Method for detecting internal defects in carbon fiber honeycomb material

Abstract

The application discloses a kind of carbon fiber honeycomb material internal defect detection method, it is related to carbon fiber honeycomb material detection field, comprising: obtaining training dataset;Build internal defect detection model, detection model is obtained by using StarFFTBlock module and ACGAFusion module to improve YOLOv8 model, StarFFTBlock module is used to extract the spatial domain feature and frequency domain feature of input data, and the input data and its spatial domain feature and frequency domain feature are fused, ACGAFusion module is used to adaptively generate weight parameters, shallow features and deep features are weighted, and the input shallow features and deep features, the weighted shallow features and deep features are fused;Train detection model;Carbon fiber honeycomb material internal defect detection is carried out using detection model.The application can significantly improve the accuracy and detection efficiency of carbon fiber honeycomb material internal defect detection.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon fiber honeycomb material detection, and in particular to a method for detecting internal defects of carbon fiber honeycomb materials. Background Art

[0002] Carbon fiber honeycomb material is a high-performance composite material with excellent properties such as light weight, high strength and high modulus. Therefore, it is widely used in aerospace, automobile, construction and other fields. Carbon fiber honeycomb material is composed of carbon fiber and honeycomb structure. It not only retains the excellent mechanical properties of carbon fiber, but also has the unique mechanical advantages of honeycomb structure. As a new type of high-strength and high-modulus material, carbon fiber itself has significantly better strength and stiffness than steel and aluminum alloy; and the honeycomb structure can give carbon fiber materials excellent energy absorption capacity and compressive properties, so that carbon fiber materials can effectively disperse stress under external loads, improve compressive strength and structural stability. However, the presence of internal defects in carbon fiber honeycomb materials may affect their service life and safety. Therefore, before using carbon fiber honeycomb materials, it is necessary to detect internal defects in carbon fiber honeycomb materials to ensure the performance and safety of carbon fiber honeycomb materials.

[0003] Currently, internal defect detection methods for carbon fiber honeycomb materials primarily rely on manual inspection and traditional image processing algorithms. However, manual inspection is inefficient and susceptible to human interference, while traditional image processing algorithms often suffer from low detection accuracy and insufficient generalization capabilities when dealing with complex backgrounds and minor defects. Summary of the Invention

[0004] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a method for detecting internal defects of carbon fiber honeycomb materials, which can significantly improve the accuracy, detection efficiency and robustness of internal defect detection of carbon fiber honeycomb materials.

[0005] The technical solutions of the present invention are as follows:

[0006] A method for detecting internal defects of carbon fiber honeycomb materials is provided, comprising:

[0007] Obtain a training data set, the training data including a speckle package phase image of internal defects of a carbon fiber honeycomb material and its corresponding defect category and location;

[0008] Constructing an internal defect detection model, wherein the internal defect detection model is obtained by replacing some C2F modules in the YOLOv8 model backbone network with StarFFTBlock modules, and replacing some Concat modules in the YOLOv8 model neck network with ACGAFusion modules, the StarFFTBlock module is used to extract spatial domain features and frequency domain features of input data, and fuse the input data and the spatial domain features and frequency domain features of the input data to obtain and output a feature map, the ACGAFusion module uses an attention mechanism to adaptively generate weight parameters according to the input shallow features and deep features, weight the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features, to obtain and output a fused feature;

[0009] training the internal defect detection model using the training data set;

[0010] The trained internal defect detection model is used to detect internal defects of carbon fiber honeycomb materials.

[0011] In some optional implementations, the training dataset is obtained by:

[0012] An experimental platform was built using laser shearing speckle interferometer;

[0013] In the experimental platform, negative pressure is applied to the test areas of multiple pre-prepared carbon fiber honeycomb materials in sequence, and a CCD camera is used to collect the speckle wrapping phase image;

[0014] The defect category and position in each collected speckle wrapped phase image are marked to obtain a plurality of training data including speckle wrapped phase images of internal defects of carbon fiber honeycomb materials and their corresponding defect categories and positions.

[0015] In some optional embodiments, the internal defect detection model includes: a backbone network, a neck network, and a head network connected in sequence;

[0016] The backbone network is used to extract features of the input image and output them to the neck network. The neck network is used to fuse feature maps of different scales input by the backbone network and output them to the head network. The head network is used to determine and output the defect category and location in the input image based on the feature maps input by the neck network.

[0017] In some optional embodiments, the backbone network includes: a first convolution unit, a second convolution unit, a first C2F module, a third convolution unit, a second C2F module, a fourth convolution unit, a first StarFFTBlock module, a fifth convolution unit, a second StarFFTBlock module and a spatial pyramid pooling module connected in sequence, the output of the second C2F module is also connected to the neck network, the output of the first StarFFTBlock module is also connected to the neck network, and the output of the spatial pyramid pooling module is connected to the neck network;

[0018] The convolution unit is used to perform convolution operations on input data, extract and output features;

[0019] The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features;

[0020] The StarFFTBlock module is used to extract the spatial domain features and frequency domain features of the input data, and fuse the input data, the spatial domain features and frequency domain features of the input data to obtain and output the feature map;

[0021] The spatial pyramid pooling module is used to fuse features at different scales of the same feature map through a connection operation and output a fusion result.

[0022] In some optional embodiments, the StarFFTBlock module includes:

[0023] A first depthwise separable convolution unit, whose input is connected to the input of the StarFFTBlock module, and whose output is connected to the input of the eighth and ninth convolution units respectively, is used to perform depthwise convolution and pointwise convolution operations on the input data;

[0024] The output of the eighth convolution unit is connected to the input of the ReLU6 activation function unit, and is used to perform a convolution operation on the input data;

[0025] The ninth convolution unit has an output connected to the input of the multiplication unit, and is used to perform a convolution operation on the input data;

[0026] The ReLU6 activation function unit, whose output is connected to the input of the multiplication unit, is used to perform nonlinear transformation processing on the input data using the ReLU6 activation function;

[0027] The multiplication unit, whose output is connected to the input of the tenth convolution unit, is used to perform element-by-element multiplication of the two inputs;

[0028] The tenth convolution unit has an output connected to the input of the second depth-wise separable convolution unit, and is configured to perform a convolution operation on the input data;

[0029] The second depthwise separable convolution unit has an output connected to the input of the addition unit, and is used to perform depthwise convolution and pointwise convolution operations on the input data;

[0030] A two-dimensional Fourier transform unit, whose input is connected to the input of the StarFFTBlock module and whose output is connected to the input of the virtual-real segmentation unit, is used to perform Fourier transform processing on the input data to obtain and output frequency domain features;

[0031] The output of the virtual-real segmentation unit is connected to the input of the eleventh convolution unit and the twelfth convolution unit respectively, and is used to separate the real part and the imaginary part of the input frequency domain feature, and output the real part and the imaginary part of the frequency domain feature to the eleventh convolution unit and the twelfth convolution unit respectively;

[0032] The eleventh convolution unit has an output connected to the input of the splicing unit, and is used to perform a convolution operation on the input data;

[0033] The output of the twelfth convolution unit is connected to the input of the splicing unit, and is used to perform a convolution operation on the input data;

[0034] The output of the concatenation unit is connected to the input of the thirteenth convolution unit, and is used to concatenate the two inputs to obtain and output a fusion feature;

[0035] The thirteenth convolution unit has an output connected to the input of the two-dimensional inverse Fourier transform unit, and is used to perform a convolution operation on the input data;

[0036] The output of the two-dimensional inverse Fourier transform unit is connected to the input of the adding unit, and is used to perform inverse Fourier transform processing on the input data, convert the frequency domain features into spatial domain features and output them;

[0037] The addition unit, whose input is also connected to the input of the StarFFTBlock module, and whose output is connected to the input of the fourteenth convolution unit, is used to add multiple inputs element by element to obtain and output fusion features;

[0038] The output of the fourteenth convolution unit is connected to the output of the StarFFTBlock module, and is used to perform a convolution operation on the input fusion features to obtain and output a feature map.

[0039] In some optional embodiments, the neck network includes: a first upsampling unit, a first ACGAFusion module, a third C2F module, a second upsampling unit, a second ACGAFusion module, a fourth C2F module, a sixth convolution unit, a first splicing unit, a fifth C2F module, a seventh convolution unit, a second splicing unit, and a sixth C2F module connected in sequence, wherein the input of the first upsampling unit is connected to the output of the spatial pyramid pooling module in the backbone network, the input of the second splicing unit is also connected to the output of the spatial pyramid pooling module in the backbone network, the input of the first ACGAFusion module is also connected to the output of the first StarFFTBlock module in the backbone network, the output of the third C2F module is also connected to the input of the first splicing unit, the input of the second ACGAFusion module is also connected to the output of the second C2F module in the backbone network, the output of the fourth C2F module is also connected to the input of the head network, the output of the fifth C2F module is also connected to the input of the head network, and the output of the sixth C2F module is connected to the input of the head network;

[0040] The upsampling unit is used to perform upsampling processing on the input data;

[0041] The ACGAFusion module is used to adaptively generate weight parameters based on the input shallow features and deep features, perform weighted processing on the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features, to obtain and output fused features;

[0042] The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features;

[0043] The convolution unit is used to perform convolution operations on input data, extract and output features;

[0044] The splicing unit is used to perform splicing operations on two inputs to obtain and output fusion features.

[0045] In some optional embodiments, the ACGAFusion module includes: a shallow feature receiving unit, a deep feature receiving unit, an ACGA module, a first multiplication unit, a second multiplication unit, an addition unit, and a fifteenth convolution unit;

[0046] The input of the shallow feature receiving unit is connected to the input of the ACGAFusion module, and the output of the shallow feature receiving unit is connected to the input of the ACGA module and the first multiplication unit respectively. The shallow feature receiving unit is used to receive the shallow features input into the ACGAFusion module and output the shallow features;

[0047] The input of the deep feature receiving unit is connected to the input of the ACGAFusion module, and the output of the deep feature receiving unit is connected to the input of the ACGA module and the second multiplication unit respectively. The deep feature receiving unit is used to receive the deep features input into the ACGAFusion module and output the deep features;

[0048] The output of the ACGA module is connected to the input of the first multiplication unit and the second multiplication unit respectively, and the ACGA module is used to adaptively generate and output weight parameters according to the shallow features and deep features of the input based on the attention mechanism;

[0049] The output of the first multiplication unit is connected to the input of the addition unit, and the first multiplication unit is used to perform element-by-element multiplication on the input shallow features and weight parameters;

[0050] The output of the second multiplication unit is connected to the input of the addition unit, and the second multiplication unit is used to perform element-by-element multiplication on the input deep features and weight parameters;

[0051] The input of the adding unit is further connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively, the output of the adding unit is connected to the input of the fifteenth convolution unit, and the adding unit is used to perform element-by-element addition on the four inputs;

[0052] The output of the fifteenth convolution unit is connected to the output of the ACGAFusion module, and the fifteenth convolution unit is used to perform a convolution operation on the input data to extract and output features.

[0053] In some optional embodiments, the ACGA module includes: an AFM module, a spatial attention module, a channel attention module, an addition unit, a splicing unit, a sixteenth convolution unit, and a Sigmoid activation function unit;

[0054] The input of the AFM module is connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively, and the output of the AFM module is connected to the input of the spatial attention module and the channel attention module respectively. The AFM module is used to perform global average pooling processing on the shallow features and the deep features, splice the two features after pooling together, and generate a weighting parameter in the range of [0,1] through a fully connected layer and a Sigmoid activation function transformation. The weighting parameter and its complement are respectively multiplied element-by-element with the shallow features and the deep features to obtain and output a fusion feature;

[0055] The output of the spatial attention module is connected to the input of the addition unit, and the spatial attention module is used to extract and fuse features of the input data in the spatial dimension;

[0056] The output of the channel attention module is connected to the input of the adding unit, and the channel attention module is used to extract features of the input data in the channel dimension;

[0057] The output of the adding unit is connected to the input of the splicing unit, and the adding unit is used to perform element-by-element addition on two inputs;

[0058] The input of the splicing unit is also connected to the output of the AFM module, and the output of the splicing unit is connected to the input of the sixteenth convolution unit. The splicing unit is used to perform a splicing operation on the two inputs to obtain and output a fusion feature;

[0059] The output of the sixteenth convolution unit is connected to the input of the Sigmoid activation function unit, and the sixteenth convolution unit is used to perform a convolution operation on the input data to extract and output features;

[0060] The output of the Sigmoid activation function unit is connected to the output of the ACGA module. The Sigmoid activation function unit is used to transform the input data using the Sigmoid activation function to generate and output a weighting parameter in the range of [0, 1].

[0061] In some optional embodiments, the internal defect detection model is trained in the following manner:

[0062] The internal defect detection model is trained by using the speckle wrapped phase image of internal defects of carbon fiber honeycomb materials in the training data in the training data set as the input of the internal defect detection model, and using the defect category and position corresponding to the input speckle wrapped phase image of internal defects of carbon fiber honeycomb materials as the output.

[0063] In some optional embodiments, the internal defect detection of the carbon fiber honeycomb material utilizes the trained internal defect detection model, and the internal defect detection of the carbon fiber honeycomb material comprises the following steps:

[0064] Obtaining a speckle wrapped phase map of the carbon fiber honeycomb material to be detected;

[0065] Inputting the speckle wrapped phase map of the carbon fiber honeycomb material to be detected into the internal defect detection model to obtain a defect category and a position corresponding to the speckle wrapped phase map of the carbon fiber honeycomb material to be detected output by the internal defect detection model.

[0066] The main advantages of the technical scheme of the present application are as follows:

[0067] The internal defect detection method of the carbon fiber honeycomb material can replace part of the C2F module in the backbone network of the YOLOv8 model with the StarFFTBlock module, retain local features in the spatial domain during feature extraction processing, effectively integrate global features through frequency domain transformation, significantly improve the feature expression ability of the model, effectively reduce the computational complexity of the model, and improve the accuracy, detection efficiency and robustness of the model during internal defect detection of the carbon fiber honeycomb material.

[0068] The internal defect detection method of the carbon fiber honeycomb material can replace part of the Concat module in the neck network of the YOLOv8 model with the ACGAFusion module, adaptively learn weights based on the attention mechanism during feature extraction and fusion processing, dynamically adjust the fusion mode of different levels of features, effectively improve the information mismatch problem caused by the difference in receptive field, promote more efficient fusion between low-level features and high-level features, improve the expression ability and performance of the model, and improve the accuracy, detection efficiency and robustness of the model during internal defect detection of the carbon fiber honeycomb material. BRIEF DESCRIPTION OF DRAWINGS

[0069] The accompanying drawings, which are included to provide a further understanding of the embodiments of the present application, constitute a part of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0070] Figure 1 The flowchart of the internal defect detection method of the carbon fiber honeycomb material provided for the embodiments of the present application;

[0071] Figure 2 The structure diagram of the internal defect detection model provided for the embodiments of the present application;

[0072] Figure 3 The internal defect detection method of the carbon fiber honeycomb material provided for the embodiments of the present application; Figure 2The schematic diagram of the structure of the StarFFTBlock module in the internal defect detection model shown;

[0073] Figure 4 for Figure 2 The schematic diagram of the structure of the ACGAFusion module in the internal defect detection model shown;

[0074] Figure 5 for Figure 4 The structural diagram of the ACGA module in the ACGAFusion module is shown. DETAILED DESCRIPTION

[0075] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0076] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0077] See also Figure 1-2 The embodiment of the present invention provides a method for detecting internal defects of carbon fiber honeycomb materials, the method comprising the following steps 1 to 4:

[0078] Step 1: Obtain training data set;

[0079] In the embodiment of the present invention, the training data set consists of multiple training data.

[0080] In an embodiment of the present invention, the training data includes a speckle wrapping phase image of internal defects of a carbon fiber honeycomb material and its corresponding defect category and position.

[0081] It should be noted that the number of training data included in the training data set is set according to the actual situation.

[0082] Step 2: Build an internal defect detection model;

[0083] In an embodiment of the present invention, an internal defect detection model is obtained by replacing part of the C2F modules in the backbone network of the YOLOv8 model with StarFFTBlock modules, and replacing part of the Concat modules in the neck network of the YOLOv8 model with ACGAFusion modules. The StarFFTBlock module is used to extract the spatial domain features and frequency domain features of the input data, and fuse the input data and the spatial domain features and frequency domain features of the input data to obtain and output a feature map. The ACGAFusion module adopts an attention mechanism to adaptively generate weight parameters according to the input shallow features and deep features, weight the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features to obtain and output a fused feature.

[0084] Step 3: Use the training dataset to train the internal defect detection model;

[0085] In an embodiment of the present invention, before training the internal defect detection model, the parameters of the internal defect detection model are initialized and set, and the internal defect detection model is trained using a training data set to update the parameters of the internal defect detection model.

[0086] Step 4: Use the trained internal defect detection model to detect internal defects of carbon fiber honeycomb materials.

[0087] In an embodiment of the present invention, based on actual detection requirements, the carbon fiber honeycomb material to be detected is determined and obtained, and the internal defect detection of the carbon fiber honeycomb material to be detected is performed using a trained internal defect detection model to determine whether the carbon fiber honeycomb material to be detected has internal defects, as well as the category and location information of the existing internal defects.

[0088] The carbon fiber honeycomb material internal defect detection method provided by an embodiment of the present invention replaces part of the C2F module in the YOLOv8 model backbone network with the StarFFTBlock module. During the feature extraction process, it can retain local features in the spatial domain and effectively integrate global features through frequency domain transformation, which can significantly improve the feature expression ability of the model and effectively reduce the computational complexity of the model, thereby improving the accuracy, detection efficiency and robustness of the model when performing internal defect detection of carbon fiber honeycomb materials; by replacing part of the Concat module in the YOLOv8 model neck network with the ACGAFusion module, during the feature extraction and fusion process, it can adaptively learn weights based on the attention mechanism, and then dynamically adjust the fusion mode of features at different levels, which can effectively improve the information mismatch problem caused by the difference in receptive fields, promote more efficient fusion between low-level features and high-level features, enhance the expression ability and performance of the model, and improve the accuracy, detection efficiency and robustness of the model when performing internal defect detection of carbon fiber honeycomb materials.

[0089] Furthermore, in an embodiment of the present invention, in step 1, a training data set is obtained by:

[0090] An experimental platform was built using laser shearing speckle interferometer;

[0091] In the experimental platform, negative pressure is applied to the test areas of multiple pre-prepared carbon fiber honeycomb materials in sequence, and a CCD camera is used to collect the speckle wrapping phase image;

[0092] The defect category and position in each collected speckle wrapped phase image are marked to obtain a plurality of training data including speckle wrapped phase images of internal defects of carbon fiber honeycomb materials and their corresponding defect categories and positions.

[0093] Furthermore, in an optional implementation of the embodiment of the present invention, in order to improve the quality of the acquired training data set and thereby improve the training efficiency and performance of the model, when acquiring the training data set, images with blurring or target occlusion problems in the acquired speckle wrapper phase image are screened out.

[0094] Furthermore, in an optional implementation of the embodiment of the present invention, in order to improve the quality of the acquired training data set and thereby improve the training efficiency and performance of the model, when acquiring the training data set, filtering processing and data enhancement are sequentially performed on each acquired speckle package phase image to obtain multiple speckle package phase images to expand the training data.

[0095] In the embodiment of the present invention, a Gaussian filter function is used to filter the speckle wrapping phase image to remove noise in the image.

[0096] In the embodiment of the present invention, the Gaussian filter function is specifically expressed as:

[0097]

[0098] Where x and y represent the horizontal and vertical coordinates of the pixels in the image, σ represents the standard deviation of the Gaussian function, and e represents a natural constant.

[0099] In the embodiment of the present invention, data enhancement includes: mirroring, rotation, distortion and Cutmix operations.

[0100] In an embodiment of the present invention, the OpenCV library and the Pillow library in Python are used to perform mirroring, rotation, distortion, and Cutmix operations on an image.

[0101] Furthermore, in an embodiment of the present invention, when acquiring a training data set, a LabelImg tool is used to label the defect categories and positions in the speckle wrapper phase image.

[0102] refer to Figure 2 ,Further, in an optional implementation of the embodiment of the present invention, in order to realize the function of the internal defect detection model defined above, the internal defect detection model includes: a backbone network, a neck network and a head network connected in sequence;

[0103] The backbone network is used to extract the features of the input image and output them to the neck network. The neck network is used to fuse the feature maps of different scales input by the backbone network and output them to the head network. The head network is used to determine and output the defect category and location in the input image based on the feature maps input by the neck network.

[0104] Furthermore, in an embodiment of the present invention, the backbone network includes: a first convolution unit, a second convolution unit, a first C2F module, a third convolution unit, a second C2F module, a fourth convolution unit, a first StarFFTBlock module, a fifth convolution unit, a second StarFFTBlock module and a spatial pyramid pooling module connected in sequence, the output of the second C2F module is also connected to the neck network, the output of the first StarFFTBlock module is also connected to the neck network, and the output of the spatial pyramid pooling module is connected to the neck network;

[0105] The convolution unit is used to perform convolution operations on input data, extract and output features;

[0106] The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features;

[0107] The StarFFTBlock module is used to extract the spatial domain features and frequency domain features of the input data, and fuse the input data, the spatial domain features and frequency domain features of the input data to obtain and output the feature map;

[0108] The spatial pyramid pooling module is used to fuse features at different scales of the same feature map through connection operations and output the fusion result.

[0109] In an embodiment of the present invention, by utilizing the backbone network defined above to extract features of the input image, the spatial relationship and contextual dependency of features at different scales can be accurately captured, and a good feature extraction effect can be achieved. In the feature extraction process, local features in the spatial domain can be retained, and global features can be effectively integrated through frequency domain transformation. Information loss during the feature extraction and transmission processes can be reduced, the feature expression capability of the model can be significantly improved, and the computational complexity of the model can be effectively reduced, thereby improving the accuracy, detection efficiency, and robustness of the model when performing internal defect detection of carbon fiber honeycomb materials.

[0110] It should be noted that in the attached Figure 2 In the figure, Conv represents the convolution unit, C2F represents the C2F module, StarFFTBlock represents the StarFFTBlock module, SPPF (Spatial Pyramid Pooling Fast) represents the spatial pyramid pooling module, K represents the number of filters or kernels, s represents the stride, specifically the step size of the filter or kernel sliding on the input, p represents padding, specifically the additional zero border added on each side of the input, n represents the number of bottleneck blocks in the C2F module, and d represents the set scaling factor.

[0111] refer to Figure 3 Furthermore, in an optional implementation of the embodiment of the present invention, in order to implement the functions of the StarFFTBlock module defined above, the StarFFTBlock module includes:

[0112] The first depthwise separable convolution unit has its input connected to the input of the StarFFTBlock module, and its output is connected to the inputs of the eighth and ninth convolution units respectively, for performing depthwise convolution and pointwise convolution operations on the input data;

[0113] The output of the eighth convolution unit is connected to the input of the ReLU6 activation function unit, which is used to perform a convolution operation on the input data;

[0114] The ninth convolution unit has its output connected to the input of the multiplication unit and is used to perform a convolution operation on the input data;

[0115] The ReLU6 activation function unit has its output connected to the input of the multiplication unit, and is used to perform nonlinear transformation processing on the input data using the ReLU6 activation function;

[0116] The multiplication unit, whose output is connected to the input of the tenth convolution unit, is used to perform element-wise multiplication of the two inputs;

[0117] a tenth convolutional unit, whose output is connected to the input of the second depthwise separable convolutional unit, for performing a convolution operation on the input data;

[0118] A second depthwise separable convolution unit, whose output is connected to the input of the addition unit, is used to perform depthwise convolution and pointwise convolution operations on the input data;

[0119] The two-dimensional Fourier transform unit has its input connected to the input of the StarFFTBlock module and its output connected to the input of the virtual-real segmentation unit. It is used to perform Fourier transform processing on the input data to obtain and output frequency domain features.

[0120] The real-virtual segmentation unit has its output connected to the input of the eleventh convolution unit and the twelfth convolution unit, respectively, for separating the real and imaginary parts of the input frequency domain features, and outputting the real and imaginary parts of the frequency domain features to the eleventh convolution unit and the twelfth convolution unit respectively;

[0121] The eleventh convolution unit has its output connected to the input of the concatenation unit and is used to perform a convolution operation on the input data;

[0122] The twelfth convolution unit has its output connected to the input of the splicing unit and is used to perform a convolution operation on the input data;

[0123] The concatenation unit, whose output is connected to the input of the thirteenth convolution unit, is used to concatenate the two inputs to obtain and output fused features;

[0124] a thirteenth convolution unit, whose output is connected to the input of the two-dimensional inverse Fourier transform unit, for performing a convolution operation on the input data;

[0125] A two-dimensional inverse Fourier transform unit, whose output is connected to the input of the addition unit, is used to perform inverse Fourier transform processing on the input data, convert the frequency domain features into spatial domain features and output them;

[0126] The addition unit, whose input is also connected to the input of the StarFFTBlock module, and whose output is connected to the input of the fourteenth convolution unit, is used to add multiple inputs element by element to obtain and output fused features;

[0127] The output of the fourteenth convolution unit is connected to the output of the StarFFTBlock module, and is used to perform a convolution operation on the input fusion features to obtain and output a feature map.

[0128] In the embodiment of the present invention, based on the structure of the StarFFTBlock module defined above, the StarFFTBlock module is divided into two processes: spatial domain processing and frequency domain processing when processing input data. In the spatial domain processing, the initial features are first extracted from the input image by depthwise separable convolution to obtain feature maps X1 and X2, and then the feature map X1 is processed by convolution and Relu6 activation function to obtain feature map X3. At the same time, the features in the feature map X2 are further extracted by convolution to obtain feature map X4, and then the feature map X3 and the feature map X4 are fused by element-by-element multiplication, and the fused feature map is convolved to obtain Spatial domain feature information; in frequency domain processing, the input image is first processed by two-dimensional Fourier transform to generate a frequency domain feature map X5, and then the real and imaginary parts of the frequency domain feature map are processed separately by convolution to double the number of channels of the feature map. Then the convolved feature map is dimensionally summed, and the frequency domain features are further extracted by convolution. Finally, the spatial information of the frequency domain feature map is restored by two-dimensional inverse Fourier transform. After completing the spatial domain processing and frequency domain processing, the feature map obtained by spatial domain processing, the feature map obtained by frequency domain processing, and the original input image of the StarFFTBlock module are fused through residual connection to obtain the final feature map.

[0129] In an embodiment of the present invention, by utilizing the StarFFTBlock module defined above to extract features of the input image, it is possible to retain local features in the spatial domain and effectively integrate global features through frequency domain transformation, thereby significantly improving the feature expression capability of the model.

[0130] It should be noted that in the attached Figure 3 In the figure, DWConv represents the depth-wise separable convolution unit, Conv represents the convolution unit, ReLU6 represents the ReLU6 activation function unit, * represents the multiplication unit, + represents the addition unit, FFT-2D represents the two-dimensional Fourier transform unit, Spilt represents the real-virtual segmentation unit, real represents the real part, imag represents the imaginary part, Concat represents the concatenation unit, IFFT-2D represents the two-dimensional inverse Fourier transform unit, K represents the number of filters or kernels, s represents the stride, specifically represents the step size of the filter or kernel sliding on the input, and p represents padding, specifically represents the extra zero border added on each side of the input.

[0131] refer to Figure 2 Furthermore, in an optional implementation of the embodiment of the present invention, in order to realize the functions of the neck network defined above, the neck network includes:

[0132] The first upsampling unit, the first ACGAFusion module, the third C2F module, the second upsampling unit, the second ACGAFusion module, the fourth C2F module, the sixth convolution unit, the first splicing unit, the fifth C2F module, the seventh convolution unit, the second splicing unit and the sixth C2F module are connected in sequence, the input of the first upsampling unit is connected to the output of the spatial pyramid pooling module in the backbone network, the input of the second splicing unit is also connected to the output of the spatial pyramid pooling module in the backbone network, the input of the first ACGAFusion module is also connected to the output of the first StarFFTBlock module in the backbone network, the output of the third C2F module is also connected to the input of the first splicing unit, the input of the second ACGAFusion module is also connected to the output of the second C2F module in the backbone network, the output of the fourth C2F module is also connected to the input of the head network, the output of the fifth C2F module is also connected to the input of the head network, and the output of the sixth C2F module is connected to the input of the head network;

[0133] The upsampling unit is used to perform upsampling processing on the input data;

[0134] The ACGAFusion module is used to adaptively generate weight parameters based on the input shallow features and deep features, perform weighted processing on the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features, to obtain and output fused features;

[0135] The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features;

[0136] The convolution unit is used to perform convolution operations on input data, extract and output features;

[0137] The splicing unit is used to perform splicing operations on two inputs to obtain and output fusion features.

[0138] In an embodiment of the present invention, by utilizing the above-defined neck network to extract and fuse the features of the input image, the spatial relationship and contextual dependency of features at different scales can be accurately captured, and a good feature extraction effect can be achieved. At the same time, in the feature fusion processing process, the weights can be adaptively learned based on the attention mechanism, and the fusion method of features at different levels can be dynamically adjusted. This can effectively improve the information mismatch problem caused by differences in receptive fields, promote more efficient fusion between low-level features and high-level features, enhance the expression ability and performance of the model, and improve the accuracy, detection efficiency and robustness of the model when detecting internal defects of carbon fiber honeycomb materials.

[0139] It should be noted that in the attached Figure 2In the figure, Upsample represents the upsampling unit, ACGAFusion represents the ACGAFusion module, and Concat represents the concatenation unit.

[0140] refer to Figure 4 Furthermore, in an optional implementation of the embodiment of the present invention, in order to implement the functions of the ACGAFusion module defined above, the ACGAFusion module includes:

[0141] Shallow feature receiving unit, deep feature receiving unit, ACGA module, first multiplication unit, second multiplication unit, addition unit and fifteenth convolution unit;

[0142] The input of the shallow feature receiving unit is connected to the input of the ACGAFusion module, and the output of the shallow feature receiving unit is connected to the input of the ACGA module and the first multiplication unit respectively. The shallow feature receiving unit is used to receive the shallow features input into the ACGAFusion module and output the shallow features;

[0143] The input of the deep feature receiving unit is connected to the input of the ACGAFusion module, and the output of the deep feature receiving unit is connected to the input of the ACGA module and the second multiplication unit respectively. The deep feature receiving unit is used to receive the deep features input into the ACGAFusion module and output the deep features;

[0144] The output of the ACGA module is connected to the input of the first multiplication unit and the second multiplication unit respectively. The ACGA module is used to adaptively generate weight parameters based on the input shallow features and deep features based on the attention mechanism and output them;

[0145] The output of the first multiplication unit is connected to the input of the addition unit, and the first multiplication unit is used to perform element-by-element multiplication of the input shallow features and weight parameters;

[0146] The output of the second multiplication unit is connected to the input of the addition unit, and the second multiplication unit is used to perform element-by-element multiplication of the input deep features and weight parameters;

[0147] The input of the addition unit is also connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively, and the output of the addition unit is connected to the input of the fifteenth convolution unit. The addition unit is used to perform element-by-element addition on the four inputs;

[0148] The output of the fifteenth convolution unit is connected to the output of the ACGAFusion module. The fifteenth convolution unit is used to perform a convolution operation on the input data, extract and output features.

[0149] In an embodiment of the present invention, based on the ACGAFusion module structure defined above, the ACGAFusion module first generates weighted parameters according to the input shallow features and deep features, multiplies the weighted parameters and their complements (1-weighted parameters) with the shallow features and deep features respectively, and then adds the weighted shallow features and deep features to the original input shallow features and deep features element by element to further enhance the expressive power of the features, and finally performs convolution processing on the added features to generate fused features with richer semantic information.

[0150] In the embodiment of the present invention, by utilizing the ACGAFusion module defined above to fuse shallow features and deep features, shallow features and deep features can be more efficiently fused, thereby improving the expression ability and performance of the model.

[0151] It should be noted that in the attached Figure 4 In the figure, Low-Level features represents shallow feature receiving units, High-Level features represents deep feature receiving units, ACGA represents ACGA module, * represents multiplication unit, + represents addition unit, Conv represents convolution unit, K represents the number of filters or kernels, and s represents stride, which specifically represents the step size of the filter or kernel sliding on the input.

[0152] refer to Figure 5 , further, in an optional implementation of the embodiment of the present invention, in order to realize the functions of the ACGA module defined above, the ACGA module includes: an AFM module, a spatial attention module, a channel attention module, an addition unit, a splicing unit, a sixteenth convolution unit and a Sigmoid activation function unit;

[0153] The input of the AFM module is connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively. The output of the AFM module is connected to the input of the spatial attention module and the channel attention module respectively. The AFM module is used to perform global average pooling on the shallow features and the deep features, concatenate the two pooled features together, and generate a weighting parameter in the range of [0,1] through a fully connected layer and a Sigmoid activation function transformation. The weighting parameter and its complement are respectively multiplied element-by-element with the shallow features and the deep features to obtain and output the fused features;

[0154] The output of the spatial attention module is connected to the input of the addition unit. The spatial attention module is used to extract and fuse features of the input data in the spatial dimension.

[0155] The output of the channel attention module is connected to the input of the addition unit. The channel attention module is used to extract features of the input data in the channel dimension.

[0156] The output of the addition unit is connected to the input of the splicing unit, and the addition unit is used to perform element-by-element addition of the two inputs;

[0157] The input of the splicing unit is also connected to the output of the AFM module, and the output of the splicing unit is connected to the input of the sixteenth convolution unit. The splicing unit is used to perform a splicing operation on the two inputs to obtain and output a fusion feature;

[0158] The output of the sixteenth convolution unit is connected to the input of the Sigmoid activation function unit. The sixteenth convolution unit is used to perform a convolution operation on the input data, extract and output features;

[0159] The output of the Sigmoid activation function unit is connected to the output of the ACGA module. The Sigmoid activation function unit is used to transform the input data using the Sigmoid activation function, generate and output a weighted parameter in the range of [0,1].

[0160] In an embodiment of the present invention, based on the structure of the ACGA module defined above, the ACGA module can adaptively generate weighted parameters based on the attention mechanism according to the shallow features and deep features of the input, thereby realizing dynamic adjustment of the fusion mode of features at different levels.

[0161] refer to Figure 5 ,Further, in an embodiment of the present invention, the spatial attention module includes: a global maximum pooling unit, a global average pooling unit, a splicing unit and a convolution unit;

[0162] The input of the global maximum pooling unit is connected to the input of the spatial attention module, and the output is connected to the input of the splicing unit to perform global maximum pooling on the input data;

[0163] The input of the global average pooling unit is connected to the input of the spatial attention module, and the output is connected to the input of the splicing unit to perform global average pooling on the input data;

[0164] The output of the concatenation unit is connected to the input of the convolution unit to perform a concatenation operation on the two inputs to obtain and output fusion features;

[0165] The output of the convolution unit is connected to the output of the spatial attention module to perform a convolution operation on the input data.

[0166] refer to Figure 5 ,Further, in an embodiment of the present invention, the channel attention module includes: a global average pooling unit and two convolution units connected in sequence;

[0167] The global average pooling unit is used to perform global average pooling on the input data, and the convolution unit is used to perform convolution operation on the input data.

[0168] In an embodiment of the present invention, based on the structure of the spatial attention module and the channel attention module defined above, it is possible to extract and fuse features of the input in the channel and spatial dimensions, so as to enhance the learning of important features in the channel and spatial dimensions and improve the feature expression ability and performance of the model.

[0169] It should be noted that in the attached Figure 5 In the figure, AFM represents AFM module, Spatial attention represents spatial attention module, Channel attention represents channel attention module, GMP represents global maximum pooling unit, GAP represents global average pooling unit, C represents concatenation unit, Conv represents convolution unit, + represents addition unit, Sigmoid represents Sigmoid activation function unit, K represents the number of filters or kernels, and s represents stride, which specifically represents the step size of the filter or kernel sliding on the input.

[0170] Furthermore, in an optional implementation of the embodiment of the present invention, the internal defect detection model is trained in the following manner:

[0171] The internal defect speckle wrapped phase image of the carbon fiber honeycomb material in the training data in the training data set is used as the input of the internal defect detection model, and the defect category and position corresponding to the input internal defect speckle wrapped phase image of the carbon fiber honeycomb material are used as the output to train the internal defect detection model.

[0172] Furthermore, in an embodiment of the present invention, a speckle-wrapped phase image of internal defects of a carbon fiber honeycomb material in training data in a training data set is used as input to an internal defect detection model, and defect categories and positions corresponding to the input speckle-wrapped phase image of internal defects of the carbon fiber honeycomb material are used as output to train the internal defect detection model, including the following steps 301 to 303:

[0173] Step 301: Input the carbon fiber honeycomb material internal defect speckle wrapped phase images in all training data into the internal defect detection model in batches, and obtain the defect category and position corresponding to each carbon fiber honeycomb material internal defect speckle wrapped phase image output by the internal defect detection model;

[0174] Step 302: Calculate a preset loss function based on the defect categories and locations in the training data and the prediction results of the defect categories and locations output by the internal defect detection model;

[0175] Step 303, determine whether the preset training stop condition is met, if so, use the current internal defect detection model as the internal defect detection model to complete the training, if not, use the loss function to update the parameters of the internal defect detection model, and return to step 301.

[0176] In the embodiment of the present invention, the loss function is specifically set according to the actual situation, for example, a cross entropy loss function is used.

[0177] In an embodiment of the present invention, the training stop condition is specifically set according to the actual situation, for example, when the number of training iterations reaches a set number of iterations or when the optimization index reaches a set threshold. Among them, the loss function can be used as the optimization index.

[0178] Furthermore, in an embodiment of the present invention, a stochastic gradient descent method is used to train and update the parameters of the internal defect detection model.

[0179] Furthermore, in an optional implementation of the embodiment of the present invention, the internal defect detection of the carbon fiber honeycomb material is performed using the trained internal defect detection model, including:

[0180] Acquire a speckle wrapped phase image of the carbon fiber honeycomb material to be inspected;

[0181] The speckle wrapped phase image of the carbon fiber honeycomb material to be detected is input into the internal defect detection model to obtain the defect type and position corresponding to the speckle wrapped phase image of the carbon fiber honeycomb material to be detected output by the internal defect detection model.

[0182] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for detecting internal defects of carbon fiber honeycomb materials, characterized in that: include: Obtain a training data set, the training data including a speckle package phase image of internal defects of a carbon fiber honeycomb material and its corresponding defect category and location; Constructing an internal defect detection model, wherein the internal defect detection model is obtained by replacing some C2F modules in the YOLOv8 model backbone network with StarFFTBlock modules, and replacing some Concat modules in the YOLOv8 model neck network with ACGAFusion modules, the StarFFTBlock module is used to extract spatial domain features and frequency domain features of input data, and fuse the input data and the spatial domain features and frequency domain features of the input data to obtain and output a feature map, the ACGAFusion module uses an attention mechanism to adaptively generate weight parameters according to the input shallow features and deep features, weight the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features, to obtain and output a fused feature; training the internal defect detection model using the training data set; Performing internal defect detection on carbon fiber honeycomb materials using the trained internal defect detection model; The StarFFTBlock module is divided into two processes when processing input data: spatial domain processing and frequency domain processing. In spatial domain processing, the initial features are first extracted from the input image by depthwise separable convolution to obtain feature maps X1 and X2, and then the feature map X1 is processed by convolution and Relu6 activation function to obtain feature map X3. At the same time, the features in the feature map X2 are further extracted by convolution to obtain feature map X4, and then the feature map X3 is fused with the feature map X4 by element-by-element multiplication, and the fused feature map is convolved to obtain spatial domain feature information; in frequency domain processing, the initial features are first extracted from the input image by depthwise separable convolution to obtain feature map X1 and X2, and then the feature map X3 is processed by depthwise separable convolution to obtain feature map X3. First, the input image is processed by two-dimensional Fourier transform to generate a frequency domain feature map X5. Then, the real and imaginary parts of the frequency domain feature map are processed separately by convolution to double the number of channels of the feature map. Then, the convolved feature map is dimensionally summed and the frequency domain features are further extracted by convolution. Finally, the spatial information of the frequency domain feature map is restored by two-dimensional inverse Fourier transform. After completing the spatial domain processing and frequency domain processing, the feature map obtained by the spatial domain processing, the feature map obtained by the frequency domain processing, and the original input image of the StarFFTBlock module are fused by residual connection to obtain the final feature map. The ACGAFusion module includes: a shallow feature receiving unit, a deep feature receiving unit, an ACGA module, a first multiplication unit, a second multiplication unit, a second addition unit and a fifteenth convolution unit; The ACGA module includes: an AFM module, a spatial attention module, a channel attention module, a third addition unit, a third splicing unit, a sixteenth convolution unit and a Sigmoid activation function unit.

2. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 1, characterized in that: Get the training dataset by: An experimental platform was built using laser shearing speckle interferometer; In the experimental platform, negative pressure is applied to the test areas of multiple pre-prepared carbon fiber honeycomb materials in sequence, and a CCD camera is used to collect the speckle wrapping phase image; The defect category and position in each collected speckle wrapped phase image are marked to obtain a plurality of training data including speckle wrapped phase images of internal defects of carbon fiber honeycomb materials and their corresponding defect categories and positions.

3. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 1, characterized in that: The internal defect detection model includes: a backbone network, a neck network and a head network connected in sequence; The backbone network is used to extract features of the input image and output them to the neck network. The neck network is used to fuse feature maps of different scales input by the backbone network and output them to the head network. The head network is used to determine and output the defect category and location in the input image based on the feature maps input by the neck network.

4. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 3, characterized in that: The backbone network includes: a first convolution unit, a second convolution unit, a first C2F module, a third convolution unit, a second C2F module, a fourth convolution unit, a first StarFFTBlock module, a fifth convolution unit, a second StarFFTBlock module and a spatial pyramid pooling module connected in sequence, wherein the output of the second C2F module is also connected to the neck network, the output of the first StarFFTBlock module is also connected to the neck network, and the output of the spatial pyramid pooling module is connected to the neck network; The convolution unit is used to perform convolution operations on input data, extract and output features; The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features; The StarFFTBlock module is used to extract the spatial domain features and frequency domain features of the input data, and fuse the input data, the spatial domain features and frequency domain features of the input data to obtain and output the feature map; The spatial pyramid pooling module is used to fuse features at different scales of the same feature map through a connection operation and output a fusion result.

5. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 4, characterized in that: The StarFFTBlock module includes: A first depthwise separable convolution unit, whose input is connected to the input of the StarFFTBlock module, and whose output is connected to the input of the eighth and ninth convolution units respectively, is used to perform depthwise convolution and pointwise convolution operations on the input data; The output of the eighth convolution unit is connected to the input of the ReLU6 activation function unit, and is used to perform a convolution operation on the input data; The ninth convolution unit has an output connected to the input of the third multiplication unit, and is used to perform a convolution operation on the input data; The ReLU6 activation function unit has an output connected to the input of the third multiplication unit, and is used to perform nonlinear transformation processing on the input data using the ReLU6 activation function; The output of the third multiplication unit is connected to the input of the tenth convolution unit, and is used to perform element-by-element multiplication on the two inputs; The tenth convolution unit has an output connected to the input of the second depth-wise separable convolution unit, and is configured to perform a convolution operation on the input data; The second depthwise separable convolution unit has an output connected to the input of the first addition unit, and is configured to perform depthwise convolution and pointwise convolution operations on the input data; A two-dimensional Fourier transform unit, whose input is connected to the input of the StarFFTBlock module and whose output is connected to the input of the virtual-real segmentation unit, is used to perform Fourier transform processing on the input data to obtain and output frequency domain features; The output of the virtual-real segmentation unit is connected to the input of the eleventh convolution unit and the twelfth convolution unit respectively, and is used to separate the real part and the imaginary part of the input frequency domain feature, and output the real part and the imaginary part of the frequency domain feature to the eleventh convolution unit and the twelfth convolution unit respectively; The eleventh convolution unit has an output connected to the input of the fourth splicing unit, and is used to perform a convolution operation on the input data; The output of the twelfth convolution unit is connected to the input of the fourth splicing unit, and is used to perform a convolution operation on the input data; The output of the fourth concatenation unit is connected to the input of the thirteenth convolution unit, and is used to perform a concatenation operation on the two inputs to obtain and output a fusion feature; The thirteenth convolution unit has an output connected to the input of the two-dimensional inverse Fourier transform unit, and is used to perform a convolution operation on the input data; The two-dimensional inverse Fourier transform unit has an output connected to the input of the first adding unit, and is used to perform inverse Fourier transform processing on the input data, convert the frequency domain features into spatial domain features and output them; The first addition unit has an input connected to the input of the StarFFTBlock module, and an output connected to the input of the fourteenth convolution unit, and is used to perform element-by-element addition on multiple inputs to obtain and output fused features; The output of the fourteenth convolution unit is connected to the output of the StarFFTBlock module, and is used to perform a convolution operation on the input fusion features to obtain and output a feature map.

6. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 4 or 5, characterized in that: The neck network includes: a first upsampling unit, a first ACGAFusion module, a third C2F module, a second upsampling unit, a second ACGAFusion module, a fourth C2F module, a sixth convolution unit, a first splicing unit, a fifth C2F module, a seventh convolution unit, a second splicing unit and a sixth C2F module connected in sequence, wherein the input of the first upsampling unit is connected to the output of the spatial pyramid pooling module in the backbone network, the input of the second splicing unit is also connected to the output of the spatial pyramid pooling module in the backbone network, the input of the first ACGAFusion module is also connected to the output of the first StarFFTBlock module in the backbone network, the output of the third C2F module is also connected to the input of the first splicing unit, the input of the second ACGAFusion module is also connected to the output of the second C2F module in the backbone network, the output of the fourth C2F module is also connected to the input of the head network, the output of the fifth C2F module is also connected to the input of the head network, and the output of the sixth C2F module is connected to the input of the head network; The upsampling unit is used to perform upsampling processing on the input data; The ACGAFusion module is used to adaptively generate weight parameters based on the input shallow features and deep features, perform weighted processing on the shallow features and deep features according to the weight parameters, and fuse the input shallow features and deep features, as well as the weighted shallow features and deep features, to obtain and output fused features; The C2F module is used to obtain and fuse features of different scales based on the input data, and obtain and output the fused features; The convolution unit is used to perform convolution operations on input data, extract and output features; The first splicing unit and the second splicing unit are both used to perform a splicing operation on two inputs to obtain and output fusion features.

7. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 6, characterized in that: The input of the shallow feature receiving unit is connected to the input of the ACGAFusion module, and the output of the shallow feature receiving unit is connected to the input of the ACGA module and the first multiplication unit respectively. The shallow feature receiving unit is used to receive the shallow features input into the ACGAFusion module and output the shallow features; The input of the deep feature receiving unit is connected to the input of the ACGAFusion module, and the output of the deep feature receiving unit is connected to the input of the ACGA module and the second multiplication unit respectively. The deep feature receiving unit is used to receive the deep features input into the ACGAFusion module and output the deep features; The output of the ACGA module is connected to the input of the first multiplication unit and the second multiplication unit respectively, and the ACGA module is used to adaptively generate and output weight parameters according to the shallow features and deep features of the input based on the attention mechanism; The output of the first multiplication unit is connected to the input of the second addition unit, and the first multiplication unit is used to perform element-by-element multiplication on the input shallow features and weight parameters; The output of the second multiplication unit is connected to the input of the second addition unit, and the second multiplication unit is used to perform element-by-element multiplication on the input deep features and weight parameters; The input of the second adding unit is further connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively, the output of the second adding unit is connected to the input of the fifteenth convolution unit, and the second adding unit is used to perform element-by-element addition on the four inputs; The output of the fifteenth convolution unit is connected to the output of the ACGAFusion module, and the fifteenth convolution unit is used to perform a convolution operation on the input data to extract and output features.

8. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 7, characterized in that: The input of the AFM module is connected to the output of the shallow feature receiving unit and the deep feature receiving unit respectively, and the output of the AFM module is connected to the input of the spatial attention module and the channel attention module respectively. The AFM module is used to perform global average pooling processing on the shallow features and the deep features, splice the two features after pooling together, and generate a weighting parameter in the range of [0,1] through a fully connected layer and a Sigmoid activation function transformation. The weighting parameter and its complement are respectively multiplied element-by-element with the shallow features and the deep features to obtain and output a fusion feature; The output of the spatial attention module is connected to the input of the third adding unit, and the spatial attention module is used to extract and fuse features of the input data in the spatial dimension; The output of the channel attention module is connected to the input of the third adding unit, and the channel attention module is used to extract features of the input data in the channel dimension; The output of the third adding unit is connected to the input of the third splicing unit, and the third adding unit is used to perform element-by-element addition on two inputs; The input of the third splicing unit is also connected to the output of the AFM module, and the output of the third splicing unit is connected to the input of the sixteenth convolution unit. The third splicing unit is used to perform a splicing operation on the two inputs to obtain and output a fusion feature; The output of the sixteenth convolution unit is connected to the input of the Sigmoid activation function unit, and the sixteenth convolution unit is used to perform a convolution operation on the input data to extract and output features; The output of the Sigmoid activation function unit is connected to the output of the ACGA module. The Sigmoid activation function unit is used to transform the input data using the Sigmoid activation function to generate and output a weighting parameter in the range of [0, 1].

9. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 1, characterized in that: The internal defect detection model is trained by: The internal defect detection model is trained by using the speckle wrapped phase image of the internal defects of the carbon fiber honeycomb material in the training data set as the input of the internal defect detection model and using the defect category and position corresponding to the input speckle wrapped phase image of the internal defects of the carbon fiber honeycomb material as the output.

10. The method for detecting internal defects of carbon fiber honeycomb materials according to claim 1, characterized in that: The method of using the trained internal defect detection model to detect internal defects of carbon fiber honeycomb materials includes: Acquire a speckle wrapped phase image of the carbon fiber honeycomb material to be inspected; The speckle wrapped phase image of the carbon fiber honeycomb material to be detected is input into the internal defect detection model to obtain the defect type and position corresponding to the speckle wrapped phase image of the carbon fiber honeycomb material to be detected output by the internal defect detection model.

Images