CT flaw detection auxiliary decision method and system based on ace-yolo instance segmentation model and medium

Abstract

The application provides a CT flaw detection auxiliary decision-making method and system of an ACE-YOLO instance segmentation model and a medium, and belongs to the technical field of computer vision and industrial non-destructive testing technology auxiliary analysis. The method comprises the following steps: S1, acquiring an industrial CT scan image dataset of a composite material cylindrical section, and labeling defects of the industrial CT scan image dataset; S2, based on a YOLOv8 network, replacing a C2f module of a backbone network of the YOLOv8 network with a CCM module, replacing a feature classification loss function BCE Loss of a detection head of a head network of the YOLOv8 network with an EMASlide Loss loss function, and constructing an ACE-YOLO instance segmentation model; S3, training the ACE-YOLO instance segmentation model through the industrial CT scan image dataset, and deploying the ACE-YOLO instance segmentation model to an edge computing terminal after pruning and knowledge distillation; and S4, acquiring an industrial CT scan image to be detected, labeling defects of the industrial CT scan image to be detected through the edge computing terminal, and generating a flaw detection report based on the labeled defects. The method can improve the accuracy of the composite material cylindrical section in the defect detection process.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer vision and industrial non-destructive testing technology auxiliary analysis, and particularly relates to a CT flaw detection auxiliary decision-making method and system of an ACE-YOLO instance segmentation model and a medium. BACKGROUND

[0002] With the rapid development of composite material forming process, high-performance composite materials are increasingly widely used in key load-bearing structures. For example, a cylindrical cylinder segment composed of multiple layers of composite materials is a typical load-bearing component in industrial equipment. However, as the service period is prolonged, the composite cylindrical cylinder segment is prone to cracks, debonding and other typical defects at the interlaminar interface. The existence of such defects can significantly weaken the overall load-bearing performance of the structure, and seriously threaten the safety and reliability of equipment operation.

[0003] Currently, for defect detection of the composite cylindrical cylinder segment, the traditional method relies on manual interpretation, and mainly targets external surface defect positioning and detection, and has low detection accuracy. When the defect is detected by a model, there is no real-time detection and interpretation capability, which is not conducive to meeting the accuracy needs of the composite cylindrical cylinder segment in the defect detection process. SUMMARY

[0004] The technical problem to be solved by the application is how to improve the accuracy of the composite cylindrical cylinder segment in the defect detection process. In view of the deficiencies of the prior art, the application provides a CT flaw detection auxiliary decision-making method and system of an ACE-YOLO instance segmentation model and a medium.

[0005] To solve the above technical problems, the technical solution adopted by the application is:

[0006] In a first aspect, the application provides a CT flaw detection auxiliary decision-making method of an ACE-YOLO instance segmentation model, comprising:

[0007] S1, obtaining an industrial CT scan image dataset of a composite cylindrical cylinder segment, and labeling defects of the industrial CT scan image dataset;

[0008] S2, based on a YOLOv8 network, replacing a C2f module of a backbone network of the YOLOv8 network with a CCM module, and replacing a feature classification loss function BCE Loss of a detection head of a head network of the YOLOv8 network with an EMASlide Loss loss function, to construct an ACE-YOLO instance segmentation model;

[0009] S3, training the ACE-YOLO instance segmentation model through the industrial CT scan image dataset, and deploying to an edge computing terminal after pruning and knowledge distillation;

[0010] S4, acquire an industrial CT scan image to be detected, label defects of the industrial CT scan image to be detected through the edge computing terminal, and generate a flaw detection report based on the labeled defects.

[0011] Compared with the prior art, the CT flaw auxiliary decision-making method of the ACE-YOLO instance segmentation model of the application has the following beneficial effects: obtaining an industrial CT scan image dataset of a composite material cylindrical section, labeling defects on the industrial CT scan image dataset, facilitating subsequent model training; constructing an instance segmentation model based on a YOLOv8 network, wherein the C2f module of the backbone network of the YOLOv8 network is replaced by a CCM module, the CCM module dynamically adjusts feature weights by using the inter-channel correlation, and then the feature expression can be enhanced by calculating the correlation between the feature map channels, so that in the subsequent feature image processing process, redundant features can be reduced, irrelevant background noise can be suppressed, the target edge and details of the feature image can be enhanced, the feature response of small targets and the feature fusion robustness can be adaptively enhanced, the precision of instance segmentation can be effectively improved, and the precision of the composite material cylindrical section in the defect detection process can be improved; on this basis, the feature classification loss function BCE Loss of the detection head of the head network of the YOLOv8 network is replaced by an EMASlide Loss loss function, and finally the ACE-YOLO instance segmentation model is constructed, wherein the EMASlide Loss loss function has a sliding average (EMA) mechanism, which can dynamically adjust the weights of difficult and easy samples, thereby reducing the overfitting of noise labels, and the dynamic threshold adjustment and EMA smoothing mechanism can alleviate the positive and negative sample imbalance problem, so that the ACE-YOLO instance segmentation model pays more attention to small target detection, thereby improving the accuracy of target detection, and the learning process can be balanced by using historical gradient information to avoid subsequent training shock, so that the learning and training process of the ACE-YOLO instance segmentation model is more stable, and the ACE-YOLO instance segmentation model can automatically pay attention to difficult samples such as small targets and occluded targets, improve the recall rate, and further effectively improve the precision of the composite material cylindrical section in the defect detection process; the ACE-YOLO instance segmentation model can be trained by using the pre-collected industrial CT scan image dataset, so that the ACE-YOLO instance segmentation model can meet the detection needs of the industrial CT scan image, after training, the ACE-YOLO instance segmentation model is pruned and compressed to realize lightweight, and the accuracy of the ACE-YOLO instance segmentation model is restored by knowledge distillation, so that the ACE-YOLO instance segmentation model not only meets the lightweight deployment requirements of the edge computing terminal, but also guarantees the accuracy, thereby guaranteeing the accuracy of the subsequent composite material cylindrical section in the defect detection process, and finally the industrial CT scan image to be detected is input into the edge computing terminal to obtain a flaw detection report after completing the defect detection and labeling, and the accurate defect detection of the composite material cylindrical section is realized.

[0012] Optionally, the process of constructing the ACE-YOLO instance segmentation model in S2 further comprises:

[0013] S21, an ASF neck network structure of an ASF-YOLO model is acquired, the neck network of the YOLOv8 network is replaced with the ASF neck network structure, a CPAM attention module of the ASF neck network structure is removed, and an image feature element-wise addition is replaced, and the pruning operation is compatible.

[0014] Optionally, the process of constructing the ACE-YOLO instance segmentation model in S2 further includes:

[0015] S22, a C3 module of the ASF neck network structure is removed and replaced with the C2f module of the YOLOv8 network.

[0016] S23, a scale sequence feature fusion SSFF module of the ASF neck network structure is removed and replaced with a dynamic scale sequence DynamicSSFF module.

[0017] Optionally, the CCM module in S2 processes the input feature map through the following steps:

[0018] S24, channel compression is performed on the input feature map to adjust the input channel number of the feature map to an intermediate channel number.

[0019] S25, channel splitting is performed on the input feature map to decompose the feature map into two parallel branches, one of which retains the original feature transmission path, and the other of which performs deep feature extraction.

[0020] S26, the feature maps of the two parallel branches are spliced in the channel dimension, and the fusion feature channel number is adjusted to a preset output channel number.

[0021] Optionally, in S2, while retaining the EMASlide Loss loss function, the detection head is replaced with an efficient multi-scale front-end shared segmentation head, which is constructed through the following steps:

[0022] S27, a target detection branch and a segmentation mask branch are set to form the efficient multi-scale front-end shared segmentation head.

[0023] S28, a Stem shared feature fusion module and two convolution branches are set to form the target detection branch, and the Stem shared feature fusion module is connected to a Bbox Loss and a Cls Loss through the two convolution branches, respectively.

[0024] S29, a convolution path is set and connected to a Seg Loss to form the segmentation mask branch.

[0025] Optionally, the loss weight expression of the EMASlide Loss loss function is as follows:

[0026] ,

[0027] Wherein, the IoU is the intersection over union of the predicted box and the real box; and the μ is a dynamic threshold.

[0028] Optionally, the generating of the flaw detection report based on the labeled defects in S4 specifically comprises:

[0029] S41, according to the labeled defects, acquiring images of the defects and acquiring position information of the same defects by means of overlapping of adjacent segmentation layer images, the defects including debonding defects and crack defects;

[0030] S42, when the defects are the debonding defects, calculating the length and width of each layer of the defects in the overlapping contour in the order of the images, and calculating the total area of the debonding defects by means of the length and width;

[0031] S43, when the defects are the crack defects, evaluating the hoop dimension of the crack defects by means of the maximum length in the multiple images of the same crack defects, and evaluating the axial dimension of the crack defects by means of the total length of the overlapping contour.

[0032] Optionally, the generating of the flaw detection report based on the labeled defects in S4 specifically further comprises:

[0033] S44, comparing the calculated size of the defects with a preset size safety threshold, and if the size of the defects is greater than the size safety threshold, generating a defect failure conclusion;

[0034] S45, generating a flaw detection report according to the defect failure conclusion.

[0035] In the second aspect, the present application further provides an ACE-YOLO instance segmentation model CT flaw detection auxiliary decision system, comprising a memory and a processor; the memory is used for storing a computer program; when the program is executed by the processor, the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method as described above is realized.

[0036] Compared with the prior art, the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision system has the same beneficial effects as the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method as described above, and will not be repeated here.

[0037] In a third aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the CT flaw detection auxiliary decision method of the ACE-YOLO instance segmentation model.

[0038] Compared with the prior art, the computer readable storage medium of the present application has the same beneficial effects as the CT flaw detection auxiliary decision method of the ACE-YOLO instance segmentation model, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0039] The present application will be further described in detail below with reference to the accompanying drawings.

[0040] Figure 1 The CT flaw detection auxiliary decision method of the ACE-YOLO instance segmentation model in the embodiment of the present application;

[0041] Figure 2 The structure diagram of the ACE-YOLO instance segmentation model in the embodiment of the present application;

[0042] Figure 3 The structure diagram of the CCM module in the embodiment of the present application;

[0043] Figure 4 The structure diagram of the dynamic scale sequence DynamicSSFF module in the embodiment of the present application;

[0044] Figure 5 The structure diagram of the efficient multi-scale front-end shared segmentation head in the embodiment of the present application;

[0045] Figure 6 The structure diagram of the Bottleneck_CABM residual unit in the embodiment of the present application;

[0046] Figure 7 The flow diagram of the defect detection in the embodiment of the present application;

[0047] Figure 8 The flow diagram of the defect adjacent segmentation layer image overlapping processing in the embodiment of the present application;

[0048] Figure 9 The flow diagram of the debonding defect size calculation in the embodiment of the present application;

[0049] Figure 10 The flow diagram of the crack defect size calculation in the embodiment of the present application;

[0050] Figure 11 The structure diagram of the CT flaw detection auxiliary decision system of the ACE-YOLO instance segmentation model in the embodiment of the present application;

[0051] Figure 12 Fig. 1 is an example diagram of defect detection by the YOLOv8 network model in an embodiment of the present application;

[0052] Figure 13 Fig. 2 is another example diagram of defect detection by the YOLOv8 network model in an embodiment of the present application;

[0053] Figure 14 Fig. 3 is still another example diagram of defect detection by the YOLOv8 network model in an embodiment of the present application;

[0054] Figure 15 Fig. 4 is an example diagram of defect detection by the ACE-YOLO instance segmentation model in an embodiment of the present application;

[0055] Figure 16 Fig. 5 is another example diagram of defect detection by the ACE-YOLO instance segmentation model in an embodiment of the present application;

[0056] Figure 17 Fig. 6 is still another example diagram of defect detection by the ACE-YOLO instance segmentation model in an embodiment of the present application. DETAILED DESCRIPTION

[0057] For a better understanding of the present application, the following further clearly sets forth the content of the present application in conjunction with embodiments, but the protection content of the present application is not limited to the following embodiments. In the following description, a large number of specific details are given in order to provide a more thorough understanding of the present application. However, it is obvious to those skilled in the art that the present application can be implemented without one or more of these details.

[0058] It should be noted that the Z-axis in the drawings represents the vertical direction, that is, the up-down position, and the positive direction of the Z-axis represents the upper side, and the negative direction of the Z-axis represents the lower side; the Y-axis in the drawings represents the horizontal direction and is designated as the front-rear position, and the positive direction of the Y-axis represents the front side, and the negative direction of the Y-axis represents the rear side; the X-axis in the drawings represents the left-right position, and the positive direction of the X-axis represents the right side, and the negative direction of the X-axis represents the left side. It should be noted that the meanings of the aforementioned Z-axis, Y-axis and X-axis are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present application.

[0059] As used herein, the term "includes" and its variants are open-ended, meaning that "includes but is not limited to"; the term "based on" means "based, at least in part, on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments." Related terms shall be construed accordingly. It should be noted that "a" or "an" entity as used herein indicates "one or more" of that entity, unless otherwise stated.

[0060] It should be noted that the modification of "one" or "multiple" mentioned in the present application is illustrative and not restrictive, and those skilled in the art should understand that unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0061] In the related art, in recent years, the YOLO (You Only Look Once) series instance segmentation model has achieved significant breakthroughs in real-time performance and accuracy. However, there are still two problems in CT cross-section defect detection of composite cylindrical segments. On the one hand, the defect size in the image is relatively small compared to the entire cylinder segment, and the scarcity of defect data sets brings challenges to the accuracy of the instance segmentation model. Although some enhancement networks and feature fusion techniques have been gradually introduced in recent years to further improve the robustness of detection, this will bring a large number of parameters to the model, resulting in a decrease in model detection speed. And different improvement strategies have different effects on the features of images in different fields, and some improvement methods for object detection are not suitable for small defect targets of some structures. On the other hand, the current YOLO network is mainly used in object detection, and its application in instance segmentation is less, and it cannot realize the subsequent quantitative analysis of defect safety after the network detects defects.

[0062] To solve the above problems, in a first aspect, an embodiment of the present application provides a CT flaw detection auxiliary decision-making method of an ACE-YOLO instance segmentation model, comprising: S1, acquiring an industrial CT scan image dataset of a composite cylindrical section, and labeling defects of the industrial CT scan image dataset; S2, based on a YOLOv8 network, replacing a C2f module of a backbone network of the YOLOv8 network with a CCM module, replacing a feature classification loss function BCE Loss of a detection head of a head network of the YOLOv8 network with an EMASlide Loss loss function, and constructing an ACE-YOLO instance segmentation model; S3, training the ACE-YOLO instance segmentation model through the industrial CT scan image dataset, and deploying the ACE-YOLO instance segmentation model to an edge computing terminal after pruning and knowledge distillation; and S4, acquiring an industrial CT scan image to be detected, labeling defects of the industrial CT scan image to be detected through the edge computing terminal, and generating a flaw detection report based on the labeled defects.

[0063] Specifically, in the embodiment, the process of acquiring the industrial CT scan image dataset of the composite cylindrical section is that the cylindrical section is continuously scanned at an axial interval of 1 mm by an industrial CT machine to obtain CT tomographic images in the hoop direction at different axial positions; the defects on the industrial CT scan image dataset are labeled, the defects in the CT tomographic images can be labeled by using labelme software, and the labels are divided into three types of cracks, debonding and hoop defects; for example, the industrial CT scan image dataset in the embodiment is formed by collecting 2678 CT images of the composite cylindrical section and labeling crack defects and debonding defects.

[0064] In the embodiment, as shown in S1 of the embodiment, Figure 1 , the industrial CT scan image dataset of the composite cylindrical section is acquired, and the defects on the industrial CT scan image dataset are labeled to facilitate subsequent model training; as shown in S2 of the embodiment, Figure 1 , the ACE-YOLO instance segmentation model is constructed based on the YOLOv8 network, wherein, as shown in S2 of the embodiment, Figure 2 and Figure 3As shown, first, the C2f module of the backbone network of the YOLOv8 network is replaced by the CCM module, the CCM module dynamically adjusts the feature weight by using the inter-channel correlation, and then the correlation between the feature map channels is calculated to enhance the feature expression, so that in the subsequent processing process of the feature image, redundant features can be reduced, irrelevant background noise can be suppressed, the target edge and details of the feature image can be enhanced, the feature response of small targets and the feature fusion robustness can be adaptively enhanced, the precision of instance segmentation can be effectively improved, and the precision of the composite material cylindrical section in the defect detection process can be improved. On this basis, the feature classification loss function BCE Loss of the detection head of the head network of the YOLOv8 network is replaced by the EMASlide Loss loss function, and finally the ACE-YOLO instance segmentation model is constructed, wherein the EMASlide Loss loss function has a sliding average (EMA) mechanism, which can dynamically adjust the weights of difficult and easy samples, thereby reducing the overfitting of noise labels. The dynamic threshold adjustment and EMA smoothing mechanism can alleviate the imbalance problem of positive and negative samples, so that the ACE-YOLO instance segmentation model pays more attention to the detection of small targets, thereby improving the accuracy of target detection. The historical gradient information can also balance the learning process, avoid subsequent training shock, and make the learning and training process of the ACE-YOLO instance segmentation model more stable. At the same time, the ACE-YOLO instance segmentation model can automatically pay attention to difficult samples such as small targets and occluded targets, improve the recall rate, and further effectively improve the precision of the composite material cylindrical section in the defect detection process. As shown in S3 in Figure 1 , the ACE-YOLO instance segmentation model can be trained by pre-collected industrial CT scan image data sets, so that the ACE-YOLO instance segmentation model can meet the detection needs of industrial CT scan images. After training is completed, the ACE-YOLO instance segmentation model is pruned and compressed to realize lightweight, and the accuracy of the ACE-YOLO instance segmentation model is restored through knowledge distillation, so that the ACE-YOLO instance segmentation model not only meets the lightweight deployment requirements of edge computing terminals, but also guarantees the accuracy, thereby guaranteeing the accuracy of the subsequent composite material cylindrical section in the defect detection process. Finally, as shown in S4 in Figure 1 , the industrial CT scan image to be detected is input into the edge computing terminal, a flaw detection report is obtained after defect detection and labeling, and accurate defect detection of the composite material cylindrical section is realized.

[0065] It should be noted that in the training process of the ACE-YOLO instance segmentation model, a data set configuration file corresponding to the data set needs to be constructed based on the industrial CT scan image data set. The paths of the training set and the validation set and the class information in the data set need to be written in the data set configuration file. In this embodiment, the environment and the basic hardware conditions required by the model are used to build the environment, and the configured training environment is cuda12.1, the deep learning framework pytorch2.2.2, Intel(R) Xeon(R) Gold 6138 CPU, 64G memory, and the GPU is NVIDIA GeForce RTX 4090D, and the video memory is 24G.

[0066] For example, in the training of this embodiment, the industrial CT scan image data set of the composite cylindrical section is used, and the images of the class labels of cracks, debonding and hoop defects are used as the training set of the model, and the training set, the validation set and the test set are divided according to 7:2:1. The training set has 1686 pictures, the validation set has 480 pictures, and the test set has 240 pictures. In the pre-configured deep learning environment, the improved ACE-YOLO instance segmentation model is imported, the parameter configuration file in the network model is modified according to the actual environment information, and then the model is trained using the processed data set. The image size is 1280*1280, the model training round is 500 rounds, the optimizer uses gradient descent, and batchsize and workers are both set to 8. During the training process, the training log is observed in real time through Tensorboard, and the weight file is saved after the training is completed

[0067] It should be noted that in the process of pruning and knowledge distillation of the ACE-YOLO instance segmentation model, the ACE-YOLO instance segmentation model generates a corresponding weight file after training, and the best.pt weight file is selected to prune and knowledge distill the weight file of the ACE-YOLO instance segmentation model, the Group Taylor structured pruning compression model is adopted, the pruning computation compression ratio is set to 2.0 times, and global pruning is adopted. The optimizer selects the SGD gradient descent, the batchsize is set to 8, the workers is set to 8, the input picture size is 1280, the iteration number of fine-tuning training is 300 rounds. The feature distillation method adopts CWD, the teacher model adopts the ACE-YOLO instance segmentation model before pruning, the feature layer selects all C2f layers in the neck network and the feature addition layer, and the weight of the feature loss is 0.4. Subsequently, the precision is restored by combining the CWD feature distillation, and then the lightweight deployment model is deployed to the edge computing terminal, the edge computing terminal adopts the Ubuntu 20.04 system, the system configuration is Jetpack 5.0.2, CUDA 11.4. The programming language is Python 3.9, and the deep learning framework adopts Pytorch 1.13.1.

[0068] Optionally, the process of constructing the ACE-YOLO instance segmentation model in S2 further includes: S21, obtaining an ASF neck network structure of the ASF-YOLO model, replacing the neck network of the YOLOv8 network with the ASF neck network structure, removing the CPAM attention module of the ASF neck network structure, and replacing it with element-wise addition of image features, and being compatible with pruning operation.

[0069] In this optional embodiment, in order to further improve the performance of the ACE-YOLO instance segmentation model, as shown in Figure 2 The ASF neck network structure of the ASF-YOLO model is obtained, the neck network of the YOLOv8 network is replaced, the CPAM attention module of the ASF neck network structure is removed, and element-wise addition of image features is replaced, and compatible with pruning operation. In this way, the Neck of YOLOv8 is replaced with ASF-Neck, which can effectively enhance multi-scale feature interaction and retain multi-scale feature fusion capability. At the same time, CPAM is removed and replaced with element-wise addition, which not only simplifies the calculation and effectively improves the detection efficiency, but also retains the spatial details of the feature map, which is suitable for small target detection and effectively improves the detection accuracy. In addition, compatible pruning operation can further compress the ACE-YOLO instance segmentation model, which is convenient for lightweight deployment of the ACE-YOLO instance segmentation model.

[0070] Optionally, the process of constructing the ACE-YOLO instance segmentation model in S2 further includes: S22, removing the C3 module of the ASF neck network structure and replacing it with the C2f module of the YOLOv8 network; S23, removing the scale sequence feature fusion SSFF module of the ASF neck network structure and replacing it with the dynamic scale sequence DynamicSSFF module.

[0071] In this optional embodiment, as shown in Figure 2 , the C3 module of the ASF neck network structure is first removed and replaced with the C2f module in YOLOv8. This setting can effectively improve the small target segmentation performance and ensure the detection accuracy; on this basis, as shown in Figure 2 and Figure 4 , the scale sequence feature fusion SSFF module of the ASF neck network structure is removed and replaced with the dynamic scale sequence DynamicSSFF module. The DynamicSSFF replaces the original Interpolate fixed interpolation method of the scale sequence feature fusion SSFF module by introducing a deformable dynamic upsampling mechanism, adopts the DySample module with adaptive learning ability, can better adapt to the feature distribution of different targets to realize adaptive alignment and fusion of multi-level features, and designs different dynamic upsampling ratios for different scale output feature maps. For example, Figure 4 , in p4 and p5 two output path distributions, 2 times and 4 times two ratios are adopted, the sampling position is dynamically adjusted through the learnable spatial offset parameter, and the sub-pixel level feature alignment is realized by combining the grouping attention mechanism and the bilinear interpolation; at the same time, the offset initialization strategy with weight constraint is introduced, the grouped feature processing is performed in the channel dimension, the multi-dimensional offset is generated through the dynamic convolution, the pixel shuffling and low-pass filtering mode selection mechanism are combined, the spatial-channel collaborative fusion framework of multi-scale features is constructed, the adaptability of the feature pyramid to the target scale change and geometric deformation is effectively improved, the semantic consistency of cross-level feature fusion is enhanced, and thus the feature representation ability in the target detection, especially in the small target scene, is optimized.

[0072] In addition, as shown in Figure 2 and Figure 4 , the SiLU activation function of the SSFF in the DynamicSSFF is removed and replaced with the combination of LeakyReLU and three-dimensional pooling layer, which can effectively improve the spatial feature extraction ability and further improve the detection accuracy.

[0073] Optionally, the CCM module in S2 processes the input feature map by the following steps: S24, channel compression is performed on the input feature map to adjust the input channel number of the feature map to an intermediate channel number; S25, channel splitting is performed on the input feature map to decompose the feature map into two parallel branches, one of which retains the original feature transmission path, and the other of which performs deep feature extraction; and S26, the feature maps of the two parallel branches are spliced in the channel dimension, and the fusion feature channel number is adjusted to a preset output channel number.

[0074] In this optional embodiment, as shown in Figure 2 and Figure 3 , the C2f module in the backbone network is replaced by a CCM module, wherein, as shown in S24 and Figure 3 , the CCM module first performs channel compression on the input feature map through a 1x1 standard convolution layer to adjust the input channel number from C in to an intermediate channel number C mid , thereby effectively reducing the calculation amount and memory occupation of subsequent convolution operations through channel compression; then, as shown in S25, the feature map is decomposed into two parallel branches through a channel splitting operation, wherein one of the parallel branches retains the original feature transmission path, and the other of the parallel branches stacks n Bottleneck_CABM residual units for deep feature extraction, so that one of the two parallel branches retains the original feature, has no calculation overhead, and can directly transmit the original feature to avoid the gradient vanishing problem in the deep network, and the other of the two parallel branches performs deep extraction to further reduce redundant calculation, extract high-order semantic features, and enhance the model expression capability. Specifically, as shown in Figure 6 , each Bottleneck unit adopts a Conv-BN-Activation-CBAM fourfold structure, specifically including 1x1 convolution dimension reduction, 3x3 deep separable convolution spatial feature extraction, 1x1 convolution dimension increase operation, and CBAM attention channel, and the gradient flow stability is maintained through cross-layer residual connection, and after the other parallel branch is subjected to n times of nonlinear transformation, the spatial dimension of the output feature map remains HxW, and the channel number remains C mid ; finally, as shown in S26, the feature maps of the two branches are spliced in the channel dimension to form fusion features with a channel number of (0.5n+1)C mid , and the fusion feature channel number is accurately adjusted to a preset output channel number Cout through a 1x1 convolution layer, and the multi-scale detection performance can be effectively improved through feature fusion, and in addition, a SiLU activation function is used to enhance the nonlinear expression capability.

[0075] Optionally, in S2, the detection head is replaced with an efficient multi-scale front-end shared segmentation head while retaining the EMASlide Loss loss function, and the efficient multi-scale front-end shared segmentation head is constructed by the following steps: S27, setting a target detection branch and a segmentation mask branch to form an efficient multi-scale front-end shared segmentation head; S28, setting a Stem shared feature fusion module and two convolution branches to form a target detection branch, and the Stem shared feature fusion module is connected to a bounding box loss Bbox Loss and a classification loss Cls Loss through the two convolution branches; S29, setting a convolution path and connecting a segmentation mask loss Seg Loss to form a segmentation mask branch.

[0076] In this optional embodiment, as shown in Figure 2 and Figure 5 , the detection head is replaced with an efficient multi-scale front-end shared segmentation head, wherein the efficient multi-scale front-end shared segmentation head is divided into two branches of target detection and segmentation mask: the first branch is a target detection branch, and two 3x3 conv convolution layers in the front end of the bounding box loss and the classification loss of the original detection head are combined into a Stem shared feature fusion module, which can reduce repeated calculation, and the Stem shared feature fusion module is composed of one efficient multi-scale enhanced convolution and one 3x3 conv convolution, and the use of a high-efficiency multi-scale enhanced convolution such as a hybrid dilated convolution (Hybrid Dilated Conv) or a dynamic convolution (Dynamic Conv) instead of an ordinary convolution can capture a larger receptive field under the same amount of calculation, thereby improving the small target detection capability. Subsequently, the Stem shared feature fusion module is divided into two branches again, and the two branches are connected to a bounding box loss Bbox Loss and a classification loss Cls Loss after passing through a conv2d convolution, thereby meeting the needs of target detection loss and avoiding task conflicts; the second branch is a segmentation mask branch, which is connected to a segmentation mask loss Seg Loss after passing through two 3x3 conv convolutions and a conv2d two-dimensional convolution, thereby meeting the needs of segmentation mask.

[0077] It should be noted that the specific process expression of the efficient multi-scale enhanced convolution is as shown in formulas (1.1) to (1.4), and each formula is as follows: Group ( x ) = rearrange( x )(1.1); F k ( x ) = ReLU(BN(Conv2d(Group( x ), K = k x k ), k =1, 3, 5, 7 (1.2); O(x ) = rearrange(stack(F1( x ), F3 ( x ), F5 ( x ), F7 ( x )))(1.3); O EMS =Conv2d(O( x ), K =1×1) (1.4). Wherein, in equation (1), Group(·) means dividing the channel dimension into 4 equal groups; rearrange(·) means performing shape reorganization operation on the input feature map; in equation (1.2), Conv2d(·) means depthwise convolution, generating corresponding weights according to the number of channels; ReLU(·) means activation function, BN(·) means batch normalization operation; F(·) means generated features, Conv2d(·) means branch convolution operation with convolution kernel of size K, each branch independently performs convolution, batch normalization BN(·) and ReLU(·) activation; in equation (1.3), stack(·) means stacking the output features of each branch along the new dimension; O(·) means weighted features, in equation (1.4), O EMS This is for the final multi-scale fusion output.

[0078] Optionally, the loss weight expression for the EMASlide Loss function is as follows:

[0079] ,

[0080] Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; μ is the dynamic threshold.

[0081] Specifically, μ The dynamic threshold is calculated using the following formula: ; where μ prev It is the value from the previous iteration. α It is the smoothing factor of EMA, and its value is usually between 0 and 1.

[0082] In this optional embodiment, EMASlide Loss is a dynamic weighted loss function that can adaptively adjust sample weights through an exponential moving average mechanism. It is particularly suitable for common problems in target detection tasks such as class imbalance, uneven distribution of easy and difficult samples, and noisy labels, thereby effectively improving detection accuracy and efficiency.

[0083] Optionally, the generating the flaw detection report based on the labeled defects in S4 specifically comprises: S41, acquiring images of the defects and acquiring position information of the same defect by means of image overlapping of adjacent segmentation layers according to the labeled defects, the defects including debonding defects and crack defects; S42, when the defect is a debonding defect, calculating the length and width of each layer of the defect in the overlapping contour in the order of the images, and calculating the total area of the debonding defect by means of the length and width; S43, when the defect is a crack defect, evaluating the circumferential size of the crack defect by means of the maximum length in the multiple images of the same crack defect, and evaluating the axial size of the crack defect by means of the total length of the overlapping contour.

[0084] In the optional embodiment, after the trained ACE-YOLO instance segmentation model can identify the defects, a defect contour overlapping method is constructed based on the image gray scale binary theory and in combination with relevant standards to determine the continuity of the defects in the CT tomographic pictures in different depth directions, and the size of the defects in the depth direction, i.e., the axial direction, is calculated by means of overlapping summation to realize automatic quantitative identification of the defects, safety evaluation and automatic generation of the report, wherein, as shown in Figure 7 , different ways are selected to identify and calculate the size of the defects according to different defect categories, and before that, as shown in Figure 8 , the position information of the same defect is acquired by means of image overlapping of adjacent segmentation layers, and then, as shown in Figure 9 and Figure 10 , when the defect is a debonding defect, the length and width of each layer of the defect in the overlapping contour are calculated in the order of the images, and the total area of the debonding defect is calculated by means of the length and width, and when the defect is a crack defect, the circumferential size of the crack defect is evaluated by means of the maximum length in the multiple images of the same crack defect, and the axial size of the crack defect is evaluated by means of the total length of the overlapping contour, so as to effectively improve the calculation accuracy and efficiency of the size of the defects.

[0085] Optionally, the generating the flaw detection report based on the labeled defects in S4 specifically further comprises: S44, comparing the calculated size of the defect with a preset size safety threshold, and if the size of the defect is greater than the size safety threshold, generating a defect failure conclusion; and S45, generating the flaw detection report according to the defect failure conclusion.

[0086] In the optional embodiment, in order to ensure the accuracy of the conclusion of the flaw detection report, as shown in Figure 7 , different safety thresholds are set for different types of defects, and when the calculated size of the defect is greater than the safety threshold, the corresponding defect failure conclusion is given, and then the flaw detection report with an accurate conclusion is generated based on the defect failure conclusion.

[0087] For example, first, the CT flaw detection picture is input into the ACE-YOLO instance segmentation model integrated in the terminal to acquire the prediction information given by the network, as shown in Figure 8As shown, all defects in the information are traversed sequentially, and their sizes are calculated. Different defect judgment and evaluation methods are used according to different defect types. To avoid redundant calculations due to network re-labeling, the defect size is calculated by overlapping adjacent segmentation layer images, and the defect information is recorded. Then, as... Figure 7 As shown, for debonding defects, a 40cm... 2 The debonding area is used as a safety threshold, and the length L of the defect in each layer of the overlapping profile is calculated sequentially. i and width D i The total area S is obtained as S = ∑S i =∑L i D i The debonding failure conclusion is given based on the threshold. Next, for crack defects, a safety threshold of 12mm is set for both axial and circumferential lengths based on the crack propagation direction. For example... Figure 9 and Figure 10 As shown, the length of circumferential defects is evaluated using the maximum length, and the length of axial cracks is evaluated using the total length of the overlapping profile. A crack failure conclusion is given based on a threshold. Finally, all defect information given in the prediction results of the ACE-YOLO instance segmentation model is traversed, and safety judgments are made according to the above methods, resulting in a failure conclusion.

[0088] It should be noted that, as Figure 11 As shown, the entire set of CT flaw detection image data to be inspected is used as system input. The system accesses the CT flaw detection host computer, which stores the flaw detection results images, via network cable interaction and FTP technology. The flaw detection images are automatically loaded, and the ACE-YOLO network model in the edge computing terminal automatically and quickly labels the location and size of relevant defects. Then, the defect logic discrimination module designed above performs calculations to obtain the final discrimination result and infer relevant conclusions. Finally, the conclusions are automatically generated into a flaw detection report and pushed to the CT flaw detection host computer, providing recommendations for flaw detection conclusions and assisting operators in making relevant judgments. This enables the uploading of intelligent analysis conclusions from CT flaw detection.

[0089] To demonstrate the beneficial effects of the ACE-YOLO instance segmentation model of this invention, the ablation experiment results will be analyzed below to verify whether each module plays a role in the model and to compare their performance. Simultaneously, comparative experiments will be conducted to compare the performance of the model after pruning and distillation. The ablation experiment results are shown in Table 1, and the pruning and distillation experiment results are shown in Table 2.

[0090] Table 1 Ablation Experiment Results

[0091] Model P(B) R(B) mAP(B) P(M) R(M) mAP(M) Params(M) FLOPs FPS YOLOv8n 0.701 0.640 0.708 0.708 0.56 0.647 3.41 12.8 315 +CCM 0.731 0.657 0.731 0.713 0.654 0.701 3.44 12.8 309 +DSFv8 0.766 0.622 0.691 0.765 0.621 0.68 3.46 13.2 256 +DSFv8+CBAM 0.707 0.712 0.76 0.699 0.702 0.74 3.49 13.2 254 +DSFv8+CBAM+EMSSegment 0.706 0.715 0.775 0.725 0.661 0.74 3.28 11 273 ACE-YOLO+EMASlide Loss 0.716 0.706 0.779 0.734 0.681 0.765 3.28 11 273

[0092] From the ablation experiment results, it can be seen that after introducing the CCM module, the detection and segmentation mAP50 are increased by 1.4% and 5.8% respectively, the information of the feature map in the channel and spatial dimensions is enhanced, the parameter quantity is increased by 0.03M, and the calculation quantity FLOPs and inference speed FPS change little. Secondly, after adding the DSFv8 neck structure, the detection decreases slightly, the segmentation mAP50 increases by about 3%, but the detection precision P(B) and segmentation precision P(M) increase by about 6%, the parameter quantity increases by 0.05M, the calculation quantity increases slightly, and the FPS decreases by 18%. However, after introducing the CCM module and the DSFv8 neck structure at the same time, the detection recall rate R(B) and the segmentation recall rate R(M) are increased by 7% and 14% respectively, the detection and segmentation mAP50 are increased by 5% and 9% respectively, which indicates that the coordinated improvement effect of the CCM module and the DSFv8 structure is greater than that of the single introduction. Since the parameter quantity of the model is large at this time, more video memory and computing resources are required during training, therefore, the EMSSegment detection head is introduced. The structure realizes the lightweight target, reduces the calculation quantity of the synergistically improved network by 2.1G, increases the FPS by 7%, and slightly increases the overall performance, realizing the lightweight of the model. After improving the loss function, the precision is increased by about 1%, and the segmentation mAP50 is increased by 2%, which indicates that the improved loss function improves the stability, enhances the generalization ability and robustness, and reduces the overfitting.

[0093] Table 2 Pruning and distillation experiment results

[0094] Model P(B) R(B) mAP(B) P(M) R(M) mAP(M) Params(M) FLOPs FPS ACE-YOLO 0.716 0.706 0.779 0.734 0.681 0.765 3.3 11 274 Pruning 0.729 0.693 0.753 0.735 0.696 0.759 1.2 5.3 374 Distillation 0.745 0.697 0.759 0.76 0.703 0.761 1.2 5.3 374

[0095] As can be seen from Table 2, after pruning, the parameter quantity is reduced by 2.1M, the calculation quantity is reduced by 5.7G, and the inference speed is increased by 36%. Due to the reduction of the parameter quantity, the detection and segmentation mAP of the model are decreased by 2.6% and 0.6% respectively. After adopting the CWD feature distillation, the detection and segmentation mAP are increased by 0.6% and 0.2% respectively, and the detection and segmentation precision P is increased by 2.9% and 2.6% respectively.

[0096] In order to demonstrate the effect achieved by the application, the improved model is combined with the original model Figures 12 to 17 for comparison. As shown in Figures 12 to 14 , due to the large span of the size of part of the defects, the original model appears repeated recognition in Figure 12 , and appears discontinuous phenomenon in Figure 13 . Correspondingly, it can be seen from the comparison chart that the detection performance of the improved ACE-YOLO instance segmentation model on the number and shape of targets is better than that of the original model.

[0097] In a second aspect, an embodiment of the present application provides an ACE-YOLO instance segmentation model CT flaw detection auxiliary decision system, comprising a memory and a processor; the memory is used for storing a computer program; when the program processor executes, the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method described above is realized.

[0098] The technical effect of the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision system in the embodiment is similar to that of the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method described above, and will not be repeated here.

[0099] It should be noted that, in the embodiment, as shown in Figure 11 The ACE-YOLO instance segmentation model CT flaw detection auxiliary decision system can integrate an edge computing terminal into a suitcase, design a unique interactive interface, i.e., a touch display, open the lid, touch the power-on key on the inner lining cover plate, so that the entire system completes booting within 20 seconds. Through the human-computer interaction UI designed on the touch display, soft shutdown can be realized, and hard shutdown can also be realized through the power key on the inner lining cover plate. After shutdown, the touch display can be quickly stored through the folding support hinge, and the upper lid is closed to complete the storage of the entire system. Then, the auxiliary decision system accesses the CT flaw detection host computer which stores flaw result images through the RJ45 interface and the network line interaction and FTP technology, scans to obtain a flaw section image, automatically loads the flaw picture, and inputs the ACE-YOLO network model in the edge computing terminal to automatically and quickly label the related defect position and size. Finally, the defect logical discrimination module designed above is operated to obtain the final discrimination result and infer the related conclusion. Finally, the conclusion is automatically formed into a flaw report and pushed to the CT flaw detection host computer to recommend the flaw conclusion, assist the operator to discriminate and obtain the related conclusion, and realize the uploading of the CT flaw intelligent analysis conclusion.

[0100] In a third aspect, an embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to realize the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method described above.

[0101] The technical effect of the computer readable storage medium in the embodiment is similar to that of the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision method described above, and will not be repeated here.

[0102] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A CT flaw detection auxiliary decision method of an ACE-YOLO instance segmentation model, characterized in that, Comprise: S1, obtain an industrial CT scan image dataset of a composite cylindrical section, and label defects in the industrial CT scan image dataset; S2, based on a YOLOv8 network, replace a C2f module of a backbone network of the YOLOv8 network with a CCM module, and replace a feature classification loss function BCE Loss of a detection head of a head network of the YOLOv8 network with an EMASlide Loss loss function, wherein the CCM module processes an input feature map through the following steps: S24, channel compression is performed on the input feature map to adjust the input channel number of the feature map to an intermediate channel number; S25, the input feature map is split in the channel dimension to decompose the feature map into two parallel branches, one of which retains the original feature transmission path, and the other of which stacks n Bottleneck_CBAM units for deep feature extraction, wherein each Bottleneck_CBAM unit adopts a Conv-BN-Activation-CBAM fourfold structure; S26, the feature maps of the two parallel branches are spliced in the channel dimension, and the fusion feature channel number is adjusted to a preset output channel number; S21, obtain an ASF neck network structure of an ASF-YOLO model, replace the neck network of the YOLOv8 network with the ASF neck network structure, remove the CPAM attention module of the ASF neck network structure, replace it with element-wise addition of image features, and be compatible with pruning operation; S27, set a target detection branch and a segmentation mask branch to form an efficient multi-scale front-end shared segmentation head; S28, set a Stem shared feature fusion module and two convolution branches to form the target detection branch, wherein the Stem shared feature fusion module is connected to a bounding box loss Bbox Loss and a classification loss ClsLoss through the two convolution branches, respectively, wherein the Stem shared feature fusion module is composed of an efficient multi-scale enhanced convolution and a 3×3 conv convolution, and the specific process expression of the efficient multi-scale enhanced convolution is as shown in formulas (1.1) to (1.4), and each formula is as follows: Group(x)=rearrange(x) (1.1); F k (x) = ReLU(BN(Conv2d(Group(x), K=k x k), k=1, 3, 5, 7 (1.2); O(x)=rearrange(stack(F1(x),F3(x),F5(x),F7(x))) (1.3); O EMS = Conv2d(O(x), K=1x1) (1.4); In the formula (1.1), Group (·) represents that the channel dimension is divided into 4 equal groups; rearrange (·) represents a shape reorganization operation on the input feature map; in the formula (1.2), Conv2d (·) represents a deep convolution, and corresponding weights are generated according to the number of channels; ReLU (·) represents an activation function, and BN (·) represents a batch normalization operation; F (·) represents generated features, and Conv2d (·) represents a branch convolution operation using a convolution kernel with a size of K, and each branch independently performs convolution, batch normalization BN (·) and ReLU (·) activation; in the formula (1.3), stack (·) represents that the output features of each branch are stacked along a new dimension; O (·) represents weighted features, and in the formula (1.4), O EMS is the final multi-scale fusion output; S29, set a convolution path and connect a segmentation mask loss Seg Loss to form the segmentation mask branch, replace the detection head with the efficient multi-scale front-end shared segmentation head while retaining the EMASlide Loss loss function, and construct an ACE-YOLO instance segmentation model; S3, train the ACE-YOLO instance segmentation model through the industrial CT scan image dataset, and deploy it to an edge computing terminal after pruning and knowledge distillation. S4, acquiring an industrial CT scan image to be detected, labeling defects of the industrial CT scan image to be detected through the edge computing terminal, and generating a flaw detection report based on the labeled defects.

2. The CT inspection auxiliary decision method of the ACE-YOLO instance segmentation model according to claim 1, wherein, The process of constructing the ACE-YOLO instance segmentation model of S2 further comprises: S22, removing the C3 module of the ASF neck network structure and replacing it with the C2f module of the YOLOv8 network; S23, removing the scale sequence feature fusion SSFF module of the ASF neck network structure and replacing it with the dynamic scale sequence DynamicSSFF module. 3.The CT flaw detection auxiliary decision method of the ACE-YOLO instance segmentation model according to claim 1 or 2, wherein, The loss weight expression of the EMASlide Loss loss function is as follows: ， Wherein, the IoU is the intersection over union of the predicted box and the real box; and the μ is a dynamic threshold.

4. The CT inspection auxiliary decision method of the ACE-YOLO instance segmentation model according to claim 1 or 2, characterized in that, The generating a flaw detection report based on the labeled defects in S4 specifically comprises: S41, according to the labeled defects, acquiring images of the defects and obtaining position information of the same defects through adjacent segmentation layer image overlap, wherein the defects include debonding defects and crack defects; S42, when the defect is the debonding defect, calculating the length and width of each layer of the defect in the overlapping contour according to the order of the images, and calculating the total area of the debonding defect through the length and width; S43, when the defect is the crack defect, evaluating the hoop dimension of the crack defect through the maximum length in the multiple images of the same crack defect, and evaluating the axial dimension of the crack defect through the total length of the overlapping contour. 5.The CT flaw detection auxiliary decision method of the ACE-YOLO instance segmentation model according to claim 4, wherein, The generating a flaw detection report based on the labeled defects in S4 specifically further comprises: S44, comparing the calculated size of the defect with a preset size safety threshold, and if the size of the defect is greater than the size safety threshold, generating a defect failure conclusion; S45, generating a flaw detection report according to the defect failure conclusion.

6. An ACE-YOLO instance segmentation model-based CT flaw detection auxiliary decision system, characterized in that, A memory and a processor; the memory is used to store a computer program; when the program is executed by the processor, the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision-making method of any one of claims 1-5 is realized.

7. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program is executed by a processor to realize the ACE-YOLO instance segmentation model CT flaw detection auxiliary decision-making method of any one of claims 1-5.

Images