant IMF components. These components have a higher proportion of speckle noise and are sent to the VMD optimization layer for secondary precise decomposition. IMF components with kurtosis values ​​below the threshold are identified as mid-to-low frequency components containing more effective signal and will be processed by the wavelet joint denoising layer.

[0143] Furthermore, the VMD optimization module 2 includes a variational problem solving submodule and an engineering flaw detection parameter optimization submodule; The variational problem solving submodule includes: By constructing a variational constraint optimization model and combining center frequency initialization with alternating direction multiplier method iteration, accurate frequency band decomposition of high-frequency IMF components is achieved. The specific execution steps are as follows: To achieve frequency band division of high-frequency IMF components, a variational optimization objective model is established. By minimizing the bandwidth and sparsity constraints of each modal component, the separation of signals in different frequency bands is achieved. The specific expression is as follows:

[0144] in, Let be the set of VMD modal components to be solved. This is the set of center frequencies corresponding to each modal component; For a partial differential operator with respect to time t; It is the Fourier frequency shift factor; It is the square of the L2 norm, used to constrain the smoothness of the modal components; It serves as a penalty factor, used to control the compactness of modal components; for The frequency domain expression; It is the L1 norm, used to constrain the sparsity of frequency domain components.

[0145] Furthermore, before iteratively solving, the center frequencies of each VMD modal component are initialized to provide an initial reference for subsequent iterations. The calculation formula is as follows:

[0146] in, Let be the initial center frequency of the k-th modal component; For the index of the modal components; The sampling frequency of the ultrasound signal; represents the number of modes in the VMD decomposition.

[0147] Furthermore, based on the initial center frequency, the alternating direction multiplier method (ADMM) is used to iteratively solve the variational optimization model. During the iteration process, the time-domain expression and frequency-domain center frequency of each modal component are updated sequentially to gradually approach the optimal solution of the variational model and realize the frequency band separation of the high-frequency IMF components.

[0148] Furthermore, after each iteration, the convergence of the solution results needs to be checked. The iteration is terminated when the convergence condition is met to ensure the stability of the decomposition results. The convergence check formula is as follows:

[0149] in, This represents the frequency domain value of the k-th modal component in the (n+1)-th iteration. This represents the frequency domain value of the k-th modal component in the nth iteration. It is an L2 norm; The convergence threshold is set when the condition is met, indicating that the current modal component has been solved.

[0150] Furthermore, the engineering flaw detection parameter optimization submodule also includes: Based on grid search verification using simulated flaw detection data of non-compact concrete, the optimal parameter combination for VMD decomposition is determined to ensure the adaptability and decomposition effect of the VMD optimization layer in engineering flaw detection scenarios.

[0151] Furthermore, the wavelet joint denoising layer module 3 includes an improved threshold function implementation submodule and a hybrid threshold strategy execution submodule; The improved threshold function implementation submodule also includes: An adaptive wavelet threshold function is constructed to overcome the limitations of traditional hard / soft threshold functions, thereby reducing amplitude distortion and edge blurring of the effective signal while achieving noise filtering.

[0152] To accommodate the nonlinear characteristics of low-frequency components in ultrasound signals, an exponentially decaying adaptive threshold function is designed, with the following expression:

[0153] in, The input is a sequence of mid-to-low frequency IMF component signals, which is the original mid-to-low frequency signal data to be denoised. This is the sign function, used to preserve the polarity (positive / negative) of the original signal and avoid phase distortion of the signal after denoising; The core logic for threshold processing is to retain the portion of the signal that exceeds the adaptive threshold when the absolute value of the signal is greater than the threshold, and otherwise set it to 0 to filter out noise. This is the threshold strength reference coefficient, used to control the initial magnitude of the threshold. Signal standard deviation is used to match the fluctuation characteristics of the input signal; for As an adaptive attenuation threshold term, the larger the signal amplitude, the smaller the threshold. This ensures that small-amplitude noise is fully filtered out while avoiding excessive compression of large-amplitude effective signals, thus solving the signal distortion problem caused by the constant threshold in traditional threshold functions.

[0154] The hybrid threshold strategy execution submodule also includes: The core parameters of wavelet decomposition are set to ensure the compatibility of the decomposition with the signal characteristics; the sym8 wavelet basis is used for decomposition, which has good compact support and symmetry, and can match the waveform characteristics of low-frequency signals in concrete flaw detection ultrasound, reducing signal energy leakage during the decomposition process; the formula for calculating the number of decomposition layers L is:

[0155] Where N is the signal length of the low-to-medium frequency IMF component. () is the rounding function. Engineering verification shows that this number of layers can achieve a balance between decomposition accuracy and computational efficiency.

[0156] Furthermore, after performing wavelet decomposition on the mid- and low-frequency IMF components, a differentiated threshold strategy is adopted based on the noise distribution characteristics of different frequency sub-bands: For high-frequency sub-bands, the SURE (Stein unbiased risk estimation) threshold is used. This threshold can be adaptively estimated using statistical methods, making it suitable for high-frequency sub-bands with complex noise distributions. It can maximize noise filtering while minimizing the signal estimation risk. For low-frequency sub-bands, the Minimax threshold is used. This threshold is the optimal threshold with a minimum maximum value, which can ensure the integrity of the effective signal in the low-frequency sub-band and avoid the loss of fine structural features of concrete due to excessively high thresholds. It is also suitable for the characteristics of high effective signal ratio and low noise ratio in the low-frequency sub-bands.

[0157] Furthermore, the wavelet coefficients after thresholding are reconstructed to recover the low-frequency signal in the time domain. The specific expression is as follows:

[0158] in, These are the wavelet coefficients after thresholding. This is the layer index for wavelet decomposition. It is a sym8 wavelet basis. Let i be the denoised signal of the i-th mid-to-low frequency IMF component, where i takes the value from 4 to 8 (corresponding to IMF4-IMF8 components).

[0159] Furthermore, the multimodal signal reconstruction and post-processing module 4 includes a high-frequency VMD component integration submodule, a full-band signal fusion submodule, and a post-processing submodule; The high-frequency VMD component integration submodule also includes: The high-frequency modal components obtained from VMD decomposition are summed along the modality number dimension, and then redundant dimensions are eliminated. The specific expression is as follows:

[0160] in, This represents the m-th low-frequency IMF component after wavelet denoising (m is 4-8). Let be the number of modes in the VMD decomposition. The result of VMD decomposition is a three-dimensional matrix with dimension 1. (N is the signal length, and 3 corresponds to 3 high-frequency IMF components;) This indicates summation along the VMD modality number dimension; This is a dimension compression operator used to eliminate redundant dimensions of length 1 after summation, ultimately... Dimensions .

[0161] The full-band signal fusion submodule also includes: After vertically concatenating the integrated high-frequency components with the low-frequency components from wavelet denoising, the summation is performed along the component dimensions to obtain the complete denoised signal. The specific expression is as follows:

[0162] in, For a one-dimensional complete time-domain denoised signal, ; The spliced ​​full-band component matrix is ​​expressed as follows:

[0163] in, This represents the m-th low-frequency IMF component after wavelet denoising (m is 4-8). It is a one-dimensional complete time-domain denoised signal.

[0164] The post-processing submodule also includes: The reconstructed one-dimensional time-domain denoised signal is restored to a two-dimensional matrix with the same dimensions as the original input image. The specific expression is as follows:

[0165] in, This is the one-dimensional time-domain denoised signal after multimodal fusion; Let H be the two-dimensional dimension of the original ultrasonic flaw detection image of concrete (H×W, where H is the image height and W is the image width); reshape() is the dimension reshaping operator. This is the two-dimensional image matrix after dimension restoration, with dimensions of H×W.

[0166] Furthermore, median filtering is performed on the dimension-restored image to eliminate residual isolated noise points, as expressed by:

[0167] in, It is a two-dimensional median filter operator with a 3×3 window. It replaces the center pixel with the median of the pixels in the window to suppress residual speckle noise and preserve defect edges. This is for the final output denoised and enhanced ultrasound image.

[0168] Furthermore, the DICOM standard access and real-time video processing module 5 includes: The DICOM file parameter reading function calls the `dicom_interface` interface. The `dicominfo` function reads the header information of the input DICOM file, and the `dicomread` function extracts the ultrasound image data. Simultaneously, it automatically reads the probe parameters from the file header information and calculates the signal sampling frequency.

[0169] The probe frequency adaptive parameter adjustment adjusts the core parameters of CEEMDAN, VMD, and wavelet denoising based on the probe center frequency (5-15MHz range).

[0170] The standard enhanced image output will complete the denoising signal after multimodal reconstruction and median filtering post-processing, and encapsulate it into a DICOM standard format image; at the same time, it supports three-dimensional ultrasound volume reconstruction based on the image, realizing three-dimensional imaging with isotropic resolution of 0.4mm, and seamlessly connecting with AI diagnostic systems.

[0171] Furthermore, the `realtime_processing` function is called to access the target video source via the `VideoReader` interface, entering a frame-by-frame reading process to acquire the original image frames of the ultrasound video one by one. For single-frame denoising, the `ultrasound_denoising` function is called for each image frame, with parameters configured by default at a 15MHz probe frequency. This sequentially completes CEEMDAN preprocessing, VMD optimization, wavelet denoising, multimodal reconstruction, and post-processing to obtain the denoised image frame.

[0172] An integrated solution was built based on the OpenCV framework. The process_frame function was defined, and tools such as pywt were integrated to realize CEEMDAN-VMD parallel computing. The processed image frames were output and displayed in real time, thus completing the real-time noise reduction and presentation of ultrasound video.

[0173] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for suppressing speckle noise in ultrasonic images used for concrete flaw detection based on multimodal signal processing, characterized in that, Includes the following steps: Step S1, the CEEMDAN preprocessing step, in which: The original ultrasonic signal from concrete flaw detection was injected with adaptive attenuated white noise multiple times and EMD decomposition was performed. The IMF component without mode mixing was obtained by mode alignment. The component classification was completed by improving the kurtosis criterion, and the high-frequency noise-dominant component and the mid-to-low frequency effective signal component were identified. Step S2, the VMD optimization step, in which: The high-frequency noise dominant component obtained in step S1 is decomposed in a second step to construct a variational constraint optimization model. The alternating direction multiplier method is used to solve the model iteratively. The frequency band of the high-frequency component is divided by combining the optimal parameters verified by engineering flaw detection data. Step S3, the wavelet joint denoising step, in which: For the effective mid-low frequency signal components obtained in step S1, wavelet decomposition is performed based on the sym8 wavelet basis and the number of decomposition layers is set; the signal is processed by an adaptive threshold function, and the coefficients of the high-frequency subband after wavelet decomposition are processed by a hierarchical hybrid threshold strategy of using SURE threshold and Minimax threshold for the low-frequency subband; the denoised mid-low frequency IMF components are reconstructed from the processed wavelet coefficients. Step S4, the multimodal signal reconstruction and post-processing step, in which: The high-frequency components processed in step S2 are summed along the modality number dimension and redundant dimensions are eliminated. They are then longitudinally spliced ​​with the denoised low- and mid-frequency IMF components obtained in step S3 and summed along the component dimensions to generate a one-dimensional complete time-domain denoised signal. This signal is then reshaped into a two-dimensional matrix with the same dimension as the original ultrasonic image. Residual isolated noise points are eliminated by two-dimensional median filtering, and a denoised and enhanced concrete ultrasonic image is output.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: Step S101, the noise injection and decomposition step, in which: The total number of noise injections is set to 100. Simultaneously, based on the characteristics of the original ultrasonic signal x, the attenuation coefficient of the adaptive noise is calculated using the following formula: in, is the attenuation coefficient for adaptive noise, and 0.2 is the fundamental noise intensity coefficient. Let x be the standard deviation of the original signal x, and t be the number of time-domain sampling points of the signal; In each loop, attenuated white noise is superimposed on the original signal to obtain a noisy signal; conventional EMD decomposition is performed on the noisy signal to obtain a set of temporary IMF components; this process is repeated 100 times to obtain 100 sets of IMF components corresponding to different noisy signals. Modal alignment is performed on all temporary IMF components. IMF components of the same level are matched among the 100 components in descending frequency order. Then, the average value of the IMF components at the same level is calculated to obtain the CEEMDAN decomposed IMF components. The specific formula is as follows: in, This represents the final k-th order IMF component; Let k be the temporary IMF component of order k obtained from the nth iteration decomposition; k is the order index of the IMF component. During the screening process of a single EMD decomposition, the continuous screening standard deviation (SD) of two adjacent screening results is monitored in real time. When SD < 0.3, it is determined that the current IMF component has been completely decomposed, and the screening iteration of that component is terminated immediately, and the decomposition of the next IMF component is carried out until all components are extracted. Step S102, the noise-driven IMF identification step, in which: For each IMF component obtained from CEEMDAN decomposition, calculate its kurtosis coefficient using the following formula: in, is the kurtosis coefficient of the i-th order IMF component; The function is used to calculate the mathematical expectation. For the i-th order IMF component; Let be the mean of the i-th order IMF component; Let be the standard deviation of the i-th order IMF component. Let be the fourth central moment of the i-th IMF component; To filter noise-dominant components, a kurtosis threshold K>3.5 is set. IMF components with kurtosis values ​​exceeding this threshold are identified as noise-dominant IMF components and sent to the VMD optimization layer for secondary precise decomposition. IMF components with kurtosis values ​​below the threshold are identified as mid-to-low frequency components containing more effective signals and will be processed by the wavelet joint denoising layer.

3. The method according to claim 1, characterized in that, Step S2 includes the following steps: Step S201, the steps for solving the variational problem, in which: A variational optimization objective model is established to minimize the bandwidth and sparsity constraints of each modal component. The specific expression is as follows: in, Let be the set of VMD modal components to be solved. This is the set of center frequencies corresponding to each modal component; For a partial differential operator with respect to time t; It is the Fourier frequency shift factor; The square of the L2 norm; As a penalty factor; for The frequency domain expression; It is an L1 norm; The center frequencies of each VMD modal component are initialized. Based on the initialized center frequencies, the variational optimization model is iteratively solved using the alternating direction multiplier method. During the iteration process, the time-domain expression and frequency-domain center frequency of each modal component are updated sequentially to gradually approach the optimal solution of the variational model. After each iteration, the convergence of the solution results is determined. The iteration terminates when the convergence condition is met. The convergence determination formula is as follows: in, This represents the frequency domain value of the k-th modal component in the (n+1)-th iteration. This represents the frequency domain value of the k-th modal component in the nth iteration. It is an L2 norm; The threshold for determining convergence; Step S202, the step of optimizing engineering flaw detection parameters, in which: Based on grid search verification using simulated flaw detection data of non-compact concrete, the optimal parameter combination for VMD decomposition is determined.

4. The method according to claim 1, characterized in that, Step S3 includes the following steps: Step S301, the step of improving the threshold function implementation, in which: Design an exponentially decaying adaptive threshold function, the specific expression of which is: in, The input is a sequence of low-to-medium frequency IMF component signals; It is a symbolic function; The core logic for threshold processing is to retain the portion of the signal that exceeds the adaptive threshold when the absolute value of the signal is greater than the threshold, and otherwise set it to 0. Signal standard deviation; For the adaptive attenuation threshold term, the larger the signal amplitude, the smaller the threshold; Step S302, the step of executing the hybrid threshold strategy, in which: The core parameters for wavelet decomposition are set, and the mid-to-low frequency IMF components are decomposed using the sym8 wavelet basis. A differentiated thresholding strategy is adopted to address the noise distribution characteristics of different frequency sub-bands: the SURE threshold is used for high-frequency sub-bands, and the Minimax threshold is used for low-frequency sub-bands. The wavelet coefficients after thresholding are then reconstructed, with the specific expression as follows: in, These are the wavelet coefficients after thresholding. This is the layer index for wavelet decomposition. It is a sym8 wavelet basis. Let i be the denoised signal of the i-th low-frequency IMF component, where i can be 4 to 8.

5. The method according to claim 1, characterized in that, Step S4 includes the following steps: Step S401, the high-frequency VMD component integration step, in which: The high-frequency modal components obtained from VMD decomposition are summed along the modality number dimension, and then redundant dimensions are eliminated. The specific expression is as follows: in, This represents the m-th low-frequency IMF component after wavelet denoising. Let be the number of modes in the VMD decomposition. The result of VMD decomposition is a three-dimensional matrix with dimension 1. ; This indicates summation along the VMD modality number dimension; For dimension compression operators; Step S402, the full-band signal fusion step, in which: After vertically splicing the integrated high-frequency components with the low-frequency components from wavelet denoising, the complete denoised signal is obtained by summing along the component dimensions. Step S403, post-processing step, in which: The reconstructed one-dimensional time-domain denoised signal is restored to a two-dimensional matrix with the same dimensions as the original input image.

6. The method according to claim 4, characterized in that, Step S3 further includes: In setting the core parameters of wavelet decomposition, the formula for calculating the number of decomposition levels L is as follows: Where N is the signal length of the low-to-medium frequency IMF component. () is the rounding function.

7. A speckle noise suppression system for ultrasonic images used in concrete flaw detection based on multimodal signal processing, characterized in that, It includes a CEEMDAN preprocessing module, a VMD optimization module, a wavelet joint denoising module, and a multimodal signal reconstruction and post-processing module; The CEEMDAN preprocessing module, in which: For the original ultrasonic signal of concrete flaw detection, adaptive attenuation white noise was injected multiple times and EMD decomposition was performed. After mode alignment and averaging, IMF components without mode mixing were obtained. Then, the improved kurtosis criterion was used to complete component classification and identify the high-frequency noise-dominant component and the mid-to-low frequency effective signal components. The VMD optimization module, in which: The high-frequency noise dominant component is decomposed in a secondary manner, a variational constraint optimization model is constructed, and the alternating direction multiplier method is used for iterative solution. The optimal parameters verified by engineering flaw detection data are combined to complete the accurate division of the frequency band of the high-frequency component. The wavelet joint denoising module, in which: For the obtained mid-to-low frequency effective signal components, wavelet decomposition is performed using the sym8 wavelet basis and the number of decomposition levels is determined. The signal is processed by an exponential decay adaptive threshold function. For the high-frequency subband after wavelet decomposition, a layered hybrid threshold strategy is used to process the coefficients, which is SURE threshold for the high-frequency subband and Minimax threshold for the low-frequency subband. The processed wavelet coefficients are then reconstructed to obtain the denoised mid-to-low frequency IMF components. The multimodal signal reconstruction and post-processing module, in which: The processed high-frequency components are summed along the modality number dimension and redundant dimensions are eliminated. They are then longitudinally concatenated with the denoised mid-to-low frequency IMF components and summed along the component dimensions to generate a one-dimensional complete time-domain denoised signal. This signal is then reshaped into a two-dimensional matrix with the same dimensions as the original ultrasonic image. Residual isolated noise points are then eliminated by two-dimensional median filtering to output a denoised and enhanced concrete ultrasonic image.

8. The system according to claim 7, characterized in that, The system also includes an initialization setup module and a DICOM standard access and real-time video processing module; The initialization setting module, in which: The ultrasound testing equipment supports the DICOM3.0 medical image transmission protocol and the probe operates in the 5-15MHz range. It receives the original ultrasound image and the probe sampling frequency as input, calculates the CEEMDAN noise standard deviation parameter, VMD adaptive penalty factor and wavelet decomposition level in sequence, and finally outputs a set of standardized parameters. The DICOM standard access and real-time video processing module, in which: The system reads the header information of the input DICOM file and extracts the ultrasound image data. It reads the probe parameters from the file header information, calculates the signal sampling frequency, and adjusts the core parameters of the CEEMDAN preprocessing module, VMD optimization module, and wavelet joint denoising module according to the read probe center frequency. It calls a function to access the target video source, enters the frame loop reading process, and acquires the original image frames of the ultrasound video frame by frame.

9. The system according to claim 8, characterized in that, The initialization settings module specifically includes: The formula for calculating the CEEMDAN noise standard deviation parameter is: in, is the L2 norm of the one-dimensional pixel sequence of the original ultrasound image, and 0.2 is the engineering-verified fundamental noise intensity coefficient. The standard deviation parameter for CEEMDAN noise; The formula for calculating the VMD adaptive penalty factor is: in, The sampling frequency of the ultrasound probe. The reference frequency coefficient is 2000, which is the optimal penalty factor reference value verified by 200 engineering flaw detection data. When the probe frequency is 10MHz, the penalty factor is the reference value 2000. When the frequency is higher or lower than 10MHz, the penalty factor increases or decreases accordingly. The VMD adaptive penalty factor; The formula for calculating the number of wavelet decomposition levels is: Where N is the signal length of the low-to-medium frequency IMF component. () is the binary logarithm of the signal length. This is the rounding function. The wavelet decomposition level; By integrating the three parameters mentioned above—CEEMDAN noise standard deviation, VMD adaptive penalty factor, and wavelet decomposition level—a complete set of initialization parameters is output.

10. The system according to claim 8, characterized in that, The DICOM standard access and real-time video processing module includes: The dicominfo function reads the header information of the input DICOM file, and the dicomread function extracts the ultrasound image data; at the same time, the probe parameters are automatically read from the header information, and the signal sampling frequency is calculated. The denoised signal, after multimodal reconstruction and median filtering, is encapsulated into a DICOM standard format image; three-dimensional ultrasound volume reconstruction is then performed based on the DICOM standard format image. The target video source is accessed through the VideoReader interface, and the frame loop reading process is entered to acquire the original image frames of the ultrasound video frame by frame. For single-frame denoising, the ultrasound_denoising function is called for each image frame, the probe frequency configuration parameter is set to 15MHz, and CEEMDAN preprocessing, VMD optimization, wavelet denoising, multimodal reconstruction and post-processing are performed in sequence to obtain the denoised image frame.

Images