Automatic detection and segmentation method based on eddy current pulse thermal imaging defects

Abstract

The invention relates to the technical field of nondestructive testing, in particular to an automatic defect detection and segmentation technology based on eddy current pulse thermal imaging, which comprises a thermogram sequence preprocessing step, a defect signal feature extraction step and an image segmentation post-processing step, in the thermogram sequence preprocessing step, an excitation peak frame image is dynamically selected through an image entropy difference, a static background is inhibited by combining image difference operation, and a geometric coil mask is generated by utilizing edge detection; and a defect signal feature extraction step: carrying out defect matrix reconstruction on the robust principal component analysis model based on the difference image by adopting an augmented Lagrange multiplier method. And an image segmentation post-processing step: designing a dual-threshold segmentation method based on reconstruction matrix local space consistency, and screening thresholds by combining an adaptive threshold of image gray level distribution and an area screening threshold of a defect candidate region. Automatic detection and segmentation of defects in eddy current pulse thermal imaging are achieved, and the purposes of defect identification and quantitative detection are achieved.

Description

Technical Field

[0001] The invention relates to the technical field of nondestructive testing and evaluation, in particular to an automatic defect detection and segmentation method based on eddy current pulse thermal imaging. Background Art

[0002] Nondestructive testing (NDT) technology is a crucial tool for controlling product quality and ensuring the safe operation of in-service equipment. Eddy current pulse thermography (ECPT), which combines eddy current and infrared thermal imaging technologies, enables rapid, contactless detection of defects of varying depths over large areas. It has been widely used in various industries, including aerospace, energy, and conductor materials.

[0003] During ECPT testing, a pulsed excitation power supply generates a short, high-frequency alternating current. This current then flows through an excitation coil and induces eddy currents in the conductive material. These eddy currents generate Joule heating within the material, and the thermal response of the material is captured by an infrared camera. The heat diffuses into the surrounding area, eventually reaching thermal equilibrium. If defects are present in the conductive material, the defective area disrupts the distribution of eddy currents and the heat diffusion process, resulting in varying temperature variations on the material surface.

[0004] Currently, ECPT detection and characterization methods are limited to manually selecting frames recorded by thermal imagers for identification and positioning. This type of method is prone to losing a large amount of data information, which can easily lead to problems such as inaccurate defect detection and positioning. Meanwhile, among existing ECPT detection and characterization methods, robust principal component analysis (RPCA) decomposes and reconstructs infrared images into low-rank matrices and sparse matrices based on the non-local autocorrelation characteristics of infrared thermal imaging, achieving separation of defects from background noise. This is a prior art method, and the specific steps can be found in the paper "Huang Y, Chen CL, Chiang CH. Analyzing a series of thermal infrared images to identify defects using a hybrid approach that combines robust principal component analysis and image segmentation [J]. NDT&E International, 2023. DOI: 10.1016 / j.ndteint.2023.102818."

[0005] Among defect characterization methods, robust principal component analysis can generate defect images, but it still has the following shortcomings. Traditional feature extraction methods have limitations in noise suppression and defect signal separation. Sparse features are easily affected by residual background noise, resulting in a high rate of missed detection of low-contrast defects. For example, principal component analysis and independent component analysis, the specific details can be found in the papers "S.Marinetti et al, Statistical analysis of IR thermographic sequences. by PCA[J]. Infrared Phys.Technol., vol. 46, no. 1, pp. 85-91, Dec. 2004 doi: 10.1016 / j.infrared.2004.03.012." and "L.Bai, B.Gao, GY.Tian, ​​WLWoo, and Y.Cheng, Spatial and Time Patterns Extraction of Eddy Current Pulsed Thermography Using Blind Source Separation[J]. IEEE Sens.J., vol. 13, no. 6, pp. 2094-2101 Jun. 2013, doi: 10.1109 / JSEN.2013.2248931". In addition, there is a lack of a suitable post-processing segmentation method to further optimize and improve the processing effect to facilitate defect extraction, identification, and quantification. Traditional image segmentation methods such as the maximum between-class variance method (OTSU) and the iterative threshold segmentation method (IT) can be found in the papers "N. Otsu, A Threshold Selection Method from Gray-Level Histograms. IEEE Trans. Syst. Man Cybern., vol. 9, no. 1, pp. 62-66, Jan. 1979, doi: 10.1109 / TSMC.1979.4310076." and "T. W. Ridler, S. Calvard, Picture Thresholding Using an Iterative Selection Method [J]. IEEE Trans. Syst. Man Cybern., vol. 8, no. 8, pp. 630-632. 1978. doi: 10.1109 / TSMC.1978.4310039." Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and propose an automatic detection and segmentation method for defects based on eddy current pulse thermal imaging to detect and segment defects, thereby facilitating the identification and quantification of defects.

[0007] To achieve the above-mentioned object, the method for automatically detecting and segmenting defects in eddy current pulse thermal imaging of the present invention comprises:

[0008] The automatic detection and segmentation method of defects based on eddy current pulse thermal imaging is characterized by comprising: a thermal image sequence preprocessing step, a defect signal feature extraction step, and an image segmentation post-processing step; wherein:

[0009] Thermal image sequence preprocessing steps: Dynamically select the excitation peak frame image through image entropy difference, combine image difference operation to suppress static background, and use edge detection to generate geometric coil mask;

[0010] Defect signal feature extraction steps: Use the augmented Lagrange multiplier method to reconstruct the defect matrix of the robust principal component analysis model based on the difference image;

[0011] Image segmentation post-processing steps: Design a dual-threshold segmentation method based on the local spatial consistency of the reconstruction matrix, by combining the adaptive threshold of the image grayscale distribution with the area screening threshold of the defect candidate area;

[0012] The thermal image sequence preprocessing step obtains a space-time thermal sequence image containing a defective conductor through an eddy current pulse thermal imaging detection system; wherein:

[0013] Calculate the image entropy EN for each frame of thermal image i and the first frame image entropy EN 1 Perform differential operation. Select the maximum peak value of image entropy difference Frame image, where: Then perform a differential operation with the initial first frame image to obtain the differential thermal image D(x,y), which is used as the subsequent processing image, where: D(x,y)=I(x,y,t max )-I(x,y,t1); I(x,y,t1) represents the first frame image; I(x,y,t max ) represents the frame image corresponding to the maximum peak value of entropy difference;

[0014] Take the first frame image I(x,y,t1) to perform Canny edge detection, extract the coil outline, and record the coil pixel coordinate set Ω in the image coil Generate a geometric coil binary mask M(x,y); where:

[0015] The defect signal feature extraction step inputs the differential thermal image D(x,y) into the robust principal component analysis model (RPCA), and uses the augmented Lagrange multiplier method (ALM) to solve the RPCA model and decompose and reconstruct it into a low-rank matrix L and a sparse matrix S.

[0016] Initialization step: set the maximum number of iterations k; convergence tolerance Tol; regularization parameter λ; Lagrange multiplier matrix Y; low-rank matrix L0 = 0, sparse matrix S0 = 0, and penalty factor μ0 = ||D|| F , parameter update rate ρ = 1.5, where: ||·|| F represents the Frobenius norm;

[0017] Iterative update step: From the number of iterations 1 to the maximum number of iterations k, perform the following iterations: Update the iterative low-rank matrix: Update the singular value decomposition intermediate matrix: Update iterative sparse matrix: Update the intermediate matrix of singular values: Update iterative Lagrange multiplier: Y k =Y k-1 +μ k-1 (DL k -S k ), update penalty factor: μ k =ρ·μ k-1 , where L k 、S k 、U k 、V k Represents the matrix of the kth iteration; M k-1 、T k-1 They represent the matrix of the k-1th iteration respectively; diag(·) represents the construction of the diagonal matrix; σ i Represents the matrix M k-1 The i-th singular value of ; sign(·) means extracting the elements in the brackets; represents the element-by-element product. Calculate the normalized decomposition error

[0018] Convergence judgment step: Repeat the above steps and update steps until the predetermined convergence condition is reached;

[0019] Convergence condition: When the normalized decomposition error

[0020] The image segmentation post-processing step comprises:

[0021] Adaptive histogram equalization is used to enhance the image details of the sparse matrix S and enhance the contrast of the defect area;

[0022] The sparse matrix S after contrast enhancement is covered and corrected with a geometric coil binary mask M(x,y), and this image matrix is ​​used as the image to be segmented. Define: Segmentation is performed by combining the adaptive threshold of the grayscale distribution of the image to be segmented with the area screening threshold of the defect candidate area.

[0023] The adaptive threshold of the image grayscale distribution and the area screening threshold of the defect candidate area include:

[0024] Calculate the pixel grayscale mean and standard deviation of the image to be segmented. The image segmentation threshold T is defined as: in: and Represent the mean and standard deviation of the input image respectively, k is an empirical constant, and k = 2 is selected. The image is segmented according to the segmentation threshold T to obtain the initial segmentation binary image;

[0025] Mark the eight-neighborhood connected domains of the initial segmented binary image and extract all candidate regions {R i}, and calculate the pixel area A of each region i , set the experience threshold A th =5, screening candidate defect areas, actual defect area D efect Expressed as: D efect ={R i |A i ≥A th}.

[0026] Beneficial effects of the present invention:

[0027] The present invention realizes effective extraction of defect features in the sparse matrix obtained by substituting the differential image into the RPCA model. Compared with the original maximum peak image without differential operation, the defect features in the sparse matrix obtained are blurred. Figure 2 (b) shown. Figure 3 This is a comparison chart of the results of common feature extraction methods and the feature extraction method in this patent method under ECPT. Figure 3 (a) is the defect feature map obtained by processing the entire thermal image sequence using the principal component analysis (PCA) method; Figure 3 (b) is the defect feature map obtained by processing the entire thermal image sequence using the independent component analysis (ICA) method;

[0028] Figure 3 (c) is the sparse matrix image obtained by the patented method after solving the robust principal component analysis on the differential image D(x,y). The comparison results show that if the commonly used feature extraction method is used, the background interference is extremely large, for example Figure 3In (a), the high-frequency noise in the coil area still remains significantly, which interferes with the subsequent segmentation process. Secondly, the extraction of defect feature details in the image is insufficient. For example Figure 3 The third defect in (b) is blurred. However, the method proposed in the present invention can effectively extract all defect features and suppress high-frequency noise interference in the coil area.

[0029] This paper designs a dual-threshold segmentation method based on the local spatial consistency of the reconstruction matrix. By combining an adaptive threshold for image grayscale distribution with an area screening threshold for defect candidate regions, it achieves automatic defect detection and segmentation in eddy current pulse thermal imaging, achieving both defect identification and quantitative detection. Figure 5 The image segmentation results of the sparse matrix S obtained in stage 2 are compared using the commonly used segmentation method and the segmentation method in this patent under ECPT. Figure 5 (a) is the segmentation result of the maximum between-class variance method (OTSU); Figure 5 (b) is the segmentation result of iterative threshold segmentation method (IT); Figure 5 (c) is the segmentation result of this patent. Comparing the segmentation results, we can see that if the common method is used to segment the defects, the background and high-frequency noise will interfere seriously, which is not conducive to the identification of defects. The automatic detection segmentation method proposed in this invention can effectively suppress the interference of background and high-frequency noise, and can accurately segment the three defects at the same time. Here, three indicators are used: as well as To objectively evaluate the performance of the segmentation method, (TP means that a defect actually exists and is correctly detected; FP means that a defect actually does not exist but is mistakenly judged as a defect; FN means that a defect actually exists but is not detected; TN means that a defect actually does not exist and is correctly judged as a non-defect). TPR reflects the detection method's ability to identify defects, FPR reflects the degree of misjudgment of non-defects by the detection method, Precision reflects the reliability of the detection results, and Recall reflects the sensitivity of the detection method. The closer TPR is to 1, the better the effect, and the lower FPR, the better the effect. In order to balance precision and recall, the F index is introduced as a comprehensive performance indicator. Among them, α represents the weight of precision and recall. A value greater than 1 means that the weight of precision is greater, and vice versa. In this patent, it is set to 1, indicating that the two are equally important. The F index is a comprehensive indicator that comprehensively considers precision and recall. The higher the value, the better the overall performance. In this embodiment, Crack1, Crack2 and Crack3 are regarded as defects, and the TPR, FPR and F index of each segmentation method are calculated. The results are shown in Table 1.

[0030] Table 1

[0031] OTSU IT Patent for this invention TPR 0.8011 0.6851 0.3591 FPR 0.7038 0.5154 0 F-score 0.0013 0.0015 0.52

[0032] As shown in Table 1, although the TPR of the two commonly used methods is higher than that of this patent, indicating that they can effectively capture defect areas, their FPR values ​​are significantly higher, indicating that although the commonly used methods can identify defects, they cannot distinguish between noise and real targets. Figure 5 (a) and Figure 5 (b) The FPR of this patent is significantly lower than that of the other two commonly used methods, indicating that the proposed method can accurately separate the noise and defect areas and remove most of the noise. Figure 5 As shown in (c), a high-quality detection and segmentation result can be obtained. The F-index, as a comprehensive performance indicator, shows that the precision and recall rates are also significantly higher than those of the other two commonly used methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 This is a flow chart of a specific embodiment of the method for automatic detection and segmentation of eddy current pulse thermal imaging defects of the present invention;

[0035] Figure 2 This is the comparison of the ablation experiment: (a) the original image of the maximum peak frame of the excitation, (b) the sparse matrix after decomposition and reconstruction of the original thermal image, (c) the difference image after the difference operation, and (d) the sparse matrix after decomposition and reconstruction of the difference image.

[0036] Figure 3 This is a comparison chart of the feature extraction effects of common feature extraction methods and the method of this patent: (a) principal component analysis, (b) independent component analysis, and (c) this patent method.

[0037] Figure 4 These are the renderings of the segmentation stage in this patent: (a) the rendering after adaptive histogram equalization, (b) the rendering of adaptive threshold segmentation under coil mask constraint coverage, and (c) the rendering of connected domain statistical analysis, that is, the final detection segmentation rendering.

[0038] Figure 5 This is a comparison chart of the common image segmentation methods and the segmentation effect of this patent: (a) Maximum Inter-Class Variance Method (OTSU), (b) Iterative Threshold Segmentation Method (IT), and (c) this patent method. DETAILED DESCRIPTION

[0039] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0040] In this embodiment, if Figure 1 As shown, the present invention includes three stages:

[0041] Stage 1: Thermal image sequence preprocessing steps:

[0042] First, a space-time thermal image sequence containing the defective conductor surface is obtained on the conductor material containing the defect through the eddy current pulse thermal imaging nondestructive testing system. The image entropy value EN is calculated for each frame of the sequence thermal image. i , and then perform a differential operation with the entropy value of the first frame image in the sequence diagram, and take the maximum entropy difference The corresponding frame image is used as a key frame, where Perform image difference operation on the corresponding key frame image and the first frame image to obtain the thermal imaging difference image D(x,y) as the subsequent processing image, such as Figure 2 (c) where D(x,y)=I(x,y,t max )-I(x,y,t1), I(x,y,t1) represents the first frame image, I(x,y,t max ) represents the frame image corresponding to the maximum peak value of entropy difference;

[0043] Secondly, in the first frame image I(x,y,t1), perform Canny edge detection to extract the coil edge contour and record the coil pixel coordinate set Ω in the image. coil Generate a geometric coil binary mask M(x,y); where, M(x,y) is reserved for use in subsequent stages.

[0044] Stage 2: Defect signal feature extraction steps:

[0045] The difference graph D(x,y) obtained in stage 1 is substituted into the robust principal component analysis model (RPCA) for decomposition and reconstruction into a low-rank matrix L and a sparse matrix S. The RPCA model is solved using the augmented Lagrange multiplier method (ALM).

[0046] The defect signal feature extraction step inputs the differential thermal image D(x,y) into the robust principal component analysis model (RPCA), and uses the augmented Lagrange multiplier method (ALM) to solve the RPCA model and decompose and reconstruct it into a low-rank matrix L and a sparse matrix S.

[0047] Initialization step: set the maximum number of iterations k; convergence tolerance Tol; regularization parameter λ; Lagrange multiplier matrix Y; low-rank matrix L0 = 0, sparse matrix S0 = 0, and penalty factor μ0 = ||D|| F , parameter update rate ρ = 1.5, where: ||·|| F represents the Frobenius norm;

[0048] Iterative update step: From the number of iterations 1 to the maximum number of iterations k, perform the following iterations: Update the iterative low-rank matrix: Update the singular value decomposition intermediate matrix: Update iterative sparse matrix: Update the intermediate matrix of singular values: Update iterative Lagrange multiplier: Y k =Y k-1 +μ k-1 (DL k -S k ), update penalty factor: μ k =ρ·μ k-1 , where L k 、S k 、U k 、V k Represents the matrix of the kth iteration; M k-1 、T k-1 They represent the matrix of the k-1th iteration respectively; diag(·) represents the construction of the diagonal matrix; σ i Represents the matrix M k-1 The i-th singular value of ; sign(·) means extracting the elements in the brackets; Represents element-wise product. Calculates the normalized decomposition error

[0049] Convergence judgment step: Repeat the above steps and update steps until the predetermined convergence condition is reached;

[0050] Convergence condition: When the normalized decomposition error

[0051] Stage 3: Image segmentation post-processing steps:

[0052] Adaptive histogram equalization is used to enhance the image details of the sparse matrix S and enhance the contrast of the defect area;

[0053] The sparse matrix S after contrast enhancement is covered and corrected with a geometric coil binary mask M(x,y), and this image matrix is ​​used as the image to be segmented. Define: Segmentation is performed by combining the adaptive threshold of the grayscale distribution of the image to be segmented with the area screening threshold of the defect candidate area.

[0054] Calculate the pixel grayscale mean and standard deviation of the image to be segmented. The image segmentation threshold T is defined as: in: and Represent the mean and standard deviation of the input image respectively, k is an empirical constant, and k = 2 is selected. The image is segmented according to the segmentation threshold T to obtain the initial segmentation binary image;

[0055] Mark the eight-neighborhood connected domains of the initial segmented binary image and extract all candidate regions {R i}, and calculate the pixel area A of each region i , set the experience threshold A th =5, screening candidate defect areas, actual defect area D efect Expressed as: D efect ={R i |A i ≥A th};

[0056] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0057] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for automatic detection and segmentation of defects based on eddy current pulse thermal imaging, characterized in that: include: Thermal image sequence preprocessing step, defect signal feature extraction step and image segmentation post-processing step; wherein: Thermal image sequence preprocessing steps: Dynamically select the excitation peak frame image through image entropy difference, combine image difference operation to suppress static background, and use edge detection to generate geometric coil mask; Defect signal feature extraction steps: Use the augmented Lagrange multiplier method to reconstruct the defect matrix of the robust principal component analysis model based on the difference image; Image segmentation post-processing steps: Design a dual-threshold segmentation method based on the local spatial consistency of the reconstruction matrix, by combining the adaptive threshold of the image grayscale distribution and the area screening threshold of the defect candidate area.

2. The automatic detection and segmentation method based on eddy current pulse thermal imaging defects according to claim 1 is characterized in that: The thermal image sequence preprocessing step obtains a space-time thermal sequence image containing a defective conductor through an eddy current pulse thermal imaging detection system; wherein: Calculate the image entropy EN for each frame of thermal image i and the first frame image entropy EN 1 Perform differential operation. Select the maximum peak value of image entropy difference Frame image, where: Then perform a differential operation with the initial first frame image to obtain the differential thermal image D(x,y), which is used as the subsequent processing image, where: D(x,y)=I(x,y,t max )-I(x,y,t1); I(x,y,t1) represents the first frame image; I(x,y,t max ) represents the frame image corresponding to the maximum peak value of entropy difference; Take the first frame image I(x,y,t1) to perform Canny edge detection, extract the coil outline, and record the coil pixel coordinate set Ω in the image coil Generate a geometric coil binary mask M(x,y); where:

3. The automatic detection and segmentation method based on eddy current pulse thermal imaging defects according to claim 1 is characterized in that: The defect signal feature extraction step inputs the differential thermal image D(x, y) into a robust principal component analysis model (RPCA), and uses the augmented Lagrange multiplier method (ALM) to solve the RPCA model and decompose and reconstruct it into a low-rank matrix L and a sparse matrix S. Initialization step: set the maximum number of iterations k; convergence tolerance Tol; regularization parameter λ; Lagrange multiplier matrix Y; The low-rank matrix L0 = 0, the sparse matrix S0 = 0, and the penalty factor μ0 = ||D|| F , parameter update rate ρ = 1.5, where: ||·|| F represents the Frobenius norm; Iterative update step: From the number of iterations 1 to the maximum number of iterations k, perform the following iterations: Update the iterative low-rank matrix: Update the singular value decomposition intermediate matrix: Update iterative sparse matrix: Update the intermediate matrix of singular values: Update iterative Lagrange multiplier: Y k =Y k-1 +μ k-1 (DL k -S k ), update penalty factor: μ k =ρ·μ k-1 , where L k 、S k 、U k 、V k Represent the matrix of the kth iteration; M k-1 、T k-1 They represent the matrix of the k-1th iteration respectively; diag(·) represents the construction of the diagonal matrix; σ i Represents the matrix M k-1 The i-th singular value of ; sign(·) means extracting the elements in the brackets; | means element-by-element multiplication. Calculate the normalized decomposition error Convergence judgment step: Repeat the above steps and update steps until the predetermined convergence condition is reached; Convergence condition: When the normalized decomposition error 4. The automatic detection and segmentation method based on eddy current pulse thermal imaging defects according to claim 1 is characterized in that: The image segmentation post-processing step comprises: Adaptive histogram equalization is used to enhance the image details of the sparse matrix S and enhance the contrast of the defect area; The sparse matrix S after contrast enhancement is covered and corrected with a geometric coil binary mask M(x,y), and this image matrix is ​​used as the image to be segmented. Define: Segmentation is performed by combining the adaptive threshold of the grayscale distribution of the image to be segmented with the area screening threshold of the defect candidate area.

5. The automatic detection and segmentation method based on eddy current pulse thermal imaging defects according to claim 4 is characterized in that: The adaptive threshold of the image grayscale distribution and the area screening threshold of the defect candidate area include: Calculate the pixel grayscale mean and standard deviation of the image to be segmented. The image segmentation threshold T is defined as: in: and Represent the mean and standard deviation of the input image respectively, k is an empirical constant, k = 2 is selected, and the image is segmented according to the segmentation threshold T to obtain the initial segmentation binary image; Mark the eight-neighborhood connected domains of the initial segmented binary image and extract all candidate regions {R i }, and calculate the pixel area A of each region i , set the experience threshold A th =5, screening candidate defect areas, actual defect area D efect Expressed as: D efect ={R i |A i ≥A th }.

Images