A method for aircraft structure crack detection based on convolutional neural network

Abstract

The present application belongs to the field of structural health monitoring, and particularly relates to a method for detecting cracks in aircraft structures based on convolutional neural networks. Step 1: Construct a crack image dataset; Step 2: Train a YOLOv5 network using the crack image dataset to obtain a YOLOv5 model; Step 3: Construct a difference map dataset; Step 4: Obtain an improved YOLOv5 network, and train the improved YOLOv5 network using the difference map dataset to obtain an improved YOLOv5 model; Step 5: Perform an overall test on the image to be detected based on the YOLOv5 model and the improved YOLOv5 model. The present application can achieve efficient and reliable crack damage detection in fatigue testing of full-scale aircraft structures.

Description

Technical Field

[0001] The present application belongs to the field of structural health monitoring, and in particular relates to a method for detecting cracks in aircraft structures based on convolutional neural networks. Background Art

[0002] Accurate and efficient crack detection during full-scale aircraft structural fatigue testing can provide an important scientific basis for improving aircraft structural design, assessing health status, and developing maintenance programs. Current crack damage detection relies primarily on visual inspection and non-destructive testing methods such as eddy current and ultrasonic testing. These methods have a high rate of missed detection, long inspection cycles, and difficulty reaching critical structural locations within confined spaces.

[0003] In recent years, deep learning-based computer vision technology has rapidly developed and has been effectively applied in numerous fields, including smart transportation, facial recognition, and industrial inspection. It can provide a reliable technical approach for automated detection of cracks in aircraft structures. However, during full-aircraft fatigue testing, crack morphologies vary significantly at different initiation locations and propagation stages, making it difficult for the network to extract effective key crack features. Furthermore, the presence of numerous interfering factors on the structural surface, such as scratches and stains that are highly similar to cracks, presents a significant detection challenge for existing deep learning algorithms.

[0004] Therefore, it is desired to have a technical solution to overcome or at least alleviate at least one of the above-mentioned deficiencies of the prior art. Summary of the Invention

[0005] The purpose of this application is to provide an aircraft structure crack detection method based on convolutional neural network to solve at least one problem existing in the prior art.

[0006] The technical solution of this application is:

[0007] A method for detecting cracks in aircraft structures based on a convolutional neural network, comprising:

[0008] Step 1: Build a crack image dataset;

[0009] Step 2: Train the YOLOv5 network using the crack image dataset to obtain a YOLOv5 model;

[0010] Step 3: Construct a difference map dataset;

[0011] Step 4: Obtain an improved YOLOv5 network, and train the improved YOLOv5 network using the difference graph dataset to obtain an improved YOLOv5 model;

[0012] Step 5: Perform an overall test on the image to be detected based on the YOLOv5 model and the improved YOLOv5 model.

[0013] In at least one embodiment of the present application, in step 1, constructing a crack image dataset includes:

[0014] S11. Using an industrial camera to continuously photograph N monitored structures in a fatigue test at fixed points until cracks initiate, obtaining N crack images containing cracks and a crack-free template image corresponding to each crack image;

[0015] S12, using Labelimg software to mark key locations where cracks initiate in the crack image, to obtain N marking files;

[0016] S13: Divide the crack image and the annotation file corresponding to the crack image into a crack image training set and a crack image test set.

[0017] In at least one embodiment of the present application, in S12, the key parts include rivets, hole edges, and chamfers.

[0018] In at least one embodiment of the present application, in S13, the crack image training set and the crack image test set are divided in a ratio of 7:3.

[0019] In at least one embodiment of the present application, in step 2, training the YOLOv5 network using the crack image dataset to obtain a YOLOv5 model includes:

[0020] The crack image training set in the crack image dataset is input into the YOLOv5 network, and network training and parameter optimization are performed until the detection accuracy of the YOLOv5 network on the crack image test set reaches a threshold, thereby obtaining a YOLOv5 model.

[0021] In at least one embodiment of the present application, in step three, constructing a difference map dataset includes:

[0022] S31, inputting N crack images into the YOLOv5 model in sequence to obtain a key part map and key part coordinate information of the crack image;

[0023] S32, performing image registration on the crack image and the corresponding template image, and cutting out a key part map from the template image according to the key part coordinate information to obtain a key part map of the template image;

[0024] S33, performing logarithmic operations and ratio operations on the key part map of the crack image and the key part map of the template image to obtain N key part difference maps;

[0025] S34, using Labelimg software to mark the cracks in the key position difference map to obtain N marking files;

[0026] S35 , dividing the key part difference map and the annotation file corresponding to the key part difference map into a difference map training set and a difference map test set.

[0027] In at least one embodiment of the present application, in S35, the difference map training set and the difference map test set are divided in a ratio of 7:3.

[0028] In at least one embodiment of the present application, in step 4, obtaining the improved YOLOv5 network includes:

[0029] Get the backbone network of the improved YOLOv5 network, specifically:

[0030] S41. Obtain a deformable convolution structure, including:

[0031] S411, after the key part difference map with a dimension of 640×640×3 is subjected to a focusing structure, a first feature map is output, where the dimension of the first feature map is 320×320×32;

[0032] S412: Input the first feature map into the offset convolution layer, output the second feature map, and obtain the offset parameter of the first feature map according to the second feature map, wherein:

[0033] The calculation model of step S412 is:

[0034]

[0035] Among them, F0 represents the index value of each pixel, F n Indicates that the pixel corresponds to the nth pixel index value in the 3×3 receptive field, n∈{1,2,3,4,5,6,7,8,9}, w1 is the convolution operation, ΔF n Represents the offset parameter corresponding to n pixel indices in the receptive field;

[0036] The convolution kernel size of the offset convolution layer is 3×3, and the number is 64. The dimension of the second feature map is 320×320×64. The first feature map includes 320×320×32×2 offset parameters, and each pixel point has one offset parameter in the x-axis and y-axis directions respectively.

[0037] S413: Add the offset parameter to the index value of each pixel in the first feature map, and perform a rounding operation to obtain a set of new pixel index values;

[0038] Obtain the pixel value of the indexed pixel point from the first feature map according to the new pixel index value, and form an offset feature map with a dimension of 320×320×32;

[0039] The offset feature map is input into a convolution layer with a convolution kernel of 3×3 and a stride of 2 to obtain a third feature map with a dimension of 160×160×64; wherein,

[0040] The calculation model of step S413 is:

[0041]

[0042] Among them, F0 represents the index value of each pixel, F n Indicates the nth pixel index value in the 3×3 receptive field corresponding to the pixel point, n∈{1,2,3,4,5,6,7,8,9}, ΔF n represents the offset parameter corresponding to the n pixel index in the receptive field, w2 is the convolution operation, s(F0+F n +ΔF n ) is the offset feature map, f(F0) represents the output feature map of the convolutional layer;

[0043] S414: Input the third feature map into a batch normalization layer and a ReLU activation layer in sequence to obtain a fourth feature map with a dimension of 160×160×64;

[0044] S42. Replace the four convolutional structures in the backbone network of the YOLOv5 network with the deformed convolutional structure to obtain the backbone network of the improved YOLOv5 network.

[0045] In at least one embodiment of the present application, in step 4, training the improved YOLOv5 network using the difference map dataset to obtain an improved YOLOv5 model includes:

[0046] S43. Input the difference map training set in the difference map data set into the improved YOLOv5 network, perform network training and parameter optimization, until the detection accuracy of the improved YOLOv5 network on the difference map test set reaches a threshold, and obtain an improved YOLOv5 model.

[0047] In at least one embodiment of the present application, in step 5, performing an overall test on the image to be detected based on the YOLOv5 model and the improved YOLOv5 model includes:

[0048] S51. Conduct first-level testing:

[0049] Input the image to be detected into the YOLOv5 model to obtain a key part map and key part coordinate information of the image to be detected;

[0050] S52. Difference map generation:

[0051] Perform image registration on the image to be detected and the template image, and cut out the key part map from the template image according to the coordinate information of the key parts to obtain the key part map of the template image;

[0052] Perform logarithmic operations and ratio operations on the key part map of the image to be detected and the key part map of the template image to obtain a key part difference map;

[0053] S53, conduct secondary testing:

[0054] The key part difference map is input into the improved YOLOv5 model. If no crack is detected, the output is "no crack". If a crack is detected, the output crack coordinate information is converted according to the coordinate information of the key part difference map in the image to be detected to obtain the coordinate information of the crack in the image to be detected.

[0055] The invention has at least the following beneficial technical effects:

[0056] The aircraft structure crack detection method based on convolutional neural network in this application, in view of the complex and changeable characteristics of aircraft structure crack morphology, improves the structure of YOLOv5 (You Only Look Once v5) network based on deformable convolution to enhance the network's ability to characterize crack characteristics, thereby having stronger robustness to crack morphological changes; in view of the high similarity between interference factors such as surface scratches and stains and cracks, a two-stage detection and difference map generation strategy is adopted in sequence to reduce interference factors in the image to be detected, thereby reducing the early warning rate of the detection results. This application can realize efficient and reliable crack damage detection in full-scale aircraft structure fatigue testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of a method for detecting cracks in aircraft structures based on a convolutional neural network according to one embodiment of the present application;

[0058] Figure 2 This is a diagram of the improved YOLOv5 network structure of one embodiment of the present application;

[0059] Figure 3 This is a schematic diagram of a deformable convolution structure according to an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the implementation of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below in conjunction with the drawings in the embodiments of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described embodiments are part of the embodiments of this application, not all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be understood as limitations on this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The embodiments of this application are described in detail below in conjunction with the drawings.

[0061] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be understood as limiting the scope of protection of this application.

[0062] The following is combined with Figures 1 to 3 This application is described in further detail.

[0063] This application provides a convolutional neural network-based aircraft structural crack detection method, which includes five aspects: constructing a crack image dataset, training a YOLOv5 network, constructing a difference map dataset, designing and training an improved YOLOv5 network, and testing the overall algorithm based on a two-stage detection and difference map generation strategy. The specific steps are as follows:

[0064] Step 1: Build a crack image dataset;

[0065] Step 2: Train the YOLOv5 network using the crack image dataset to obtain the YOLOv5 model;

[0066] Step 3: Construct a difference map dataset;

[0067] Step 4: Obtain an improved YOLOv5 network, train the improved YOLOv5 network using the difference graph dataset, and obtain an improved YOLOv5 model;

[0068] Step 5: Based on the YOLOv5 model and the improved YOLOv5 model, perform an overall test on the image to be detected.

[0069] The aircraft structure crack detection method based on convolutional neural network of this application, in step 1, constructs a crack image dataset, specifically including:

[0070] S11. Using an industrial camera to continuously photograph N monitored structures in a fatigue test at fixed points until cracks initiate, obtaining N crack images containing cracks and a crack-free template image corresponding to each crack image;

[0071] S12, using Labelimg software to mark key locations where cracks initiate in the crack image, and obtaining N marking files;

[0072] S13. Divide the crack images and the annotation files corresponding to the crack images into a crack image training set and a crack image test set.

[0073] The aircraft structure crack detection method based on convolutional neural networks of the present application first uses an industrial camera to continuously shoot fixed-point images of N monitored structures in component-level fatigue tests and full-scale fatigue tests until cracks initiate, thereby obtaining N images containing cracks, and retaining a corresponding crack-free image for each crack image as a template image of the monitored structure; next, since cracks generally initiate at key locations with stress concentration, such as rivets, hole edges, and chamfers, Labelimg software is used to annotate the positions and categories of the above-mentioned key locations in the crack images, obtaining N annotated files; finally, the captured crack images and their corresponding annotated files are divided into training and test sets according to a 7:3 ratio.

[0074] In a preferred embodiment of the present application, in step 2, the YOLOv5 network is trained using a crack image dataset to obtain a YOLOv5 model, including:

[0075] The crack image training set in the crack image dataset is input into the YOLOv5 network, and network training and parameter optimization are performed until the detection accuracy of the YOLOv5 network on the crack image test set reaches a threshold, and the YOLOv5 model is obtained.

[0076] This convolutional neural network-based aircraft structural crack detection method utilizes an existing YOLOv5 network consisting of four main components: an input terminal, a backbone network, a feature fusion network, and a prediction network. A training set of crack image data is fed into the YOLOv5 network, where network training and parameter optimization are performed until the YOLOv5 network's detection accuracy on the test set reaches a specified threshold, effectively outputting key features and coordinates such as rivets and hole edges in the image to be inspected.

[0077] In a preferred embodiment of the present application, in step 3, constructing a difference map dataset includes:

[0078] S31, inputting N crack images into the YOLOv5 model in sequence to obtain a key part map and key part coordinate information of the crack image;

[0079] S32, performing image registration on the crack image and the corresponding template image, and cutting out a key part map from the template image according to the key part coordinate information to obtain a key part map of the template image;

[0080] S33, performing logarithmic operations and ratio operations on the key part map of the crack image and the key part map of the template image to obtain N key part difference maps;

[0081] S34. Labelimg software is used to mark the cracks in the difference map of key parts to obtain N marking files;

[0082] S35. Divide the key part difference map and the annotation files corresponding to the key part difference map into a difference map training set and a difference map test set.

[0083] In this embodiment, the difference map training set and the difference map test set are also divided in a ratio of 7:3.

[0084] In this application's convolutional neural network-based aircraft structural crack detection method, as in step 2, the YOLOv5 network consists of four main components: the input, the backbone network, the feature fusion network, and the prediction network. The backbone network is composed of a focus structure, a convolutional structure, a cross-stage local structure, and a spatial pyramid pooling structure. To optimize the network's ability to extract crack features, the YOLOv5 backbone network was improved.

[0085] In this embodiment, in step 4, obtaining an improved YOLOv5 network includes:

[0086] Get the backbone network of the improved YOLOv5 network, specifically:

[0087] S41. Obtain a deformable convolution structure, including:

[0088] Take the input difference image 640×640×3 as an example, Figure 3 As shown:

[0089] S411, after the key part difference map with a dimension of 640×640×3 is subjected to the focusing structure, a first feature map with a dimension of 320×320×32 is output; the feature map is used as the input feature map of the deformable convolution structure;

[0090] S412: Input the first feature map into the offset convolution layer, output the second feature map, and obtain the offset parameter of the first feature map according to the second feature map, wherein:

[0091] The calculation model of step S412 is:

[0092]

[0093] Among them, F0 represents the index value of each pixel, F n Indicates that the pixel corresponds to the nth pixel index value in the 3×3 receptive field, n∈{1,2,3,4,5,6,7,8,9}, w1 is the convolution operation, ΔF n Represents the offset parameter corresponding to n pixel indices in the receptive field;

[0094] The convolution kernel size of the offset convolution layer is 3×3, and the number of kernels is 64. The dimension of the second feature map is 320×320×64. The first feature map includes 320×320×32×2 offset parameters. Each pixel has one offset parameter in the x-axis and y-axis directions respectively.

[0095] S413: Add the offset parameter to the index value of each pixel in the first feature map, and perform a rounding operation to obtain a new set of pixel index values;

[0096] Obtain the pixel value of the indexed pixel point from the first feature map according to the new pixel index value, and form an offset feature map with a dimension of 320×320×32;

[0097] The offset feature map is input into the convolution layer with a convolution kernel of 3×3 and a stride of 2 to obtain a third feature map with a dimension of 160×160×64; where,

[0098] The calculation model of step S413 is:

[0099]

[0100] Among them, F0 represents the index value of each pixel, F n Indicates the nth pixel index value in the 3×3 receptive field corresponding to the pixel point, n∈{1,2,3,4,5,6,7,8,9}, ΔF n represents the offset parameter corresponding to the n pixel index in the receptive field, w2 is the convolution operation, s(F0+F n +ΔF n ) is the offset feature map, f(F0) represents the output feature map of the convolutional layer;

[0101] S414, input the third feature map into the batch normalization layer and the ReLU activation layer in sequence to obtain a fourth feature map with a dimension of 160×160×64;

[0102] S42. Replace the four convolutional structures in the backbone network of the YOLOv5 network with deformable convolutional structures to obtain the backbone network of the improved YOLOv5 network. In this embodiment, the first three deformable convolutional structures used are deformable convolutional structure-1, as mentioned above, and the step size of the convolutional layer is 2; the fourth deformable convolutional structure used is deformable convolutional structure-2, and the step size of the convolutional layer is 1 to avoid excessive downsampling leading to loss of crack information. The improved YOLOv5 network structure is as follows: Figure 2 shown.

[0103] Finally, the improved YOLOv5 network is trained using the difference graph dataset to obtain the improved YOLOv5 model, including:

[0104] S43. Input the difference map training set in the difference map dataset into the improved YOLOv5 network, perform network training and parameter optimization, until the detection accuracy of the improved YOLOv5 network on the difference map test set reaches a threshold, and obtain an improved YOLOv5 model.

[0105] In the aircraft structural crack detection method based on a convolutional neural network of this application, in step 5, a comprehensive test is performed on the image to be detected based on the YOLOv5 model and the improved YOLOv5 model, including:

[0106] S51. Conduct first-level testing:

[0107] Input the image to be detected into the YOLOv5 model to obtain the key part map and key part coordinate information of the image to be detected;

[0108] S52. Difference map generation:

[0109] Perform image registration on the image to be detected and the template image, and cut out the key part map from the template image according to the coordinate information of the key parts to obtain the key part map of the template image;

[0110] Perform logarithmic operations and ratio operations on the key part map of the image to be detected and the key part map of the template image to obtain a key part difference map;

[0111] S53, conduct secondary testing:

[0112] The key part difference map is input into the improved YOLOv5 model. If no crack is detected, the output is "no crack". If a crack is detected, the output crack coordinate information is converted according to the coordinate information of the key part difference map in the image to be detected to obtain the coordinate information of the crack in the image to be detected.

[0113] The aircraft structure crack detection method based on convolutional neural network in this application improves the backbone network of YOLOv5 (You Only Look Once v5) through a deformable convolution structure, so that the network's receptive field can be dynamically adjusted according to the appearance characteristics of the crack being detected, thereby enhancing the network's ability to characterize crack characteristics, making it more robust to the size and morphological changes of aircraft structure cracks, and effectively improving the accuracy of crack detection. Through the first-level detection, the key areas in the image where cracks may initiate are obtained, and then the second-level detection is performed based on the key areas for crack detection, which can significantly narrow the detection range and improve the detection efficiency. Interference factors such as surface scratches and stains are originally located on the surface of the structure, that is, they exist in both the image to be detected and the template image, while cracks initiate during the fatigue test and may only exist in the image to be detected. Through the difference map generation strategy, the same information of the detection image and the template image, that is, interference factors such as scratches and stains that are highly similar to the crack, can be removed, and only the difference information between the two, that is, the crack information, can be retained, thereby reducing the warning rate of the detection result.

[0114] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting cracks in aircraft structures based on convolutional neural networks, characterized in that: include: Step 1: Build a crack image dataset; Step 2: Train the YOLOv5 network using the crack image dataset to obtain a YOLOv5 model; Step 3: Construct a difference map dataset; Step 4: Obtain an improved YOLOv5 network, and train the improved YOLOv5 network using the difference graph dataset to obtain an improved YOLOv5 model; Step 5: Perform an overall test on the image to be detected based on the YOLOv5 model and the improved YOLOv5 model; In step 4, obtaining the improved YOLOv5 network includes: Get the backbone network of the improved YOLOv5 network, specifically: S41. Obtain a deformable convolution structure, including: S411, after the key part difference map with a dimension of 640×640×3 is subjected to a focusing structure, a first feature map is output, where the dimension of the first feature map is 320×320×32; S412: Input the first feature map into the offset convolution layer, output the second feature map, and obtain the offset parameter of the first feature map according to the second feature map, wherein: The calculation model of step S412 is: Among them, F0 represents the index value of each pixel, F n Indicates that the pixel point corresponds to the nth pixel index value in the 3×3 receptive field, n∈{1,2,3,4,5,6,7,8,9}, w1 is the convolution operation, s(F0+F n ) is the first characteristic diagram, ΔF n Represents the offset parameter corresponding to n pixel indices in the receptive field; The convolution kernel size of the offset convolution layer is 3×3, and the number is 64. The dimension of the second feature map is 320×320×64. The first feature map includes 320×320×32×2 offset parameters, and each pixel point has one offset parameter in the x-axis and y-axis directions respectively. S413: Add the offset parameter to the index value of each pixel in the first feature map, and perform a rounding operation to obtain a set of new pixel index values; Obtain the pixel value of the indexed pixel point from the first feature map according to the new pixel index value, and form an offset feature map with a dimension of 320×320×32; The offset feature map is input into a convolution layer with a convolution kernel of 3×3 and a stride of 2 to obtain a third feature map with a dimension of 160×160×64; wherein, The calculation model of step S413 is: Among them, F0 represents the index value of each pixel, F n Indicates the nth pixel index value in the 3×3 receptive field corresponding to the pixel point, n∈{1,2,3,4,5,6,7,8,9}, ΔF n represents the offset parameter corresponding to the n pixel index in the receptive field, w2 is the convolution operation, s(F0+F n +ΔF n ) is the offset feature map, f(F0) represents the output feature map of the convolutional layer; S414: Input the third feature map into a batch normalization layer and a ReLU activation layer in sequence to obtain a fourth feature map with a dimension of 160×160×64; S42. Replace the four convolutional structures in the backbone network of the YOLOv5 network with the deformed convolutional structure to obtain the backbone network of the improved YOLOv5 network.

2. The aircraft structure crack detection method based on convolutional neural network according to claim 1 is characterized in that: In step 1, constructing a crack image dataset includes: S11. Using an industrial camera to continuously photograph N monitored structures in a fatigue test at fixed points until cracks initiate, obtaining N crack images containing cracks and a crack-free template image corresponding to each crack image; S12, using Labelimg software to mark key locations where cracks initiate in the crack image, to obtain N marking files; S13: Divide the crack image and the annotation file corresponding to the crack image into a crack image training set and a crack image test set.

3. The aircraft structure crack detection method based on convolutional neural network according to claim 2 is characterized in that: In S12, the key parts include rivets, hole edges and chamfers.

4. The aircraft structure crack detection method based on convolutional neural network according to claim 2, characterized in that: In S13, the crack image training set and the crack image test set are divided in a ratio of 7:

3.

5. The aircraft structure crack detection method based on convolutional neural network according to claim 2, characterized in that: In step 2, the YOLOv5 network is trained using the crack image dataset to obtain a YOLOv5 model, including: The crack image training set in the crack image dataset is input into the YOLOv5 network, and network training and parameter optimization are performed until the detection accuracy of the YOLOv5 network on the crack image test set reaches a threshold, thereby obtaining a YOLOv5 model.

6. The aircraft structure crack detection method based on convolutional neural network according to claim 5, characterized in that: In step 3, constructing a difference map dataset includes: S31, inputting N crack images into the YOLOv5 model in sequence to obtain a key part map and key part coordinate information of the crack image; S32, performing image registration on the crack image and the corresponding template image, and cutting out a key part map from the template image according to the key part coordinate information to obtain a key part map of the template image; S33, performing logarithmic operations and ratio operations on the key part map of the crack image and the key part map of the template image to obtain N key part difference maps; S34, using Labelimg software to mark the cracks in the key position difference map to obtain N marking files; S35 , dividing the key part difference map and the annotation file corresponding to the key part difference map into a difference map training set and a difference map test set.

7. The aircraft structure crack detection method based on convolutional neural network according to claim 6, characterized in that: In S35, the difference map training set and the difference map test set are divided into a ratio of 7:

3.

8. The aircraft structure crack detection method based on convolutional neural network according to claim 7, characterized in that: In step 4, the improved YOLOv5 network is trained using the difference map dataset to obtain an improved YOLOv5 model, including: S43. Input the difference map training set in the difference map data set into the improved YOLOv5 network, perform network training and parameter optimization, until the detection accuracy of the improved YOLOv5 network on the difference map test set reaches a threshold, and obtain an improved YOLOv5 model.

9. The aircraft structure crack detection method based on convolutional neural network according to claim 8, characterized in that: In step 5, the image to be detected is tested as a whole according to the YOLOv5 model and the improved YOLOv5 model, including: S51. Conduct first-level testing: Input the image to be detected into the YOLOv5 model to obtain a key part map and key part coordinate information of the image to be detected; S52. Difference map generation: Perform image registration on the image to be detected and the template image, and cut out the key part map from the template image according to the coordinate information of the key parts to obtain the key part map of the template image; Perform logarithmic operations and ratio operations on the key part map of the image to be detected and the key part map of the template image to obtain a key part difference map; S53, conduct secondary testing: The key part difference map is input into the improved YOLOv5 model. If no crack is detected, the output is "no crack". If a crack is detected, the output crack coordinate information is converted according to the coordinate information of the key part difference map in the image to be detected to obtain the coordinate information of the crack in the image to be detected.

Images