High-precision industrial CT (computed tomography) size measurement method used under diffuse metal artifact working condition

Abstract

The invention relates to a high-precision industrial CT (Computed Tomography) size measurement method used under a diffuse metal artifact working condition, and belongs to the field of nondestructive testing and image processing. According to the method, a three-dimensional CT image of a detected workpiece is obtained, after slicing processing is carried out, the Poisson distribution noise reduction technology is adopted to reduce image noise, and a double-threshold method is adopted to remove highlight metal artifacts. Then, through binarization, corrosion expansion and opening and closing operation processing, a primary rough edge is extracted, and fine correction is carried out; and the adaptability of the model is enhanced by using a data enhancement technology. Through training of a high-precision conditional adversarial generative network, an image with fine edge information is obtained. And finally, key parameters such as straightness, inner diameter, outer diameter, convex curvature radius, concave curvature radius, tube wall thickness and thickness standard deviation of the tubular material are calculated and output. According to the method, the precision and reliability of industrial CT size measurement can be remarkably improved under the condition of diffuse metal artifacts.

Description

Technical Field

[0001] The present invention relates to the field of imaging and image processing, and more specifically, to a high-precision industrial CT dimension measurement method for diffuse metal artifact working conditions. Background Art

[0002] In recent years, X-ray computed tomography has been widely used in the field of industrial nondestructive testing. By performing projection scans of the object at multiple different angles and reconstructing the projection data using computational methods, the three-dimensional internal structure information of the object can be obtained. However, when the workpiece being inspected contains metal materials with high atomic numbers, metal artifacts will seriously interfere with the image quality, causing the traditional edge extraction algorithm to fail, and the measurement of the geometric size parameters of the workpiece becomes particularly difficult. To address this type of problem, many patents have designed methods based on image post-processing to perform secondary processing on images with metal artifacts. For example, the patent with publication number CN116543063A designs a method for removing metal artifacts from CT images. By constructing an initial adaptive iterative learning model based on wavelet transform, the CT image is decomposed by using the target adaptive iterative learning model optimized by the optimization objective function, the area of ​​metal artifacts is calculated and the metal artifacts in the CT image are removed to obtain a CT image with reduced metal artifacts; and the patent with publication number CN116228916A proposes using a metal artifact suppression method to perform image enhancement or suppression processing on a metal artifact image to obtain a predicted image; performing intensity distribution calculation on the predicted image to obtain an intensity distribution; performing correlation calculation on the intensity distribution and an external memory module to obtain a feature map; and inputting the feature map into a decoder to obtain a reconstruction result without metal artifacts. The patent invention with the publication number: provides a method for unsupervised correction of metal artifacts in CT images based on Transformer, constructs an unsupervised correction network for metal artifacts based on the CycleGAN architecture, trains the unsupervised correction network for metal artifacts with the training data set Z1, and obtains a trained unsupervised correction network for metal artifacts; the CT image with metal artifacts to be corrected is input into the trained generator GAB to obtain the corresponding CT image without metal artifacts. However, although this type of metal artifact removal method based on image post-processing can alleviate the impact of metal artifacts on image quality to a certain extent, it does not add any prior knowledge or information. Most of the information comes from estimation. Once a more complex metal artifact is designed, such as diffuse metal artifacts, it is difficult to effectively handle. Considering that the workpiece measured by industrial CT often has certain processing information of geometric contour dimensions in advance, if such prior information can be added, it can provide effective supplementary information for traditional edge extraction methods and prevent edge distortion caused by diffuse metal artifacts.Therefore, the image edge information with prior knowledge is used as a supplement and together with the traditional edge extraction method as a training label to train the conditional adversarial generative network model, and the test result image of the model that meets the standard is measured. This can better solve the problem that the size parameter information is difficult to measure in the diffuse metal artifact image scene. A high-precision industrial CT size measurement method for diffuse metal artifact conditions is proposed. By supplementing the prior geometric contour parameters and training the conditional adversarial generative network under the background of data enhancement, an efficient, accurate and sufficiently robust workpiece geometric contour size measurement method is achieved, further improving the accuracy of industrial non-destructive testing. Summary of the invention

[0003] The purpose of the present invention is to provide an industrial CT geometric dimension parameter measurement method that can adapt to diffuse metal artifact working conditions, and solve the problem that it is difficult to accurately measure the geometric dimensions of a workpiece under diffuse metal artifact working conditions. The present invention uses the prior contour parameters in industrial production as a supplement to the traditional edge extraction method, thereby obtaining a CT image with high-quality fine edge contour information. In addition, the data enhancement method and conditional adversarial generative network are used for training to ensure the robustness of the model, which provides the possibility for CT precision measurement under such complex working conditions.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows, and the flow chart is as shown in the attached Figure 1 As shown:

[0005] S1: Acquire a three-dimensional CT image and perform slicing processing on the three-dimensional CT image;

[0006] S2: preprocessing the cross-sectional image of the original workpiece CT, including Poisson distribution noise reduction;

[0007] S3: Use the “double threshold method” to remove the “highlight metal artifact” part while retaining the effective information;

[0008] S4: Binarization of the processed image;

[0009] S5 operates the image using the methods of corrosion, dilation and opening and closing operations to obtain a primary rough edge;

[0010] S6: Use prior information to correct the image edge, including the size information in the manufacturing process, the parameter information of the appearance measurement, the collective contour information of straight lines, arcs, etc., to repair the parts that obviously do not meet the requirements in the rough edge extraction, so as to obtain advanced fine edge information;

[0011] S7: The image seat label image with high-level fine edge information obtained in step S6 and the unprocessed CT slice image obtained in step S1 are used as the original image, and the images at corresponding positions in the two are paired to form a training data set, and data enhancement is performed at the same time, specifically including the following steps:

[0012] S7.1: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d based on the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously translate the paired images in the range of [x_0±(ad) / 2, y_0±(ad) / 2). Repeat this process multiple times and add the translated images to the training data set to ensure the image position adaptability in model training.

[0013] S7.2: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d according to the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously scale the paired images. The circumscribed circle radius transformation range of the workpiece to be measured is: [0, d / 2+min((a-x_0),|x_0-a|)]. Repeat multiple times and add the scaled images to the training data set to ensure the image size adaptability in model training.

[0014] S7.3: Rotate the paired images by angle θ_k=k×10°, (k=1,2,3,…,36), repeat multiple times, and add the rotated images to the training dataset to ensure image rotation adaptability in model training;

[0015] S8: Divide the data set obtained in step S7 into training, verification and test data sets according to a predetermined ratio;

[0016] S9: Use step S8 to divide the training and validation data sets to train the "high-precision" conditional adversarial generation network;

[0017] S10: Use the test data set divided in step S8 to test the model. If the effect is satisfactory, use new unlabeled data for testing and obtain an image with fine edge information.

[0018] S11: Calculate the average centroid of the image with fine edge information obtained in step S10, and calculate the straightness of the tubular material according to the centroids of multiple images;

[0019] S12: Calculate the inner and outer diameters of the workpiece according to the image with fine edge information obtained in step S10;

[0020] S13: Calculating the convex curvature radius and the concave curvature radius of the workpiece according to the image with fine edge information obtained in step S10;

[0021] S14: Calculate the tube wall thickness, thickness standard deviation, and key point thickness according to the image with fine edge information obtained in step S10;

[0022] S15: Output the measurement parameters of S11-S14 and draw a result graph with the measurement parameters.

[0023] Compared with the traditional edge extraction method of CT images, the present invention has the following advantages:

[0024] (1) High-precision edge extraction: The accuracy of edge extraction can be significantly improved by the prior information correction method in step S6. Rough edges are repaired using the size information and appearance measurement parameters in the manufacturing process to obtain more detailed edge information.

[0025] (2) Applicable to diffuse metal artifact scenarios: When faced with complex working conditions of diffuse metal artifacts, the CT image processing method of the present invention has significant advantages over traditional methods. The "double threshold method" is used to effectively remove highlight metal artifacts, reducing the interference of artifacts on the image while retaining important structural information. Combined with the prior information correction in step S6, edge information can be accurately repaired and extracted, and high precision can be maintained even when the artifact has a large impact. In addition, the multiple data enhancement techniques in step S7 improve the robustness of the model to artifacts and image deformation, ensuring the adaptability and universality of the model. Finally, through the precise measurement of steps S11 to S14, reliable geometric parameters can be provided under complex working conditions, ensuring the efficiency and accuracy of the entire detection process.

[0026] (3) Strong robustness: The various data enhancement techniques in step S7 (including translation, scaling, and rotation) can improve the generalization ability of the model, making the model more adaptable to different positions, sizes, and rotation angles, thereby improving the robustness and accuracy of the model.

[0027] (4) Comprehensive measurement capabilities: In steps S11 to S14, various geometric parameters are calculated for images with fine edge information, including straightness, inner and outer diameters, radius of curvature, and wall thickness, etc., to provide comprehensive workpiece measurement information and meet various needs in practical applications. The result graph in step S15 can intuitively display the measurement parameters, making it easier for users to understand and analyze the test results. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a high-precision industrial CT dimension measurement method for diffuse metal artifact conditions;

[0029] Figure 2 Example of geometric measurement results of diffuse metal artifacts. DETAILED DESCRIPTION

[0030] Taking the measurement of a cross-helical tube workpiece containing high atomic number metal as an example, a high-precision industrial CT size measurement method for diffuse metal artifact working conditions is designed. Starting from the acquisition of the three-dimensional CT image volume reconstructed by the FDK algorithm, the CT image with fine edge information is obtained through operations such as slicing, data preprocessing, primary edge extraction, and advanced edge extraction. After data enhancement methods and conditional adversarial generation network training, the geometric dimensions of the workpiece are finally measured, thereby realizing accurate non-destructive detection of the workpiece geometric parameters under complex working conditions of diffuse metal artifacts.

[0031] In this real-time example, it is assumed that the size of the workpiece to be measured is: 1024^3, and the workpiece to be measured is a three-dimensional tubular structure of a cross spiral. The specific implementation method is as follows:

[0032] S1: Acquire a three-dimensional CT image and slice the three-dimensional CT image. First, use a CT scanner to perform three-dimensional imaging of the target workpiece. Ensure that the resolution and quality of the scan meet the requirements of subsequent analysis. The acquired three-dimensional CT image data is usually stored in the form of voxels (volume pixels). Next, use specialized image processing software to slice the three-dimensional CT image. This step divides the three-dimensional volume data into multiple two-dimensional cross-sectional images, and each slice corresponds to a specific depth layer. The interval and thickness of the slices should be adjusted according to the analysis requirements to ensure that sufficient detail information can be obtained.

[0033] S2: Preprocess the cross-sectional images of the original workpiece CT, including Poisson distribution noise reduction. The two-dimensional cross-sectional images obtained by slicing are preprocessed to improve the effect of subsequent processing. The preprocessing steps include noise removal and image enhancement. Here, Poisson distribution noise reduction technology is used to reduce the noise in the image, which usually appears in CT images due to low counts. By applying an appropriate denoising algorithm (such as Gaussian filtering or Poisson denoising filter), the image can be smoothed and important detail information can be retained. This step helps to reduce the interference of noise on subsequent analysis.

[0034] S3: Use the "double threshold method" to remove the "highlight metal artifact" part while retaining valid information. When processing the preprocessed image, the "double threshold method" is applied to identify and remove the highlight metal artifact. Metal artifacts usually produce highlight areas in CT images, which may interfere with the analysis of the actual target. By setting two thresholds: a low threshold and a high threshold, the highlight areas in the image (the part exceeding the high threshold) can be marked as artifacts and removed during image processing. This can clean up the artifact area while retaining useful information, improving image quality and the accuracy of subsequent analysis.

[0035] S4: Binarize the processed image. Binarize the image after denoising and artifact removal. Binarization is the process of converting the pixel values ​​of an image into only two states (such as black and white). In this process, a threshold is set, and the area with pixel values ​​above the threshold is set to white (foreground), and the area below the threshold is set to black (background). Binarization can effectively separate the target features in the image from the background, laying the foundation for subsequent morphological processing.

[0036] S5: Operate the image using the methods of erosion, dilation, and opening and closing operations to obtain primary rough edges. Apply morphological operations on the binarized image to extract the primary edges of the target. First, perform an erosion operation to locally erode the image using structural elements to remove small noise points and connected areas. Next, apply a dilation operation to expand the foreground area to connect broken edges. Then, perform an opening operation (erosion followed by dilation) and a closing operation (dilation followed by erosion) to further smooth the edges and remove noise. After these operations, the rough edges in the image can be obtained, which serves as the basis for subsequent more accurate edge detection and feature extraction.

[0037] S6: Use prior information to correct the image edge, including the size information in the manufacturing process, the parameter information of the appearance measurement, the collective contour information of straight lines, arcs, etc., to repair the parts that obviously do not meet the requirements in the rough edge extraction, so as to obtain advanced fine edge information;

[0038] S7: The image seat label image with high-level fine edge information obtained in step S6 and the unprocessed CT slice image obtained in step S1 are used as the original image, and the images at corresponding positions in the two are paired to form a training data set, and data enhancement is performed at the same time, specifically including the following steps:

[0039] S7.1: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d based on the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously translate the paired images in the range of [x_0±(ad) / 2, y_0±(ad) / 2). Repeat this process multiple times and add the translated images to the training data set to ensure the image position adaptability in model training.

[0040] S7.2: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d according to the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously scale the paired images. The circumscribed circle radius transformation range of the workpiece to be measured is: [0, d / 2+min((a-x_0),|x_0-a|)]. Repeat multiple times and add the scaled images to the training data set to ensure the image size adaptability in model training.

[0041] S7.3: Rotate the paired images by angle θ_k=k×10°, (k=1,2,3,…,36), repeat multiple times, and add the rotated images to the training dataset to ensure image rotation adaptability in model training;

[0042] S8: Divide the data set obtained by step S7 into a training set, a validation set, and a test set according to a preset ratio. The specific ratio may be set according to specific needs, such as 70% training set, 15% validation set, and 15% test set. Ensure that each data set contains different types of image transformations to maintain the representativeness and diversity of the data set.

[0043] S9: Use the training set and validation set divided in step S8 to train a "high-precision" conditional generative adversarial network (CGAN). The training goal of the network is to improve the accuracy of image restoration and generation, so as to accurately extract fine edge information. A "high-precision" conditional generative adversarial network (CGAN) will be trained using the training set and validation set divided in step S8. Specifically, it is necessary to prepare training data first, and pair the image with high-level fine edge information generated in step S7 with the unprocessed original CT slice image in step S1 to form a training data set. Data preprocessing includes image standardization and data enhancement to improve the robustness of the model. Next, a conditional generative adversarial network (CGAN) is constructed, in which the generator accepts the original image and conditional information (fine edge contour information) and outputs the restored image; the discriminator is responsible for distinguishing the generated image from the real image. During the training process, the generator optimizes the image restoration quality through adversarial loss and reconstruction loss, while the discriminator improves the ability to judge the authenticity of the image through cross entropy loss. By alternately training the generator and the discriminator, the adversarial effect of the model can be continuously improved.

[0044] S10: Test the model: Use the test data set divided in step S8 to test the trained model and evaluate the performance of the model. After the model training is completed, we will use the validation set in step S8 to evaluate the model. Check the fineness and accuracy of the generated images by calculating the quality indicators of the generated images, such as the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR). At the same time, adjust the hyperparameters and architecture of the model according to the verification results to optimize the performance. During the training process, an early stopping mechanism should also be introduced to stop the training when the model performance no longer improves to avoid overfitting. Finally, use the trained CGAN model to process the unlabeled data to generate images with fine edge information, and actually measure these images to verify the application effect and actual performance of the model. If the test effect meets the standard, it means that the model can accurately extract fine edge information. Process unlabeled data: Input the newly acquired unlabeled data into the model to obtain images with fine edge information. These images are used for subsequent actual measurements and analysis.

[0045] S11: In order to evaluate the straightness and manufacturing quality of tubular materials, we first need to calculate the center of mass of each image. The image with fine edge information obtained according to step S10 is first binarized to ensure that the edge part is white and the rest is black. Then, the center of mass calculation formula is used to weighted average the white pixels in each image to obtain the center of mass position of the image. Then, this process is repeated for all images to calculate the center of mass position of each image. Finally, the average of these center of mass is calculated to obtain the average center of mass position of the tubular material. By analyzing the deviation of the center of mass of each image from the average center of mass, the straightness and geometric consistency of the material can be evaluated, thereby comprehensively evaluating the manufacturing quality.

[0046] S12: In order to accurately measure the inner and outer diameters of the workpiece, the image with fine edge information obtained in step S10 is first used. The processing steps include: binarizing the image to ensure that the edges are clearly discernible. Then, edge detection techniques such as Hough transform are used to extract the inner and outer contours of the workpiece. After the inner and outer contours are extracted, the inner and outer diameters of the workpiece are calculated by fitting a circular model. For each extracted contour, the circle is fitted using the least squares method, and the inner and outer diameters are derived from the radius of the fitted circle. Finally, the circular diameter obtained by fitting is used as the measurement value of the inner and outer diameters. These precise dimensional measurements help determine the compliance and functional suitability of the workpiece, ensuring that it meets design and quality standards.

[0047] S13: Analyze the image obtained in step S10 to calculate the convex curvature radius and concave curvature radius of the workpiece. These curvature values ​​are used to evaluate the shape characteristics of the workpiece surface, which are important for matching and adaptation in engineering applications.

[0048] S14: Calculate the tube wall thickness, thickness standard deviation, and key point thickness according to the image with fine edge information obtained in step S10;

[0049] S15: Output the measurement parameters of S11-S14, and draw an example diagram of the geometric measurement results of the diffuse metal artifact workpiece with the measurement parameters, as shown in the attached figure. Figure 2 shown.

[0050] The present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A high-precision industrial CT dimension measurement method for diffuse metal artifact conditions, characterized in that: The steps include: S1: Acquire a three-dimensional CT image and perform slicing processing on the three-dimensional CT image; S2: preprocessing the cross-sectional image of the original workpiece CT, including Poisson distribution noise reduction; S3: Use the "double threshold method" to remove the "highlight metal artifact" part while retaining the effective information; S4: Binarization of the processed image; S5 operates the image using the methods of corrosion, dilation and opening and closing operations to obtain a primary rough edge; S6: Use prior information to correct the image edge, including the size information in the manufacturing process, the parameter information of the appearance measurement, the collective contour information of straight lines, arcs, etc., to repair the parts that obviously do not meet the requirements in the rough edge extraction, so as to obtain advanced fine edge information; S7: The image seat label image with high-level fine edge information obtained in step S6 and the unprocessed CT slice image obtained in step S1 are used as the original image, and the images at corresponding positions in the two are paired to form a training data set, and data enhancement is performed at the same time. S8: Divide the data set obtained in step S7 into training, verification and test data sets according to a predetermined ratio; S9: Use step S8 to divide the training and validation data sets to train the "high-precision" conditional adversarial generation network; S10: Use the test data set divided in step S8 to test the model. If the effect is satisfactory, use new unlabeled data for testing and obtain an image with fine edge information. S11: Calculate the average centroid of the image with fine edge information obtained in step S10, and calculate the straightness of the tubular material according to the centroids of multiple images; S12: Calculate the inner and outer diameters of the workpiece according to the image with fine edge information obtained in step S10; S13: Calculating the convex curvature radius and the concave curvature radius of the workpiece according to the image with fine edge information obtained in step S10; S14: Calculate the tube wall thickness, thickness standard deviation, and key point thickness according to the image with fine edge information obtained in step S10; S15: Output the measurement parameters of S11-S14 and draw a result graph with the measurement parameters.

2. The high-precision industrial CT size measurement method for diffuse metal artifact conditions according to claim 1, characterized in that: The following steps are involved: S7.1: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d based on the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously translate the paired images in the range of [x_0±(ad) / 2, y_0±(ad) / 2). Repeat this process multiple times and add the translated images to the training data set to ensure the image position adaptability in model training. S7.2: Determine the image center position (x_0, y_0) and the circumscribed circle diameter d according to the fine edge contour. Assume that the image size is a*a. Use the random number method to simultaneously scale the paired images. The circumscribed circle radius transformation range of the workpiece to be measured is: [0, d / 2+min((a-x_0),|x_0-a|)]. Repeat multiple times and add the scaled images to the training data set to ensure the image size adaptability in model training. S7.3: Rotate the paired images by angle θ_k=k×10°, (k=1,2,3,…,36), repeat multiple times, and add the rotated images to the training dataset to ensure image rotation adaptability in model training.

Images