Defect detection image reconstruction method, device and equipment of additive manufacturing component and medium

Abstract

The application relates to a defect detection image reconstruction method, device, equipment and medium of an additive manufacturing assembly, wherein a polarized image of a workpiece to be detected manufactured by a selective laser melting technology is subjected to a Stokes characteristic analysis to obtain multiple defect images, each defect image is input into an image reconstruction network, image reconstruction is performed on the defect images, and high-resolution reconstructed images corresponding to the defect images are obtained, wherein the image reconstruction network is a trained generator, the generator comprises an input layer, multiple residual blocks and an output layer connected in sequence, PReLU and Hardswish are used as activation functions in the input layer, and Tanh is used as an activation function in the output layer. The method can obtain multiple high-resolution reconstructed defect images at different angles, so that subsequent efficient detection can be performed on different surface defects.

Description

Technical Field

[0001] This application relates to the field of additive manufacturing technology, and in particular to a method, apparatus, equipment and medium for defect detection image reconstruction of additive manufacturing components. Background Technology

[0002] Additive manufacturing (AM), also known as 3D printing, is a highly competitive manufacturing technology characterized by low cost and high manufacturing freedom. This technology manufactures parts by adding material layer by layer. Unlike traditional material-removing manufacturing methods, additive manufacturing can produce parts with complex geometries and high precision. Due to its unique advantages, additive manufacturing technology is widely used in aerospace, military, medical equipment, energy, and automotive manufacturing, among other fields.

[0003] However, additive manufacturing technologies, especially Selective Laser Melting (SLM), face numerous challenges in practical applications. In SLM, a laser beam selectively melts metal powder, building parts layer by layer. Defects in this process are closely related to the mechanical and physical properties of the final part. Common defects include porosity, cracks, unfused areas, and residual stress, which can severely affect the strength, toughness, and reliability of the part.

[0004] SLM (Surface Mount Technology) technology generates extreme environments during processing, including high temperatures, intense light, and powder splashes, posing significant challenges to defect detection. Traditional detection methods struggle to accurately and effectively identify and analyze defects under such harsh conditions. Furthermore, the surfaces of metal workpieces manufactured using SLM typically have high reflectivity, leading to overexposure of pixels in the image sensor. This can obscure crucial defect information, rendering it undetectable. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, device, and medium for reconstructing defect detection images of additive manufacturing components that can obtain clear defect detection images, in order to address the aforementioned technical problems.

[0006] A defect detection image reconstruction method for additive manufacturing components, the method comprising:

[0007] The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology.

[0008] Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images;

[0009] Each defective image is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as the activation function.

[0010] In one embodiment, the polarization angles of the acquired multiple polarization images are 0°, 45°, 90° and 135°, respectively.

[0011] In one embodiment, the step of performing Stokes characteristic analysis based on each of the polarization images to obtain multiple defect images includes:

[0012] Stokes feature analysis was performed on multiple polarization images to obtain multiple Stokes parameter images;

[0013] The polarization degree image and polarization angle image are obtained by calculating based on the Stokes parameter image;

[0014] The Stokes parameter image, polarization degree image, and polarization angle image are used as the defect image.

[0015] In one embodiment, the Stokes parameter image serving as the defect image employs an intensity difference image of horizontally and vertically polarized light, as well as an intensity difference image of light polarized in two diagonal directions.

[0016] In one embodiment, when training the image reconstruction network, the image reconstruction network includes a generator to be trained and an adversarial unit, and employs the following loss function:

[0017]

[0018] In the above formula, This represents the content loss for the generator. The losses incurred in response to the confrontation are expressed by the following formulas:

[0019]

[0020]

[0021] In the above formula, This indicates that the generator is based on the input image I. LR The reconstructed image obtained by reconstruction, Indicates input image I LR The label image, This represents the reconstructed image obtained by the generator. The probability of being identified as a real image is given. The size of the image is represented as H×W×C, and the real-valued tensor is rH×rW×C. H represents the image height, W represents the image width, rH and rW represent the height and width scaling ratios, respectively, and x and y are the output pixel coordinates in the HR space.

[0022] In one embodiment, in the generator:

[0023] The input layer includes a convolutional layer and a PReLU layer connected in sequence;

[0024] Each residual block includes a first convolutional block, a second convolutional block, and a third convolutional block connected in sequence. Each convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer. The activation function layer in each convolutional block uses PreLU and Hardswish as activation functions, respectively.

[0025] The output layer comprises a convolutional layer, a batch normalization layer, an Elementwise Sun activation function layer, two subpixel convolutional blocks, a convolutional layer, and a Tanh activation function layer connected in sequence.

[0026] In one embodiment, the Hardswish activation function is expressed as:

[0027]

[0028] This application also provides a defect detection image reconstruction apparatus for additive manufacturing components, the apparatus comprising:

[0029] The polarization image acquisition module is used to acquire polarization images corresponding to multiple polarization angles of the workpiece under test, which is manufactured using selective laser melting technology.

[0030] The defect image acquisition module is used to perform Stokes characteristic analysis based on each polarization image to obtain multiple defect images;

[0031] A high-resolution image reconstruction module is used to input each of the defective images into an image reconstruction network to reconstruct the defective images and obtain high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as activation function.

[0032] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0033] The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology.

[0034] Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images;

[0035] Each defective image is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as the activation function.

[0036] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0037] The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology.

[0038] Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images;

[0039] Each defective image is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as the activation function.

[0040] The aforementioned defect detection image reconstruction method, apparatus, equipment, and medium for additive manufacturing components utilize Stokes' characteristics analysis on the polarization image of a workpiece manufactured using selective laser melting (SLM) technology to obtain multiple defect images. Each defect image is then input into an image reconstruction network to reconstruct a high-resolution image for each defect. The image reconstruction network is a pre-trained generator comprising an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish activation functions, while the output layer uses Tanh activation. This method can obtain multiple high-resolution reconstructed defect images from different angles, facilitating efficient detection of various surface defects. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating a defect detection image reconstruction method for a material manufacturing component in one embodiment.

[0042] Figure 2 This is a schematic diagram of the generator network architecture in one embodiment;

[0043] Figure 3 This is a schematic diagram of the discriminator network architecture in one embodiment;

[0044] Figure 4 This is a schematic diagram of the detection of scratches and holes under polarized light angles of 0°, 45°, 90° and 135° in an experiment.

[0045] Figure 5 This is a schematic diagram of solving the transformed polarization angle image Aop, polarization degree image Dop, and Stokes vector images S1 and S2 in an experiment.

[0046] Figure 6 This is a schematic diagram of the image detection results for scratch and hole types Aop, Dop, S1, and S2 in an experiment.

[0047] Figure 7 This is a schematic diagram of the reconstruction of the scratch-type polarization angle image Aop, polarization degree image Dop, and Stokes vector images S1 and S2 after processing with SRGAN-H in an experiment.

[0048] Figure 8 This is a schematic diagram of the reconstruction of the hole-type polarization angle image Aop, polarization degree image Dop, and Stokes vector images S1 and S2 after processing with SRGAN-H in an experiment.

[0049] Figure 9 This is a schematic diagram illustrating the evaluation results of scratch and hole images after SRGAN-H processing in an experiment.

[0050] Figure 10 This is a structural block diagram of a defect detection image reconstruction device for a sub-material manufacturing component in one embodiment;

[0051] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] In existing technologies, when detecting surface defects in metal workpieces manufactured using SLM technology, image detection often obscures much defect information due to the high reflectivity of the metal surface during image acquisition, leading to inaccurate and inefficient defect detection. This application addresses this issue by... Figure 1 As shown, a defect detection image reconstruction method for additive manufacturing components is provided, including the following steps:

[0054] Step S100: Obtain polarization images corresponding to multiple polarization angles of the workpiece to be tested, which is manufactured using selective laser melting technology.

[0055] Step S110: Perform Stokes characteristic analysis based on each polarization image to obtain multiple defect images.

[0056] Step S120: Each defect image is input into the image reconstruction network to reconstruct the defect images and obtain high-resolution reconstructed images of each defect image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks and an output layer connected in sequence. PReLU and Hardswish are used as activation functions in the input layer and Tanh is used as activation function in the output layer.

[0057] In this application, the defect detection and characterization in the SLM process were studied. By introducing polarization technology into defect detection, the defect detection capability of SLM was improved. At the same time, a high-resolution image reconstruction network was introduced to reconstruct defect images with different defect capabilities, so that the reconstructed image for defect detection can clearly show different types of defects on the surface of the metal workpiece, thereby improving the efficiency and accuracy of defect detection.

[0058] Specifically, considering that existing methods for detecting surface defects in workpieces rely solely on measuring acoustic, optical, and thermal sensors, resulting in low signal accuracy, significant limitations, and insufficient information content, making it difficult to detect surface defects in complex environments, this method proposes detection based on polarization images. By optimizing the optical structure of the imaging system, clear defect detection images are obtained. Furthermore, polarization-based defect detection facilitates the extraction of information such as texture structure, surface material, and surface roughness from the polarization information of the target, effectively improving the accuracy and reliability of the detection.

[0059] In step S100, during the process of acquiring the polarization image of the workpiece to be tested, multiple sets of polarization images are acquired by changing the polarization angle. The polarization angles of the acquired multiple polarization images are 0°, 45°, 90°, and 135°, respectively.

[0060] Furthermore, after obtaining multiple polarization images, Stokes characteristic analysis is used to obtain Stokes vector, degree of polarization, and polarization angle images, thereby achieving the purpose of suppressing high reflection on the metal surface, improving the contrast of the defect area in the target image, and highlighting the defect contour information.

[0061] In this embodiment, Stokes feature analysis is performed on each polarization image to obtain multiple defect images, including: performing Stokes feature analysis on multiple polarization images to obtain multiple Stokes parameter images; calculating the degree of polarization image and the polarization angle image based on the Stokes parameter images; and finally, using the Stokes parameter image, the degree of polarization image, and the polarization angle image as the defect images.

[0062] Specifically, the Stokes vector method is used to represent polarized light. The Stokes vector method can be used to represent the polarization state information and intensity of polarized light, indicating that any polarization state of light can be represented by four Stokes vectors, defined as:

[0063]

[0064] In formula (1), S(x,y) is the Stokes vector, represented by a column vector composed of the four parameter images S0, S1, S2, and S3, where (x,y) are spatial coordinates. Specifically, the four parameter images S0, S1, S2, and S3 represent the total intensity image, the intensity difference image between horizontally and vertically polarized light, the intensity difference image between polarized light in the two diagonal directions (45° and 135°), and the intensity difference image between right-handed and left-handed polarized light, respectively.

[0065] According to the Stokes vector method, any polarized light beam can be represented by the polarization angle image Aop, the degree of polarization image Dop, and the polarization ellipticity. To indicate:

[0066]

[0067] At this point, we can also obtain that, during polarized light imaging, the Stokes vector of the incident light is represented by S(x,y), and the Stokes vector of the outgoing light is:

[0068]

[0069] In this embodiment, the Stokes parameter image used as the defect image is the intensity difference image of horizontal and vertical polarized light and the intensity difference image of two diagonally polarized light, namely the two Stokes parameter images S1 and S2.

[0070] In this embodiment, a polarized light image acquisition system for a workpiece under test is also proposed, including the workpiece under test, an LED light source, a linear polarizer, an electric rotating frame, a CMOS camera, and a computer. When acquiring polarized images using this system, the workpiece under test is illuminated by the LED light source. The light is reflected from the workpiece surface and then captured by the CMOS image sensor through the polarizer. The surface of the metal workpiece under test exhibits high reflectivity. The computer controls the rotating frame to adjust the angle of the polarizer, and four sets of defect detection images (S1, S2, Aop, and Dop) with different polarization angles are acquired and solved using the Stokes vector method.

[0071] In the aforementioned polarized light image acquisition system for the workpiece under test, after acquiring polarized images from multiple angles, the computer equipment further performs step S110 to obtain a defect image. In other embodiments, the system may acquire only polarized images from multiple angles. This method is not limited to the aforementioned polarized light image acquisition system; other polarized light image acquisition systems can also be used. Furthermore, the parameters of each component in the aforementioned polarized light image acquisition system are adjusted according to the specific application scenario and are not limited to the parameter settings described above.

[0072] For the four defect images S1, S2, Aop, and Dop obtained from the detection, the next step will be to improve the quality of the detected images at the super-resolution software algorithm level, and to extract information such as the regional contours and basic morphology of the defects more accurately.

[0073] In step S120, the SRGAN-H image reconstruction network model is used to perform effective super-resolution reconstruction of the defect images of the SLM process. This image reconstruction network utilizes an adversarial network of generator and discriminator to achieve super-resolution reconstruction.

[0074] The training process of the image reconstruction network is essentially training a generator that can successfully deceive the discriminator. The resulting generator becomes the core of the image reconstruction network; the discriminator is merely used to assist the generator's training. Through mutual generation and discrimination, a high-quality SR image that conforms to human visual perception is trained and generated.

[0075] In the game between the generator and the discriminator, the generator's goal is to generate SR images that can confuse the discriminator as much as possible, i.e., to minimize the probability that the discriminator will classify its generated images as fake; while the discriminator's goal is to distinguish between real and fake images as much as possible, i.e., to maximize the probability of correct classification. This is achieved by analyzing the discriminator... and generator Alternating optimization to solve the minimization problem is represented as:

[0076]

[0077] A generator G was trained using formula (4) to deceive a differentiable discriminant D, which is used to distinguish between SR images and real images. This method allows the generator to be trained to produce images highly similar to real images.

[0078] In this embodiment, when training the image reconstruction network, the image reconstruction network includes a generator to be trained and an adversarial unit, and adopts the following loss function:

[0079]

[0080] In formula (5), This represents the content loss for the generator. The losses incurred in response to the confrontation are expressed by the following formulas:

[0081]

[0082] In formulas (6) and (7), This indicates that the generator is based on the input image I. LR The reconstructed image obtained by reconstruction, Indicates input image I LR The label image, This represents the reconstructed image obtained by the generator. The probability of being identified as a real image is given. The size of the image is represented as H×W×C, and the real-valued tensor is rH×rW×C. H represents the image height, W represents the image width, rH and rW represent the height and width scaling ratios, respectively, and x and y are the output pixel coordinates in the HR space.

[0083] In this embodiment, as Figure 2 As shown, in the generator: the input layer includes a convolutional layer and a PReLU layer connected in sequence; each residual block includes a first convolutional block, a second convolutional block, and a third convolutional block connected in sequence; each convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer; the activation function layers in each convolutional block use PreLU and Hardswish as activation functions, respectively; the output layer includes a convolutional layer, a batch normalization layer, an Elementwise Sun activation function layer connected in sequence, two sub-pixel convolutional blocks, a convolutional layer, and a Tanh activation function layer.

[0084] Specifically, the generator uses a series of SRResNet modules as its core, employs three convolutional layers with 3×3 kernels and 64 feature maps, and introduces ParametricReLU and Hardswish as activation functions to overcome the edge blurring and gradient vanishing problems caused by the convolutional layers. Then, a batch normalization layer is added.

[0085] Specifically, in terms of activation function selection, the Swish activation function often performs better than ReLU in deeper models. For example, simply replacing ReLU with Swish units can improve the top-1 classification accuracy on ImageNet by 0.9%. Compared to the Swish function, Hardswish has advantages such as better numerical stability and faster computation speed. Therefore, this method introduces the Hardswish activation function to further improve model accuracy and computation speed. The Hardswish activation function can be expressed as follows:

[0086]

[0087] Furthermore, after the residual module ends, the image is processed through convolutional layers and batch normalization layers, and a subpixel convolution module is used to magnify the image by 4 times.

[0088] In the image output stage, the generator network introduces a convolutional layer and Tanh as activation functions to normalize the output to the [-1,1] interval. This reduces the differences between pixels, prevents gradient explosion during neural network training, and accelerates the convergence speed of the model.

[0089] In this embodiment, the discriminator contains multiple 3×3 filter kernels, with a total of 512 kernels. These kernels are activated using LeakyReLU, preferably with α = 0.2, and the entire max-pooling network is avoided for solving the maximization problem of equation (4). The resulting 512 feature maps are activated by two dense layers and the activation functions LeakyReLU and Sigmoid, ultimately yielding a possible classification of real and fake images. The network structure is as follows: Figure 3 As shown.

[0090] In this embodiment, four sets of defect detection images with different polarization angles (S1, S2, Aop, and Dop) were obtained. Then, a super-resolution algorithm was used to represent these images. First, the SRResNet model was trained to accelerate the convergence speed of the SRGAN-H network. Then, the SRGAN-H model was trained in an adversarial process between the generator and the discriminator to obtain a super-resolution reconstruction model that conforms to human visual observation. Finally, this trained model was used to reconstruct the detected images, thus completing the representation of the SLM defect images.

[0091] In this paper, the effectiveness of the proposed method (this method) is also demonstrated through experiments.

[0092] First, quantitative evaluation metrics are used: to quantitatively evaluate the quality of the detected and characterized images, V-channel Value is used to assess the quality of defect detection images during the SLM process, and PSNR, SSIM, and SD are used to assess the quality of super-resolution reconstruction of defect images.

[0093] PSNR stands for Peak Signal-to-Noise Ratio, a metric used to measure image quality. It is typically used to assess the similarity between a reconstructed image and the original image, and can be expressed by the following formula:

[0094]

[0095] In the above formula, MSE represents the mean squared error. The larger the PSNR result, the higher the similarity between the reconstructed image and the original image, and the smaller the image distortion.

[0096] SSIM stands for Structural Similarity Index, a metric used to measure image quality. It considers factors such as brightness, contrast, and structure to assess the similarity between two images and can be expressed by the following formula:

[0097]

[0098] In the above formula, μ x μ y This represents the mean of x and y. σ represents the variance of x and y. xy C1 and C2 represent the covariance; C1 and C2 are constants. The closer the SSIM result is to 1, the higher the similarity between the reconstructed image and the original image in terms of structure, brightness, and contrast, that is, the better the quality of the reconstructed image.

[0099] The standard deviation (SD) usually refers to the standard deviation of the brightness or color distribution of an image. It is a statistical measure of the degree of difference between image pixels and can be represented by equation (11). The larger the SD result, the smaller the difference between pixels.

[0100]

[0101] Next, four sets of defect detection images, representing scratches (#1) and holes (#2), were acquired at polarization angles of 0°, 45°, 90°, and 135°, respectively. Figure 4 As shown.

[0102] The polarization angle image Aop, polarization degree image Dop, and Stokes vector images S1 and S2 of the detected image are calculated according to formulas (2) and (3). 2，Each as Figures 5 to 6 As shown.

[0103] Furthermore, the luminance channel values ​​(V-channel values) of Aop, Dop, S1, and S2 are evaluated. For example... Figure 6 As shown, Stokes vector images S1 and S2 represent the linear polarization information of the defect image. Compared with the intensity image, they have a more uniform grayscale distribution, highlighting the edge contours of some key defect areas, but also losing some detail information. In contrast, the degree of polarization image and the polarization angle image have clearer detail information of the defect area compared to Stokes vector images S1 and S2. This demonstrates that the defect detection method based on Stokes properties is effective in suppressing high radioactivity in metals, improving contrast, and highlighting defect contours.

[0104] Next, using the method presented in this paper, SR (high-resolution image) reconstruction is performed on the Aop, Dop, S1 and S2 images respectively to further enhance their detail and edge information.

[0105] In this experiment, 25,372 images from COCO2014 were randomly selected as the training set, while the test set consisted of widely used BSD100, Set5, Set14, and the Aop, Dop, S1, and S2 image sets of scratch (#1) and hole (#2) defects detected by the above detection systems.

[0106] All experiments were conducted using a 4x scaling factor between low-resolution and high-resolution images, which is equivalent to reducing the image pixel count by a factor of 16. For each mini-batch, 16 96×96 HR sub-images were randomly cropped from different training images. The training of the SRGAN-H model (the image reconstruction model mentioned above) was based on the training of SRResNet based on MSE to accelerate the convergence of SRGAN-H, with a learning rate of η = 10. -4 The residual modules are all set to 16. The number of iterations for SRResNet is set to 130, while the number of iterations for SRGAN-H is set to 50. In SRGAN-H, the generator and discriminator are updated alternately to achieve model training.

[0107] In the experiment, widely accepted datasets BSD100, Set5, Set14, and scratch (#1) and hole (#2) image datasets were selected to evaluate the SRGAN-H reconstruction performance. The test metrics were the average PSNR and SSIM of the entire test set, and the results are shown in Table 1. It can be seen that the SRGAN-H model performs well in reconstructing SLM defect images, and this improvement is suitable for SR reconstruction of SLM defect images. Specifically, the PSNR of scratch images is higher than that of hole images, indicating lower distortion in SR reconstruction of scratch images; the SSIM of hole images is higher than that of scratch images, indicating higher similarity to the original image in terms of brightness, contrast, and structure in SR reconstruction of hole images.

[0108] Table 1 SR Evaluation of Test Set

[0109]

[0110] To verify the SRGAN-H model's performance in super-resolution reconstruction of single images, the experiment randomly cropped portions of the Aop, Dop, S1, and S2 images (scratch and hole categories) obtained above and used them as samples for super-resolution reconstruction. Simultaneously, the experiment compared the original SRGAN model, Nearest, and Bicubic models, demonstrating the significant role of the SRGAN-H model in improving image quality. Figure 7 As shown, for SLM process defect images with scratches, the image quality is improved in terms of PSNR / SSIM / SD after processing by the SRGAN-H model, and the indexes are better than other super-resolution processing methods.

[0111] By combining and reconstructing the four sets of images, from Figure 7 As can be seen from the data, the Stokes vector image S1 has the best reconstruction effect in terms of scratch-type images. Its PNSR value is higher, SSIM is closer to 1, and SD value is smaller. This proves that after SR processing, the image is closer to the original image, and the image distortion is lower. Its brightness, contrast, etc. are closer to the original image, and the image quality is higher.

[0112] For images of defects in the SLM process, such as holes, processing with the SRGAN-H model also improved image quality in terms of PSNR / SSIM / SD metrics. The results are as follows: Figure 8 As shown in the figure, for hole-type images, the polarization angle image Aop has the best reconstruction effect, with a higher PSNR value, a SSIM closer to 1 and a smaller SD value, indicating that its distortion level is lower. The SR image is more consistent with the original image in terms of contrast, brightness and similarity.

[0113] from Figure 9It can be seen that SRGAN-H outperforms the SRGAN / Nearest / Bicubic methods in PSNR / SSIM / SD evaluations. For both scratch-type (#1) and hole-type (#2) images, its reconstruction results are the best, and it also shows advantages in PSNR and SSIM evaluations compared to the other three algorithms. The SD values ​​of SRGAN-H are all lower than the other three methods, indicating that the image reconstructed by SRGAN-H is more similar to the original image.

[0114] This indicates that the polarization angle image, polarization degree image, and Stokes image obtained through polarization technology, after SRGAN-H super-resolution reconstruction, exhibit more prominent edge information and richer detail. This compensates for the shortcomings of polarization detection technology, resulting in a more complete characterization of defects in the SLM process.

[0115] In the aforementioned image reconstruction method for defect detection of additive manufacturing components, Stokes properties are introduced into the defect detection system. A polarization channel imaging system is specifically constructed, and multiple sets of polarization images are obtained by changing the polarization angle. The Stokes vector, degree of polarization, and polarization angle images are then obtained. The obtained images are then evaluated using the brightness channel value (V-channel Value). The results show that the contrast of the defect area, defect contour information, and high reflectance suppression effect of the target image are significantly improved, demonstrating the importance of introducing polarization technology into the defect detection system. Experimental results show that this method can effectively extract and characterize surface defects such as cracks, scratches, and pores, providing solutions and references for defect detection and processing parameter optimization in laser additive manufacturing workpieces. Furthermore, this paper also uses the SRGAN-H model, i.e., the target reconstruction network, to perform SR reconstruction on the obtained target images. In the evaluation of scratch and pore test sets, the PSNR / SSIM values ​​are significantly improved, verifying the effectiveness of the image reconstruction method. Meanwhile, the experiment also performed SR reconstruction on single images of scratches and holes, highlighting the edge features of the defect area and enriching the detailed information of the image, making the characterization of SLM process defects more complete.

[0116] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0117] In one embodiment, such as Figure 10 As shown, a defect detection image reconstruction device for additive manufacturing components is provided, comprising: a polarization image acquisition module 200, a defect image acquisition module 210, and a high-resolution image reconstruction module 220, wherein:

[0118] The polarization image acquisition module 200 is used to acquire polarization images corresponding to multiple polarization angles of the workpiece under test, which is manufactured using selective laser melting technology.

[0119] The defect image acquisition module 210 is used to perform Stokes characteristic analysis based on each polarization image to obtain multiple defect images;

[0120] The high-resolution image reconstruction module 220 is used to input each of the defective images into the image reconstruction network, and to reconstruct the defective images to obtain high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as activation function.

[0121] Specific limitations regarding the defect detection image reconstruction device for additive manufacturing components can be found in the limitations of the defect detection image reconstruction method for additive manufacturing components described above, and will not be repeated here. Each module in the aforementioned defect detection image reconstruction device for additive manufacturing components can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0122] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a defect detection image reconstruction method for additive manufacturing components. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0123] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0124] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0125] The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology.

[0126] Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images;

[0127] Each defective image is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as the activation function.

[0128] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0129] The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology.

[0130] Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images;

[0131] Each defective image is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer uses PReLU and Hardswish as activation functions, and the output layer uses Tanh as the activation function.

[0132] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0133] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0134] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for reconstructing defect images of additively manufactured components, characterized in that, The method includes: The polarization images corresponding to multiple polarization angles of the workpiece under test are acquired. The workpiece under test is manufactured using selective laser melting technology. Stokes characteristic analysis was performed on each of the polarization images to obtain multiple defect images; Each of the defective images is input into an image reconstruction network to reconstruct the defective images, thereby obtaining high-resolution reconstructed images of each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer includes a convolutional layer and a PReLU layer connected in sequence; Each residual block includes a first convolutional block, a second convolutional block, and a third convolutional block connected in sequence. Each convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer. The activation function layer in each convolutional block uses PReLU and Hardswish as activation functions, respectively. The output layer includes a convolutional layer, a batch normalization layer, an Elementwise Sum activation function layer, two sub-pixel convolutional blocks, a convolutional layer, and a Tanh activation function layer connected in sequence. The polarization angles of the acquired multiple polarization images are 0°, 45°, 90° and 135°, respectively; The Stokes characteristic analysis performed on each of the polarization images to obtain multiple defect images includes: Stokes feature analysis was performed on multiple polarization images to obtain multiple Stokes parameter images; The polarization degree image and polarization angle image are obtained by calculating based on the Stokes parameter image; The Stokes parameter image, polarization degree image, and polarization angle image are used as the defect image.

2. The defect detection image reconstruction method according to claim 1, characterized in that, The Stokes parameter image used as the defect image is an image of the intensity difference between horizontally and vertically polarized light, as well as an image of the intensity difference between two diagonally polarized light.

3. The defect detection image reconstruction method according to claim 1 or 2, characterized in that, When training the image reconstruction network, the image reconstruction network includes a generator to be trained and an adversarial unit, and employs the following loss function: In the above formula, This represents the content loss for the generator. The adversarial losses for the countermeasures are expressed by the following formulas: In the above formula, This indicates that the generator is based on the input image. The reconstructed image obtained by reconstruction, Indicates the input image The label image, This represents the reconstructed image obtained by the generator. The probability of being classified as a real image, where the image size is represented as... The real-valued tensor is , Indicates the image height. Indicates the image width. and These represent the height and width scaling ratios, respectively. , These are the output pixel coordinates in the HR space.

4. The defect detection image reconstruction method according to claim 3, characterized in that, The Hardswish activation function is expressed as: 。 5. A defect detection image reconstruction device for an additive manufacturing component, characterized in that, The defect detection image reconstruction method according to any one of claims 1 to 4, the apparatus comprising: The polarization image acquisition module is used to acquire polarization images corresponding to multiple polarization angles of the workpiece under test, which is manufactured using selective laser melting technology. The defect image acquisition module is used to perform Stokes characteristic analysis based on each polarization image to obtain multiple defect images; A high-resolution image reconstruction module is used to input each of the defective images into an image reconstruction network to reconstruct the defective images and obtain high-resolution reconstructed images corresponding to each defective image. The image reconstruction network is a trained generator, which includes an input layer, multiple residual blocks, and an output layer connected in sequence. The input layer includes a convolutional layer and a PReLU layer connected in sequence. Each residual block includes a first convolutional block, a second convolutional block, and a third convolutional block connected in sequence. Each convolutional block includes a convolutional layer, a batch normalization layer, and an activation function layer. The activation function layer in each convolutional block uses PReLU and Hardswish as activation functions, respectively. The output layer includes a convolutional layer, a batch normalization layer, an Elementwise Sum activation function layer, two sub-pixel convolutional blocks, a convolutional layer, and a Tanh activation function layer connected in sequence.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images