Optical fiber coil defect detection and evaluation method and system fusing industrial CT and deep learning

Abstract

The invention relates to the technical field of nondestructive testing and computer vision, in particular to an optical fiber coil defect detection and evaluation method and system fusing industrial CT and deep learning. Performing three-dimensional imaging on the optical fiber coil based on industrial CT to obtain an original image; performing uniform discretization sampling on the original image in the circumferential direction to obtain a two-dimensional cross section image; expanding the two-dimensional cross section image, and constructing an expanded image data set; according to the method, the Center Net target detector is optimized, the content awareness feature pyramid network is adopted, deep semantic features and shallow high-resolution features are fused step by step, and each layer focuses on detecting targets of different scales. A lightweight attention module is added, the attention focusing capability of human vision is simulated, and the feature signal-to-noise ratio of the detector in a small target scene is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of nondestructive testing and computer vision, and in particular to a method and system for detecting and evaluating defects in optical fiber wire packages by combining industrial CT and deep learning. BACKGROUND

[0002] Guided optical fiber wire packages are the core transmission carriers of modern optical fiber guidance systems, and they play a key role in establishing a high-reliability two-way data channel between the control platform and the aircraft. Relying on the inherent low transmission loss, wide spectral bandwidth, and strong anti-electromagnetic interference characteristics of optical fibers, the system significantly improves guidance accuracy and battlefield environmental adaptability, gives the aircraft the ability to "lock after launch", supports over-the-horizon operations, and has tactical advantages such as multi-task compatibility, high efficiency, and "man-in-the-loop" precision guidance.

[0003] In actual manufacturing processes, guided optical fiber wire packages face serious challenges in the winding, curing, and storage stages. Problems such as uneven wire diameter, winding tension control deviation, lag angle deviation, uneven adhesive coating, and mismatched thermal expansion coefficients of composite materials can easily induce multi-scale structural defects such as cracks, interlayer voids, and end face extrusion. Research shows that when the unevenness of the optical fiber wire diameter exceeds 3%, the fibers at the edge of the wire package will be extruded and overflowed, causing local stress concentration and structural instability. Such defects not only directly damage the physical integrity of the wire package and reduce long-term storage stability, but also can cause dynamic attitude misalignment during high-speed release, leading to a sharp increase in optical fiber microbend loss. It is particularly worth noting that the residual stress accumulation effect caused by the difference in the thermal expansion coefficient of the material can further amplify the risk of structural deformation during the release process, forming a reliability degradation chain throughout the "manufacturing-storage-release" cycle.

[0004] Traditional machine vision methods, such as feature extraction techniques based on grayscale and threshold segmentation, have been applied in surface detection of concrete and bearings, but they are extremely sensitive to image lighting conditions and noise interference. Hand-designed features are difficult to adapt to the complex texture background of optical fiber wire packages, and they cannot penetrate the surface to detect internal defects. With the rapid development of machine vision technology and deep learning, image-based defect detection technology has made significant progress. Deep learning methods automatically learn features through convolutional neural networks (CNN), significantly improving detection performance. However, existing models consume a lot of training time and computing resources when processing large-scale data sets, and may miss some subtle defect features. For example, FasterR-CNN and other region proposal-based models are prone to miss and false detection when detecting small targets, while YOLO and other single detection models struggle to balance detection accuracy and speed. Therefore, there is an urgent need for an efficient and high-precision detection model for optical fiber wire package defect detection. SUMMARY

[0005] In view of the problems mentioned in the prior art, the application proposes a fiber cable defect detection and evaluation method and system combining industrial CT and deep learning, which realizes high-precision defect detection, positioning and multi-dimensional feature quantitative evaluation. The application optimizes the detection algorithm, effectively improves the detection precision and efficiency of the internal micro defects of the fiber cable.

[0006] In order to achieve the above-mentioned purpose, the application adopts the following technical solutions: The application proposes a fiber cable defect detection and evaluation method combining industrial CT and deep learning, which includes the following steps: S1, based on industrial CT, three-dimensional imaging of the fiber cable is performed to obtain an original image; S2, the original image is uniformly discretely sampled in the circumferential direction to obtain a two-dimensional cross-sectional image; the two-dimensional cross-sectional image is expanded to obtain an expanded image dataset; S3, the expanded image dataset is input into a CenterNet target detector to obtain a defect detection result; the CenterNet target detector includes a backbone network, a neck network and a head network, the neck network uses a feature pyramid network to replace the original up-sampling stacking structure in the CenterNet target detector, and a content-aware module is inserted into the top-down path of the feature pyramid network; the head network increases a lightweight dual-attention mechanism; S4, the defect class and coordinates in the defect detection result are extracted, and based on the coordinates and the core tube coating boundary position of the fiber cable, the number of layers and turns of the defect position are calculated; S5, different evaluation methods are selected according to the defect class to obtain an evaluation result.

[0007] As a further improvement of the application, the process of S2 includes: The original image is uniformly discretely sampled along the circumferential direction, and the step length is set to 1°; starting from 0°, a series of two-dimensional cross-sectional images are obtained by cross-sectional slicing of the three-dimensional image every 1°; The series of two-dimensional cross-sectional images are preprocessed to obtain processed images; The processed images are stored in order to obtain an image database; The image database is sequentially subjected to rotation transformation, scaling transformation, flip transformation, sharpening and dark top hat processing to obtain an expanded image dataset.

[0008] As a further improvement of the application, the processing process of the CenterNet target detector in S3 includes: The feature pyramid network extracts features from the image dataset to obtain shallow semantic feature maps and deep semantic feature maps; The deep semantic feature map is up-sampled to obtain an up-sampled feature map; wherein the up-sampled feature map is input to a content-aware module for processing to obtain a processed feature map; the processed feature map is fused with the shallow feature map to obtain a multi-scale feature pyramid containing multi-level fused feature maps; The fused feature map of any layer in the multi-scale feature pyramid is input to a lightweight double-path attention mechanism to perform global average pooling and global maximum pooling on the fused feature map to obtain a channel attention feature vector; The channel attention feature vector is input to a multi-layer perception to generate a channel attention weight; the channel attention weighted feature map is obtained based on the channel attention weight; The channel attention weighted feature map is subjected to a convolution operation to generate a spatial attention weight, and a spatial attention weighted feature map is obtained based on the spatial attention weight; The channel attention weighted feature map and the spatial attention weighted feature map are combined to obtain a final attention feature map; The final attention feature map is input to a detection head to obtain a defect detection result.

[0009] As a further improvement of the present application, the up-sampled feature map is input to a content-aware module for processing to obtain a processed feature map, comprising: The up-sampled feature map is input to the content-aware module to learn the local feature relationship in the up-sampled feature map through a feature reorganization layer to generate a feature index map; According to the feature index map, the up-sampling process is adaptively adjusted to obtain a reorganized feature map; The reorganized feature map is input to a feature enhancement layer to obtain a processed feature map.

[0010] As a further improvement of the present application, the defect class and coordinates in the defect detection result are extracted in S4, comprising: The image in the defect detection result is sequentially subjected to color enhancement, morphological-based boundary box coordinate extraction, and color threshold classification to obtain the defect class and coordinates.

[0011] As a further improvement of the present application, the number of layers and turns of the defect location are calculated based on the coordinates and the fiber optic line jacket tube coating boundary position in S4, comprising: The two-dimensional cross-sectional image in S2 is converted into a gray-scale image, and the gray-scale image is subjected to global adaptive threshold segmentation; The binarized image is subjected to denoising processing; According to the aspect ratio and area size of the core tube surface coating on the CT cross-sectional image, the connected domain extraction condition is set; PCA is used to fit an approximate quadrilateral of the core tube surface coating, and a quadrilateral boundary is drawn. According to the coordinates, the vertical distance from the center point of the defect detection frame to the upper boundary of the coating and the horizontal distance from the defect center to the nearest side boundary are calculated; The layer number and the turn number of the defect position are calculated according to the following formula: (1) (2) In the formula, n1 represents the layer number where the defect center is located; z1 represents the turn number where the defect center is located; θ represents the taper angle of the core tube; d represents the average diameter of the optical fiber; h1 represents the vertical distance from the defect center to the upper boundary of the coating; and w1 represents the horizontal distance from the defect center to the nearest side boundary.

[0012] As a further improvement of the application, the process of S5 is: The coordinates are projected into the two-dimensional cross-sectional image to extract the corresponding local image; The local image is processed according to different defect categories; if the defect is an extrusion defect, local adaptive threshold segmentation is used to calculate the number of connected domains conforming to the cross section of the optical fiber, and the local image is marked; if the defect is a crack defect, global threshold segmentation is used to calculate the crack area, and the local image is marked; if the defect is an interlayer void defect, local threshold segmentation is used to count the voxels to calculate the void volume, and the local image is marked; The spatial structure of the defect is analyzed to evaluate the severity and impact of the defect.

[0013] The application proposes a fiber cable defect detection and evaluation system combining industrial CT and deep learning, comprising: An imaging module for three-dimensional imaging of the fiber cable based on industrial CT to obtain an original image; An expansion module for uniformly discretizing and sampling the original image in the circumferential direction to obtain a two-dimensional cross-sectional image; and expanding the two-dimensional cross-sectional image to construct an expanded image dataset; A detection module for inputting the expanded image dataset into a CenterNet target detector to obtain a defect detection result; the CenterNet target detector includes a backbone network, a neck network, and a head network, the neck network uses a feature pyramid network to replace the original up-sampling stacking structure in the CenterNet target detector, and a content perception module is inserted into the top-down path of the feature pyramid network; the head network increases a lightweight dual-channel attention mechanism before the classification and regression branch; A calculation module for extracting the defect category and coordinates in the defect detection result, and calculating the layer number and the turn number of the defect position based on the coordinates and the position of the core tube coating boundary of the fiber cable; An evaluation module is configured to select different evaluation methods according to the defect categories, and obtain evaluation results.

[0014] The application provides a fiber cable defect detection and evaluation device combining industrial CT and deep learning, which comprises a processor and a memory.

[0015] The application provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to realize the fiber cable defect detection and evaluation method combining industrial CT and deep learning.

[0016] Compared with the prior art, the application has the following technical effects: The application provides a technical means for multi-dimensional positioning and feature quantization evaluation of defects in the target detection post-processing stage through geometric topology analysis and voxel counting algorithm, so that researchers can more deeply understand the nature, structure and cause of defects, and improve the efficiency of subsequent maintenance and quality control.

[0017] The application optimizes the CenterNet target detector according to the defect characteristics of the fiber cable, including adopting a content-aware feature pyramid network to gradually fuse deep semantic features and shallow high-resolution features, so that each level focuses on detecting targets of different scales. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 This is a flowchart of the embodiment; Figure 2 This is a defect category and coordinate diagram of the embodiment; Figure 3 This is a defect category and coordinate effect diagram of the embodiment; Figure 4 This is a core tube coating boundary position diagram of the fiber cable of the embodiment; Figure 5 This is an extrusion defect diagram of the embodiment; Figure 6 This is a crack defect diagram of the embodiment; Figure 7 This is an interlayer hole defect diagram of the embodiment. DETAILED DESCRIPTION

[0019] The application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the application, but not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for the convenience of description, not all the structures.

[0020] Referring to Figure 1 The application proposes a fiber cable defect detection and evaluation method combining industrial CT and deep learning, including the following steps: S1, based on industrial CT, three-dimensional imaging of the fiber cable is performed to obtain an original image; S2, the original image is uniformly discretely sampled in the circumferential direction to obtain a two-dimensional cross-sectional image; the two-dimensional cross-sectional image is augmented to construct an augmented image dataset; S3, the augmented image dataset is input into a CenterNet target detector to obtain a defect detection result; the CenterNet target detector includes a backbone network, a neck network and a head network, the neck network uses a feature pyramid network to replace the original up-sampling stacking structure in the CenterNet target detector, and a content-aware module is inserted in the top-down path of the feature pyramid network; the head network increases a lightweight dual-attention mechanism; S4, the defect class and coordinates in the defect detection result are extracted, and based on the coordinates and the core tube coating boundary position of the fiber cable, the number of layers and turns at the defect position are calculated; S5, different evaluation methods are selected according to the defect class to obtain an evaluation result.

[0021] The application will be further described in detail below with reference to the accompanying drawings and embodiments: Step 1, first place the fiber cable on the rotating platform of the industrial CT scanner, adjust to a stable state, set the voltage, current, exposure time and other parameters, turn on the X-ray source, make the X-ray penetrate the fiber cable, the detector receives the signal and converts it into digital data; then start the rotating platform to drive the fiber cable to complete 360° rotation, acquire a two-dimensional projection image every certain angle, form a sequence containing internal structure attenuation information; then use the FDK algorithm of cone beam geometry for three-dimensional reconstruction, first filter the projection data to enhance the high frequency component and reduce the fuzziness, then back-project the data to the three-dimensional voxel space; finally integrate all back-projected data, calculate the voxel attenuation coefficient, generate the original image of the fiber cable, realize non-destructive three-dimensional imaging, and provide a basis for subsequent operations.

[0022] Step 2, the aligned positioned original image is uniformly discretely sampled in the circumferential direction with a step of 1°, and a series of two-dimensional cross-sectional images are obtained by taking a cross section every 1° starting from 0°; then normalization, denoising and smoothing operations are sequentially performed on each cross-sectional image to improve image quality and enhance the accuracy of subsequent processing; finally, the processed two-dimensional cross-sectional images are stored in order to lay a data foundation for defect detection and analysis.

[0023] Step 3, the two-dimensional cross-sectional images are subjected to rotation transformation to generate 90°, 180° and 270° rotated images to simulate different observation angles; at the same time, scaling transformation is performed on the two-dimensional cross-sectional images to generate 0.5 times and 2 times scale images to reflect defect features of different scales; then flip transformation is performed to generate left-right and up-down flipped images to increase data diversity and improve the generalization ability of the model; then 1.5 times sharpening processing is performed on the images to highlight edge and detail features and enhance the contrast between defects and background; finally, dark top hat processing technology is used to highlight dark area details and improve the feature visibility of images in dark light conditions to construct an expanded image dataset.

[0024] Step 4, in the neck network integrated feature pyramid network module (FPN, Feature Pyramid Network) of the CenterNet target detector, first feature extraction is performed on the image dataset to obtain shallow semantic feature maps and deep semantic feature maps, the shallow feature maps record details with high resolution, and the deep feature maps carry semantic information with low resolution.

[0025] The deep semantic feature maps are upsampled to obtain upsampled feature maps; the upsampled feature maps are input into the content-aware module for processing to obtain processed feature maps; the processed feature maps and the shallow feature maps are fused to enhance the richness of feature expression, obtaining a multi-scale feature pyramid, and the target detection task is performed through each level in the multi-scale feature pyramid, so that the model can detect defect targets of different scales at the same time.

[0026] Step 5, the upsampled feature maps are input into the content-aware (CARAFE Content-Aware ReAssembly of Features) module, the feature reorganization layer learns the local feature relationships in the upsampled feature maps to generate a feature index map; according to the feature index map, the upsampled process is adaptively adjusted to reorganize the feature maps to retain more details and structural information; finally, the reorganized feature maps are input into the feature enhancement layer to optimize semantic and detail information, improve the quality of the feature maps, and obtain the processed feature maps.

[0027] Step 6, first determine the position of adding a lightweight double-channel attention mechanism (CBAMConvolutional Block Attention Module) in the detection head of the CenterNet target detector; perform channel attention calculation on the fused feature map of any layer in the multi-scale feature pyramid, and use global average and maximum pooling to obtain a channel attention feature vector.

[0028] Input the attention feature vector into the MLP (Multi-Layer Perceptron) to generate channel attention weights, enhance key feature channel weights and suppress background responses, and obtain a channel attention weighted feature map; then perform spatial attention calculation, and perform convolution on the channel attention processed feature map to generate spatial attention weights, focus on the target area and weaken the interference pixels, and obtain a spatial attention weighted feature map; finally, combine the channel attention weighted feature map and the spatial attention weighted feature map to obtain a final attention feature map, which is used for subsequent target detection tasks to improve the feature signal-to-noise ratio.

[0029] Step 7, optimize the output module of the CenterNet detector, adjust the position of the bounding box drawing and improve the color contrast of different categories to prevent label text from blocking the bounding box, and ensure that the color contrast of the bounding box of each category is high, as shown in Figure 2 .

[0030] Based on the optimized detection image, the morphological processing method of MATLAB is used to automatically identify the boundary box region and extract its pixel coordinates.

[0031] For the three types of defects, set the mask threshold interval of the corresponding color, determine the defect type according to the matching of the color feature in the bounding box with the preset threshold, and test the sample wire package using the morphological recognition and color threshold classification algorithm to extract the defect coordinates and obtain the left upper corner and right lower corner coordinates of the defect bounding box, as shown in Figure 3 .

[0032] Step 8, convert the two-dimensional cross-sectional image in step two to a grayscale image and perform global adaptive threshold segmentation; denoise the binary image to remove small connected domains; set the connected domain extraction conditions according to the aspect ratio and area of the core tube coating; use the PCA method to fit the approximate quadrilateral of the core tube coating and draw the boundary, as shown in Figure 4 .

[0033] According to the coordinates, calculate the vertical distance from the center point of the defect detection frame to the upper boundary of the coating and the horizontal distance from the defect center to the nearest side boundary; Calculate the layer number and turn number of the defect location according to the following formula: (1) (2) In the formula, n1 represents the layer number where the defect center is located; z1 represents the turn number where the defect center is located; θ represents the taper angle of the core tube; d represents the average diameter of the optical fiber; h1 represents the vertical distance from the defect center to the upper boundary of the coating; and w1 represents the horizontal distance from the defect center to the nearest side boundary.

[0034] Step 9, project the coordinates of the defect into a two-dimensional cross-sectional image, and extract a local image.

[0035] According to the defect category, select an image processing method: for extrusion defects, use local adaptive threshold segmentation (Niblack) for binaryzation processing, optimize the image quality through denoising and smoothing the boundary, then calculate the connected domain attributes, set the roundness and area threshold of the fiber cross section, screen the connected domains meeting the conditions and count the number, and mark on the local image, as shown in Figure 5 , the leftmost Original image in Figure 5 is a two-dimensional cross-sectional image, the Binarized image in the middle is a binaryzation image after step 8 processing, and the Marked image is a marked extrusion defect image.

[0036] For crack defects, use global adaptive threshold algorithm (OTSU) for segmentation. After image segmentation, the binaryzation image is inverted, with the black part as the foreground. Then set the area threshold, identify and screen the connected domains, so as to extract the crack region and mark its area on the image, as shown in Figure 6 , the Original image and the Binarized image in Figure 6 are the same as above, and will not be elaborated here, and the Marked image is a marked crack defect image.

[0037] For interlayer hole defects, use local adaptive threshold segmentation (Niblack). After image segmentation, the binaryzation image also needs to be inverted, with the black part as the foreground. Set the area threshold of the interlayer hole, identify and screen the connected domains, identify the interlayer hole, and mark its area on the image, as shown in Figure 7 , the Original image and the Binarized image in Figure 7 are the same as above, and will not be elaborated here, and the Marked image is a marked interlayer hole defect image.

[0038] Finally, combine geometric topological analysis and voxel counting algorithm to analyze the spatial structure of the defect and evaluate the severity and potential impact of the defect.

[0039] Based on the same inventive concept, the application also provides a fiber cable defect detection and evaluation system combining industrial CT and deep learning. Since the principle of solving problems of the fiber cable defect detection and evaluation system is similar to the aforementioned method of detecting and evaluating fiber cable defects combining industrial CT and deep learning, the implementation of the fiber cable defect detection and evaluation system combining industrial CT and deep learning can refer to the implementation of the method of detecting and evaluating fiber cable defects combining industrial CT and deep learning, and the repeated parts will not be described here.

[0040] In a specific implementation, the fiber cable defect detection and evaluation system provided by the application embodiment specifically comprises: An imaging module is configured to perform three-dimensional imaging of the fiber cable based on the industrial CT to obtain an original image. An expansion module is configured to perform uniform discrete sampling in the circumferential direction on the original image to obtain a two-dimensional cross-sectional image, and expand the two-dimensional cross-sectional image to construct an expanded image dataset. A detection module is configured to input the expanded image dataset into a CenterNet target detector to obtain a defect detection result. The CenterNet target detector comprises a backbone network, a neck network and a head network. The neck network replaces the original up-sampling stacking structure in the CenterNet target detector with a feature pyramid network, and inserts a content perception module in the top-down path of the feature pyramid network. The head network increases a lightweight dual-path attention mechanism before the classification and regression branch. A calculation module is configured to extract the defect category and coordinates in the defect detection result, and calculate the layer number and turn number of the defect location based on the coordinates and the core tube coating boundary position of the fiber cable. An evaluation module is configured to select different evaluation methods according to the defect category to obtain an evaluation result.

[0041] Correspondingly, the application embodiment also provides a fiber cable defect detection and evaluation device combining industrial CT and deep learning, comprising a processor and a memory. When the processor executes the computer program stored in the memory, the method of detecting and evaluating fiber cable defects combining industrial CT and deep learning provided by the application embodiment is realized.

[0042] For more specific processes of the above method, reference can be made to the corresponding contents disclosed in the foregoing embodiments, which will not be described here.

[0043] Correspondingly, the application embodiment also provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to realize the aforementioned method of detecting and evaluating fiber cable defects combining industrial CT and deep learning provided by the application embodiment.

[0044] The various embodiments described in this specification are presented purely by way of example and for purposes of illustration only. Each of the various embodiments highlights a different aspect or feature of the other embodiments, and each of the various embodiments can stand on its own as an application independently of the other embodiments. For the systems, devices, storage media disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple and the relevant parts are referred to the description of the methods.

[0045] Those skilled in the art will further appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various examples disclosed herein are described generally in terms of their functionality, without reference to the corresponding

[0046] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0047] Finally, it should be noted that the terms "first" and "second" and the like are used merely to distinguish one element from another, and do not necessarily indicate a physical or chronological priority of one element over another. Furthermore, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. The terms "includes," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. The terms "includes," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0048] The fusion industrial CT and deep learning optical fiber line package defect detection and evaluation method and system provided by the present application are described in detail above, and the principles and implementation modes of the present application are described in this paper. The above example is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as a limitation of the present application.

Claims

1. A method for detecting and evaluating defects of fiber optic cable by fusing industrial CT and deep learning, characterized in that, The method comprises the following steps: S1, based on industrial CT, three-dimensional imaging of the optical fiber cable is performed to obtain an original image; S2, the original image is uniformly discretely sampled in the circumferential direction to obtain a two-dimensional cross-sectional image; The two-dimensional cross-sectional image is expanded to obtain an expanded image dataset; S3, the expanded image dataset is input into a CenterNet target detector to obtain a defect detection result; the CenterNet target detector comprises a backbone network, a neck network and a head network, the neck network adopts a feature pyramid network to replace the original up-sampling stacking structure in the CenterNet target detector, and a content-aware module is inserted into the top-down path of the feature pyramid network; the head network increases a lightweight double-channel attention mechanism; S4, the defect category and coordinates in the defect detection result are extracted, and based on the coordinates and the core tube coating boundary position of the optical fiber cable, the layer number and the number of turns of the defect position are calculated; S5, different evaluation methods are selected according to the defect category to obtain an evaluation result. 2.The method of claim 1, wherein, The process of S2 comprises: The original image is uniformly discretely sampled in the circumferential direction, and the step length is set to 1°; starting from 0°, a cross-sectional slice of the three-dimensional image is obtained every 1° to obtain a series of two-dimensional cross-sectional images; The series of two-dimensional cross-sectional images are preprocessed to obtain processed images; The processed images are stored in sequence to obtain an image database; The image database is sequentially subjected to rotation transformation, scaling transformation, flip transformation, sharpening treatment and dark top hat treatment to obtain an expanded image dataset. 3.The method of claim 1, wherein, The processing process of the CenterNet target detector in S3 comprises: The feature pyramid network extracts features from the image dataset to obtain shallow semantic feature maps and deep semantic feature maps; The deep semantic feature maps are up-sampled to obtain up-sampled feature maps; wherein the up-sampled feature maps are input into a content-aware module for processing to obtain processed feature maps; the processed feature maps and the shallow feature maps are fused to obtain a multi-scale feature pyramid containing multi-level fused feature maps; The fused feature maps of any layer in the multi-scale feature pyramid are input into a lightweight double-channel attention mechanism to perform global average pooling and global maximum pooling on the fused feature maps to obtain a channel attention feature vector; The channel attention feature vector is input into a multi-layer perceptron to generate a channel attention weight; the channel attention weighted feature map is obtained based on the channel attention weight; The channel attention weighted feature map is subjected to convolution operation to generate a spatial attention weight, and a spatial attention weighted feature map is obtained based on the spatial attention weight; The channel attention weighted feature map and the spatial attention weighted feature map are combined to obtain a final attention feature map; The final attention feature map is input into a detection head to obtain a defect detection result.

4. The method according to claim 3, wherein, The up-sampled feature map is input into the content-aware module for processing to obtain a processed feature map, comprising: The up-sampled feature map is input into the content-aware module to learn the local feature relationship in the up-sampled feature map through a feature reorganization layer to generate a feature index map; According to the feature index map, an up-sampling process is adaptively adjusted to obtain a reorganized feature map; The reorganized feature map is input into a feature enhancement layer to obtain a processed feature map.

5. The method according to claim 1, wherein the method is characterized by, The S4 includes: The S4 includes:

6. The method according to claim 1, wherein the method is characterized by, The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: （1） （2） The S4 includes:

7. The method according to claim 1, wherein the method is characterized by, The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes:

8. A fiber line package defect detection and evaluation system fusing industrial CT and deep learning, characterized in that, The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The S4 includes: The evaluation module is used to select different evaluation methods based on the defect category and obtain the evaluation results.

9. A fiber line package defect detection and evaluation device fusing industrial CT and deep learning, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the fiber optic cable defect detection and evaluation method that integrates industrial CT and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the fiber optic cable defect detection and evaluation method integrating industrial CT and deep learning as described in any one of claims 1 to 7.

Images