An ultrasonic TFM detection image intelligent correction method and device fusing physical priori and a lightweight network
Patent Information
- Application Number
- CN202610689242.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-19
- Publication Date
- 2026-09-11
AI Technical Summary
[0008]本申请实施例的目的在于提供一种融合物理先验与轻量化网络的超声TFM检测图像智能校正方法及装置，以解决现有技术中图像校正方法违背声学物理特性、依赖标注数据、抗噪能力差且开销大，难以校正无编码器手工扫查造成超声TFM图像采样失真的技术问题
[0019]The beneficial effects of this application are as follows: This application provides an intelligent correction method and device for ultrasonic TFM inspection images that integrates physical priors and lightweight networks. It acquires raw ultrasonic C-scan images obtained through manual scanning without an encoder, preprocesses and normalizes them to obtain normalized images, and performs multi-scale discrete wavelet decomposition to achieve frequency domain decoupling of defect acoustic response, mechanical deformation edges, and speckle noise. This effectively isolates the interference of grain scattering noise on distortion feature extraction, improving the accuracy and noise resistance of feature extraction. Simultaneously, a lightweight one-dimensional constraint network is constructed and trained using a zero-sample domain to obtain a mapping prediction model, eliminating the need for real data annotation. This model performs forced global dimensionality reduction compression along the width axis on the fused two-dimensional feature tensor obtained by concatenating the normalized image and the two-dimensional feature tensor, outputting only a one-dimensional transverse coordinate mapping sequence representing the mechanical scanning displacement law. This avoids the width-direction artifacts and defect topology tearing problems that are prone to occur in traditional two-dimensional generative networks, ensuring the physical consistency of the correction process. Furthermore, the mapping prediction model constructed in this application uses the ReLU activation function and bias term to jointly construct a forced monotonically increasing one-dimensional lateral coordinate mapping sequence, and is trained using a loss function that incorporates acoustic physics priors. Even in the absence of local texture features, it can still converge in a direction that conforms to the propagation law of ultrasound beams, significantly improving the model's generalization and correction accuracy. Finally, the original ultrasound C-scan image is reversed spatially resampled based on the one-dimensional lateral coordinate mapping sequence, accurately restoring the spatial sampling offset caused by scanning speed fluctuations, thus restoring the defect geometry to its physical reality. This significantly reduces computational overhead, enables end-side deployment, and ensures the geometric fidelity and acoustic consistency of the corrected ultrasound image.
Smart Images
Figure CN122736896A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of nondestructive testing technology, and more specifically, it relates to an intelligent correction method and device for ultrasonic TFM test images that integrates physical priors and lightweight networks. Background Technology
[0002] Totally Focused Ultrasonic (TFM) provides extremely high-resolution images of internal defects. In continuous scanning imaging, TFM typically relies heavily on mechanical position encoders to record probe displacement in real time to maintain uniformity in image spatial sampling. However, in real-world industrial non-destructive testing environments, there are extremely common scenarios where encoders are unsuitable or limited. For example, encoders are difficult to install in confined, narrow spaces (such as fillet welds on pipe fittings or densely packed tube bundles); wheels are prone to suspending or slipping when scanning complex, irregular curved surfaces (such as reducing pipes or irregularly shaped forgings); and precision photoelectric encoders are easily damaged in high-temperature, highly corrosive, or underwater testing environments.
[0003] In the aforementioned scenarios, inspectors are often forced to resort to manual scanning without an encoder. However, due to factors such as arm muscle fatigue and abrupt changes in contact surface coupling damping (e.g., encountering weld reinforcement or dried coupling agent), the probe's scanning speed is prone to stuttering, rapid advancement, or local stagnation. This non-uniform displacement directly leads to severe nonlinear sampling distortion (stretching or compression distortion) in the ultrasonic TFM image along the scanning direction, resulting in misjudgments of defect size and location (e.g., cracks, lack of fusion), severely weakening the quantitative evaluation capability of ultrasonic TFM technology in harsh engineering environments.
[0004] In recent years, some studies have attempted to introduce deep learning methods based on two-dimensional convolutional neural networks (CNNs) or generative adversarial networks (GANs) to correct such distorted images, but these methods have the following significant drawbacks when applied in practical engineering: First, it violates the physical properties of acoustics: existing two-dimensional networks treat ultrasound images as ordinary visual matrices, ignoring the physical law that the lateral probe motion and the longitudinal ultrasound time-of-flight (TOF) are orthogonal and independent. This makes it highly susceptible to erroneously altering the longitudinal width coordinates during pixel reconstruction, leading to artifacts and tearing of the defective topology.
[0005] Secondly, it heavily relies on real-world labeled data: training existing models requires massive amounts of distorted-undistorted image pairs. Obtaining precisely aligned paired data in industrial settings is extremely costly, resulting in very poor generalization ability of the model when faced with unknown scanning habits or workpieces.
[0006] In addition, it has poor noise resistance and high computational cost: existing networks are unable to effectively extract mechanical deformation features under strong ultrasonic speckle noise (such as scattering from coarse-grained materials); and the large number of parameters in the massive two-dimensional model makes it impossible to achieve real-time end-side correction on portable non-destructive testing instruments with limited computing power.
[0007] Therefore, there is an urgent need to provide a new intelligent image correction method to solve the above problems. Summary of the Invention
[0008] The purpose of this application is to provide an intelligent correction method and device for ultrasound TFM detection images that integrates physical priors and lightweight networks, so as to solve the technical problems in the prior art where image correction methods violate acoustic physical characteristics, rely on labeled data, have poor noise resistance and high overhead, and are difficult to correct the sampling distortion of ultrasound TFM images caused by manual scanning without encoders.
[0009] To achieve the above objectives, a first aspect of this application provides an intelligent correction method for ultrasound TFM detection images that integrates physical priors and lightweight networks, comprising the following steps: Construct and train a one-dimensional constraint network to obtain a mapping prediction model; The original ultrasound C-scan images obtained by manual scanning without encoder are acquired, and after preprocessing and normalization, normalized images are obtained. Low-frequency approximate components and high-frequency detail components are extracted by multi-scale discrete wavelet decomposition and then stitched together to obtain a two-dimensional feature tensor. The normalized image and the two-dimensional feature tensor are concatenated to obtain a fused two-dimensional feature tensor, which is then input into the mapping prediction model to reduce the dimension along the width axis to a one-dimensional horizontal coordinate mapping sequence. The training is optimized by combining the acoustic physics prior loss function. Based on the remapping function of the one-dimensional horizontal coordinate mapping sequence, the original ultrasound C-scan image is inversely spatially resampled to obtain the corrected ultrasound image.
[0010] Preferably, the process of training a one-dimensional constraint network includes: extracting distortion-free ultrasonic acoustic response images as standard defect templates, introducing a manual scanning dynamics model that integrates ergonomic low-frequency sinusoidal modulation and Markov random walk laws, and generating a non-uniform coordinate increment sequence characterizing the transverse scanning speed fluctuation of the probe. A training set is constructed by applying physical deformation and imaging speckle perturbation to a standard defect template based on a non-uniform coordinate increment sequence. Combined with a physical prior loss function broadened based on the ultrasonic B-scan point diffusion function theory, a one-dimensional constraint network is trained in the zero-sample domain to obtain a mapping prediction model.
[0011] Preferably, the one-dimensional constraint network includes a two-dimensional feature extraction layer, an anisotropic pooling layer, and a one-dimensional mapping regression layer; A two-dimensional feature extraction layer is used to extract two-dimensional feature tensors; Anisotropic pooling layers are used to globally compress and reduce the dimensionality of the two-dimensional feature tensor along the width axis of the probe array to obtain a one-dimensional feature sequence. The one-dimensional mapping regression layer is used to estimate the displacement deformation of the one-dimensional feature sequence to obtain a one-dimensional lateral coordinate mapping sequence.
[0012] Preferably, before stitching to obtain a two-dimensional feature tensor, it is necessary to upsample the low-frequency approximation component and the high-frequency detail component, align the spatial resolution of the low-frequency approximation component and the high-frequency detail component with the normalized image, and obtain the aligned physical feature components.
[0013] Preferably, the normalized image is concatenated with the aligned physical feature components to obtain a two-dimensional feature tensor, as shown in the formula: ; In the formula, For a two-dimensional feature tensor, For splicing, For normalized images, This is a low-frequency approximation component. For vertical high-frequency components, and These represent the number of sampling points in the horizontal and vertical directions of the image, respectively.
[0014] Preferably, the process of obtaining a one-dimensional lateral coordinate mapping sequence includes: reducing the dimensionality of the fused two-dimensional feature tensor to a one-dimensional sequence, outputting the predicted lateral coordinate increments, performing cumulative integration, and normalizing to the global physical scan length to obtain a one-dimensional lateral coordinate mapping sequence.
[0015] Preferably, the acoustic physics prior loss function formula is: ; In the formula, For the overall loss function, For coordinate mapping loss function, The coordinate increment loss function, This is a physical penalty term for the equivalent width of the main lobe based on B-scan simulation. , , These are dynamic weighting coefficients.
[0016] Preferably, the formula for obtaining the corrected ultrasound image is: ; In the formula, The corrected C-scan image, The grid consists of equally spaced physical coordinates. The coordinates of pixels with fixed equal spacing along the width direction of the probe array. This is the original, uncorrected ultrasound C-scan image. It is the inverse function of the monotonic mapping function.
[0017] Preferably, the formula for constructing the mapping prediction model is: ; In the formula, It is a horizontal mapping function. Sampling points The incremental sequence at that point, This is the index of discrete sampling points along the lateral scanning direction.
[0018] A second aspect of this application provides an intelligent correction device for ultrasound TFM detection images that integrates physical priors and lightweight networks, including: Ultrasonic phased array equipment, flaw detector main unit, and display; Ultrasonic phased array equipment is used to acquire raw ultrasonic C-scan images of the inside of a workpiece under manual scanning conditions without a position encoder, and transmit them to the flaw detector host. The flaw detector host includes a memory and a processor. The memory and processor are connected in communication and are used to store the computer program pre-burned in and the mapping prediction model that has been trained offline. The processor receives the original ultrasound C-scan image, preprocesses and normalizes it to obtain a normalized image, extracts low-frequency approximate components and high-frequency detail components through multi-scale discrete wavelet decomposition, and concatenates them to obtain a two-dimensional feature tensor. The normalized image and the two-dimensional feature tensor are concatenated to obtain a fused two-dimensional feature tensor, which is input into the mapping prediction model to reduce the dimension along the width axis to a one-dimensional horizontal coordinate mapping sequence. The model is then optimized and trained using an acoustic-physical prior loss function. Based on the remapping function of the one-dimensional horizontal coordinate mapping sequence, the original ultrasound C-scan image is inversely spatially resampled to obtain a corrected ultrasound image. A display is used to receive and present the corrected ultrasound images in real time.
[0019] The beneficial effects of this application are as follows: This application provides an intelligent correction method and device for ultrasonic TFM inspection images that integrates physical priors and lightweight networks. It acquires raw ultrasonic C-scan images obtained through manual scanning without an encoder, preprocesses and normalizes them to obtain normalized images, and performs multi-scale discrete wavelet decomposition to achieve frequency domain decoupling of defect acoustic response, mechanical deformation edges, and speckle noise. This effectively isolates the interference of grain scattering noise on distortion feature extraction, improving the accuracy and noise resistance of feature extraction. Simultaneously, a lightweight one-dimensional constraint network is constructed and trained using a zero-sample domain to obtain a mapping prediction model, eliminating the need for real data annotation. This model performs forced global dimensionality reduction compression along the width axis on the fused two-dimensional feature tensor obtained by concatenating the normalized image and the two-dimensional feature tensor, outputting only a one-dimensional transverse coordinate mapping sequence representing the mechanical scanning displacement law. This avoids the width-direction artifacts and defect topology tearing problems that are prone to occur in traditional two-dimensional generative networks, ensuring the physical consistency of the correction process. Furthermore, the mapping prediction model constructed in this application uses the ReLU activation function and bias term to jointly construct a forced monotonically increasing one-dimensional lateral coordinate mapping sequence, and is trained using a loss function that incorporates acoustic physics priors. Even in the absence of local texture features, it can still converge in a direction that conforms to the propagation law of ultrasound beams, significantly improving the model's generalization and correction accuracy. Finally, the original ultrasound C-scan image is reversed spatially resampled based on the one-dimensional lateral coordinate mapping sequence, accurately restoring the spatial sampling offset caused by scanning speed fluctuations, thus restoring the defect geometry to its physical reality. This significantly reduces computational overhead, enables end-side deployment, and ensures the geometric fidelity and acoustic consistency of the corrected ultrasound image. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application, 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A schematic diagram of the overall process of an intelligent correction method for ultrasonic TFM detection images that integrates physical priors and lightweight networks, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the hardware architecture of an intelligent correction device for ultrasonic TFM detection images that integrates physical priors and lightweight networks, provided as an embodiment of this application.
[0022] In the diagram: 200. Flaw detector main unit; 201. Ultrasonic phased array device; 202. Processor; 203. Memory; 204. Display. Detailed Implementation
[0023] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.
[0024] Please see Figure 1 The first embodiment of this application provides an intelligent correction method for ultrasound TFM detection images that integrates physical priors and lightweight networks, comprising: S1: Construct and train a one-dimensional constraint network to obtain a mapping prediction model.
[0025] First, a one-dimensional constraint network adapted to the dual-axis orthogonal physical characteristics of ultrasound C-scan is constructed. The specific construction process involves building a lightweight convolutional neural network architecture consisting of a two-dimensional feature extraction layer, an anisotropic pooling layer, and a one-dimensional mapping regression layer. The network extracts a two-dimensional feature tensor through the two-dimensional feature extraction layer, then performs global compression and dimensionality reduction along the probe array width axis through the anisotropic pooling layer, converting the two-dimensional feature tensor into a one-dimensional feature sequence. Subsequently, the one-dimensional mapping regression layer estimates displacement deformation only along the one-dimensional dimension of the lateral manual scan. This network architecture structurally constrains the spatial coordinates of the probe array width dimension to remain constant, strictly conforming to the objective law that the physical spacing between array elements within the probe is fixed and independent of the lateral manual scan speed fluctuations.
[0026] Next, distortion-free ultrasonic acoustic response images are extracted as standard defect templates. A manual scanning dynamics model that integrates ergonomic low-frequency sinusoidal modulation and Markov random walk laws is introduced to generate a non-uniform coordinate increment sequence characterizing the transverse scanning speed fluctuation of the probe.
[0027] A training set is constructed by applying physical deformation and imaging speckle perturbation to a standard defect template based on incremental sequences. Combined with a physical prior loss function broadened based on the ultrasonic B-scan point diffusion function theory, a one-dimensional constraint network is trained in the zero-shot domain to obtain a trained mapping prediction model.
[0028] Specifically, in the training phase of the one-dimensional constrained network, the training data is generated by a manual scanning dynamics model incorporating ergonomics, and the steps are as follows: The distortion-free acoustic response image from the standard test block is used as the standard defect template. ,in For physically real uniform spatial coordinates, These are the coordinates of pixels with fixed equal spacing along the width direction of the probe array.
[0029] To realistically simulate the mechanical characteristics of manual scanning operations, such as rapid propulsion, local stagnation, and periodic jerking caused by hand muscle fatigue and uneven damping of the coupling agent at the contact surface, a dynamic state equation integrating low-frequency sinusoidal modulation and Markov random walk principles is introduced to generate sampling points. Positive horizontal coordinate increment sequence at the location The formula is: ; In the formula, Based on the shift step size, and These represent the amplitude and frequency, respectively, which characterize the periodic force fluctuations of the human arm. To simulate Markov walk noise caused by local stuttering and sudden changes in sliding friction, To ensure that the scanning direction is irreversible, a very small physical constant is required.
[0030] Based on sampling points Incremental sequence at the location Construct a forced monotonically increasing transverse mapping function : ; In the formula, It is a horizontal mapping function. This is the index of discrete sampling points along the lateral scanning direction.
[0031] Furthermore, ultrasound imaging perturbations, including amplitude fluctuations and speckle variations, are superimposed on this data to generate training data with lateral non-uniform distortion. : ; In the formula, A standard defect template image without distortion. It is a horizontal mapping function. These are the discrete pixel spatial coordinates (i.e., the distortion generation axis) along the manual scanning direction. The coordinates of pixels with fixed equal spacing along the width direction of the probe array. The Rayleigh speckle noise perturbation is introduced. Through the above dynamic simulation, this application completely eliminates the dependence on real scan coordinate annotation data, achieves high generalization training of the network under zero-shot conditions, and obtains a fully trained mapping prediction model.
[0032] S2: Obtain the original ultrasound C-scan image obtained by manual scanning without an encoder. After preprocessing and normalization, a normalized image is obtained. Low-frequency approximate components and high-frequency detail components are extracted by multi-scale discrete wavelet decomposition and then stitched together to obtain a two-dimensional feature tensor.
[0033] Raw ultrasound C-scan images containing non-uniform spatial sampling distortion were acquired under manual scanning conditions without an encoder. , is represented as: ; In the formula, These are the discrete pixel spatial coordinates (i.e., the distortion generation axis) along the manual scanning direction. The coordinates of pixels with fixed equal spacing along the width direction of the probe array. and These represent the number of sampling points in the horizontal and vertical directions of the image, respectively.
[0034] Acquire raw ultrasound C-scan images Then, amplitude envelope extraction and threshold truncation operations are sequentially performed on the original ultrasound C-scan image to eliminate the phase reversal phenomenon of the ultrasound bipolar radio frequency signal and suppress extreme bright speckle noise in the background area to obtain a normalized image.
[0035] Specifically, firstly, the original ultrasound C-scan image... Perform a Hilbert transform along the width axis of the probe array to extract the amplitude envelope of its analytical signal. The formula is: ; In the formula, For amplitude envelope, This is the original ultrasound C-scan image. For the longitudinal width axis The one-dimensional Hilbert transform, It is the imaginary unit.
[0036] Subsequently, the amplitude envelope was adjusted based on adaptive upper and lower limits. After threshold truncation and normalization, the pixel values are strictly mapped to... Normalized image of the interval The formula is: ; In the formula, and These represent the upper and lower limits of amplitude cutoff used to suppress extreme high-brightness speckle and filter underlying electrical noise, respectively.
[0037] Next, the normalized image Multi-scale discrete wavelet decomposition is performed to decouple acoustic features from physical properties in the frequency domain: low-frequency approximate components characterizing the macroscopic acoustic response profile of defects are extracted, as well as high-frequency detail components characterizing the local geometric stretching or compression edges caused by non-constant scanning, thereby effectively isolating mechanical deformation information from ultrasonic speckle noise during the feature extraction stage.
[0038] Specifically, for normalized images A two-dimensional discrete wavelet decomposition (2D-DWT) with L layers is performed to obtain low-frequency approximate components and high-frequency detail components at various scales. Its core physical significance lies in using wavelet transform to separate the disordered speckle noise caused by grain scattering from the true defect master response in the frequency domain, enabling subsequent networks to accurately pinpoint the lateral edge stretching and compression features caused by manual non-uniform scanning. The decomposition process is represented as follows: ; In the formula, To preserve the macroscopic acoustic response profile of the defect, the first Low-frequency approximation components of the layer, These represent the high-frequency detail components in the horizontal, vertical, and diagonal directions, respectively. represents the wavelet basis functions. Among them, the vertical high-frequency components... It represents edge information that exists along the width direction but changes drastically along the scanning direction, and is the core feature for capturing the lateral mechanical distortion rate.
[0039] Because multi-scale two-dimensional discrete wavelet decomposition involves downsampling of the spatial dimension, the extracted low-frequency approximate components... With vertical high frequency components The spatial resolution is lower than that of the original normalized image. Therefore, before stitching, it is necessary to perform an upsampling operation to restore the resolution to the same size as the normalized image, so as to achieve strict alignment of spatial dimensions and obtain the aligned physical feature components.
[0040] Normalized image The physical feature components aligned with the upsampled data (i.e., the retained low-frequency approximation components) and vertical high frequency components The concatenation is performed along the channel dimension. This operation effectively removes high-frequency speckle noise in the horizontal and diagonal directions, ultimately forming a two-dimensional fusion feature tensor used as input to the one-dimensional constrained network. ; In the formula, It is a two-dimensional feature tensor.
[0041] S3: The normalized image and the two-dimensional feature tensor are concatenated to obtain the fused two-dimensional feature tensor. This is then input into the mapping prediction model to reduce the dimension along the width axis to a one-dimensional horizontal coordinate mapping sequence. The training is optimized by combining the acoustic physics prior loss function. Based on the remapping function of the one-dimensional horizontal coordinate mapping sequence, the original ultrasound C-scan image is inversely spatially resampled to obtain the corrected ultrasound image.
[0042] Normalized image and its separated multi-scale two-dimensional feature tensors Channel splicing yields a fused two-dimensional feature tensor, which is then input into the mapping prediction model trained in S1. This mapping prediction model is essentially a physically inspired lightweight convolutional neural network architecture. Instead of directly generating a two-dimensional corrected image, it leverages the physical property that the transverse mechanical distortion of the probe and the longitudinal ultrasonic time of flight are independently orthogonal to force global dimensionality reduction compression of the input two-dimensional feature tensor along the longitudinal acoustic beam propagation axis (i.e., the width direction of the probe array), outputting only a one-dimensional transverse coordinate mapping sequence representing the mechanical scanning displacement along the transverse scanning space axis.
[0043] Specifically, the mapping prediction model reduces the dimensionality of the fused two-dimensional feature tensor to a one-dimensional sequence through an anisotropic global pooling layer along the z-axis, and outputs the predicted lateral coordinate increment. To ensure that the physical space mapping does not collapse or backtrack, the predicted lateral coordinate increment is... Perform cumulative integration to construct the predictive transverse mapping function. The formula is: ; In the formula, the ReLU activation function and the bias Together, they ensure the forced monotonically increasing nature of the mapping sequence. This will predict the lateral mapping function. Normalized to the global physical scan length, the final one-dimensional lateral coordinate mapping sequence is obtained. .
[0044] Furthermore, to ensure that the predicted one-dimensional lateral coordinate mapping sequence during network training simultaneously meets the requirements of numerical accuracy, procedural rationality, and physical consistency, the loss function for training the mapping prediction model innovatively incorporates an acoustic physics prior loss function, resulting in an overall loss function. Represented as: ; In the formula, For the overall loss function, For coordinate mapping loss function, The coordinate increment loss function, This is a physical penalty term for the equivalent width of the main lobe based on B-scan simulation. , , These are dynamic weighting coefficients.
[0045] ; ; ; In the formula, This represents the total number of sampling points along the manual scanning direction. To predict the lateral mapping function, For the true lateral mapping function, This is the one-dimensional horizontal coordinate mapping sequence predicted by the network. This represents the tiny physical distance actually moved laterally between two adjacent samples (two adjacent frames). This is a physical penalty term for the equivalent width of the main lobe based on B-scan simulation. This represents the physical width of the defect reconstructed under the current prediction mapping. The ideal main lobe width is calculated using ultrasonic diffraction theory. This physical loss forces the network to converge in a direction that conforms to the physical laws of ultrasonic beam propagation, even in the absence of local texture features.
[0046] Furthermore, a forced monotonically increasing remapping function is constructed based on a one-dimensional transverse coordinate mapping sequence. Based on this remapping function, the discrete distorted pixels along the scanning direction in the original ultrasound C-scan image are subjected to inverse coordinate system interpolation transformation. The spatial sampling offset caused by the scanning speed fluctuation is reset, and a corrected C-scan image is obtained that restores the geometric shape of the defect to its physical reality.
[0047] Specifically, the original ultrasound C-scan image is reversed spatially resampled based on a one-dimensional transverse coordinate mapping sequence to obtain the corrected C-scan image. The formula is: ; In the formula, The grid consists of equally spaced physical coordinates. This is the inverse function of the monotonic mapping function. In practical engineering deployments, monotonicity-preserving cubic Hermite spline interpolation (PCHIP) is used to complete the coordinate inverse offset, thereby resetting the spatial sampling offset. Since the mapping prediction model in this application only learns one-dimensional mappings, it eliminates the risks of width-direction artifacts and defective topology tearing commonly found in traditional two-dimensional generative adversarial networks (GANs).
[0048] Please see Figure 2 The present application provides a smart correction device for ultrasonic TFM detection images that integrates physical priors and lightweight networks, comprising: an ultrasonic phased array device 201, a flaw detector host 200, and a display 204.
[0049] The ultrasonic phased array device 201 includes a multi-element ultrasonic phased array probe and a multi-channel transceiver front end, which is used to acquire the original ultrasonic C-scan image inside the workpiece under the condition of manual scanning without a position encoder, and transmit it to the flaw detector host 200.
[0050] The flaw detector host 200 includes a memory 203 and a processor 202, wherein the memory 203 (e.g., a non-volatile storage medium such as Flash, RAM, or solid-state drive) is communicatively connected to the processor 202. The memory 203 contains a pre-programmed computer program and a mapping prediction model that has been trained offline.
[0051] The processor 202 is the intelligent processing core of the device. Specifically, it can be an embedded edge computing unit with built-in hardware accelerators (such as FPGA, DSP, or ARM architecture chip with NPU). It executes the mapping prediction model and inverse spatial resampling in real time on the edge through a forced dimensionality reduction architecture. The processor 202 executes the program in memory 203 to preprocess the received raw ultrasound C-scan image and perform multi-scale discrete wavelet decomposition. Then, it calls the mapping prediction model to perform forced dimensionality reduction prediction to obtain a one-dimensional lateral coordinate mapping sequence. Finally, it performs PCHIP interpolation to complete the physical coordinate reset of the distorted TFM image and obtain the corrected ultrasound image.
[0052] Display 204 is used to receive the corrected ultrasound images and to render and present distortion-free and width-artifact-free corrected ultrasound images to the inspectors in real time.
[0053] With the above hardware topology, since the mapping prediction model called by the processor 202 only performs one-dimensional sequence prediction, the consumption of hardware computing power and memory bandwidth is greatly reduced. This allows the device to achieve high frame rate end-side real-time intelligent correction on the low-power hardware of portable flaw detectors without relying on bulky high-end graphics cards (GPUs).
[0054] Example 1: Online detection and intelligent image correction process.
[0055] This embodiment takes the non-destructive testing of carbon steel welds, which is common in actual engineering, as an example to illustrate the online calibration workflow of this application under manual scanning without an encoder.
[0056] When the inspector manually scans the weld surface with a handheld ultrasonic phased array probe, the scanning speed fluctuates due to arm muscle tremors and uneven friction from the coupling agent. The portable instrument simultaneously acquires raw ultrasonic C-scan images exhibiting severe tensile / compression distortion.
[0057] The portable instrument's built-in processor first performs acoustic response-based preprocessing, extracts the amplitude envelope using Hilbert transform, and sets a lower limit for removing underlying electrical noise. and the upper limit of suppressing high-brightness speckle Normalize the signal to The interval eliminates phase reversal of bipolar signals.
[0058] Subsequently, the processor performs L-layer (preferably 3 layers) two-dimensional discrete wavelet decomposition on the normalized image. Since ultrasonic speckle noise is mostly concentrated in the diagonal high-frequency components, this step only extracts the low-frequency approximation components and the vertical high-frequency components, thereby physically separating the macroscopic contour features of the defect from the high-frequency speckle noise and enhancing noise resistance.
[0059] After feature extraction, the system calls a pre-trained mapping prediction model pre-loaded in internal memory. This model receives the concatenated two-dimensional feature tensor, compresses it along the width axis using a global pooling layer, and directly outputs a one-dimensional coordinate mapping sequence representing the probe's true lateral displacement. .
[0060] Finally, the system uses monotonic cubic Hermite spline interpolation (PCHIP) to perform inverse transverse resampling on the original ultrasound C-scan image, moving pixels only laterally to perfectly preserve the original width information of the defect, and outputs the morphology-corrected distortion-free ultrasound image on the monitor screen in real time.
[0061] Example 2: Zero-shot offline training method based on physical constraints.
[0062] Traditional deep learning struggles to acquire ultrasound pairing data with and without distortion. This application provides an unsupervised offline training method that relies entirely on physical simulation and dynamic models.
[0063] First, the ideal acoustic response of a standard test block is extracted as a standard template. Then, a dynamic state equation is constructed based on ergonomics. : ; In this embodiment, a basic shift step size is set. Approximately 0.5 times the spacing between probe elements, the periodic amplitude is set to target the human muscle fatigue cycle. fluctuation frequency The range is 0.5Hz to 2.0Hz; This is used to simulate the sudden jerking or jamming that occurs when the probe encounters weld reinforcement or dried-out coupling agent.
[0064] Based on the above equations, tens of thousands of distortion datasets that closely approximate real manual scanning were generated and fed into the mapping prediction model for training. During the training phase, an innovative PSF-based acoustic physical penalty term was introduced. When network prediction errors lead to over-compression of defects, the reconstructed defect width... It will be less than the ultrasonic diffraction limit. The loss function will impose a huge penalty, thus forcing the network to output a deformation field that conforms to the acoustic physical limits, achieving zero-sample high-fidelity training.
[0065] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0066] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for intelligent correction of ultrasound TFM detection images integrating physical priors and lightweight networks, characterized in that, Includes the following steps: Construct and train a one-dimensional constraint network to obtain a mapping prediction model; The original ultrasound C-scan images obtained by manual scanning without encoder are acquired, and after preprocessing and normalization, normalized images are obtained. Low-frequency approximate components and high-frequency detail components are extracted by multi-scale discrete wavelet decomposition and then stitched together to obtain a two-dimensional feature tensor. The normalized image and the two-dimensional feature tensor are concatenated to obtain a fused two-dimensional feature tensor. This fused tensor is then input into the mapping prediction model to reduce the dimension along the width axis to a one-dimensional horizontal coordinate mapping sequence. The model is then optimized and trained using an acoustic physics prior loss function. Based on the remapping function of the one-dimensional horizontal coordinate mapping sequence, the original ultrasound C-scan image is inversely spatially resampled to obtain the corrected ultrasound image.
2. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The process of training the one-dimensional constraint network includes: extracting distortion-free ultrasonic acoustic response images as standard defect templates, introducing a manual scanning dynamics model that integrates ergonomic low-frequency sinusoidal modulation and Markov random walk laws, and generating a non-uniform coordinate increment sequence that characterizes the transverse scanning speed fluctuation of the probe. The training set is constructed by applying physical deformation and imaging speckle perturbation to the standard defect template based on the non-uniform coordinate increment sequence, and by combining the physical prior loss function broadened based on the ultrasonic B-scan point diffusion function theory, the one-dimensional constraint network is trained in the zero-sample domain to obtain the mapping prediction model.
3. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The one-dimensional constraint network includes a two-dimensional feature extraction layer, an anisotropic pooling layer, and a one-dimensional mapping regression layer; The two-dimensional feature extraction layer is used to extract two-dimensional feature tensors; The anisotropic pooling layer is used to globally compress and reduce the dimensionality of the two-dimensional feature tensor along the width axis of the probe array to obtain a one-dimensional feature sequence. The one-dimensional mapping regression layer is used to estimate the displacement deformation of the one-dimensional feature sequence to obtain the one-dimensional lateral coordinate mapping sequence.
4. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, Before stitching together to obtain the two-dimensional feature tensor, it is necessary to upsample the low-frequency approximation component and the high-frequency detail component, and align the spatial resolution of the low-frequency approximation component and the high-frequency detail component with the normalized image to obtain the aligned physical feature components.
5. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 4, characterized in that, The normalized image is concatenated with the aligned physical feature components to obtain the two-dimensional feature tensor, as shown in the formula: ； In the formula, For a two-dimensional feature tensor, For splicing, For normalized images, This is a low-frequency approximation component. For vertical high-frequency components, and These represent the number of sampling points in the horizontal and vertical directions of the image, respectively.
6. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The process of obtaining the one-dimensional lateral coordinate mapping sequence includes: reducing the dimensionality of the fused two-dimensional feature tensor to a one-dimensional sequence, outputting the predicted lateral coordinate increments, performing cumulative integration, and normalizing to the global physical scan length to obtain the one-dimensional lateral coordinate mapping sequence.
7. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The formula for the acoustic physics prior loss function is: ； In the formula, For the overall loss function, For coordinate mapping loss function, The coordinate increment loss function, This is a physical penalty term for the equivalent width of the main lobe based on B-scan simulation. , , These are dynamic weighting coefficients.
8. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The formula for obtaining the corrected ultrasound image is: ； In the formula, The corrected C-scan image, The grid consists of equally spaced physical coordinates. The coordinates of pixels with fixed equal spacing along the width direction of the probe array. This is the original, uncorrected ultrasound C-scan image. It is the inverse function of the monotonic mapping function.
9. The intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in claim 1, characterized in that, The formula for constructing the mapping prediction model is as follows: ； In the formula, It is a horizontal mapping function. Sampling points The incremental sequence at that point, This is the index of discrete sampling points along the lateral scanning direction.
10. An intelligent correction device for ultrasound TFM detection images integrating physical priors and lightweight networks, applied to the intelligent correction method for ultrasound TFM detection images integrating physical priors and lightweight networks as described in any one of claims 1-9, characterized in that, include: Ultrasonic phased array equipment, flaw detector main unit, and display; The ultrasonic phased array device is used to acquire raw ultrasonic C-scan images of the inside of a workpiece under manual scanning conditions without a position encoder, and transmit them to the flaw detector host. The flaw detector host includes a memory and a processor. The memory is communicatively connected to the processor and is used to store a pre-burned computer program and a mapping prediction model that has been trained offline. The processor is configured to receive the original ultrasound C-scan image, preprocess and normalize it to obtain a normalized image, extract low-frequency approximation components and high-frequency detail components through multi-scale discrete wavelet decomposition, and concatenate them to obtain a two-dimensional feature tensor; concatenate the normalized image and the two-dimensional feature tensor to obtain a fused two-dimensional feature tensor, input it into the mapping prediction model to reduce the dimension along the width axis to a one-dimensional horizontal coordinate mapping sequence, optimize the training by combining the acoustic physics prior loss function, and perform inverse spatial resampling on the original ultrasound C-scan image according to the remapping function of the one-dimensional horizontal coordinate mapping sequence to obtain a corrected ultrasound image; The display is used to receive the corrected ultrasound image and present the corrected ultrasound image in real time.