Intelligent segmentation method for enhancing defect-containing basalt fiber composite material image

Abstract

The invention relates to an intelligent segmentation method for enhancing an image of a defect-containing basalt fiber composite material, and the method comprises the following steps: 1), collecting an XCT digital image, randomly selecting a part of the image as a real image set, marking yarns, cracks and bubbles in the image, and generating a label set corresponding to the real image set; 2) selecting a plurality of vertexes of a real yarn contour to carry out polygon smooth fitting to generate a random virtual yarn contour, putting the random virtual yarn contour into a pseudo label image according to real yarn arrangement, and adding bubble and crack characteristics to obtain a pseudo label set; 3) converting the pseudo label into a pseudo image by using a DP-GAN style conversion network, and performing yarn edge reconstruction by using a Palette network model; and 4) mixing the real training set and the pseudo training set, inputting the mixture into a Mask DINO segmentation network for training, and reasoning the XCT original image by using the training weight after the training is ended to obtain a segmentation result. The method can enhance the image data, can effectively improve the network segmentation capability, and saves the manual marking cost.

Description

Technical Field

[0001] The invention relates to an intelligent segmentation method for enhancing defective basalt fiber composite images, which can complete automatic data enhancement of features such as yarns, bubbles and cracks in three-dimensional woven basalt fiber composite images and belongs to the technical field of image processing. Background Art

[0002] Three-dimensional woven basalt fiber composite material is a new type of high-performance composite material that uses basalt fiber as reinforcement and resin as matrix, and is compounded through a three-dimensional weaving process.

[0003] However, in practical applications, the microstructure of three-dimensional woven basalt fiber composites is complex. In order to more comprehensively understand the relationship between the microstructure damage evolution and macroscopic mechanical properties of composite materials, it is necessary to use deep learning segmentation algorithms to spatially divide the internal phases of composite materials. However, it is costly and difficult to obtain a large amount of annotated data. Therefore, it is of great theoretical and practical significance to perform data enhancement based on real data sets and improve the segmentation ability of internal phases.

[0004] Therefore, in order to solve the above technical problems, it is necessary to provide an innovative intelligent segmentation method for enhancing defective basalt fiber composite images to overcome the above defects in the prior art. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent segmentation method for enhancing defective basalt fiber composite images, which generates a pseudo label set from a partial real image set obtained by XCT technology, and then combines the DP-GAN style transfer network and the Palette repair network to generate a pseudo image set with smooth transition characteristics, thereby realizing data enhancement of the XCT image segmentation data set of three-dimensional woven basalt fiber composites, and realizing accurate segmentation of the fiber composite material at the single yarn level.

[0006] To achieve the above object, the technical solution adopted by the present invention is: an intelligent segmentation method for enhancing defective basalt fiber composite images, which comprises the following steps: 1) Using XCT scanning technology to collect digital images of basalt fiber composites, constructing an original data set, screening based on these original data, selecting a part of them as a real image set, painting different colors on the yarn area, crack area, and bubble area in the real image for manual annotation, and generating a real label set corresponding to the real image set; 2) Based on the above real label set, six vertices are selected from each contour of the yarn as polygons, and the pseudo yarn contour is generated by smooth fitting; then, it is randomly placed on a blank label to form a pseudo label image; then, ellipses of different shapes and angles are randomly selected to simulate initial bubbles; finally, pseudo cracks are generated on the yarn cross section using Perlin noise, and then the pseudo label set is derived; 3) The pseudo-label image is converted into a pseudo-image using the DP-GAN style transfer network, and then the Palette network model is used to reconstruct the yarn edge to obtain a pseudo-image set after the yarn edge is reconstructed; 4) The real image set and the real label set are combined into a real training set, the pseudo image set and the pseudo label set are combined into a pseudo training set, and the two training sets are combined into a mixed training set and put into the Mask DINO segmentation network for training; after the training, the .pth weight file of the trained network model is used to infer the XCT original data set, divide the yarn area, crack area, and bubble area in the image, and attach a color mask to obtain the segmentation result.

[0007] The intelligent segmentation method of the reinforced defective basalt fiber composite material image of the present invention is further as follows: the step 1) is specifically: 1-1), collect basalt fiber composite material samples through XCT scanning, obtain XCT raw data, and select part of the tomography data to construct a real image set; 1-2), in the real image set, the yarn area, crack area, and bubble area are manually labeled with different colors, and the colors can be randomly selected, and the same type of area can use the same color, thereby generating a real labeled image set; 1-3), the smeared areas in the above-mentioned real labeled image set become connected areas with the same grayscale value. Connected areas with different features express different grayscale values. The Boolean operation in the OpenCV library is used to extract the labels, and the grayscale values ​​of different features are reassigned. The 24-bit image is converted into an 8-bit grayscale image to obtain a real label set containing matrix, yarn, crack, and bubble features. The image is an 8-bit .png file, and the grayscale standard is: background: 0, matrix: 1, yarn: 2, crack: 3, bubble: 4. Different grayscale areas in the image represent different feature divisions.

[0008] The intelligent segmentation method of the reinforced defective basalt fiber composite image of the present invention is further as follows: in the step 1-1), the basalt fiber composite material is specifically a composite material produced by three-dimensional four-way four-step weaving and packaging by VARTM process; the yarn used for weaving is composed of 1200Tex basalt fiber bundles.

[0009] The intelligent segmentation method of the reinforced defective basalt fiber composite material image of the present invention is further as follows: in the step 1-1), the XCT raw data is stored in .raw format and 16 bits in size; then the 16-bit data is reduced to 8 bits in Avizo software, and after rotation correction, cropping, and contrast enhancement, 1350 8-bit .png format tomography raw data are obtained, and 111 data are randomly selected to construct a .png format real image set.

[0010] The intelligent segmentation method of the reinforced defective basalt fiber composite material image of the present invention is further as follows: the step 2) is specifically: 2-1), using the real yarn profile as a template, selecting the upper and lower two vertices and the left and right four vertices as polygons, and smoothly fitting to generate a random virtual yarn profile; 2-2), placing the randomly generated virtual yarn profile into a blank label to form a pseudo label image; 2-3), using the OpenCV library to generate ellipses of different sizes, angles, and shapes to simulate bubbles in real samples, and randomly placing them into the pseudo-label image of step 2-2) to avoid the virtual yarn contour; 2-4), use the Perlin noise library to generate virtual cracks of different lengths, curvatures, and branch numbers, and place them in the virtual yarn contour of the pseudo-label image; finally, a pseudo-label set containing yarn, bubbles, and cracks is obtained, in which the grayscale value standard is the same as the real label set.

[0011] The intelligent segmentation method of the reinforced defective basalt fiber composite image of the present invention is further as follows: in the step 2-1), according to the six vertex coordinates, a closed hexagon is obtained by series connection, and after adding noise points around the hexagonal lines, the noise point disturbance amplitude is controlled, that is, the amplitude of the noise points around the hexagon, and the noise point lines are fitted and smoothed using Bezier curves to obtain a smooth, non-angular virtual yarn profile.

[0012] The intelligent segmentation method of the reinforced defective basalt fiber composite material image of the present invention is further as follows: the step 3) is specifically: 3-1), using the DP-GAN style transfer network to convert pseudo-label images into pseudo-images; 3-2), using the pseudo label set after expansion and erosion of the image to perform contour Boolean subtraction operation to obtain a pseudo edge mask image; 3-3), use the Palette network model to reconstruct the yarn edges of the pseudo image; the Palette network detects and optimizes the edges in the pseudo image through multi-scale feature extraction and fusion, and finally generates a set of pseudo images after yarn edge reconstruction.

[0013] The intelligent segmentation method of the enhanced defective basalt fiber composite material image of the present invention is further as follows: the yarn edge reconstruction is specifically: the pseudo image is covered with the pseudo edge mask described in 3-2), and the Palette network model regenerates the masked yarn edge area according to the pseudo image features.

[0014] The intelligent segmentation method of the reinforced defective basalt fiber composite material image of the present invention can also be: the step 4) is specifically: 4-1), the real image set and the real label set are combined into a real training set, the pseudo image set and the pseudo label set are combined into a pseudo training set, and the two training sets are combined into a mixed training set and put into the Mask DINO segmentation network for training. The network divides the feature area in the corresponding image according to the label guidance, and then compares the segmentation result with the label image to calculate the loss function; through continuous cycle iteration, a segmentation model weight file with high accuracy is finally obtained; 4-2), after the training is completed, the trained network model weight .pth file is used to infer the entire .png format XCT original image, divide the different areas in the original image, and mark them into color masks to obtain the segmentation result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The intelligent segmentation method of the enhanced defective basalt fiber composite image of the present invention utilizes different deep learning networks such as the DP-GAN style transfer network and the Palette network, which can automatically generate pseudo training sets for data enhancement, and realize the accurate segmentation of the fiber composite material at the level of a single yarn, and provide data support for the subsequent combination of finite element analysis and other methods to simulate and improve the various properties of the composite material, optimize the composite material formula and process, greatly improve the accuracy of the three-dimensional model of the composite material, and have good generalization ability and application prospects.

[0016] 2. The intelligent segmentation method of enhanced defective basalt fiber composite images of the present invention introduces the DP-GAN style transfer network and the Palette repair network, which can effectively alleviate the problems of difficult annotation of XCT image segmentation data sets and high production costs of three-dimensional woven basalt fiber composites.

[0017] 3. The intelligent segmentation method of the enhanced defective basalt fiber composite material image of the present invention is trained using the MaskDINO segmentation network, which can effectively improve the network segmentation capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the intelligent segmentation method of the enhanced defect-containing basalt fiber composite material image of the present invention.

[0019] Figure 2 It is a schematic diagram of a virtual contour generated after polygon smoothing fitting according to the vertices of the yarn contour of the real image in step 2) of the present invention.

[0020] Figure 3 This is a working principle diagram of the DP-GAN style transfer network in step 3) of the present invention.

[0021] Figure 4 This is a working principle diagram of the Palette network in step 3) of the present invention.

[0022] Figure 5 It is a grayscale comparison diagram between the pseudo image generated in the present invention and the real image.

[0023] Figure 6 It is the XCT original image obtained by the present invention through the XCT scanning technology.

[0024] Figure 7 This is an image after data enhancement and re-segmentation by the present invention. DETAILED DESCRIPTION

[0025] Please refer to the instruction manual Figure 1 As shown, it is an intelligent segmentation method for enhancing defective basalt fiber composite material images of the present invention, which includes the following steps: 1) Collect digital images of basalt fiber composite materials through XCT scanning technology, as shown in the attached manual. Figure 6 As shown in the figure, the original data set is constructed. Based on these original data, a part of them is selected as the real image set, and the yarn area, crack area, and bubble area in the real image are painted with different colors for manual annotation to generate a real label set corresponding to the real image set.

[0026] The basalt fiber composite material is specifically a composite material produced by three-dimensional four-way four-step weaving and then encapsulating by VARTM process. The yarn used for weaving is composed of 1200Tex basalt fiber bundles.

[0027] Specifically, step 1) is implemented by: 1-1), collect basalt fiber composite samples through XCT scanning, obtain XCT raw data, and select part of the tomography data to construct a real image set.

[0028] XCT (X-ray Computed Tomography) is a three-dimensional imaging technology that uses X-rays to penetrate an object and reconstruct its internal structure through a computer. Through XCT scanning, the microstructure of a three-dimensional woven basalt fiber composite material can be obtained, including fiber distribution, fiber orientation, and existing defects (cracks, bubbles).

[0029] In this embodiment, the XCT raw data is stored in .raw format and 16 bits in size; then the 16-bit data is reduced to 8 bits in Avizo software, and after rotation correction, cropping, and contrast enhancement, 1350 8-bit .png format tomography raw data are obtained, and 111 data are randomly selected to construct a .png format real image set.

[0030] 1-2), in the real image set, the yarn area, crack area, and bubble area are manually labeled with different colors. The colors can be randomly selected, and the same color can be used for the same type of area, thereby generating a real labeled image set.

[0031] 1-3), the smeared areas in the above real labeled image set become connected areas with the same grayscale value. Connected areas with different features express different grayscale values. The Boolean operation in the OpenCV library is used to extract the labels, and the grayscale values ​​of different features are reassigned. Specifically, the 24-bit image is converted to an 8-bit grayscale image to obtain a real label set containing matrix, yarn, crack, and bubble features. The image is an 8-bit .png file, and the grayscale standard is: background: 0, matrix: 1, yarn: 2, crack: 3, bubble: 4. Different grayscale areas in the image represent different feature divisions.

[0032] 2) Based on the above real label set, six vertices are selected from each contour of the yarn as polygons, and the pseudo yarn contour is generated by smooth fitting. Then the pseudo yarn contour is randomly placed on the blank label to form a pseudo label image. Then, ellipses of different shapes and angles are randomly selected to simulate the initial bubbles. Finally, pseudo cracks are generated on the yarn cross section using Perlin noise, and then the pseudo label set is derived.

[0033] The step 2) is specifically as follows: 2-1), as shown in the instruction manual Figure 2 As shown in the figure, the real yarn contour is used as a template, the upper and lower two vertices and the left and right four vertices are selected as polygons, and the random virtual yarn contour is generated by smooth fitting. Specifically, according to the coordinates of the six vertices, a closed hexagon is obtained after series connection. After adding noise points around the hexagonal lines, the noise disturbance amplitude is controlled, that is, the amplitude of the noise points around the hexagon, which can make the contour more diverse. However, too large a disturbance amplitude will cause the virtual yarn contour to be distorted, and too small a disturbance amplitude will result in a high similarity between the generated virtual yarn contours. Afterwards, the Bezier curve is used to fit and smooth the noise point lines to obtain a smooth, non-angular virtual yarn contour.

[0034] 2-2), a randomly generated virtual yarn profile is placed into a blank label to form a pseudo label image.

[0035] 2-3), using the OpenCV library to generate ellipses of different sizes, angles, and shapes to simulate bubbles in real samples, and randomly placing them into the pseudo-label image of step 2-2) to avoid the virtual yarn contour.

[0036] 2-4), use the Perlin noise library to generate virtual cracks of different lengths, curvatures, and branch numbers, and place them in the virtual yarn contour of the pseudo-label image; finally, a pseudo-label set containing yarn, bubbles, and cracks is obtained, in which the grayscale value standard is the same as the real label set.

[0037] 3) The DP-GAN style transfer network is used to convert the pseudo-label image into a pseudo-image, and then the Palette network model is used to reconstruct the yarn edge to obtain a pseudo-image set after yarn edge reconstruction.

[0038] Among them, as shown in the attached manual Figure 3 As shown, the DP-GAN (Blended Diffusion Model) style transfer network is an image style transfer model based on a generative adversarial network (GAN), which can convert pseudo-label images in three-dimensional woven basalt fiber composites into pseudo images with the style of real XCT images.

[0039] The specific steps of this step are: 3-1), the DP-GAN style transfer network is used to convert the pseudo-label image into a pseudo-image. During the conversion process, the DP-GAN network retains the yarn, bubble and crack structures in the pseudo-label, and at the same time integrates the style features of the real XCT image to generate a pseudo-image similar to the real XCT image.

[0040] 3-2), using the pseudo label set after expansion and erosion of the image to perform contour Boolean subtraction operation to obtain a pseudo edge mask image.

[0041] 3-3), as shown in the instruction manual Figure 4 As shown in the figure, the Palette network model is used to reconstruct the yarn edge of the pseudo image. The Palette repair network is a deep learning-based image repair technology that aims to repair the problem of uneven yarn interface transition in the pseudo image generated by DP-GAN, making the transition from yarn to other features in the pseudo image more natural and close to the real image. The Palette network detects and optimizes the edges in the pseudo image through multi-scale feature extraction and fusion, and finally generates a pseudo image set after the yarn edge is reconstructed, as shown in the attached manual. Figure 5 As shown, it is a grayscale comparison between the pseudo image and the real image in the pseudo image set. It can be seen that the grayscale distribution of the real and fake images is highly overlapped, proving that the pseudo image has a high degree of realism.

[0042] The yarn edge reconstruction is specifically as follows: the pseudo image is covered with the pseudo edge mask described in 3-2), and the Palette network model regenerates the masked yarn edge area according to the pseudo image features. Through the above edge reconstruction, the problem of uneven transition of the interface area between different features in the pseudo image is solved.

[0043] 4), the real image set and the real label set are combined into a real training set, the pseudo image set and the pseudo label set are combined into a pseudo training set, and the two training sets are combined into a mixed training set and put into the Mask DINO segmentation network for training; after the training, the .pth weight file of the trained network model is used to infer the XCT original data set, divide the yarn area, crack area, and bubble area in the image, and add a color mask to obtain the segmentation result, that is, the image after data enhancement and re-segmentation is obtained, as shown in the attached manual Figure 7 shown.

[0044] The step 4) is specifically as follows: 4-1), the real image set and the real label set are composed of the real training set, the pseudo image set and the pseudo label set are composed of the pseudo training set, and the two training sets are composed of the mixed training set and put into the Mask DINO segmentation network for training. The network divides the feature area in the corresponding image according to the label guidance, and then compares the segmentation result with the label image to calculate the loss function; through continuous cycle iteration, the segmentation model weight file with higher accuracy is finally obtained.

[0045] 4-2), after the training is completed, the trained network model weight .pth file is used to infer the entire .png format XCT original image, divide the different areas in the original image, and mark them into color masks to obtain the segmentation result.

[0046] In summary, the intelligent segmentation method of the image of the reinforced defective basalt fiber composite material of the present invention obtains the image of the relevant fiber composite material through XCT scanning to obtain a real image set, and obtains the real label image by extracting it through an algorithm after manual marking; then, the yarn features of the real label image are used to fit the virtual yarn contour by an algorithm, and cracks are added by Perlin noise and ellipses are randomly placed to simulate bubble features to obtain a pseudo label set; then, the DP-GAN style conversion network is used to convert the pseudo label set into a pseudo image set, and the yarn edge is reconstructed on the yarn cross-sectional area in the pseudo image set through the Palette repair network to obtain a pseudo image set after the yarn edge is reconstructed; finally, the real image set plus the real label set constitute a real training set, the pseudo image set and the pseudo label set constitute a pseudo training, and the real training set and the pseudo training set constitute a mixed training set and are put into the MaskDINO segmentation network for training. After the training is completed, the model weight is obtained, and the XCT original data is segmented by reasoning using its weight to obtain the segmented three-dimensional reconstructed model.

[0047] The intelligent segmentation method for enhancing defective basalt fiber composite images of the present invention can automatically generate pseudo training sets for data enhancement by combining different deep learning networks, and introduces DP-GAN style conversion network and Palette repair network. These networks can effectively alleviate the problems of difficult labeling and high production cost of XCT image segmentation data sets of three-dimensional woven basalt fiber composites, and realize accurate segmentation of single yarn level of fiber composite materials, which provides data support for subsequent simulation and improvement of various properties of composite materials by combining finite element analysis and other methods, and optimization of composite material formula and process, greatly improves the accuracy of three-dimensional models of composite materials, and has good generalization ability and application prospects.

[0048] The above specific implementation methods are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent segmentation method for enhancing defective basalt fiber composite images, characterized by: The steps include: 1) Using XCT scanning technology to collect digital images of basalt fiber composites, constructing an original data set, screening based on these original data, selecting a part of them as a real image set, painting different colors on the yarn area, crack area, and bubble area in the real image for manual annotation, and generating a real label set corresponding to the real image set; 2), based on the above-mentioned real label set, six vertices are selected from each contour of the yarn to form a polygon, and a pseudo yarn contour is generated by smooth fitting; Then, random placement is performed on a blank label to form a pseudo-label image; then, ellipses of different shapes and angles are randomly selected to simulate initial bubbles; finally, pseudo cracks are generated on the yarn cross section using Perlin noise, and then a pseudo-label set is derived; 3) The pseudo-label image is converted into a pseudo-image using the DP-GAN style transfer network, and then the Palette network model is used to reconstruct the yarn edge to obtain a pseudo-image set after the yarn edge is reconstructed; 4) The real image set and the real label set are combined into a real training set, the pseudo image set and the pseudo label set are combined into a pseudo training set, and the two training sets are combined into a mixed training set and put into the Mask DINO segmentation network for training; After the training, the .pth weight file of the trained network model is used to infer the XCT original data set, divide the yarn area, crack area, and bubble area in the image, and add a color mask to obtain the segmentation result.

2. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 1, characterized in that: The step 1) is specifically as follows: 1-1), collect basalt fiber composite material samples through XCT scanning, obtain XCT raw data, and select part of the tomography data to construct a real image set; 1-2), in the real image set, the yarn area, crack area, and bubble area are manually labeled with different colors, and the colors can be randomly selected, and the same type of area can use the same color, thereby generating a real labeled image set; 1-3), the smeared areas in the above-mentioned real labeled image set become connected areas with the same grayscale value. Connected areas with different features express different grayscale values. The Boolean operation in the OpenCV library is used to extract the labels, and the grayscale values ​​of different features are reassigned. The 24-bit image is converted into an 8-bit grayscale image to obtain a real label set containing matrix, yarn, crack, and bubble features. The image is an 8-bit .png file, and the grayscale standard is: background: 0, matrix: 1, yarn: 2, crack: 3, bubble:

4. Different grayscale areas in the image represent different feature divisions.

3. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 2, characterized in that: In the step 1-1), the basalt fiber composite material is specifically a composite material produced by three-dimensional four-way four-step weaving and then encapsulating by VARTM process; the yarn used for weaving is composed of 1200Tex basalt fiber bundles.

4. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 2, characterized in that: In the step 1-1), the XCT raw data is stored in .raw format and 16 bits in size; then the 16-bit data is reduced to 8 bits in Avizo software, and after rotation correction, cropping, and contrast enhancement, 1350 8-bit .png format tomography raw data are obtained, and 111 data are randomly selected to construct a .png format real image set.

5. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 1, characterized in that: The step 2) is specifically as follows: 2-1), using the real yarn profile as a template, selecting the upper and lower two vertices and the left and right four vertices as polygons, and smoothly fitting to generate a random virtual yarn profile; 2-2), placing the randomly generated virtual yarn profile into a blank label to form a pseudo label image; 2-3), using the OpenCV library to generate ellipses of different sizes, angles, and shapes to simulate bubbles in real samples, and randomly placing them into the pseudo-label image of step 2-2) to avoid the virtual yarn contour; 2-4), use the Perlin noise library to generate virtual cracks of different lengths, curvatures, and branch numbers, and place them in the virtual yarn contour of the pseudo-label image; finally, a pseudo-label set containing yarn, bubbles, and cracks is obtained, in which the grayscale value standard is the same as the real label set.

6. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 5, characterized in that: In the step 2-1), a closed hexagon is obtained by connecting the six vertex coordinates in series. After adding noise points around the hexagonal lines, the noise disturbance amplitude, that is, the amplitude of the noise points around the hexagon, is controlled. The noise point lines are fitted and smoothed using a Bezier curve to obtain a smooth, non-angular virtual yarn profile.

7. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 1, characterized in that: The step 3) is specifically as follows: 3-1), using the DP-GAN style transfer network to convert pseudo-label images into pseudo-images; 3-2), using the pseudo label set after expansion and erosion of the image to perform contour Boolean subtraction operation to obtain a pseudo edge mask image; 3-3), use the Palette network model to reconstruct the yarn edges of the pseudo image; the Palette network detects and optimizes the edges in the pseudo image through multi-scale feature extraction and fusion, and finally generates a set of pseudo images after yarn edge reconstruction.

8. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 7, characterized in that: The yarn edge reconstruction is specifically as follows: the pseudo image is covered with the pseudo edge mask described in 3-2), and the Palette network model regenerates the masked yarn edge area according to the pseudo image features.

9. The intelligent segmentation method of reinforced defective basalt fiber composite material image according to claim 1, characterized in that: The step 4) is specifically as follows: 4-1), the real image set and the real label set are combined into a real training set, the pseudo image set and the pseudo label set are combined into a pseudo training set, and the two training sets are combined into a mixed training set and put into the Mask DINO segmentation network for training. The network divides the feature area in the corresponding image according to the label guidance, and then compares the segmentation result with the label image to calculate the loss function; through continuous cycle iteration, a segmentation model weight file with high accuracy is finally obtained; 4-2), after the training is completed, the trained network model weight .pth file is used to infer the entire .png format XCT original image, divide the different areas in the original image, and mark them into color masks to obtain the segmentation result.

Images