A 3D reconstruction method for damaged fiber composite images based on deep learning

Abstract

The present invention relates to a three-dimensional reconstruction method of damaged fiber composite images based on deep learning, which obtains images of related fiber composite materials through XCT scanning, obtains original data sets, and generates a label image set by marking features in the image; then, the label image is input into a deep learning network for training to obtain an instance segmentation result, and then the warp and weft are separated by an image processing algorithm to obtain warp contour slices and weft contour slices; then, a dynamic damaged yarn three-dimensional reconstruction algorithm combining intersection-over-combination ratio with centroid tracking is used to normalize the grayscale of the same yarn in all slices; finally, the normalized continuous image is input into the software to obtain an accurate three-dimensional reconstruction model of the warp and weft, and then the three-dimensional accurate reconstruction models of the warp and weft are merged to obtain an accurate three-dimensional reconstruction model of a single yarn level. The present invention significantly improves the accuracy and robustness of reconstruction, and has good generalization ability and application prospects.

Description

Technical Field

[0001] The present invention relates to a deep learning-based three-dimensional reconstruction method for damaged fiber composite images, which completes the accurate three-dimensional reconstruction of damaged yarn models in fiber composite materials and belongs to the field of image processing technology. Background Art

[0002] Fiber composites are widely used in aerospace, automobile manufacturing, construction engineering and energy fields due to their excellent strength, stiffness and corrosion resistance. In actual use, yarn damage inside composite materials has a crucial impact on their overall performance. As the main load-bearing structure of composite materials, the occurrence and development of yarn damage is directly related to the mechanical properties, service life and safety of the material. Although the instance segmentation results assisted by deep learning can realize the identification of damaged warp and weft yarns, it is difficult to achieve accurate three-dimensional reconstruction of damaged warp and weft yarns, and often there are problems such as yarn modeling errors, yarn missing and yarn incompleteness. Accurate three-dimensional reconstruction is the prerequisite for statistical analysis of sample yarn microstructure parameters. Therefore, accurate three-dimensional reconstruction of fiber composite materials, especially the realization of efficient and accurate yarn tracking, has important scientific significance and engineering application value.

[0003] Most of the existing yarn tracking algorithms are based on the K nearest neighbor algorithm, intersection-over-union ratio, semantic result yarn separation and instance conversion methods. These methods are affected by the image noise level and segmentation method, thus showing shortcomings such as poor generalization ability, weak anti-interference ability, and low robustness. At the same time, since the yarn instance segmentation results can only realize the recognition and distinction of warp and weft yarns, it is impossible to achieve accurate three-dimensional reconstruction of warp and weft yarns, which causes troubles in optimizing the formulation and process of fiber composite materials.

[0004] Therefore, in order to solve the above problems, it is necessary to provide an innovative deep learning-based three-dimensional reconstruction method for damaged fiber composite images to overcome the above-mentioned defects in the prior art. Summary of the invention

[0005] The purpose of the present invention is to provide a three-dimensional reconstruction method of damaged fiber composite images based on deep learning, which performs deep learning instance segmentation on a continuous image set obtained based on XCT technology, tracks the segmentation results and normalizes the grayscale of the same yarn in each frame of the image, thereby achieving accurate three-dimensional reconstruction of damaged fiber composite materials at the level of a single yarn.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a three-dimensional reconstruction method of damaged fiber composite material image based on deep learning, which comprises the following steps:

[0007] 1) Collect XCT digital images of damaged fiber composite materials to obtain the original data set; then manually annotate the yarns in the image to generate a label set;

[0008] 2) Input the label set and the corresponding image into the deep learning instance segmentation network for training, and use the weights obtained from the training to infer the entire original data set to obtain continuous instance segmentation result slices; then, use image Boolean operations to separate the warp and weft yarns to obtain warp yarn contour slices and weft yarn contour slices;

[0009] 3) Apply the 3D reconstruction algorithm to the slices of the warp and weft contours to track the yarn contours, and perform grayscale normalization on the contours of the same yarn in each slice;

[0010] 4) The three-dimensional precise reconstruction of the warp and weft yarns is achieved by slicing the warp and weft yarn contours after grayscale normalization, and the three-dimensional precise reconstruction models of the warp and weft yarns are merged to obtain the three-dimensional precise reconstruction model of the entire fiber composite material.

[0011] The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning of the present invention is further as follows: the step 1) is specifically:

[0012] 1-1), the original data of damaged fiber composite materials are obtained by XCT scanning;

[0013] 1-2), a data set was composed of 300 randomly selected XCT digital images, and the different types of yarns in the data set were manually marked using Avizo software. The same type of yarn was painted with a uniform grayscale value, and the label set was exported after marking.

[0014] The deep learning-based three-dimensional reconstruction method of damaged fiber composite images of the present invention is further as follows: the fiber composite material is specifically a non-curled biaxial warp-woven fiber composite material; the yarn is specifically divided into warp yarn or weft yarn.

[0015] The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning of the present invention is further as follows: the step 2) is specifically:

[0016] 2-1), the 300 XCT data sets and the corresponding manually annotated label sets are combined into a training set to train the deep learning instance segmentation network model;

[0017] 2-2), using the trained model weights to infer the entire XCT original data set to obtain continuous instance segmentation result slices;

[0018] 2-3), using image Boolean operations to separate the warp and weft yarns in the continuous instance segmentation images to obtain continuous warp yarn contour slices and weft yarn contour slices.

[0019] The three-dimensional reconstruction method of damaged fiber composite material image based on deep learning of the present invention is further as follows: in the step 3), the three-dimensional reconstruction algorithm is specifically:

[0020] 3-1), selecting a seed image without noise points and containing all the yarn contours to be tracked from the continuous image data set;

[0021] 3-2), extract all yarn contours in the seed image to update the initial contour library; and calculate the centroid of each yarn contour to establish a centroid library;

[0022] 3-3), establish a monitoring module to count the number of times a certain yarn number cannot match a qualified contour on continuous slices, and each scan of a slice is considered a traversal;

[0023] 3-4) Read all slices in grayscale 8-bit format according to the slice naming order;

[0024] 3-5), call each contour in the contour library and the corresponding centroid coordinates in the centroid library in turn, use the centroid coordinates of the current yarn number to locate on the scanned slice, set a field of view moment with a height of H and a width of W, and extract the yarn contour within the field of view moment in the scanned slice;

[0025] 3-6), traverse all the contours detected in step 3-5), and calculate the intersection and union ratio with the contour of the current yarn number in the contour library; when a detected contour exceeds the set standard threshold R, the current contour is determined to be the contour of the current yarn number on this scanned slice, and its grayscale is normalized to the grayscale value specified by the current yarn number;

[0026] 3-7), when the detected contour meets the standard threshold R, it means that the contour similarity is extremely high, and this contour information is written into the contour of the current yarn number in the contour library for overwriting and updating, and the centroid of the current yarn number in the centroid library is updated at the same time;

[0027] 3-8), traverse the contour library again to process the contours that have not been successfully tracked and matched, that is, the contour of a certain yarn in the contour library is not updated after completing the traversal of a slice, and the unupdated yarn contour in the contour library is still drawn when outputting the image;

[0028] 3-9), according to the number of times the marked yarn profile has not been updated, the monitor module is used to determine the triggering of multi-level threshold judgment to determine the warning level of the marked yarn;

[0029] 3-10), define a monitor record update function. When the detected yarn contour meets a threshold requirement in the multi-level threshold, the function removes the locked yarn number and cancels the warning. At the same time, it updates the contour library and draws the updated contour in the output image, but does not update the centroid library. The centroid update is only performed when the contour meets the standard threshold R.

[0030] The method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning of the present invention is further as follows: Steps 3-8) are specifically:

[0031] 3-8-1), in addition to the standard intersection-and-union ratio threshold R, set the n-level intersection-and-union ratio threshold R 1 , R 2 , R 3 ……R n , which corresponds to the warning area M 1 、M 2 、M 3 ……M n ;

[0032] 3-8-2), if the maximum intersection ratio reaches the maximum threshold R, the contour matching is considered successful, and the contour library and centroid library information of this yarn number are updated;

[0033] 3-8-3), if the maximum intersection-over-union ratio cannot meet the threshold R but reaches the threshold R 1 When the current tracking yarn number has been identified by the corresponding warning area M in step 3-8-1) 1 Record, update its contour library information, but do not update the centroid library; if it does not meet the requirements, the threshold is downgraded for judgment. If it meets a certain threshold R n , the operation is the same as above; when all profiles cannot meet the set minimum threshold, the yarn number is skipped and the number of times the current yarn profile is not updated is added to the monitor to increase the warning level M n , and then continue to track the next target.

[0034] The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning of the present invention is further as follows: in the step 3-8-1), R>R 1 >R 2 >R 3 ……>R n .

[0035] The three-dimensional reconstruction method of damaged fiber composite material image based on deep learning of the present invention is further as follows: Steps 3-9) are specifically:

[0036] According to step 3-8-1) set n-level warning area M 1 、M 2 、M 3 ……Mn If the number of profile monitor records reaches the set number, the corresponding threshold R is activated. n , add the current yarn number to the corresponding level warning area M n ;

[0037] When the number of times the target in the warning area is not updated increases, the warning level M n Gradually increase, the corresponding threshold R n The values ​​are progressively reduced to facilitate the profile update to its target yarn number.

[0038] The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning of the present invention is further as follows: the specific working process of the monitor record update function in step 3-10 is as follows:

[0039] According to the input, the contour information corresponding to the yarn number in the contour library is updated, and the updated contour information is drawn in the output image with an 8-bit gray value; the yarn number is 1, 2, 3...n; if the contour exists but is not obvious when drawing with a gray 8-bit image, the gray value needs to be increased. After increasing the basic gray value to 80, the yarn contour is drawn starting with a gray value of 81 corresponding to yarn number 1, and so on;

[0040] After the corresponding yarn number is updated in the contour library and the contour is drawn on the output graph, it is checked whether the yarn number exists in the multi-level warning area. If it exists, the record of the current yarn number on the monitor is cleared to remove the warning, indicating that the target contour is successfully tracked.

[0041] The method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning of the present invention may also be: the step 4) is specifically:

[0042] 4-1), inputting the continuous tracking result images into Avizo 3D imaging software for 3D reconstruction to obtain accurate 3D reconstruction models of yarn clusters with single yarn level accuracy;

[0043] 4-2), the warp and weft yarn model sections are aligned through Avizo software to achieve model merging and obtain a complete single yarn-level accurate three-dimensional reconstruction model of the fiber composite material.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The deep learning-based three-dimensional reconstruction method of damaged fiber composite images of the present invention utilizes a deep learning network and a three-dimensional reconstruction algorithm to track the contour of the yarn, thereby achieving accurate three-dimensional reconstruction of a single yarn level of the fiber composite material, and providing data support for the subsequent combination of finite element analysis and other methods to simulate and improve the various properties of the composite material, and optimize the composite material formulation and process.

[0046] 2. The deep learning-based 3D reconstruction method of damaged fiber composite images of the present invention combines the intersection-over-union method with the center of mass tracking, and adds a monitor and a multi-level threshold method to solve the problem of yarn contour mutation after damage, and automatically completes the damaged yarn tracking task with excellent accuracy and efficiency.

[0047] 3. The deep learning-based three-dimensional reconstruction method of damaged fiber composite images of the present invention improves the accuracy of the three-dimensional reconstruction model of the composite material, and significantly improves the accuracy and robustness of the reconstruction, and has good generalization ability and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning of the present invention.

[0049] Figure 2 It is a flow chart of the three-dimensional reconstruction algorithm in step 3) of the present invention.

[0050] Figure 3 It is a working principle diagram of the monitor module in step 3) of the present invention.

[0051] Figure 4 This is the original image of the damaged fiber composite material obtained by XCT scanning in the present invention.

[0052] Figure 5 It is a schematic diagram of the three-dimensional damaged yarn reconstruction model obtained in step 4) of the present invention. DETAILED DESCRIPTION

[0053] Please refer to the instruction manual Figure 1 To Attachment Figure 5 As shown, it is a three-dimensional reconstruction method of damaged fiber composite material image based on deep learning of the present invention, which includes the following steps:

[0054] 1) Collect XCT (X-ray computed tomography) digital images of damaged fiber composites to obtain the original data set; then manually annotate the yarns in the image to generate a label set.

[0055] Specifically, the step 1) is:

[0056] 1-1), obtaining raw data of the damaged fiber composite material by XCT scanning. In this embodiment, the fiber composite material is specifically a non-crimped biaxial warp knitted fiber composite material.

[0057] 1-2), a dataset is formed by randomly selecting 300 XCT digital images, and the different types of yarns in the dataset are manually marked using Avizo software. The different types of yarns are specifically divided into warp yarns or weft yarns. The same type of yarn is painted with a uniform grayscale value, but it is necessary to ensure that each yarn contour area in the same type of yarn is independent and does not adhere to each other. After the marking is completed, the label set is exported.

[0058] 2) Input the label set and the corresponding image into the deep learning instance segmentation network for training, and use the weights obtained from the training to infer the entire original data set to obtain continuous instance segmentation result slices; then, use image Boolean operations to separate the warp yarn and weft yarn to obtain warp yarn contour slices and weft yarn contour slices.

[0059] Specifically, 2-1), the 300 XCT data sets and the corresponding manually annotated label sets are combined into a training set to train the deep learning instance segmentation network model.

[0060] 2-2), use the trained model weights to infer the entire XCT original data set to obtain continuous instance segmentation result slices.

[0061] 2-3), using image Boolean operations to separate the warp and weft yarns in the continuous instance segmentation images to obtain continuous warp yarn contour slices and weft yarn contour slices.

[0062] 3) Apply the three-dimensional reconstruction algorithm to the slices of the warp and weft contours to track the yarn contours, and normalize the grayscale of the contours belonging to the same yarn in each slice.

[0063] The three-dimensional reconstruction algorithm is specifically:

[0064] 3-1), a seed image without noise points and containing all the yarn contours to be tracked is selected from the continuous image data set.

[0065] 3-2), extract all yarn contours in the seed image to update the initial contour library; and calculate the center of mass of each yarn contour to establish a center of mass library.

[0066] 3-3) Establish a monitor module to record the number of times each yarn profile in the profile library has not been updated after several consecutive traversals. Use a dictionary to count the number of times a yarn number cannot be matched to a qualified profile on consecutive slices. Each scan of a slice is considered a traversal.

[0067] 3-4) Read all slices in grayscale 8-bit format in the order of slice naming.

[0068] 3-5), call each contour in the contour library and the corresponding center of mass coordinates in the center of mass library in turn, use the center of mass coordinates of the current yarn number to locate on the scanned slice, set a field of view moment with a height of H and a width of W, and extract the yarn contour within the field of view moment in the scanned slice.

[0069] Among them, the field of view moment with height H and width W can be customized according to the circumscribed moment of the target contour, and its specific impact is:

[0070] (a) Limit the search range to avoid indiscriminately searching for contour targets on the entire slice, greatly improving computational efficiency;

[0071] (b) Reduce the size of the field of view: reduce the target matching area, effectively exclude irrelevant contours, and improve matching accuracy. It is suitable for scenes with limited target movement range or complex background;

[0072] (c) Increase the size of the field of view: expand the target matching area, increase the probability of successful matching during target movement, improve the continuity and robustness of tracking, and is suitable for scenes with a large target displacement range;

[0073] (d) A boundary protection mechanism is provided to ensure that the field of view moment is always within the image range and to prevent the field of view moment from exceeding the boundary and causing the target outline to be clipped.

[0074] 3-6), traverse all the contours detected in step 3-5), and calculate the intersection and union ratio with the contour of the current yarn number in the contour library; when a detected contour exceeds the set standard threshold R, the current contour can be determined as the contour of the current yarn number on this scanned slice, and its grayscale is normalized and adjusted to the grayscale value specified by the current yarn number.

[0075] 3-7), when the detected contour meets the standard threshold R, it means that the contour similarity is extremely high, and this contour information is written into the contour of the current yarn number in the contour library for overwriting and updating, and the centroid of the current yarn number in the centroid library is updated at the same time.

[0076] 3-8), traverse the contour library again and process the contours that have not been successfully tracked and matched, that is, the contour of a certain yarn in the contour library is not updated after completing the traversal of a slice. In order to avoid the influence of unqualified contours, the unupdated yarn contours in the contour library are still drawn when outputting the image to ensure that the contour will not be interrupted or mutated, and the yarn number is marked and recorded in the monitor module.

[0077] In this step, 3-8-1), in addition to the standard intersection-and-union ratio threshold R, set the n-level intersection-and-union ratio threshold R 1 , R 2 , R 3 ……R n , and R>R 1 >R2 >R 3 ……>R n , which corresponds to the warning area M 1 、M 2 、M 3 ……M n It should be noted that the threshold level N and the specific value can be adjusted according to the actual task requirements.

[0078] 3-8-2) If the maximum intersection ratio reaches the maximum threshold R, the contour matching is considered successful, and the contour library and centroid library information of this yarn number are updated.

[0079] 3-8-3), if the maximum intersection-over-union ratio cannot meet the threshold R but reaches the threshold R 1 When the current tracking yarn number has been identified by the corresponding warning area M in step 3-8-1) 1 Record, update its contour library information, but do not update the centroid library; if it does not meet the requirements, the threshold is downgraded for judgment. If it meets a certain threshold R n , the operation is the same as above; when all profiles cannot meet the set minimum threshold, the yarn number is skipped and the number of times the current yarn profile is not updated is added to the monitor to increase the warning level M n , and then continue to track the next target.

[0080] 3-9), according to the number of times the marked yarn profile has not been updated, the monitor module is used to determine the triggering of multi-level threshold judgment to determine the warning level of the marked yarn.

[0081] Specifically, according to the n-level warning area M set in step 3-8-1) 1 、M 2 、M 3 ……M n Taking the third-level warning as an example, when a yarn number contour in the contour library has not been updated for more than 2 consecutive slices, the warning level is M 1 , when more than 5 slices are not updated, it is promoted to M 2 , when more than 10 slices are not updated, it is promoted to M 3 ,The warning level and number of values ​​are not fixed and can be adjusted according to the actual task needs.

[0082] If the profile monitor records a certain number of times, the corresponding threshold R is activated. n , add the current yarn number to the corresponding level warning area M n .

[0083] When the number of times the target in the warning area is not updated increases, the warning level M n Gradually increase, the corresponding threshold R nThe values ​​are progressively reduced to facilitate the profile update to its target yarn number.

[0084] 3-10), define a monitor record update function. When the detected yarn contour meets a threshold requirement in the multi-level threshold, the function removes the locked yarn number and cancels the warning. At the same time, it updates the contour library and draws the updated contour in the output image, but does not update the centroid library. The centroid update is only performed when the contour meets the standard threshold R.

[0085] The specific working process of the monitor record update function is as follows:

[0086] According to the input, the contour information corresponding to the yarn number in the contour library is updated, and the updated contour information is drawn in the output image with an 8-bit grayscale value; the yarn numbers are 1, 2, 3...n; if the contour exists but is not obvious when drawn with a grayscale 8-bit image, the grayscale value needs to be increased. After increasing the basic grayscale value to 80, the yarn contour is drawn starting with a grayscale value of 81 corresponding to yarn number 1, and so on.

[0087] After the corresponding yarn number is updated in the contour library and the contour is drawn on the output graph, it is checked whether the yarn number exists in the multi-level warning area. If it exists, the record of the current yarn number on the monitor is cleared to remove the warning, indicating that the target contour is successfully tracked.

[0088] 4) The three-dimensional precise reconstruction of the warp and weft yarns is achieved by slicing the warp and weft yarn contours after grayscale normalization, and the three-dimensional precise reconstruction models of the warp and weft yarns are merged to obtain the three-dimensional precise reconstruction model of the entire fiber composite material.

[0089] The specific steps of this step are:

[0090] 4-1), inputting the continuous tracking result images into Avizo 3D imaging software for 3D reconstruction to obtain accurate 3D reconstruction models of yarn clusters with single yarn level accuracy.

[0091] 4-2), the warp and weft yarn model sections are aligned through Avizo software to achieve model merging and obtain a complete single yarn-level accurate three-dimensional reconstruction model of the fiber composite material.

[0092] In summary, the deep learning-based three-dimensional reconstruction method of damaged fiber composite images of the present invention obtains images of relevant fiber composite materials through XCT scanning to obtain original data sets, and manually marks features in the images through Avizo software to generate a labeled image set; then, the labeled images are input into a deep learning network for training; then, continuous XCT images are input into the trained deep learning network to obtain instance segmentation results, and then the warp yarns and weft yarns are separated by an image processing algorithm to obtain warp yarn contour slices and weft yarn contour slices; then, the dynamic damaged yarn three-dimensional reconstruction algorithm that combines the intersection ratio with the centroid tracking algorithm is used to normalize the grayscale of the same yarn in all slices; finally, the normalized continuous images are input into the Avizo software to obtain accurate three-dimensional reconstruction models of the warp yarns and weft yarns, and then the three-dimensional accurate reconstruction models of the warp yarns and weft yarns are merged to obtain an accurate three-dimensional reconstruction model of the damaged fiber composite material at the level of a single yarn.

[0093] The deep learning-based three-dimensional reconstruction method of damaged fiber composite images of the present invention utilizes a deep learning network and a three-dimensional reconstruction algorithm to track the yarn contour, thereby achieving accurate three-dimensional reconstruction at the level of a single yarn of the fiber composite material. This provides data support for the subsequent simulation and improvement of various properties of the composite material by combining finite element analysis and other methods, and for optimizing the formulation and process of the composite material. This greatly improves the accuracy of the three-dimensional reconstruction model of the composite material, and has good generalization ability and application prospects.

[0094] 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. A three-dimensional reconstruction method for damaged fiber composite images based on deep learning, characterized in that: The steps include: 1) Collect XCT digital images of damaged fiber composite materials to obtain the original data set; then manually annotate the yarns in the image to generate a label set; 2) Input the label set and the corresponding image into the deep learning instance segmentation network for training, and use the weights obtained from the training to infer the entire original data set to obtain continuous instance segmentation result slices; Then, the warp yarn and the weft yarn are separated by using the image Boolean operation to obtain the warp yarn contour slice and the weft yarn contour slice; 3) Apply the 3D reconstruction algorithm to the slices of the warp and weft contours to track the yarn contours, and perform grayscale normalization on the contours of the same yarn in each slice; The three-dimensional reconstruction algorithm is specifically: 3-1), select a seed image without noise and containing all the yarn contours to be tracked in the continuous image data set; 3-2), extract all yarn contours in the seed image to update the initial contour library; and calculate the centroid of each yarn contour to establish a centroid library; 3-3), establish a monitoring module to count the number of times a certain yarn number cannot match a qualified contour on continuous slices, and each scan of a slice is considered a traversal; 3-4) Read all slices in grayscale 8-bit format according to the slice naming order; 3-5), call each contour in the contour library and the corresponding centroid coordinates in the centroid library in turn, use the centroid coordinates of the current yarn number to locate on the scanned slice, set a field of view moment with a height of H and a width of W, and extract the yarn contour within the field of view moment in the scanned slice; 3-6), traverse all the contours detected in step 3-5), and calculate the intersection and union ratio with the contour of the current yarn number in the contour library; when a detected contour exceeds the set standard threshold R, the current contour is determined to be the contour of the current yarn number on this scanned slice, and its grayscale is normalized to the grayscale value specified by the current yarn number; 3-7), when the detected contour meets the standard threshold R, it means that the contour similarity is extremely high, and this contour information is written into the contour of the current yarn number in the contour library for overwriting and updating, and the centroid of the current yarn number in the centroid library is updated at the same time; 3-8), traverse the contour library again to process the contours that have not been successfully tracked and matched, that is, the contour of a certain yarn in the contour library is not updated after completing the traversal of a slice, and the unupdated yarn contour in the contour library is still drawn when outputting the image; 3-9), according to the number of times the marked yarn profile has not been updated, the monitor module is used to determine the triggering of multi-level threshold judgment to determine the warning level of the marked yarn; 3-10), define a monitor record update function. When the detected yarn profile meets a threshold requirement in the multi-level threshold, the function removes the locked yarn number and releases the warning. At the same time, the profile library is updated and the updated profile is drawn in the output image, but the centroid library is not updated. The centroid update is only performed when the profile meets the standard threshold R; Wherein, the steps 3-8) are specifically as follows: 3-8-1), in addition to the standard intersection-and-combination ratio threshold R, set n levels of intersection-and-combination ratio thresholds R1, R2, R3, ... R n , which corresponds to the warning areas M1, M2, M3...M n ; 3-8-2), if the maximum intersection ratio reaches the maximum threshold R, the contour matching is considered successful, and the contour library and centroid library information of this yarn number are updated; 3-8-3), if the maximum intersection ratio cannot meet the threshold R but reaches the threshold R1, and the current tracking yarn number has been recorded by the corresponding warning area M1 in step 3-8-1), update its profile library information, but do not update the centroid library; if it does not meet the threshold, the threshold is downgraded for judgment. If it meets a certain threshold R n , the operation is the same as above; when all profiles cannot meet the set minimum threshold, the yarn number is skipped and the number of times the current yarn profile is not updated is added to the monitor to increase the warning level M n , and then continue to track the next target; 4) The three-dimensional precise reconstruction of the warp and weft yarns is achieved by slicing the warp and weft yarn contours after grayscale normalization, and the three-dimensional precise reconstruction models of the warp and weft yarns are merged to obtain the three-dimensional precise reconstruction model of the entire fiber composite material.

2. The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning according to claim 1, characterized in that: The step 1) is specifically as follows: 1-1), the original data of damaged fiber composite materials are obtained by XCT scanning; 1-2), a data set was composed of 300 randomly selected XCT digital images, and the different types of yarns in the data set were manually marked using Avizo software. The same type of yarn was painted with a uniform grayscale value, and the label set was exported after marking.

3. The method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning according to claim 2, characterized in that: The fiber composite material is specifically a non-crimped biaxial warp knitted fiber composite material; the yarn is specifically divided into warp yarn or weft yarn.

4. The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning according to claim 2, characterized in that: The step 2) is specifically as follows: 2-1), a dataset consisting of 300 XCT digital images and the corresponding manually annotated label set are used as a training set to train the deep learning instance segmentation network model; 2-2), use the trained model weights to infer the entire XCT original data set to obtain continuous instance segmentation result slices; 2-3), use image Boolean operations to separate the warp and weft yarns in continuous instance segmentation result slices to obtain continuous warp yarn contour slices and weft yarn contour slices.

5. The method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning according to claim 1, characterized in that: In the step 3-8-1), R>R1>R2>R3...>R n .

6. The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning according to claim 1, characterized in that: The steps 3-9) are specifically as follows: According to the n-level warning areas M1, M2, M3...M set in step 3-8-1) n If the number of profile monitor records reaches the set number, the corresponding threshold R is activated. n , add the current yarn number to the corresponding level warning area M n ; When the number of times the target in the warning area is not updated increases, the warning level M n Gradually increase, the corresponding threshold R n The values ​​are progressively reduced to facilitate the profile update to its target yarn number.

7. The method for three-dimensional reconstruction of damaged fiber composite material images based on deep learning according to claim 1, characterized in that: The specific working process of the monitor record update function in step 3-10) is as follows: According to the input, the contour information corresponding to the yarn number in the contour library is updated, and the updated contour information is drawn in the output image with an 8-bit gray value; the yarn numbers are 1, 2, 3, ... n; when drawing with a gray 8-bit image, the contour exists but is not obvious, and the gray value needs to be increased. After increasing the basic gray value to 80, the yarn contour is drawn starting with a gray value of 81 corresponding to yarn number 1, and so on; After the corresponding yarn number is updated in the contour library and the contour is drawn on the output graph, it is checked whether the yarn number exists in the multi-level warning area. If it exists, the record of the current yarn number on the monitor is cleared to remove the warning, indicating that the target contour is successfully tracked.

8. The method for three-dimensional reconstruction of damaged fiber composite images based on deep learning according to claim 1, characterized in that: The step 4) is specifically as follows: 4-1), input the continuous tracking result images into Avizo 3D imaging software for 3D reconstruction to obtain accurate 3D reconstruction models of yarn clusters with single yarn level accuracy; 4-2), the warp and weft yarn model sections are aligned through Avizo software to achieve model merging and obtain a complete single yarn-level accurate three-dimensional reconstruction model of the fiber composite material.

Images