Composite material defect detection method and system fusing infrared imaging and deep learning

Abstract

The invention discloses a composite material defect detection method and system fusing infrared imaging and deep learning, and the method comprises the steps: preparing a defect-containing composite material plate, obtaining an original infrared thermal image data set of the defect-containing composite material plate, carrying out the preprocessing, training the defect-containing data set through the processed data set, and carrying out the recognition of the defect-containing data set. Obtaining a neural network containing defect characteristics of the composite material; and establishing a multi-index evaluation system, obtaining a defect contour, and marking a defect type and a defect area. By combining infrared nondestructive testing with an artificial intelligence method, batch detection of internal defects of the composite material is realized, errors caused by manual detection are reduced, and the accuracy and efficiency of damage detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of material internal defect detection, and specifically to a composite material defect detection method and system that integrates infrared imaging and deep learning. Background Art

[0002] Carbon fiber is widely used in various fields due to its excellent mechanical properties and structural strength. However, composite structures will inevitably produce some defects during the preparation and service process. These defects may directly or indirectly affect the mechanical properties of the composite materials. In severe cases, they may cause structural failure and trigger major safety accidents. Infrared thermal imaging technology has the advantages of non-contact, visualization, and a wide detection range. It is widely used in non-destructive testing of composite materials. Depending on whether there is an external excitation source, infrared thermal imaging technology can be divided into two types: one is passive infrared thermal imaging technology, which uses the infrared characteristics of the object itself for direct imaging and detection; the other is active infrared thermal imaging technology, which uses an external excitation source to realize non-destructive detection of defects through infrared image acquisition, image processing and defect information extraction through the difference in heat dissipation between the defective area and the main material. Therefore, the core purpose of infrared non-destructive testing is image recognition and defect quantification. Based on the previous non-destructive testing of composite materials containing defective composite materials, the following problems still exist:

[0003] (I) Due to the characteristics of infrared thermal imaging, it is more susceptible to environmental influences and is easily disturbed by noise. The edge information of infrared images is blurred and difficult to quantify accurately. (II) Manual recognition has large differences and is easily affected by subjective factors of the inspectors. Only a small amount of defect classification and calculation can be performed, making it difficult to accurately and quickly quantitatively detect composite material damage. (III) Artificial intelligence algorithms can quickly realize damage identification and segmentation, but currently in the field of non-destructive testing, there is still a lack of batch and accurate calculation of composite material defect areas, and the accuracy of irregular shaped defect detection needs to be further improved. Summary of the Invention

[0004] The purpose of the present invention is to provide a composite material defect detection method and system that integrates infrared imaging and deep learning, so as to overcome the shortcomings of the existing technology that are easily interfered by noise, the edge information of the infrared image is inaccurate, and the manual recognition accuracy and efficiency are low.

[0005] In order to achieve the above objectives, the present invention provides and technical solutions are implemented:

[0006] In a first aspect, the present invention provides a composite material defect detection method integrating infrared imaging and deep learning, comprising:

[0007] S1, preparing a composite material plate containing defects, obtaining an original infrared thermal image sequence dataset of the composite material plate containing defects, and performing preprocessing to obtain a processed infrared thermal image sequence dataset;

[0008] S2, building a defect detection neural network, which is trained on an infrared thermal image sequence data set to obtain a neural network capable of detecting composite material defects;

[0009] S3, inputting the image to be inspected into the neural network capable of detecting composite material defects, and making predictions on the image to be inspected based on the neural network capable of detecting composite material defects, traversing each prediction result, and performing defect instance segmentation on the image to be inspected to obtain a defect image;

[0010] S4, expressing the pixel coordinates of the closed polygon outline of the defect image segmentation area in a matrix form to obtain a closed polygon in the matrix form;

[0011] S5, divide the closed polygon in matrix form into n triangles, calculate the area of each triangle by vector cross product, accumulate the areas of all triangles and take the absolute value to obtain the original area of the defect area;

[0012] S6, calculating the number of pixels in the defect outline and determining an area coefficient, multiplying the area coefficient by the original area to obtain the actual defect area, and obtaining a detected image;

[0013] S7, draw the defect outline on the detected image, mark the defect type and defect area, and output a visual image.

[0014] Furthermore, S1 specifically includes: the defects of the defective composite material plate include regular geometric defects, irregular geometric defects and composite defects.

[0015] Furthermore, in S1, an original infrared thermal image sequence is collected by an infrared imaging detection system to construct an original infrared thermal image sequence dataset, and the original infrared thermal image sequence dataset is expanded and preprocessed. The infrared imaging detection system includes a thermal excitation source, an infrared thermal imager, and an infrared acquisition system. The process of collecting the original infrared thermal image sequence dataset by the infrared imaging detection system includes:

[0016] The defective composite material plate is non-contact loaded using a thermal excitation source. The infrared thermal imager captures the surface temperature evolution of the defective sample, and the infrared acquisition system obtains the original infrared thermal image sequence data set of the defective composite material plate.

[0017] The preprocessing process includes: removing the background of the original infrared thermal image sequence data set and using non-local mean denoising to suppress noise interference and reduce pseudo defects to obtain the preprocessed infrared thermal image sequence data set.

[0018] Furthermore, the expansion of the original infrared thermal image sequence data set in S1 includes rotation, flipping and mirroring.

[0019] Furthermore, the S2 specifically includes:

[0020] The preprocessed infrared thermal image sequence dataset includes training set, validation set and test set;

[0021] Labelme software is used to mark the defect edge positions in the infrared thermal image sequence data set and classify them according to the defect type, generating corresponding defect feature information for each independent defect in the infrared thermal image;

[0022] By integrating infrared thermal images and annotated defect feature information, a defect dataset is constructed, wherein the defect dataset includes a defect training dataset and a defect verification dataset, and a defect detection neural network is trained using the defect training dataset;

[0023] After verification using the defect verification dataset, the test set was used for testing to obtain a neural network that can detect defects in composite materials.

[0024] Furthermore, the neural network capable of detecting composite material defects includes a Backbone layer, a Neck layer, and a Head layer;

[0025] The Backbone layer consists of five convolution modules, four C2f modules, and one SPPF module. The convolution modules extract defect features from the input images with different features. The C2f modules process and fuse the extracted defect features. The SPPF modules perform pooling on the defect features, integrating defect features of different scales and outputting a fixed-size feature map.

[0026] The Neck layer fuses and enhances the fixed-size feature map by using the multi-scale fusion of the feature pyramid network and the path aggregation network to obtain the defect feature information and obtain the infrared feature map;

[0027] The head layer processes the infrared feature map through the convolution layer and introduces a nonlinear activation function to improve the expression ability of the infrared feature map. It extracts the infrared feature map bounding box, processes each bounding box and segmentation mask, obtains the defect contour, calculates the contour area, draws the mask contour on the image, and outputs a visual image.

[0028] Furthermore, the S4 is specifically:

[0029] The vertex coordinates of the closed polygon of the defect segmentation area are arranged in counterclockwise order as follows:

[0030] p i =(x i,y i )(i=1,2,3,...,n);

[0031] Among them, p i is the vertex coordinate, x i ,y i is the pixel coordinate;

[0032] Define the vertex coordinates p n+1 =p1; that is, x n+1 =x1 and y n+1 =y1 The coordinate set of the contour points can be expressed in matrix form:

[0033] Furthermore, the S6 is specifically as follows:

[0034] Place the known real defect specimen in the infrared imaging detection system, the defect area A phys The image resolution is ρ x ×ρ y ;

[0035] The physical sizes corresponding to horizontal and vertical pixels are and

[0036] Then the area coefficient λ=δ x ×δ y , then the actual area after adjustment is A phys =λ×A px ;

[0037] Among them, A px is the original area.

[0038] Furthermore, the ratio of the training set, the validation set, and the test set in the infrared thermal image sequence dataset is 8:1:1.

[0039] In a second aspect, the present invention provides a composite material defect detection system that integrates infrared imaging and deep learning, comprising:

[0040] Acquisition module: prepares a composite material plate containing defects, obtains the original infrared thermal image sequence data set of the composite material plate containing defects, and performs preprocessing to obtain a processed image data set;

[0041] Training module: Build a defect detection neural network. The defect detection neural network obtains a neural network that can detect composite material defects through the training data set;

[0042] Instance segmentation module: Input the image to be inspected into the neural network capable of detecting composite material defects, make predictions on the image to be inspected based on the neural network capable of detecting composite material defects, traverse each prediction result, and perform defect instance segmentation on the image to be inspected to obtain a defect image;

[0043] Conversion module: represents the pixel coordinates of the closed polygon outline of the image segmentation area to be detected in matrix form to obtain a closed polygon in matrix form;

[0044] Area calculation module: Divide the closed polygon in matrix form into n triangles. The area of each triangle is calculated by vector cross product. The areas of all triangles are accumulated and the absolute value is taken to obtain the original area of the defect area. The area coefficient is determined and multiplied by the original area to obtain the actual defect area.

[0045] Output module: Draw defect outlines on the detected image, mark the defect type and defect area, and output a visual image.

[0046] Compared with the prior art, the present invention has the following beneficial technical effects:

[0047] The present invention provides a composite material defect detection method that integrates infrared imaging and deep learning, which can detect internal defects of composite materials in batches, reduce errors caused by manual detection and environmental noise interference, and improve the accuracy and efficiency of damage detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of a composite material defect detection method that integrates infrared imaging and deep learning in an embodiment of the present invention.

[0049] Figure 2 Schematic diagram of the defect detection neural network structure in an embodiment of the present invention.

[0050] Figure 3 Schematic diagram of defect instance segmentation in an embodiment of the present invention.

[0051] Figure 4 Schematic diagram of output defect area in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 should fall within the scope of protection of the present invention.

[0053] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0054] See Figure 1 The present invention proposes a composite material defect detection method that integrates infrared imaging and deep learning, which specifically includes the following steps:

[0055] S1, preparing a composite material plate containing defects, obtaining an original infrared thermal image sequence dataset of the composite material plate containing defects, and preprocessing the original infrared thermal image sequence dataset to obtain a processed infrared thermal image sequence dataset, wherein the processed infrared thermal image sequence dataset includes a training set, a validation set, and a test set;

[0056] In order to fully simulate the different types of defects and damage modes of composite plates, a defect dataset of different shapes and sizes was created. The defect dataset includes regular geometric defects, irregular geometric defects, and composite defects, including rectangles, triangles, trapezoids, diamonds, circles, ovals, butterflies, hearts, I-shapes, C-shapes, H-shapes, and U-shapes. The above twelve different types of defects were prefabricated using a 3D printer.

[0057] An infrared imaging detection system is used to capture raw infrared thermal image sequences and construct a raw infrared thermal image sequence dataset. The infrared imaging detection system consists of a thermal excitation source, an infrared thermal imager, an infrared acquisition system, and an image processing unit. The thermal excitation source can be a halogen lamp, laser, flash lamp, or hot air. Non-contact loading is performed on a composite material plate containing defects. The infrared thermal imager captures the thermal field evolution of the defect surface, generating a raw infrared thermal image sequence dataset. This non-contact loading method avoids physical damage to the composite plate while simultaneously utilizing temperature differences to detect defects. Preprocessing includes background removal and noise reduction. Background removal highlights defect areas, while noise reduction improves image quality and reduces noise interference with subsequent analysis, resulting in a clearer infrared thermal image dataset. In this example, non-local means denoising is used to suppress noise and batch process the infrared thermal image sequence.

[0058] The infrared thermal image sequence data set obtained after the preprocessing is expanded, and the expansion of the infrared thermal image sequence data set includes rotating, flipping and mirroring the infrared thermal image sequence data set.

[0059] S2, building a defect detection neural network, which is trained on an infrared thermal image sequence data set to obtain a neural network capable of detecting composite material defects;

[0060] Specifically, Labelme software is used to annotate the defect edge positions in the infrared thermal image sequence dataset and classify them according to different defect types, such as regular geometric defects, irregular geometric defects, or composite defects;

[0061] After labeling is completed, Labelme software generates corresponding defect feature information for each independent defect in the infrared thermal image;

[0062] By integrating infrared thermal images and annotated defect feature information, a defect dataset is constructed, wherein the defect dataset includes a defect training dataset and a defect verification dataset, and a defect detection neural network is trained using the defect training dataset;

[0063] After verification by the defect verification data set, the test set is used for testing to obtain a neural network that can detect composite material defects; optionally, the defect detection neural network selects a YOLOv8 deep learning detection network.

[0064] S3, inputting the image to be inspected into the neural network capable of detecting composite material defects, and making predictions on the image to be inspected based on the neural network capable of detecting composite material defects, traversing each prediction result, and performing defect instance segmentation on the image to be inspected to obtain a defect image;

[0065] S4, expressing the pixel coordinates of the closed polygon outline of the defect image segmentation area in a matrix form to obtain a closed polygon in the matrix form;

[0066] S5, divide the closed polygon in matrix form into n triangles, calculate the area of each triangle by vector cross product, accumulate the areas of all triangles and take the absolute value to get the original area of the defect area

[0067] The vertex coordinates of the closed polygon of the defect image segmentation area are arranged in counterclockwise order as follows:

[0068] p i =(x i ,y i )(i=1,2,3,...,n);

[0069] Among them, p i is the vertex coordinate, x i ,y iis the pixel coordinate;

[0070] To ensure the integrity of the contour, define the vertex coordinates p n+1 =p1; that is, x n+1 =x1 and y n+1 =y1 The coordinate set of the contour points can be expressed in matrix form:

[0071] The defect area segmentation algorithm can satisfy any geometric shape and provide a geometric basis for subsequent area calculation.

[0072] S6, the closed polygon is divided into n triangles, the area of each triangle is calculated by the vector cross product, the areas of all triangles are accumulated and the absolute value is taken to obtain the original area of the defect area; the area of the i-th triangle is The original area A of the defect area can be obtained by adding up the areas of all triangles and taking the absolute value. px ,Right now Since the contour vertices are arranged counterclockwise, reflecting the directionality of the area, the absolute value is taken to be the unified total area, which is expressed in pixels. 2 (px 2 ) is used as the unit, reflecting the geometric range of the defect in the image.

[0073] S7, calculating the number of pixels in the defect outline, determining the area coefficient, and multiplying the area coefficient by the original area to obtain the actual defect area;

[0074] Place the known real defect specimen in the infrared imaging detection system, and the original defect area A px The image resolution is ρ x ×ρ y ;

[0075] The physical sizes corresponding to horizontal and vertical pixels are and

[0076] Then the area adjustment coefficient λ=δ x ×δ y , then the actual area after adjustment is A phys =λ×A px ;

[0077] Among them, A phys is the actual area.

[0078] like Figure 2 As shown, the image to be inspected is input into a defect detection neural network containing composite material defect features for inspection. The inspection neural network specifically includes: Backbone layer, Neck layer and Head layer;

[0079] The Backbone layer consists of five convolution modules, four C2f modules and one SPPF module. The convolution module extracts features from defect training dataset images with different characteristics through convolution operations, and gradually extracts feature information from low-level to high-level in the image, such as edges and textures; the C2f module further processes and fuses these features to enhance the expressiveness of the features; the SPPF module performs pooling operations on the extracted features, integrates features of different scales, and outputs feature information of a fixed size for processing by subsequent network layers.

[0080] The Neck layer utilizes a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN) to process different feature maps. The FPN starts with the deep features output by the Backbone layer and upsamples them layer by layer. During this process, the upsampled results of each layer are fused with the feature maps of the corresponding stage in the Backbone layer. The PAN uses a bottom-up approach to fuse the lower-level feature maps processed by the FPN with the higher-level feature maps, resulting in enhanced infrared feature information.

[0081] The Head layer receives the infrared feature information output by the Neck layer and further processes it through a convolutional module. It also introduces nonlinear activation functions such as Sigmoid, Tanh, and ReLU. These functions enhance the network's nonlinear representation capabilities, enabling the neural network to learn more complex feature relationships. After convolution and activation functions, the Head layer extracts bounding box information from the infrared feature map. These bounding boxes locate the defect within the infrared feature map. Simultaneously, the Head layer performs segmentation masking on the area corresponding to each bounding box, extracting the defect outline and calculating its area. The mask outline is then drawn on the image and output as a visual image of the defect.

[0082] After each round of training, the defect detection neural network uses a validation set to evaluate its performance. This validation set is fed into the trained neural network, which then identifies and predicts defects in the images in the validation set. The predictions are then compared with the ground truth annotations in the validation set. The neural network's performance on the defect detection task is measured by calculating a series of evaluation metrics, such as accuracy, recall, and mean average precision.

[0083] The test set is fed into a defect detection neural network with proven performance. The trained neural network then performs instance segmentation on the images in the test set, further accurately segmenting each detected defect at the pixel level to obtain detailed defect contour information, including various feature information such as category, confidence level, and shape characteristics. Furthermore, based on the number of pixels in the defect contour and the conversion relationship between image pixels and actual size, the trained neural network can accurately predict the defect area and output the prediction results as a visual image. The visual image then uses different colors or lines to draw the defect contour, clearly demonstrating the defect's location and shape in the composite material panel. Key information, such as the predicted defect area, is also annotated, allowing operators to intuitively understand the composite material panel's defect status and assess its quality.

[0084] Example:

[0085] Multiple groups of composite material plates containing different types of defects were produced to simulate various defects that may occur in actual applications.

[0086] A thermal excitation source consisting of two 500W halogen lamps was used to perform non-contact loading on the fabricated defective composite plate. The stable output of the halogen lamps ensured uniform heating of the specimen surface, providing stable thermal excitation conditions for subsequent defect detection through temperature changes. The infrared thermal imager in the infrared thermal imaging system captured the temperature changes on the surface of the defective composite plate at a sampling frequency of 25Hz. Throughout the entire process, the infrared thermal imager operated continuously, acquiring a large amount of raw infrared thermal image sequence data. The raw infrared thermal image sequence dataset from the 2s-24s heating phase was selected as input for deep learning and preprocessed. This phase covers the process from the initial heating of the defective composite plate to a relatively stable state and contains rich information about the thermal field containing the defect. After image processing, the raw infrared thermal image sequence dataset from the 2s-24s heating phase was divided into training, validation, and test sets. The training, validation, and test sets in the infrared thermal image sequence dataset had a ratio of 8:1:1. Infrared thermal image sequences spanning 2s to 24s were selected from the training set for annotation. The defect edges were accurately labeled and classified according to defect type (regular geometric defects, irregular geometric defects, or composite defects). After annotation, Labelme software generated the corresponding labeled defect information file, which was then integrated with the corresponding infrared thermal image sequence to construct a defect dataset.

[0087] A defect detection neural network based on YOLOv8 deep learning was constructed. The constructed defect training dataset was input into the YOLOv8 neural network for training. YOLOv8 was used to identify and predict defects in the images in the validation set. The predicted results were then compared with the real-world annotation information in the validation set to obtain a neural network capable of detecting defects in composite materials.

[0088] Use a neural network that can detect composite material defects to predict the image, traverse each prediction result, and perform instance segmentation on the image to be inspected, including extracting defect bounding boxes, segmenting defects, and classifying defects;

[0089] The closed polygon is divided into n triangles. The area of each triangle is calculated by the vector cross product. The areas of all triangles are accumulated and the absolute value is taken to obtain the original area of the defect area.

[0090] For a closed polygon, the polygon is divided into n triangles, and the area of each triangle is calculated by the vector cross product. The area of the i-th triangle is The original area A of the defect area can be obtained by adding up the areas of all triangles and taking the absolute value. px ,Right now Since the contour vertices are arranged counterclockwise, the absolute value is taken to be the uniform total area, which is expressed in pixels. 2 (px 2 ) is used as the unit, reflecting the geometric range of the defect in the image.

[0091] The actual defect area is calculated by multiplying the area coefficient with the original area. The size of the original infrared thermal image sequence dataset is 80mm×80mm×3mm, and the pixel size of the preprocessed infrared thermal image sequence dataset is 640×640. The area coefficient is 0.015625.

[0092] Place the known real defect specimen in the infrared imaging detection system, the defect area A phys The image resolution is ρ x ×ρ y ;

[0093] The physical sizes corresponding to horizontal and vertical pixels are and

[0094] Then the area coefficient λ=δ x ×δ y , then the actual area after adjustment is A phys =λ×A px ;

[0095] Among them, A px is the original area.

[0096] Draw defect outlines on the detected image, mark the defect type and defect area, and output the defect area image.

[0097] like Figure 3 As shown, it is the output defect segmentation image.

[0098] like Figure 4 As shown, this is the output defect area image.

[0099] As shown in Table 1 below, the composite material defect detection system and method that integrates infrared imaging and deep learning provided by the present invention can predict twelve defect types (including irregular shapes) in practical applications, and can efficiently detect the shape and area of composite material defects, with an area detection error of less than 8%.

[0100] Table 1

[0101]

[0102] The present invention also provides a composite material defect detection system that integrates infrared imaging and deep learning, comprising:

[0103] Acquisition module: prefabricates a composite material plate containing defects, obtains the original infrared thermal image data set of the composite material plate containing defects, and preprocesses it to obtain a processed data set;

[0104] Training module: Builds an intelligent defect detection algorithm and trains a neural network using preprocessed data sets to obtain a defect detection neural network containing defect information.

[0105] Instance segmentation module: Input the image to be inspected into the defect detection neural network, detect the image to be inspected based on the defect detection neural network, traverse each detection result, and perform instance segmentation on the image to be inspected;

[0106] Conversion module: represents the pixel coordinates of the closed polygon outline of the image segmentation area to be detected in matrix form to obtain a closed polygon in matrix form;

[0107] Area calculation module: Divide the closed polygon in matrix form into n triangles. The area of each triangle is calculated by vector cross product. The areas of all triangles are accumulated and the absolute value is taken to obtain the original area of the defect area. The area coefficient is determined and multiplied by the original area to obtain the actual defect area.

[0108] Output module: Draw defect outlines on undetected images, mark defect types and defect areas, and output visual images.

[0109] The present invention reduces environmental noise and equipment interference through noise reduction processing, thereby improving the quality of original infrared thermal image data, using deep learning methods to classify and segment infrared thermal images of defective composite materials, processing defect data through extracted features, further obtaining defect features and quantitatively detecting composite material defects to achieve visualization effects. The present invention can reduce environmental noise interference and errors caused by manual detection, and can significantly improve the efficiency and accuracy of defect detection.

[0110] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A composite material defect detection method integrating infrared imaging and deep learning, characterized in that: The following steps are involved: S1, preparing a composite material plate containing defects, obtaining an original infrared thermal image sequence dataset of the composite material plate containing defects, and performing preprocessing to obtain a processed infrared thermal image sequence dataset; S2, constructing a defect detection neural network, preparing a composite material plate containing defects using the defect detection neural network, obtaining an original infrared thermal image sequence dataset of the composite material plate containing defects and preprocessing it to obtain a processed infrared thermal image sequence dataset, and obtaining a neural network capable of detecting composite material defects by training the infrared thermal image sequence dataset; S3, inputting the image to be inspected into the neural network capable of detecting composite material defects, and making predictions on the image to be inspected based on the neural network capable of detecting composite material defects, traversing each prediction result, and performing defect instance segmentation on the image to be inspected to obtain a defect image; S4, expressing the pixel coordinates of the closed polygon outline of the defect image segmentation area in a matrix form to obtain a closed polygon in the matrix form; S5, divide the closed polygon in matrix form into n triangles, calculate the area of each triangle by vector cross product, accumulate the areas of all triangles and take the absolute value to obtain the original area of the defect area; S6, calculating the number of pixels in the defect outline and determining an area coefficient, multiplying the area coefficient by the original area to obtain the actual defect area, and obtaining a detected image; S7, draw the defect outline on the detected image, mark the defect type and defect area, and output a visual image.

2. The composite material defect detection method integrating infrared imaging and deep learning according to claim 1, characterized in that: S1 specifically includes: the defects of the defective composite material plate include regular geometric defects, irregular geometric defects and composite defects.

3. The composite material defect detection method integrating infrared imaging and deep learning according to claim 1, characterized in that: In S1, an original infrared thermal image sequence is collected by an infrared imaging detection system to construct an original infrared thermal image sequence dataset, and the original infrared thermal image sequence dataset is expanded and preprocessed. The infrared imaging detection system includes a thermal excitation source, an infrared thermal imager, and an infrared acquisition system; The process of collecting raw infrared thermal image sequence datasets through the infrared imaging detection system includes: The defective composite material plate is non-contact loaded using a thermal excitation source. The infrared thermal imager captures the surface temperature evolution of the defective sample, and the infrared acquisition system obtains the original infrared thermal image sequence data set of the defective composite material plate. The preprocessing process includes: removing the background of the original infrared thermal image sequence data set and using non-local mean denoising to suppress noise interference and reduce pseudo defects to obtain the preprocessed infrared thermal image sequence data set.

4. The composite material defect detection method integrating infrared imaging and deep learning according to claim 1, characterized in that: The expansion of the original infrared thermal image sequence data set in S1 includes rotation, flipping and mirroring.

5. The composite material defect detection method integrating infrared imaging and deep learning according to claim 2, characterized in that: The S2 specifically includes: The preprocessed infrared thermal image sequence dataset includes training set, validation set and test set; Labelme software is used to mark the defect edge positions in the infrared thermal image sequence data set and classify them according to the defect type, generating corresponding defect feature information for each independent defect in the infrared thermal image; By integrating infrared thermal images and annotated defect feature information, a defect dataset is constructed, wherein the defect dataset includes a defect training dataset and a defect verification dataset, and a defect detection neural network is trained using the defect training dataset; After verification using the defect verification dataset, the test set was used for testing to obtain a neural network that can detect defects in composite materials.

6. The composite material defect detection method integrating infrared imaging and deep learning according to claim 5, characterized in that: The neural network capable of detecting composite material defects includes a Backbone layer, a Neck layer and a Head layer; The Backbone layer consists of five convolution modules, four C2f modules, and one SPPF module. The convolution modules extract defect features from the input images with different features. The C2f modules process and fuse the extracted defect features. The SPPF modules perform pooling on the defect features, integrating defect features of different scales and outputting a fixed-size feature map. The Neck layer fuses and enhances the fixed-size feature map by using the multi-scale fusion of the feature pyramid network and the path aggregation network to obtain the defect feature information and obtain the infrared feature map; The head layer processes the infrared feature map through the convolution layer and introduces a nonlinear activation function to improve the expression ability of the infrared feature map. It extracts the infrared feature map bounding box, processes each bounding box and segmentation mask, obtains the defect contour, calculates the contour area, draws the mask contour on the image, and outputs a visual image.

7. The composite material defect detection method integrating infrared imaging and deep learning according to claim 1, characterized in that: The S4 is specifically: The vertex coordinates of the closed polygons in the defect map segmentation area are arranged in counterclockwise order as follows: i =(x i ,y i )(i=1,2,3,...,n); Among them, p i is the vertex coordinate, x i ,y i is the pixel coordinate; Define the vertex coordinates p n+1 =p1; that is, x n+1 =x1 and y n+1 =y1 The coordinate set of the contour points can be expressed in matrix form:

8. The composite material defect detection method integrating infrared imaging and deep learning according to claim 1, characterized in that: The S6 is specifically: The known actual defect specimen area A phys Placed in the infrared imaging detection system, the image resolution is ρ x ×ρ y ; The physical sizes corresponding to horizontal and vertical pixels are and Then the area adjustment coefficient λ=δ x ×δ y , then the actual area after adjustment is A phys =λ×A px ; Among them, A px is the original area.

9. The composite material defect detection method integrating infrared imaging and deep learning according to claim 5, characterized in that: The ratio of the training set, validation set, and test set in the infrared thermal image sequence dataset is 8:1:

1.

10. A composite material defect detection system integrating infrared imaging and deep learning, characterized in that: include: Acquisition module: prepares a composite material plate containing defects, obtains the original infrared thermal image sequence data set of the composite material plate containing defects, and performs preprocessing to obtain a processed infrared thermal image sequence data set; Training module: Build a defect detection neural network. The defect detection neural network is trained on infrared thermal image sequence data sets to obtain a neural network that can detect composite material defects. Instance segmentation module: Input the image to be inspected into the neural network capable of detecting composite material defects, make predictions on the image to be inspected based on the neural network capable of detecting composite material defects, traverse each prediction result, and perform defect instance segmentation on the image to be inspected to obtain a defect image; Conversion module: represents the pixel coordinates of the closed polygon outline of the image segmentation area to be detected in matrix form to obtain a closed polygon in matrix form; Area calculation module: Divide the closed polygon in matrix form into n triangles. The area of each triangle is calculated by vector cross product. The areas of all triangles are accumulated and the absolute value is taken to obtain the original area of the defect area. The area coefficient is determined and multiplied by the original area to obtain the actual defect area. Output module: Draw defect outlines on the detected image, mark the defect type and defect area, and output a visual image.

Images