A micro-defect active thermal imaging space encoding super-resolution detection method and system

Abstract

The application provides a kind of micro defect active thermal imaging space encoding super-resolution detection method and system, belong to nondestructive testing and image processing technical field, it can partially solve the problem that the resolution of long-wave infrared imaging is insufficient in prior art, leading to fuzzy micro defect boundary, inaccurate positioning and poor noise resistance.The method comprises: applying external excitation to the sample to be detected and collecting low-resolution thermal imaging data; based on a multi-scale reconstruction strategy, the low-resolution thermal imaging data is recovered by region sparse based on the suppression of reconstruction noise, and a LASSO model is used to fuse to obtain a reconstructed image; the reconstructed image is input into a neural network for recovery optimization, and an output super-resolution thermal imaging image is obtained; threshold segmentation is performed on the super-resolution thermal imaging image to realize defect positioning, and the defect position information and / or size information are output.The application has the advantages of obvious resolution improvement, strong noise resistance, high positioning accuracy and easy engineering integration.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing and image processing technology, specifically relating to a spatially coded super-resolution detection method and system for active thermal imaging of micro-defects. Background Technology

[0002] With the increasing demands for reliability in composite materials, metals, and PCB structures in aerospace, rail transportation, and electronic packaging industries, the need for non-destructive testing (NDT) of micro-defects (such as microcracks, micropores, debonding, and inclusions) has significantly increased. Active thermal imaging NDT, due to its advantages of being non-contact, highly efficient, and widely applicable, has become a crucial method for detecting surface and near-surface defects in materials. However, in micro-defect detection scenarios, the spatial resolution is often insufficient due to limitations such as the pixel size of the focal plane array and imaging distance of long-wave infrared cameras. This leads to blurred defect boundaries, large size estimation errors, and even missed detections. Furthermore, the lateral thermal diffusion after excitation heating further smooths the temperature gradient, causing the thermal response characteristics of micro-defects to be obscured by the background, reducing defect distinguishability and location accuracy.

[0003] While existing technologies such as micro-scanning and optical super-resolution can improve spatial resolution, they typically require complex mechanical structures or high-cost optical components, making system implementation difficult and stability requirements challenging, thus hindering engineering deployment. Deep learning-based super-resolution methods, although capable of improving image clarity to some extent, rely heavily on large amounts of high-quality labeled data and often "predict" high-resolution details from low-resolution images, lacking support from real structural information. Furthermore, they exhibit insufficient generalization ability when facing different materials, noise levels, and thermal diffusion conditions, easily leading to artifacts or misjudgments. Therefore, we propose a spatially encoded super-resolution detection method and system for active thermal imaging of micro-defects. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method and system for active thermal imaging spatial coding super-resolution detection of micro-defects.

[0005] This invention provides a spatially encoded super-resolution detection method for active thermal imaging of micro-defects, comprising the following steps: S1: Apply external excitation to the sample to be tested and collect thermal imaging observation data to obtain low-resolution thermal imaging data containing defect response information; S2: Reconstruct the low-resolution thermal imaging data to obtain a high-resolution reconstructed image of the low-resolution thermal imaging data; S3: Input the reconstructed image into a neural network for image restoration and optimization, and output a super-resolution thermal imaging image of the reconstructed image; S4: Based on the super-resolution thermal imaging image, locate the defect and output the location information and / or size information of the defect in the super-resolution thermal imaging image.

[0006] Furthermore, in step S1, the external excitation is optical excitation, and the thermal imaging observation data is acquired by an infrared camera.

[0007] Specifically, in step S1, the signal of the external excitation is modulated using a spatial coding method, so that the external excitation acts on the surface of the sample to be detected with a preset spatial coding pattern, and the low-resolution thermal imaging observation data corresponding to the spatial coding pattern is collected.

[0008] Specifically, the infrared camera operates in the 7~14μm band and has a sampling frame rate of no less than 30fps.

[0009] Preferably, the spatial encoding is achieved by a digital micromirror device (DMD), and the optical excitation is near-infrared laser excitation.

[0010] Specifically, in step S2, the reconstruction process includes multi-scale reconstruction, which divides the low-resolution thermal imaging data into multiple regions of different sizes and performs sparse restoration on each region to obtain a multi-scale reconstruction result.

[0011] Furthermore, the different sized regions include 1×1, 3×3, and 5×5 regions, which are divided using a sliding window method, with the step size of the sliding window ranging from 1 to 2 pixels.

[0012] Furthermore, the sparse recovery is solved using the LASSO model, and the reconstruction noise is suppressed by the multi-scale reconstruction method. The results of the multi-scale reconstruction are fused by a weighted vector, and the weighted vector is adjusted according to the noise intensity of the low-resolution thermal imaging data.

[0013] Specifically, in step S3, the neural network is an encoder-decoder structure network, which includes a U-Net network, a network with residual modules, and a network with an attention mechanism. The input of the neural network is the reconstructed image or an enhanced image generated from the reconstructed image.

[0014] Another aspect of the present invention provides a spatially encoded super-resolution detection system for active thermal imaging of micro-defects, comprising: The excitation and acquisition module is used to apply the external excitation to the sample to be detected and acquire the low-resolution thermal imaging data; The reconstruction module is used to reconstruct the low-resolution thermal imaging data to obtain the reconstructed image; A neural network module is used to restore and optimize the reconstructed image and output the super-resolution thermal imaging image; and The defect localization module is used to locate defects based on the super-resolution thermal imaging image and output the location information or size information of the defects in the super-resolution thermal imaging image.

[0015] The beneficial effects of this invention are as follows: This invention achieves super-resolution imaging of low-resolution thermal images by combining spatial coding excitation acquisition with reconstruction and network recovery optimization, making the boundaries of micro-defects clearer; multi-scale sparse reconstruction and neural network joint optimization can effectively reduce the blurring caused by thermal diffusion and improve the accuracy of defect localization and size assessment; the introduction of multi-scale fusion and adaptive weight adjustment of noise intensity enhances the reconstruction stability under different noise conditions; training data is constructed based on a physical constraint model and noise enhancement is performed to reduce the dependence on a large amount of real high-resolution labeled data and improve engineering applicability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a spatially encoded super-resolution detection method for active thermal imaging of micro-defects according to a specific embodiment of the present invention. Figure 2 This image shows the defect boundary recognition effect of an aluminum alloy standard part after threshold segmentation, based on a specific embodiment of the micro-defect active thermal imaging spatial coding super-resolution detection method of the present invention. Figure 3 This is a schematic diagram of the network structure of a spatially encoded super-resolution detection method for active thermal imaging of micro-defects according to a specific embodiment of the present invention. Figure 4 This is a schematic diagram of the active thermal imaging super-resolution detection device for a micro-defect active thermal imaging spatial coding super-resolution detection method according to a specific embodiment of the present invention. Figure 5 The images show the effect of an aluminum alloy standard part (containing a 1mm×1mm micro-defect) before and after processing by the active thermal imaging spatial coding super-resolution detection method for micro-defects according to a specific embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a specific embodiment of the present invention provides a spatially encoded super-resolution detection method for active thermal imaging of micro-defects, comprising the following steps: S1: Apply external excitation to the sample to be tested and collect thermal imaging observation data to obtain low-resolution thermal imaging data containing defect response information; S2: Reconstruct the low-resolution thermal imaging data to obtain a high-resolution reconstructed image of the low-resolution thermal imaging data; S3: Input the reconstructed image into the neural network for image restoration and optimization, and output a super-resolution thermal image of the reconstructed image; S4: Defect localization based on super-resolution thermal imaging images, outputting the location and / or size information of the defects in the super-resolution thermal imaging images.

[0019] Specifically, in step S3, based on the equivalent heat flux model, the temperature field distribution of micro-defects of different shapes is simulated by frequency domain calculation to generate a training dataset with physical constraints, and multi-intensity noise is introduced for enhancement. The equivalent heat flux model is derived based on the heat conduction equation and the surface temperature field is calculated by triple Fourier transform (x, y, t domains). The simulation time for a single set of thermal imaging data is ≤0.67s. In the noise enhancement step, two types of Gaussian noise with PSNR of 19.6dB and 12.96dB are introduced to simulate noise interference in actual testing and improve the generalization ability of the model. Noise enhancement is achieved by superimposing Gaussian noise on the training samples and using peak signal-to-noise ratio (PSNR) to characterize the noise intensity. PSNR represents the ratio of the strength of the signal (ideal image) to the noise (error). In step S2, a multi-scale compressed sensing strategy is used for the collected low-resolution data to divide the region into 1×1, 3×3, and 5×5 different sizes for sparse restoration, and the LASSO model is combined to suppress reconstruction noise. In this step, regions of different sizes are divided using a sliding window with a step size of 1. The reconstruction process is achieved by pairing the measurement matrix (Input1) and the mask matrix (Input2). The reconstruction results at each scale are fused using weighted vectors (b1, b2, b3). The weight configuration is dynamically adjusted according to the noise level of the observed data (increasing the weight of small-scale regions in high-noise scenarios).

[0020] The construction of the training dataset includes the following process: (1) Establish the defect shape matrix D(x, y), where the defect region takes the value of 1 and the non-defect region takes the value of 0; (2) Set the surface heat flux q(x, y, t) according to the excitation mode to be simulated, and superimpose the defect shape matrix D(x, y) with the heat flux q(x, y, t) to obtain the equivalent heat flux q_tol(x, y, t); (3) Substitute the equivalent heat flux q_tol(x, y, t) into the frequency domain expression of the equivalent heat flux model, and obtain the frequency domain temperature field distribution by performing triple Fourier transform on the spatial domain x, y and the time domain t, and further obtain the time domain temperature field T_tol(x, y, t) by triple inverse Fourier transform. (4) Normalize the temperature field data to obtain noise-free high-resolution temperature response samples; (5) Gaussian noise of different intensities is superimposed on the normalized temperature response samples so that the PSNR between the superimposed samples and the original samples are 19.6dB and 12.96dB, respectively, to construct training samples with multiple noise levels, thereby improving the robustness of the neural network to actual collected noise; the training dataset is divided into training set, validation set and test set, preferably in a ratio of 8:1:1; the number of samples in each dataset is preferably 550, and different defect shapes include but are not limited to rectangular defects, circular defects, strip defects and irregular defects; The multi-scale compressed sensing reconstruction process may include the following steps: (1) Define the low-resolution thermal imaging observation data obtained by spatial coding as measurement matrix Input1, wherein Input1 contains the observation results corresponding to m coding acquisitions; (2) Define the coding mask corresponding to the spatial coding pattern as the mask matrix Input2, wherein Input2 and Input1 correspond one-to-one in the acquisition sequence frame by frame; (3) Divide Input1 into regions with three scales: 1×1, 3×3, and 5×5. Use a sliding window to traverse the regions pixel by pixel to cover them. (4) For each scale region, construct the corresponding sparse recovery problem, and solve it through the LASSO model to obtain the sparse vector β, so as to obtain the reconstruction result at that scale; (5) The reconstruction results at each scale are denoted as S1(n, H, W), where n is the scale number, and H and W are the height and width of the reconstructed image, respectively. (6) The multi-scale reconstruction results are weighted and fused using a weighted vector b=[b1, b2, b3] to obtain the fused reconstructed image; The preferred method for dynamic adjustment of noise level is to obtain the background area of ​​Input1 through statistical estimation. Specifically, the noise intensity can be characterized by calculating the variance, standard deviation, or signal-to-noise ratio of the background area. When the noise intensity is high, the weight b1 corresponding to the small-scale area is increased, and when the noise intensity is low, the weight b3 corresponding to the large-scale area is increased to balance the ability to preserve details and the ability to suppress noise.

[0021] Furthermore, in step S3, the multi-scale reconstruction result is used as the input to the upsampling layer of the U-Net neural network. The image features are optimized by a recovery network composed of ResNet and an attention mechanism to suppress thermal diffusion blurring. Before the network input, the gray-scale contrast of the defects is amplified by using the formula Y=arctan(up(In3))exp(μ*S2) (μ is the enhancement coefficient) based on low-resolution interpolated data (In3). The network training process uses PSNR as the evaluation index, and the training time is approximately 166.12s to ensure that the model converges to the optimal state quickly.

[0022] Furthermore, in step S4, threshold segmentation and IOU calculation are performed on the super-resolution image output by the network to locate the boundaries of micro-defects and output high-resolution detection results. For color thermal imaging data, the above process can be performed independently on the temperature pseudo-color channel and then merged for output. Before output, the extra boundaries introduced by zero padding need to be cropped to restore the original image size (e.g., reconstructing 7×7 low-resolution data into 28×28 high-resolution data to achieve 4× super-resolution). Threshold segmentation includes extracting the defect region using fixed threshold segmentation or adaptive threshold segmentation, preferably using the Otsu adaptive thresholding method to obtain the segmentation threshold. Morphological processing is performed on the segmentation results to remove isolated noise and smooth the defect boundaries. Morphological processing includes, but is not limited to, opening operations, closing operations, or connected component filtering. The defect localization results include the coordinates of the bounding box of the defect, the coordinates of the defect center, and the defect size parameters, including the defect length, width, or area. By converting the pixel area of ​​the segmented region with the camera spatial resolution, the actual size information of the defect can be obtained. The IOU calculation is used to evaluate the degree of overlap between the segmentation results and the high-resolution ground truth. The higher the IOU, the higher the defect localization accuracy. Preferably, the IOU is not less than 80%.

[0023] Based on the above basic implementation method, in step S1, the external excitation is optical excitation, and the thermal imaging observation data is acquired by an infrared camera; in step S1, the signal of the external excitation is modulated by spatial coding so that the external excitation acts on the surface of the sample to be detected with a preset spatial coding pattern, and low-resolution thermal imaging observation data corresponding to the spatial coding pattern is acquired.

[0024] Specifically, the spatial coding pattern includes random coding pattern, pseudo-random coding pattern, or structured coding pattern; preferably, a two-dimensional Bernoulli random coding matrix is ​​used as the spatial coding mask to enhance the ability to supplement the true structural information of micro-defects.

[0025] Furthermore, spatial coding acquisition is achieved through synchronous control, that is, when each frame of spatial coding pattern is loaded onto the coding device, the infrared camera is triggered to perform synchronous acquisition, thereby ensuring a one-to-one correspondence between Input1 and Input2 in the time series.

[0026] In one specific implementation, the infrared camera operates in the 7~14μm band and has a sampling frame rate of no less than 30fps; spatial coding is achieved through a digital micromirror device (DMD), and optical excitation is near-infrared laser excitation.

[0027] In this embodiment, a 20W, 808nm near-infrared laser is used as the excitation source. After beam expansion by an 8× Galilean beam expander, spatial coding modulation at the 1200×800 pixel level is achieved by a DMD. Simultaneously, a long-wave infrared camera (7-14μm band, 640×480 pixels) is controlled to acquire observation data at an imaging distance of 40cm. The excitation time is set to 2s and the sample cooling time is set to 5s to ensure that the data contains complete defect thermal response information.

[0028] Furthermore, the sample to be tested is preferably a metallic or composite material sample, and the defect types include, but are not limited to, holes, cracks, debonding, or inclusions.

[0029] In another specific embodiment, in step S2, the reconstruction process includes multi-scale reconstruction, which divides the low-resolution thermal imaging data into multiple regions of different sizes and performs sparse restoration on each region to obtain the multi-scale reconstruction result; the regions of different sizes include 1×1, 3×3, and 5×5 regions, and the regions of different sizes are divided using a sliding window method, with the step size of the sliding window ranging from 1 to 2 pixels.

[0030] Specifically, the sliding window step size is preferably 1 pixel to ensure that the reconstruction result covers all pixel positions and avoids sampling omissions; when it is necessary to improve the reconstruction speed, the step size can be set to 2 pixels to reduce the amount of computation.

[0031] Furthermore, to ensure continuity at the boundaries of the reconstructed region, weighted averaging or weighted fusion can be used to smooth the overlapping areas of the windows, thereby reducing block artifacts and improving the overall consistency of the reconstructed image.

[0032] In another specific embodiment, sparse recovery is solved using the LASSO model, and reconstruction noise is suppressed by multi-scale reconstruction. The results of multi-scale reconstruction are fused by weighted vectors, and the weighted vectors are adjusted according to the noise intensity of the low-resolution thermal imaging data. In step S3, the neural network is an encoder-decoder structure network, which includes a U-Net network, a network with residual modules, and a network with attention mechanisms. The input of the neural network is the reconstructed image or the enhanced image generated from the reconstructed image.

[0033] Furthermore, the weighted vector configuration methods include: in high-noise scenarios, b=[1, 2, 3] is used to increase the proportion of small-scale reconstruction results, thereby enhancing the ability to recover local details; in low-noise scenarios, b=[3, 2, 1] is used to increase the proportion of large-scale reconstruction results, thereby enhancing the overall structural consistency and smoothness.

[0034] Furthermore, neural networks can select different structural forms according to the application scenario. Without changing the input and output interfaces, the U-Net structure can be replaced with other encoder-decoder networks to achieve further recovery and optimization of multi-scale reconstruction results, thereby improving the clarity of micro-defect boundaries and the accuracy of localization.

[0035] In one specific embodiment, another aspect of the present invention provides a micro-defect active thermal imaging spatial coding super-resolution detection system, which is suitable for implementing the above-described micro-defect active thermal imaging spatial coding super-resolution detection method, comprising: The excitation and acquisition module is used to apply external excitation to the sample to be tested and acquire low-resolution thermal imaging data; the reconstruction module is used to reconstruct the low-resolution thermal imaging data to obtain a reconstructed image; the neural network module is used to restore and optimize the reconstructed image and output a super-resolution thermal imaging image; and the defect localization module is used to locate defects based on the super-resolution thermal imaging image and output the location or size information of the defect in the super-resolution thermal imaging image.

[0036] In this embodiment, the excitation and acquisition module includes an excitation source, a spatial coding device, a synchronization controller, and an infrared camera. The synchronization controller is used to achieve temporal synchronization between the loading of the spatial coding pattern and the acquisition by the infrared camera. The reconstruction module includes a sparse recovery unit and a multi-scale fusion unit to complete multi-scale compressed sensing reconstruction and weighted fusion. The neural network module includes a model loading unit and an inference calculation unit to complete super-resolution image output. The defect localization module includes a threshold segmentation unit, a connected component analysis unit, and a parameter output unit to output defect boundary coordinates and size information.

[0037] In one specific implementation, such as Figure 3 , Figure 4 , Figure 5As shown, the sample to be tested was obtained, with the aluminum alloy standard part having a size of 20mm×20mm (including two types of defects: 1mm×1mm and 2mm×4mm); the sample was fixed on a room temperature test bench (25℃), ensuring that the defect area was facing the infrared camera lens; a two-dimensional Bernoulli random coding matrix (Input2, size m×H×W, where m is the number of coding acquisitions, and H and W are the image height and width, respectively) was loaded onto the DMD; the near-infrared laser was irradiated onto the sample surface after coding; the infrared camera simultaneously acquired low-resolution data after coding excitation (Input1, size m×H / 4×W / 4), and each acquisition was stored as a 16-bit grayscale image.

[0038] In this embodiment, the original image size H=W=28 pixels (corresponding to an actual 6mm×6mm area), and the low-resolution data size is 7×7 pixels; based on the equivalent heat flux model, the defect shape matrix D(x,y) (defect area D=1, non-defect area D=0) and the surface heat flux q(x,y,t) are input, and the following formula is used to calculate:

[0039] Calculate the superimposed equivalent heat flux Substitute the formula to calculate the temperature field in the frequency domain, and then perform a triple inverse Fourier transform to obtain the temperature field in the time domain. :

[0040] After normalizing the temperature data, Gaussian noise of PSNR=19.6dB and 12.96dB was added respectively to generate three datasets (each containing 550 samples, divided into training, validation and test sets in an 8:1:1 ratio).

[0041] Specifically, Input1 and Input2 are paired, and regions are divided into 1×1, 3×3, and 5×5 areas. The sparse vector β is solved using the LASSO model to obtain the reconstruction results S1 (n,H,W) at each scale (n=3, corresponding to 3 scales). The reconstruction results are fused using a weighted vector b=[1,2,3] (for high-noise scenes) or b=[3,2,1] (for low-noise scenes) to obtain a preliminary denoised high-resolution image. For the fused reconstructed image S2, based on the low-resolution interpolated data In3 (7×7 data is enlarged to 28×28 through bilinear interpolation), the values ​​are substituted into the formula Y=arctan(up) The algorithm (In3)exp(μ*S2) calculates the enhanced image; where μ is set to 1.0 to balance defect contrast and noise suppression. The network backbone consists of a ResNet block (3 layers) and an attention block (1 layer). The input is the enhanced image Y, and the output is the optimized 28×28 super-resolution image. The network is trained using the Adam optimizer with an initial learning rate of 1e-4, converging after 100 iterations. Both training and inference support GPU acceleration. The Otsu adaptive thresholding method is used to extract defect regions, and the IOU value between the segmentation result and the high-resolution ground truth is calculated, such as... Figure 2 The diagram shows the assessment of defect location accuracy.

[0042] To aid in a better understanding of the present invention, a more comprehensive and specific embodiment is described, in which the present invention provides a spatially encoded super-resolution detection method for active thermal imaging of micro-defects, comprising the following steps: S1: Apply external excitation to the sample to be tested and collect thermal imaging observation data to obtain low-resolution thermal imaging data containing defect response information; S2: Reconstruct the low-resolution thermal imaging data to obtain a high-resolution reconstructed image of the low-resolution thermal imaging data; S3: Input the reconstructed image into the neural network for image restoration and optimization, and output a super-resolution thermal image of the reconstructed image; S4: Defect localization based on super-resolution thermal imaging images, outputting the location and / or size information of the defects in the super-resolution thermal imaging images.

[0043] In this embodiment, in step S1, the external excitation is optical excitation, and the thermal imaging observation data is acquired by an infrared camera.

[0044] Specifically, in step S1, the signal of the external excitation is modulated using spatial coding, so that the external excitation acts on the surface of the sample to be detected with a preset spatial coding pattern, and low-resolution thermal imaging observation data corresponding to the spatial coding pattern are collected; the working band range of the infrared camera is 7~14μm, and the sampling frame rate of the infrared camera is not less than 30fps; spatial coding is implemented by a digital micromirror device (DMD), and the optical excitation is near-infrared laser excitation; in step S2, the reconstruction processing includes multi-scale reconstruction, which divides the low-resolution thermal imaging data into multiple regions of different sizes and performs sparse recovery on each region to obtain the multi-scale reconstruction result; different sizes The inch region includes 1×1, 3×3, and 5×5 regions. Different sized regions are divided using a sliding window method, with a step size of 1 to 2 pixels. Sparse recovery is solved using the LASSO model, and reconstruction noise is suppressed by multi-scale reconstruction. The results of multi-scale reconstruction are fused using weighted vectors, and the weighted vectors are adjusted according to the noise intensity of the low-resolution thermal imaging data. In step S3, the neural network is an encoder-decoder structure network, which includes a U-Net network, a network with residual modules, and a network with attention mechanisms. The input of the neural network is the reconstructed image or an enhanced image generated from the reconstructed image.

[0045] Specifically, on the other hand, an active thermal imaging spatial coding super-resolution detection system for micro-defects is provided. This system is suitable for implementing the aforementioned active thermal imaging spatial coding super-resolution detection method for micro-defects, and includes: The excitation and acquisition module is used to apply external excitation to the sample to be tested and acquire low-resolution thermal imaging data; the reconstruction module is used to reconstruct the low-resolution thermal imaging data to obtain a reconstructed image; the neural network module is used to restore and optimize the reconstructed image and output a super-resolution thermal imaging image; and the defect localization module is used to locate defects based on the super-resolution thermal imaging image and output the location or size information of the defect in the super-resolution thermal imaging image.

[0046] In summary, the embodiments disclosed herein have at least the following technical effects: This invention effectively supplements the true structural information of defects by spatially encoding and modulating external stimuli and collecting thermal imaging observation data, combined with reconstruction processing and neural network recovery optimization, thereby reducing the problem of blurred defect boundaries caused by insufficient resolution of long-wave infrared cameras and realizing high-resolution imaging and clear presentation of micro-defects. This invention employs sparse recovery and denoising constraints in the reconstruction stage, and restores and optimizes image features in the neural network stage. This can suppress the temperature field smoothing effect caused by lateral thermal diffusion, enhance the contrast of defect thermal response, and improve the distinguishability and positioning accuracy of defect boundaries. This invention employs a multi-scale regional sparse restoration strategy and fuses the reconstruction results at each scale using a weighted vector. The weights can be dynamically adjusted according to the noise intensity of the low-resolution thermal imaging data, thereby obtaining stable and reliable reconstruction results under different noise levels and significantly reducing reconstruction noise and artifacts. This invention constructs a training dataset with physical constraints based on an equivalent heat flux model and introduces multi-intensity noise enhancement, which can complete model training even in the absence of a large amount of real high-resolution labeled data, thereby improving the feasibility and promotion value of super-resolution detection methods. This invention uses reconstructed images (or enhanced images) as input to a neural network for recovery optimization, avoiding the problem of false textures and structural deviations caused by directly predicting high-resolution details from low-resolution images. This makes the super-resolution output more consistent with the real defect structure and enhances the model's generalization ability under different materials and different defect morphologies. This invention can achieve super-resolution detection based on conventional infrared cameras through spatial coding excitation and algorithm reconstruction, without the need for complex mechanical micro-scanning devices or high-cost optical super-resolution hardware, thus reducing system costs and implementation difficulty, and is suitable for rapid deployment in industrial sites.

[0047] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A spatially encoded super-resolution detection method for micro-defects using active thermal imaging, characterized in that, Includes the following steps: S1: Apply external excitation to the sample to be tested and collect thermal imaging observation data to obtain low-resolution thermal imaging data containing defect response information; S2: Reconstruct the low-resolution thermal imaging data to obtain a high-resolution reconstructed image of the low-resolution thermal imaging data; S3: Input the reconstructed image into a neural network for image restoration and optimization, and output a super-resolution thermal imaging image of the reconstructed image; S4: Based on the super-resolution thermal imaging image, locate the defect and output the location information and / or size information of the defect in the super-resolution thermal imaging image.

2. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 1, characterized in that, In step S1, the external excitation is optical excitation, and the thermal imaging observation data is acquired by an infrared camera.

3. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 2, characterized in that, In step S1, the signal of the external excitation is modulated using a spatial coding method, so that the external excitation acts on the surface of the sample to be detected with a preset spatial coding pattern, and the low-resolution thermal imaging observation data corresponding to the spatial coding pattern is collected.

4. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 2, characterized in that, The infrared camera operates in the 7~14μm band and has a sampling frame rate of no less than 30fps.

5. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 3, characterized in that, The spatial encoding is achieved through a digital micromirror device (DMD), and the optical excitation is near-infrared laser excitation.

6. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 1, characterized in that, In step S2, the reconstruction process includes multi-scale reconstruction, which divides the low-resolution thermal imaging data into multiple regions of different sizes and performs sparse restoration on each region to obtain a multi-scale reconstruction result.

7. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 6, characterized in that, The different sized regions include 1×1, 3×3, and 5×5 regions. The different sized regions are divided using a sliding window method, and the step size of the sliding window ranges from 1 to 2 pixels.

8. The micro-defect active thermal imaging spatial coding super-resolution detection method according to claim 6, characterized in that, The sparse recovery is solved using the LASSO model, and the reconstruction noise is suppressed by the multi-scale reconstruction method. The results of the multi-scale reconstruction are fused by a weighted vector, which is adjusted according to the noise intensity of the low-resolution thermal imaging data.

9. The micro-defect active thermal imaging spatial coding super-resolution detection method according to any one of claims 1 to 8, characterized in that, In step S3, the neural network is an encoder-decoder structure network, which includes a U-Net network, a network with residual modules, and a network with an attention mechanism. The input of the neural network is the reconstructed image or an enhanced image generated from the reconstructed image.

10. A spatially coded super-resolution detection system for active thermal imaging of micro-defects, characterized in that, include: The excitation and acquisition module is used to apply the external excitation to the sample to be detected and acquire the low-resolution thermal imaging data; The reconstruction module is used to reconstruct the low-resolution thermal imaging data to obtain the reconstructed image; The neural network module is used to restore and optimize the reconstructed image and output the super-resolution thermal imaging image; as well as The defect localization module is used to locate defects based on the super-resolution thermal imaging image and output the location information or size information of the defects in the super-resolution thermal imaging image.

Images