A noise reduction method and system for infrared thermal images of composite materials

Abstract

This invention discloses a noise reduction method and system for infrared thermal images of composite materials, belonging to the field of image processing technology. The method involves acquiring the original infrared thermal image sequence of the composite material, fitting a preset background fitting surface model using the gray values ​​of the pixels in the original infrared thermal image, and obtaining the optimal fitting coefficients. Based on the difference between the gray values ​​of the pixels in the original infrared thermal image and the predicted gray values ​​of corresponding points in the thermal image background predicted by the background fitting surface model, an error function is determined. During the fitting process, the fitting coefficients of the background fitting surface model are adjusted to minimize the error function, resulting in the optimal fitting coefficients. The best-fitting thermal image background is determined based on the optimal fitting coefficients, and this background is removed from the original infrared thermal image to obtain the denoised infrared thermal image. This method can reconstruct the spatial distribution characteristics of background noise, achieving the noise reduction objective.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a noise reduction method and system for infrared thermal images of composite materials. Background Technology

[0002] Advanced composite materials are characterized by their light weight, high specific strength and specific modulus, good ductility, corrosion resistance, heat insulation, sound insulation, vibration reduction and high (low) temperature resistance, and are widely used in aerospace, shipbuilding, automotive industry, petrochemical, wind power generation and military industry.

[0003] However, due to the special manufacturing process (such as uneven resin impregnation and curing stress) and service conditions (such as impact and vibration), composite laminates are prone to unavoidable defects such as delamination, debonding, and fiber breakage in the interlayer interface area. These defects significantly reduce the material's performance and reliability. The generation and accumulation of these defects will seriously affect the structural integrity or safety and reliability of composite products or core components, and may even lead to major safety accidents. Because these defects often exhibit internal hidden characteristics, conventional visual inspection methods are insufficient for effective characterization. Therefore, rapid detection and identification of defects are crucial for ensuring the safe use of composite materials.

[0004] Infrared thermography, a non-destructive testing technology, has been effectively applied in the detection of defects in composite materials due to its wide range of applications, large detection area, fast detection speed, non-contact measurement, and safety and reliability. However, due to the special structure of composite materials, they are easily affected by factors such as uneven heating of the sample surface and changes in emissivity during detection. This results in infrared thermogram sequences with problems such as low resolution, low contrast, and low signal-to-noise ratio, which makes subsequent defect identification and quantitative analysis difficult. Currently, pulse phase analysis and principal component analysis methods for processing infrared thermogram sequences still suffer from noise interference in the processed infrared image sequences.

[0005] In summary, due to the special nature of composite material structures, the pulsed infrared thermal imaging detection of composite laminates is affected by many factors, such as infrared thermal imager lens reflection, uneven heating of the excitation source, edge effect of the test block, and ambient noise, resulting in poor quality of the processed infrared thermal images. Summary of the Invention

[0006] To address the problems existing in the aforementioned fields, this invention proposes a noise reduction method and system for infrared thermal images of composite materials. The fitted background surface model uses polynomial, Gaussian, and hyperbolic terms to characterize the low-frequency components of the background temperature field, temperature anomalies caused by local heat sources, and boundary heat dissipation effects, respectively. This fully considers the influence of thermal noise interference, lens reflection, and edge effects on the background temperature field. By determining the error function and minimizing it, the optimal thermal image background is determined by adjusting the fitting coefficients of the background fitted surface model. The optimal thermal image background is then removed from the original infrared thermal image to achieve noise reduction.

[0007] To address the aforementioned technical problems, this invention discloses a noise reduction method for infrared thermal images of composite materials, comprising the following steps:

[0008] Obtain the original infrared thermal image sequence of the composite material;

[0009] The grayscale values ​​of the pixels in the original infrared thermal image are used to fit a preset background fitting surface model to obtain the optimal fitting coefficients. The background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by local heat sources, and the boundary heat dissipation effect.

[0010] During the fitting process, the optimal fitting coefficients of the background fitting surface model are obtained by adjusting the fitting coefficients of the background fitting surface model with the goal of minimizing the error function. The error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model.

[0011] Based on the optimal fitting coefficient, the best-fitting thermal image background is determined by the background fitting surface model. The best-fitting thermal image background is then removed from the original infrared thermal image to obtain the denoised infrared thermal image.

[0012] Preferably, the preset background fitting surface model is:

[0013] ;

[0014] In the formula, This is the background fitting function for the original infrared thermal image. A The coefficients are to be fitted. The polynomial basis terms characterizing the low-frequency components of the background temperature field; A Gaussian bulge term to describe temperature anomalies caused by local heat sources; The hyperbolic concave term is used to simulate the boundary heat dissipation effect;

[0015] ;

[0016] ;

[0017] ;

[0018] In the formula, The coefficients are to be fitted. x , y This indicates the location coordinates.

[0019] Preferably, the error function is:

[0020] Obtain the grayscale values ​​of the pixels in the original infrared thermal image and the predicted grayscale values ​​of the corresponding points in the thermal image background predicted by the background fitted surface model;

[0021] The difference between the grayscale value of a pixel in the original infrared thermal image and the grayscale prediction value of the corresponding point in the background of the thermal image predicted by the background fitted surface model is called the residual.

[0022] The error function is defined as follows: by summing the squared residuals of all non-defective pixels.

[0023] ;

[0024] In the formula, I ( i , j The standardized infrared thermal image at pixel points ( i , j The grayscale value at () To fit the background surface model with the coefficients to be fitted A coordinates below x and y The predicted grayscale value at that location.

[0025] Preferably, the optimal fitting coefficients of the obtained background fitted surface model are obtained by iteratively updating the minimum value of the error function through the trust region algorithm, thus obtaining the optimal solution that minimizes the error function.

[0026] Preferably, the minimum value of the error function is iteratively updated using the trust region algorithm to obtain the optimal solution that minimizes the error function, specifically including:

[0027] Define the trust region of the current iteration point as:

[0028] ;

[0029] Error function S ( A At the extreme points, it is equivalent to a quadratic function, and a trust region quadratic model is constructed through quadratic approximation. To solve for the step size s k :

[0030] ;

[0031] ;

[0032] In the formula, Error function S ( A At the current iteration point A k gradient at; H k Error function S ( A At the iteration point A k The Hessian matrix at that location; It is the Euclidean norm;

[0033] Error function S ( A ) in the k The actual decrease in step is:

[0034] ;

[0035] Trust Region Quadratic Model Function The predicted decrease is:

[0036] ;

[0037] Define ratio for:

[0038] ;

[0039] pass Measuring the quadratic model function Sum of error functions S ( A The degree of approximation, and also based on Determine whether to update the trust region radius for the next iteration. ;

[0040] Iterate step by step until the convergence condition is met. When the iteration point is updated, the trust region corresponding to the iteration point is determined, and the optimal solution that minimizes the error function is obtained.

[0041] Preferably, the according to Determine whether to update the trust region radius for the next iteration. Specifically, it includes:

[0042] Initialize the fitting coefficients And trust region radius And set convergence conditions. ;

[0043] When the number is reached k During the step, the gradient is calculated. and the Hessian matrix H k ,judge Does it meet the convergence condition? If satisfied, output A = A k To obtain the optimal fitting coefficients, otherwise, by solving the trust region quadratic model, the th... k Step length s k Calculate the ratio ;

[0044] like A value less than 0 indicates that the error function value is increasing, which is opposite to the optimal solution objective. k = k +1, the iteration point remains unchanged. A k+1 = A k and reduce the trust region radius. Resolve the trust region quadratic model;

[0045] like A value greater than 0 indicates that the error function value is decreasing, which is consistent with the optimal solution objective. Continue iteration, letting... k = k +1, iteration point ;

[0046] When 0 < When <0.25, reduce the trust region radius. ;

[0047] When 0.25 < When <0.75, the trust region radius remains unchanged. ;

[0048] when When the value is greater than 0.75, increase the trust region radius. And recalculate the gradient. and the Hessian matrix and judge Does it meet the convergence condition? To obtain the optimal fitting coefficients.

[0049] Preferably, determining the best-fitting thermal image background specifically includes:

[0050] By minimizing the optimal solution of the error function, the minimum value of the predicted value corresponding to the unfitted coefficient of the background fitting surface model at the pixel point of the original infrared thermal image is obtained.

[0051] Based on the position coordinates corresponding to the minimum value and the optimal fitting coefficient, the background fitting surface model corresponding to the optimal fitting coefficient is obtained;

[0052] The best-fitting thermal image background is determined based on the background fitting surface model corresponding to the optimal fitting coefficient.

[0053] Preferably, it also includes a noise reduction system for infrared thermal images of composite materials, comprising:

[0054] The image acquisition module is used to acquire the original infrared thermal image sequence of the composite material;

[0055] The background fitting module is used to fit a preset background fitting surface model with the gray values ​​of the pixels in the original infrared thermal image to obtain the optimal fitting coefficients. The background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by the local heat source, and the boundary heat dissipation effect.

[0056] The fitting coefficient optimization module is used to adjust the fitting coefficients of the background fitting surface model during the fitting process, with the goal of minimizing the error function, to obtain the optimal fitting coefficients of the background fitting surface model; the error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model.

[0057] The image background denoising module is used to determine the best-fitting thermal image background based on the optimal fitting coefficient and the background fitting surface model, and remove the best-fitting thermal image background from the original infrared thermal image to obtain the denoised infrared thermal image.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] This invention proposes a noise reduction method for infrared thermal images of composite materials. A background fitting surface model of a multimodal substrate is fitted, using polynomial, Gaussian, and hyperbolic terms to characterize the low-frequency components of the background temperature field, temperature anomalies caused by local heat sources, and boundary heat dissipation effects, respectively. This method fully considers the influence of thermal noise interference, lens reflection, and edge effects on the background temperature field, thereby reconstructing the spatial distribution characteristics of background noise. To obtain the optimal fitting coefficients of the background fitting surface model, an error function is defined, with the goal of minimizing the error function. This minimizes the sum of squared residuals between the predicted grayscale values ​​of corresponding points in the thermal image background and the actual grayscale values ​​of pixels in the infrared thermal image. By adjusting the fitting coefficients of the background fitting surface model and obtaining the minimum value of the error function, the background fitting surface is iteratively updated to obtain the best-fitting thermal image background. The best-fitting thermal image background is then removed from the original infrared thermal image to achieve noise reduction. Attached Figure Description

[0060] Figure 1 This is a flowchart of the noise reduction method for infrared thermal images of composite materials proposed in this invention;

[0061] Figure 2 A flowchart illustrating the calculation of fitting coefficients for the trust region algorithm provided in this invention;

[0062] Figure 3 This is a schematic diagram of a flat-bottomed hole defect in a GFRP laminate provided in an embodiment of the present invention;

[0063] Figure 4 Representative frames of the infrared image sequence provided in this embodiment of the invention;

[0064] Figure 5 This refers to the temperature sampling regions of interest in the defective and non-defective areas provided in the embodiments of the present invention.

[0065] Figure 6 Temperature and temperature difference curves of defective and non-defective regions provided in embodiments of the present invention;

[0066] Figure 7 The 360th frame of the original thermal image provided in this embodiment of the invention;

[0067] Figure 8 The fitting background provided for the embodiments of the present invention;

[0068] Figure 9 Infrared thermal images with background correction provided in embodiments of the present invention;

[0069] Figure 10 The PPT processing result of the original infrared sequence image provided in the embodiments of the present invention;

[0070] Figure 11 This is a PPT processing result of an infrared sequence image after background removal provided in an embodiment of the present invention;

[0071] Figure 12 The PCA processing result of the original infrared sequence image provided in the embodiments of the present invention;

[0072] Figure 13 The PCA processing result of the background-removed infrared sequence image provided in the embodiment of the present invention. Detailed Implementation

[0073] The following will refer to the appendices in the embodiments of the present invention. Figures 1-13 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0074] Example

[0075] This embodiment addresses the thermal noise issues of uneven heating of the excitation source and lens reflection in infrared thermal imaging. The present invention proposes a method for denoising infrared sequence images using composite materials, such as... Figure 1 As shown, it includes the following steps:

[0076] S1: Obtain the original infrared thermal image sequence of the composite material;

[0077] S2: Fit the preset background fitting surface model using the gray values ​​of the pixels in the original infrared thermal image to obtain the optimal fitting coefficients; the background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by the local heat source, and the boundary heat dissipation effect.

[0078] S3: During the fitting process, the optimal fitting coefficient of the background fitting surface model is obtained by adjusting the fitting coefficient of the background fitting surface model with the goal of minimizing the error function; the error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model.

[0079] S4: Based on the optimal fitting coefficient, determine the best-fitting thermal image background through the background fitting surface model, remove the best-fitting thermal image background from the original infrared thermal image, and obtain the denoised infrared thermal image.

[0080] Specifically, in step S2, the mathematical expression for the preset background fitting surface model is:

[0081] ;

[0082] In the formula, This is the background fitting function for the original infrared thermal image.A The coefficients are to be fitted. The polynomial basis terms characterizing the low-frequency components of the background temperature field; A Gaussian bulge term to describe temperature anomalies caused by local heat sources; This is a hyperbolic concave term used to simulate the boundary heat dissipation effect.

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, The coefficients are to be fitted. x , y This indicates the location coordinates.

[0087] In step S3, to find the optimal fitting coefficients of the background fitting surface model, such that the sum of squared residuals between the predicted grayscale values ​​of corresponding points in the thermal image background and the actual grayscale values ​​of pixels in the infrared thermal image is minimized, the error function is defined as follows:

[0088] ;

[0089] In the formula, I ( i , j The standardized infrared thermal image at pixel points ( i , j The grayscale value at () To fit the background surface model with the coefficients to be fitted A coordinates below x and y The predicted grayscale value at that location.

[0090] like Figure 2 As shown, the optimal solution that minimizes the error function is obtained through iterative updates using the trust region algorithm, specifically including:

[0091] Define the trust region of the current iteration point as:

[0092] ;

[0093] Error function S ( A At the extreme points, it is equivalent to a quadratic function, and a trust region quadratic model is constructed through quadratic approximation. To solve for the step size s k :

[0094] ;

[0095] ;

[0096] In the formula, Error function S ( A At the current iteration point A k gradient at; H k Error function S ( A At the iteration point A k The Hessian matrix at that location; It is the Euclidean norm;

[0097] Error function S ( A ) in the k The actual decrease in step is:

[0098] ;

[0099] Trust Region Quadratic Model Function The predicted decrease is:

[0100] ;

[0101] Define ratio for:

[0102] ;

[0103] pass Measuring the quadratic model function Sum of error functions S ( A The degree of approximation, and also based on Determine whether to update the trust region radius for the next iteration. ;

[0104] Iterate step by step until the convergence condition is met. When the iteration point is updated, the trust region corresponding to the iteration point is determined, and the optimal solution that minimizes the error function is obtained.

[0105] according to Determine whether to update the trust region radius for the next iteration. Specifically, it includes:

[0106] Initialize the fitting coefficients And trust region radius And set convergence conditions. ;

[0107] When the number is reached k During the step, the gradient is calculated. and the Hessian matrix H k ,judge Does it meet the convergence condition? If satisfied, output A = A k To obtain the optimal fitting coefficients, otherwise, by solving the trust region quadratic model, the th... k Step length s k Calculate the ratio ;

[0108] like A value less than 0 indicates that the error function value is increasing, which is opposite to the optimal solution objective. k = k +1, the iteration point remains unchanged. A k+1 = A k and reduce the trust region radius. Resolve the trust region quadratic model;

[0109] like A value greater than 0 indicates that the error function value is decreasing, which is consistent with the optimal solution objective. Continue iteration, letting... k = k +1, iteration point ;

[0110] When 0 < When <0.25, reduce the trust region radius. ;

[0111] When 0.25 < When <0.75, the trust region radius remains unchanged. ;

[0112] when When the value is greater than 0.75, increase the trust region radius. And recalculate the gradient. and the Hessian matrix and judge Does it meet the convergence condition? To obtain the optimal fitting coefficients.

[0113] In step S4, the best-fitting thermal image background is determined, specifically including:

[0114] By minimizing the optimal solution of the error function, the minimum value of the predicted value corresponding to the unfitted coefficient of the background fitting surface model at the pixel point of the original infrared thermal image is obtained.

[0115] Based on the position coordinates corresponding to the minimum value and the optimal fitting coefficient, the background fitting surface model corresponding to the optimal fitting coefficient is obtained;

[0116] The best-fitting thermal image background is determined based on the background fitting surface model corresponding to the optimal fitting coefficient.

[0117] This invention also proposes a noise reduction system for infrared thermal images of composite materials, comprising:

[0118] The image acquisition module is used to acquire the original infrared thermal image sequence of the composite material;

[0119] The background fitting module is used to fit a preset background fitting surface model with the gray values ​​of the pixels in the original infrared thermal image to obtain the optimal fitting coefficients. The background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by the local heat source, and the boundary heat dissipation effect.

[0120] The fitting coefficient optimization module is used to adjust the fitting coefficients of the background fitting surface model during the fitting process, with the goal of minimizing the error function, to obtain the optimal fitting coefficients of the background fitting surface model; the error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model.

[0121] The image background denoising module is used to determine the best-fitting thermal image background based on the optimal fitting coefficient and the background fitting surface model, and remove the best-fitting thermal image background from the original infrared thermal image to obtain the denoised infrared thermal image.

[0122] The noise reduction method proposed in this invention reconstructs the spatial distribution characteristics of background noise by fitting a multimodal substrate background fitting surface model, and then subtracts the fitted background surface from the original image to reduce background noise interference.

[0123] Experimental Analysis

[0124] Experimental test block

[0125] The most common defect in GFRP is delamination, which is often simulated by artificial defects such as flat-bottomed holes. The designed test block is as follows: Figure 3 As shown, the test block measures 250 mm × 180 mm × 4 mm and contains 25 defects in 5 rows and 5 columns. The defect diameters, from largest to smallest, are 20 mm, 16 mm, 12 mm, 8 mm and 4 mm, respectively, and the defect depths, from deepest to shallowest, are 3.5 mm, 3.0 mm, 2.5 mm, 2.0 mm and 1.5 mm, respectively. Detailed information on the defects is shown in Table 1.

[0126] Table 1. Information on flat-bottomed hole defects in GFRP laminates

[0127]

[0128] Infrared sequence thermal image acquisition

[0129] The experiment used a German VarioCAM® HD980 uncooled long-wave thermal imager with a response wavelength range of 7.5–14 µm, a resolution of 1024 × 768 pixels, a temperature resolution of 50 mK, and a temperature detection accuracy of ±1.5 K. The excitation source consisted of four 500 W halogen lamps with a total power of 2000 W. The excitation time was set to 8 s, the acquisition frequency to 15 Hz, and the number of frames acquired to 900. The resulting raw infrared thermal image sequence is shown below. Figure 4 As shown.

[0130] The infrared thermal image sequence fully records the dynamic evolution of the temperature field of the specimen after pulse excitation, from... Figure 4 It can be seen that: Figure 4 Figure (a) shows the evolution result of frame 1. In the initial stage (frame 1), it is obviously affected by environmental noise. Figure (b) shows the evolution result of frame 70. At frame 70, shallow defects are initially visible but their outlines are blurred. Figure (c) shows the evolution result of frame 120. At the end of the excitation (frame 120), the peak thermal contrast is reached, and 9 defects are clearly visible, with the smallest and shallowest defects being distinguishable. Figure (d) shows the evolution result of frame 300. In the middle stage (frame 300), most defects continue to be visible, while the deepest defects are not visible. Figure (e) shows the evolution result of frame 600. In the later stage (frame 600), thermal diffusion causes small aperture defects to disappear. Figure (f) shows the evolution result of frame 900. From frame 900, it can be seen that large aperture defects also become blurred due to thermal diffusion.

[0131] To reflect the temperature change process in the infrared thermal image sequence, regions of interest are set in both defective and non-defective regions within the image sequence, such as... Figure 5 As shown. Based on the average temperature data of the sampling area, the history curves of surface temperature change over time at both defective and non-defective locations on the specimen are plotted as follows. Figure 6 As shown in Figure (a), the horizontal axis t / s represents time, and the vertical axis T / ℃ represents temperature. Further, the history curve of temperature difference change over time is obtained; see [reference needed]. Figure 6 In Figure (b), the horizontal axis t / s represents time, and the vertical axis... This indicates the change in temperature difference.

[0132] Depend on Figure 6As shown in Figure (b), the temperature of different parts of the specimen surface exhibits a pattern of first rising and then falling over time. However, neither the rising nor falling phase is monotonous; rather, it is a fluctuating rise or fall. This indicates that the temperature field is significantly affected by noise, which inevitably impacts the quality of thermal imaging. Therefore, noise reduction is needed to improve the quality of infrared thermal images.

[0133] Image performance comparison before and after background removal

[0134] Background fitting can be performed on each frame of the original infrared thermal image sequence. Taking frame 360 ​​as an example, ... Figure 7 As shown, Figure 7 Figures (a) and (b) in the figure are the original two-dimensional image and the original three-dimensional temperature field distribution map, respectively. The background is fitted using the multimodal trust region method to obtain the following results: Figure 8 The results shown Figure 8 Figures (a) and (b) show the fitted background two-dimensional image and the fitted background temperature field three-dimensional distribution map, respectively. Subtracting the fitted background yields the corrected infrared thermal image, i.e., the denoised infrared thermal image. Figure 9 As shown, Figure 9 Figures (a) and (b) in the figure are the corrected two-dimensional image and the three-dimensional temperature field distribution map, respectively.

[0135] Peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and entropy were selected as image evaluation metrics. The imaging quality of images before and after background removal was compared. The data are shown in Table 2. After removing the fitted background, all performance metrics were significantly improved. Specifically, the PSNR increased by 13.07%, the RMSE decreased significantly by 58.07%, and the entropy increased by 21.72%.

[0136] Table 2 Comparison of image quality before and after background removal (taking frame 360 ​​as an example)

[0137]

[0138] Pulse Phase Thermography (PPT), also known as Fourier Transform, is a thermal image processing method developed by combining the advantages of pulsed thermal excitation and modulated thermal excitation. In pulsed thermal excitation, the excitation thermal wave contains multiple frequency components, while modulated thermal excitation uses only a single excitation frequency each time. By organically combining the two and analyzing the spectral response of an object at different frequencies under the thermal excitation pulse using Fourier transform, the pulse phase method is formed. This method inherits the high-efficiency detection characteristics of pulsed thermal imaging technology and effectively suppresses environmental noise and surface non-uniformity interference through frequency domain signal analysis.

[0139] A sequence of N infrared thermal images, for any pixel in each frame of the infrared thermal image ( x , y Perform a Fourier transform on the temperature signal:

[0140] ;

[0141] In the formula, T K For the first K Pixels on a frame image ( x , y Temperature value at ( ); n The index is the result of frequency discretization. j The imaginary unit; , These are the real and imaginary parts of the transformed complex number, respectively.

[0142] By calculating the real part and the virtual part To obtain the amplitude of its Fourier transform. A n With phase :

[0143] ;

[0144] ;

[0145] like Figure 10 Figures (a) and (b) show the amplitude and phase maps obtained from the original sequence images through PPT processing. The PPT processing of the original infrared thermal image sequence yields the amplitude and phase maps at a frequency of 0.2 Hz as follows: Figure 11 As shown, Figure 11 Figures (a) and (b) in the figure are the amplitude map and phase map of the image sequence after background removal, respectively. Figure 11 It can be seen that the amplitude image has relatively less noise and a high contrast between defect-free and defect-free areas, which is beneficial for defect detection, but the influence of uneven heating still exists; while the phase image contains most of the noise, and defect information is masked, making it less effective than the amplitude image. For infrared thermal image sequences, the background noise contained in each frame of the infrared thermal image changes differently over time. Therefore, it is necessary to perform background-based multimodal trust region fitting on each frame of the thermal image, and subtract the fitted background from the original infrared thermal image to obtain a new infrared thermal image sequence. Through analysis... Figure 11Figures (a) and (b) show that the amplitude and phase maps after background removal can detect four more defects with the smallest apertures. Furthermore, the phase map after background removal can more clearly distinguish the four deepest holes. Table 3 shows a comparison of the image performance indicators of the phase maps after PPT processing of the sequence images before and after background removal. After background removal, all performance indicators of the pulse phase map are improved to varying degrees: peak signal-to-noise ratio increases by 1.95%, root mean square error decreases significantly by 53.70%, and information entropy increases by 14.79%.

[0146] Table 3 Comparison of phase image performance metrics of PPT before and after background removal

[0147]

[0148] Principal Component Analysis (PCA), a classic unsupervised linear dimensionality reduction algorithm, essentially projects the original high-dimensional feature space to a low-dimensional subspace through orthogonal transformation. This method is based on eigenvalue decomposition of the covariance matrix, selecting the eigenvector corresponding to the largest eigenvalue as the new basis, achieving dimensionality compression while maximizing the preservation of the original data variance. PCA can effectively extract key features, eliminate noise redundancy, and significantly improve the efficiency and analytical results of high-dimensional data processing.

[0149] Assume the sample vector to be evaluated is composed of n If the data consists of observable variables and the evaluation process satisfies the stochastic condition, then a set of mutually orthogonal coordinate axes needs to be constructed in the original high-dimensional data space. The sample data is then centered based on these coordinate axes. The general model can be expressed as follows:

[0150] ;

[0151] In the formula, Y 1. Y 2. Y m For the first principal component, the second principal component, and the third principal component... m principal component; a ij These are the column vectors of the corresponding coefficient matrix; x 1. x 2. x m It is a common factor.

[0152] PCA processing was performed on the original infrared image sequence to obtain the first to third principal components (PCA1~PCA3). The results are as follows: Figure 12As shown in Figures (a), (b), and (c), PCA1 mainly presents the feature information of the defect area, but there is a severe heat accumulation phenomenon caused by uneven heating. The outline of shallower defects is clear and easy to identify, while the detection of deeper defects in the fourth column is affected. In PCA2 and PCA3, the contrast in the fourth column is even lower, which is less conducive to defect identification. To reduce noise interference, background fitting based on multimodal trust region is used. After removing the fitted background, a new sequence image is obtained, and PCA is applied to the new sequence image. The results are shown in Figure (c). Figure 13 As shown in Figures (a), (b), and (c), comparing the PCA processing results before and after background removal, it is subjectively clear that the contrast of all three principal components is improved, and the noise is significantly reduced, especially... Figure 13 Figure (a) shows that PCA1 can detect four minimum apertures, an increase of three compared to before background removal. Table 4 shows the comparison of the image performance indicators of the first principal component of the sequence images before and after background removal. After background removal, all performance indicators of the pulse phase diagram are improved to varying degrees, with the peak signal-to-noise ratio increasing by 4.37%, the root mean square error decreasing significantly by 52.12%, and the information entropy increasing by 2.37%.

[0153] Table 4 Comparison of the performance indicators of the first principal component image of principal component analysis before and after background removal.

[0154]

[0155] Experiments show that background removal significantly improves image performance, with all performance indicators showing marked improvements. Taking the 360th frame of the infrared sequence image of a flat-bottomed hole defect in glass fiber composite material as an example, after removing the fitted background, the peak signal-to-noise ratio (PSNR) increased by 13.07%, the root mean square error (RMSE) decreased significantly by 58.07%, and the information entropy increased by 21.72%. Furthermore, regardless of whether pulse phase or principal component analysis was used, the performance indicators of the thermal image sequence after background removal were improved to varying degrees. For example, after background removal, the PNR of the pulse phase method image increased by 1.95%, the RMS error decreased significantly by 53.70%, and the information entropy increased by 14.79%. After background removal, the PNR of the first principal component method image increased by 4.37%, the RMS error decreased significantly by 52.12%, and the information entropy increased by 2.37%, with all performance indicators of the first principal component method showing the best results after background removal.

[0156] In summary, this embodiment demonstrates the effectiveness of the proposed infrared sequence image denoising method through comparative experiments. Whether evaluating single-frame images or the resulting images from sequence image processing, removing the background noise from the proposed method improves various performance indicators (peak signal-to-noise ratio, root mean square error, and entropy), especially the root mean square error, which is reduced by more than 50%. This indicates that the background fitting noise removal method based on the multimodal trust region is an effective denoising method.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0158] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A noise reduction method for infrared thermal images of composite materials, characterized in that, Includes the following steps: Obtain the original infrared thermal image sequence of the composite material; The grayscale values ​​of the pixels in the original infrared thermal image are used to fit a preset background fitting surface model to obtain the optimal fitting coefficients. The background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by local heat sources, and the boundary heat dissipation effect. During the fitting process, the optimal fitting coefficients of the background fitting surface model are obtained by adjusting the fitting coefficients of the background fitting surface model with the goal of minimizing the error function. The error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model. Based on the optimal fitting coefficient, the best-fitting thermal image background is determined by the background fitting surface model. The best-fitting thermal image background is removed from the original infrared thermal image to obtain the denoised infrared thermal image. The preset background fitting surface model is: ； In the formula, This is the background fitting function for the original infrared thermal image. A The coefficients are to be fitted. The polynomial basis terms characterizing the low-frequency components of the background temperature field; A Gaussian bulge term to describe temperature anomalies caused by local heat sources; The hyperbolic concave term is used to simulate the boundary heat dissipation effect; ； ； ； In the formula, The coefficients are to be fitted. x , y Indicates location coordinate information; The error function is: Obtain the grayscale values ​​of the pixels in the original infrared thermal image and the predicted grayscale values ​​of the corresponding points in the thermal image background predicted by the background fitted surface model; The difference between the grayscale value of a pixel in the original infrared thermal image and the grayscale prediction value of the corresponding point in the background of the thermal image predicted by the background fitted surface model is called the residual. The error function is defined as follows: by summing the squared residuals of all non-defective pixels. ； In the formula, I ( i , j The standardized infrared thermal image at pixel points ( i , j The grayscale value at () To fit the background surface model with the coefficients to be fitted A coordinates below x and y The predicted grayscale value at that location; The optimal fitting coefficients of the obtained background fitted surface model are obtained by iteratively updating the minimum value of the error function through the trust region algorithm, thus obtaining the optimal solution that minimizes the error function. The step of iteratively updating the minimum value of the error function using the trust region algorithm to obtain the optimal solution that minimizes the error function specifically includes: Define the trust region of the current iteration point as: ； Error function S ( A At the extreme points, it is equivalent to a quadratic function, and a trust region quadratic model is constructed through quadratic approximation. To solve for the step size s k : ； ； In the formula, Error function S ( A At the current iteration point A k gradient at; H k Error function S ( A At the current iteration point A k The Hessian matrix at that location; It is the Euclidean norm; Error function S ( A ) in the k The actual decrease in step is: ； Trust Region Quadratic Model Function The predicted decrease is: ； Define ratio for: ； pass Measuring the quadratic model function Sum of error functions S ( A The degree of approximation, and also based on Determine whether to update the trust region radius for the next iteration. ; Iterate step by step until the convergence condition is met. When the iteration point is updated, the trust region corresponding to the iteration point is determined, and the optimal solution that minimizes the error function is obtained.

2. The noise reduction method for infrared thermal images of composite materials according to claim 1, characterized in that, According to Determine whether to update the trust region radius for the next iteration. Specifically, it includes: Initialize the fitting coefficients And trust region radius And set convergence conditions. ; When the number is reached k During the step, the gradient is calculated. and the Hessian matrix H k ,judge Does it meet the convergence condition? If satisfied, output A = A k To obtain the optimal fitting coefficients, otherwise, by solving the trust region quadratic model, the th... k Step length s k Calculate the ratio ; like A value less than 0 indicates that the error function value is increasing, which is opposite to the optimal solution objective. k = k +1, the iteration point remains unchanged. A k+1 = A k and reduce the trust region radius. Resolve the trust region quadratic model; like A value greater than 0 indicates that the error function value is decreasing, which is consistent with the optimal solution objective. Continue iteration, letting... k = k +1, iteration point ; When 0 < When <0.25, reduce the trust region radius. ; When 0.25 < When <0.75, the trust region radius remains unchanged. ; when When the value is greater than 0.75, increase the trust region radius. And recalculate the gradient. and the Hessian matrix and judge Does it meet the convergence condition? To obtain the optimal fitting coefficients.

3. The noise reduction method for infrared thermal images of composite materials according to claim 2, characterized in that, The determination of the optimal thermal image background specifically includes: By minimizing the optimal solution of the error function, the minimum value of the predicted value corresponding to the unfitted coefficient of the background fitting surface model at the pixel point of the original infrared thermal image is obtained. Based on the position coordinates corresponding to the minimum value and the optimal fitting coefficient, the background fitting surface model corresponding to the optimal fitting coefficient is obtained; The best-fitting thermal image background is determined based on the background fitting surface model corresponding to the optimal fitting coefficient.

4. A noise reduction system for infrared thermal images of composite materials, characterized in that, include: The image acquisition module is used to acquire the original infrared thermal image sequence of the composite material; The background fitting module is used to fit a preset background fitting surface model with the gray values ​​of the pixels in the original infrared thermal image to obtain the optimal fitting coefficients. The background fitting surface model includes terms that characterize the low-frequency component of the background temperature field of the composite material, the temperature anomaly caused by the local heat source, and the boundary heat dissipation effect. The fitting coefficient optimization module is used to adjust the fitting coefficients of the background fitting surface model during the fitting process, with the goal of minimizing the error function, to obtain the optimal fitting coefficients of the background fitting surface model; the error function represents the difference between the gray value of the pixel in the original infrared thermal image and the gray value prediction of the corresponding point in the thermal image background predicted by the background fitting surface model. The image background denoising module is used to determine the best-fitting thermal image background based on the optimal fitting coefficient and the background fitting surface model, and remove the best-fitting thermal image background from the original infrared thermal image to obtain the denoised infrared thermal image. The preset background fitting surface model is as follows: ； In the formula, This is the background fitting function for the original infrared thermal image. A The coefficients are to be fitted. The polynomial basis terms characterizing the low-frequency components of the background temperature field; A Gaussian bulge term to describe temperature anomalies caused by local heat sources; The hyperbolic concave term is used to simulate the boundary heat dissipation effect; ； ； ； In the formula, The coefficients are to be fitted. x , y Indicates location coordinate information; The error function is: Obtain the grayscale values ​​of the pixels in the original infrared thermal image and the predicted grayscale values ​​of the corresponding points in the thermal image background predicted by the background fitted surface model; The difference between the grayscale value of a pixel in the original infrared thermal image and the grayscale prediction value of the corresponding point in the background of the thermal image predicted by the background fitted surface model is called the residual. The error function is defined as follows: by summing the squared residuals of all non-defective pixels. ； In the formula, I ( i , j The standardized infrared thermal image at pixel points ( i , j The grayscale value at () To fit the background surface model with the coefficients to be fitted A coordinates below x and y The predicted grayscale value at that location; The optimal fitting coefficients of the obtained background fitted surface model are obtained by iteratively updating the minimum value of the error function through the trust region algorithm, thus obtaining the optimal solution that minimizes the error function. The step of iteratively updating the minimum value of the error function using the trust region algorithm to obtain the optimal solution that minimizes the error function specifically includes: Define the trust region of the current iteration point as: ； Error function S ( A At the extreme points, it is equivalent to a quadratic function, and a trust region quadratic model is constructed through quadratic approximation. To solve for the step size s k : ； ； In the formula, Error function S ( A At the current iteration point A k gradient at; H k Error function S ( A At the current iteration point A k The Hessian matrix at that location; It is the Euclidean norm; Error function S ( A ) in the k The actual decrease in step is: ； Trust Region Quadratic Model Function The predicted decrease is: ； Define ratio for: ； pass Measuring the quadratic model function Sum of error functions S ( A The degree of approximation, and also based on Determine whether to update the trust region radius for the next iteration. ; Iterate step by step until the convergence condition is met. When the iteration point is updated, the trust region corresponding to the iteration point is determined, and the optimal solution that minimizes the error function is obtained.

Images