A coating appearance defect detection system for aerospace components

Abstract

This invention discloses a system for detecting surface defects in the coatings of aerospace components. It comprises a high-definition industrial camera, an auxiliary light source, an image analysis server, a data management system, a network controller, and a signal indicator. The high-definition industrial camera can acquire images of the coating appearance of aerospace components in real time and transmit these images to the image analysis server via a wired network. The image analysis server performs real-time defect analysis on the coating appearance images acquired by the industrial camera using an intelligent recognition algorithm platform and feeds back defective coating appearance images to the data management system in real time. The data management system stores, manages, and displays the analyzed image data and sends coating defect signals to the signal indicator via the network controller. This invention avoids interference from human factors, greatly reducing the significant manpower and time investment required for long-term visual inspections by inspectors, and ensuring the stability and consistency of coating appearance inspection for aerospace components.

Description

Technical Field

[0001] This invention relates to a coating appearance defect detection system for aerospace parts, which is mainly applicable to the image recognition of coating appearance defects after spraying aerospace parts. Background Technology

[0002] With the emergence of automated painting production lines, various new processes and equipment are constantly being updated and improved, leading to the rapid development of surface treatment technology. Correspondingly, the corresponding inspection technology also needs to be automated. In the inspection of coating appearance defects, traditional methods rely on visual inspection. Manual inspection requires workers to be at fixed positions, visually observing the product to determine the presence of defects. Prolonged inspection can easily cause eye damage, and the results are susceptible to influences such as the inspector's emotional state, skill level, judgment standards, and individual differences. This makes it impossible to guarantee the consistency and stability of coating appearance defect detection for parts, resulting in fluctuating and inconsistent product quality.

[0003] With the rapid development of computer vision technology, intelligent detection of coating appearance defects based on deep learning provides a feasible technical solution for automated inspection of coating quality. High-definition imaging equipment is used to acquire and process images of the coating appearance of aerospace components, establishing a database of coating appearance defect images with different defect types. For image training, a convolutional neural network model is built, and a reasonable optimized loss function is designed. By inputting images into the convolutional neural network model for forward inference, loss function error calculation, and error backpropagation, the weights of the convolutional neural network are continuously updated, resulting in a deep neural network model for defect recognition in coating appearance images of aerospace components. This enables automated extraction and recognition of coating defects in coating appearance images.

[0004] The current reliance on manual visual inspection for the appearance quality inspection of aerospace component coatings leads to problems such as poor stability, low efficiency, and large errors in coating quality inspection, which seriously affects the efficiency of coating quality inspection for aerospace components. Furthermore, visual inspection methods cannot be matched with automated coating processes. Therefore, it is necessary to adopt a new generation of artificial intelligence image recognition technology for automated inspection of coating appearance defects of aerospace components. Summary of the Invention

[0005] To overcome the limitations of visual inspection in judging the quality of coatings on aircraft parts, this invention provides a coating appearance defect detection system for aircraft parts, which avoids coating quality detection problems caused by personal subjective factors and realizes real-time automated detection of coating appearance defects on aircraft parts.

[0006] The technical solution adopted by this application to solve its technical problem is:

[0007] A system for detecting surface defects in the coating of aerospace components includes a high-definition industrial camera, an auxiliary light source, an image analysis server, a data management system, a network controller, and a signal indicator. The high-definition industrial camera can acquire images of the coating appearance of aerospace components in real time and transmit these images to the image analysis server via a wired network. The image analysis server is equipped with an intelligent algorithm platform for recognizing surface defects in the coating appearance of aerospace components. The image analysis server performs real-time defect analysis on the coating appearance images acquired by the industrial camera and feeds back defective coating appearance images to the data management system in real time. The data management system stores, manages, and displays the analyzed image data and sends coating defect signals to the signal indicator via the network controller. The system includes the following steps:

[0008] Step 1: Use a high-definition industrial camera to take multiple images of the coating appearance of aerospace parts, so that the captured images cover the entire coating surface of the aerospace parts, and import all images into an image analysis server. Manually screen the captured coating appearance images to identify coating defect images, classify and locate the coating defects in each coating appearance defect image, and establish an image database of coating appearance defects of aerospace parts.

[0009] Step 2: Build a deep neural network model for coating appearance defect recognition based on the coating appearance defect image database;

[0010] Step 3: Train the deep neural network model using the coating appearance defect image database until the training loss value of the deep neural network model converges to obtain the final coating appearance image defect recognition model. During training, add image preprocessing module and data augmentation module to improve the diversity of coating appearance defect image database and enable the coating appearance defect recognition model to achieve better generalization.

[0011] Step 4: Capture images of the coating appearance of aircraft parts in real time, use a coating appearance image defect recognition model to identify coating defects in the images, and feed back the recognition results, including the defect type and coordinates, to the data management system in real time. The network controller will then send the defect signal to the signal prompter.

[0012] Furthermore, the intelligent recognition algorithm platform for coating appearance defects includes an image preprocessing module, a data augmentation module, and a deep neural network model for coating appearance image defect recognition. The image preprocessing module is used to scale training images to a uniform size and normalize the images. The image data augmentation module is used to expand the size of the coating appearance defect image database and improve the recognition accuracy of the deep neural network model. The deep neural network model for coating appearance image defect recognition includes a convolutional neural network architecture YOLOv5-OBB, a channel attention module, a spatial attention module, and a feature pyramid nonlocal block, which is used to perform multi-scale target classification and localization of defects in real-time acquired coating appearance images, obtain the final defect detection result, and meet the accuracy requirements for on-site coating appearance image defect detection.

[0013] Furthermore, in step 1, when classifying and locating the coating defects in each coating appearance defect image, a corresponding label document information is generated to record the coordinate position of the defect in the image, the defect type, and other information; the aerospace component coating appearance defect image database consists of coating appearance defect images and corresponding label documents.

[0014] Furthermore, in step 2, when building a deep neural network model for coating appearance image defect recognition, the deep neural network model uses the coating appearance defect image database as the input data of the network model. The deep neural network model is built based on the target recognition YOLOv5-OBB architecture. Channel attention module, spatial attention module, and feature pyramid nonlocal block are designed and embedded in the deep neural network model YOLOv5-OBB to improve the feature representation ability of the deep neural network model for the target area of ​​coating appearance defects. This enables the deep neural network model to extract defect features containing multi-scale and effective target information, thereby improving the detection capability of different types of defects.

[0015] The design process of the channel attention module includes the following steps:

[0016] To aggregate the spatial information of each feature channel, mean pooling and max pooling are employed. These two pooling methods collect different important features about the target. Then, a fractional weighted network is used to further integrate the two pooled features, given a feature map F∈R. C×H×W As input, mean pooling features are obtained simultaneously. With max pooling features To compute channel attention features, average pooling is performed on the features. With max pooling features Cascade generation of an effective feature expressor This serves as the input tensor to the score weighting network, which consists of multiple convolutional layers used to generate channel attention scaling features S. channel ∈R C×1×1 To reduce computation, the sizes of the two convolutional kernels are set to W1∈R. C / r×C×2×1 and W2∈R C×C / r×1×1 Where r is the reduction ratio, the channel attention scaling feature S is then calculated using the sigmoid activation function. channel The channel attention scaling feature is calculated as follows:

[0017]

[0018] In the formula, SWN represents the score-weighted network, Concat represents the cascade function, AvgPool represents average pooling, MaxPool represents maximum pooling, and σ is the sigmoid activation function. Represents convolution operation;

[0019] The design process of the spatial attention module includes the following steps:

[0020] Inspired by the encoder-decoder network architecture, a spatial attention module was constructed to compute the spatial attention score map S. spatial ∈R 1×H×W This is used to highlight the effective spatial location information of the target, given the input feature F′∈R. C×H×W To reduce computation, average pooling and max pooling operations are performed along the channel axis to obtain the average pooling feature. and max pooling features Then average pooling features and max pooling features Cascade to obtain input features Instead of directly inputting the feature F′ into the spatial attention module, which consists of an encoder and a decoder, the encoder is constructed using 3×3 convolutional operations and pooling operations. The decoder is constructed using upsampling operations, 2×2 convolutional operations, and a sigmoid activation function to generate the spatial attention score map S. spatial ∈R 1×H×W .

[0021] The pyramid nonlocal block is composed of a feature pyramid pooling (FPP) module and feature matrix operations, and the process is as follows:

[0022] Given an input feature map F∈R 2C×H×W As input, it is first fed into the convolutional layer to generate three new feature maps in parallel. Feature F θ and F gInput the feature pyramid pooling FPP module to generate matrix M θ M g ∈R C×135 The calculation is as follows:

[0023] M θ =FPP(F θ ),M g =FPP(F g )

[0024] Feature F φ Dimension deformation and transpose into a matrix Where N = H × W represents the number of pixels, and the similarity matrix M s ∈R N×135 It can be obtained through matrix multiplication and softmax operations, calculated as follows:

[0025] M S =Softmax(M φ T ×M θ )

[0026] Matrix M g and similarity matrix M S Perform matrix multiplication between the transposes of the matrix to produce the output matrix M. g,s ∈R C×N The calculation is as follows:

[0027] M g,S =M g ×M S T

[0028] Where M g,S Size deformation F g,s ∈R C×H×W Then feature F g,S The input convolutional layer W is used to perform convolution to obtain the input features F∈R. 2C×H×W Features F1 with the same output dimension are then multiplied by the scaling parameter α and summed element-wise with the input features F to obtain the final output features F as follows. O ∈R 2C×H×W The calculation is as follows:

[0029]

[0030] in Represents matrix multiplication. This represents element-wise summation, where the parameter α is initialized to zero and changes as the deep neural network is trained.

[0031] The design process of the Feature Pyramid Pooling (FPP) module includes the following steps: n represents the width or height of the output feature size of the pooling layer. In the Feature Pyramid Pooling (FPP) module, the sizes of the four output features are 1×1, 3×3, 5×5 and 10×10, respectively. Then, the four output feature maps are flattened and concatenated for subsequent matrix multiplication.

[0032] Furthermore, in step 3, when training the deep neural network model using the coating appearance defect image database,

[0033] loss=λ1L cls +λ2L obj +λ3L loc +λ4L ang

[0034] In the formula, L cls L obj L loc L ang These are the classification loss value, target loss value, localization loss value, and angle classification loss value, respectively. λ1, λ2, λ3, and λ4 are hyperparameters used to balance these four items. An image preprocessing module and a data augmentation module are added during training. The image preprocessing module includes scaling the acquired raw images to a 640×640 pixel size and performing normalization processing. The data augmentation module includes adjusting the brightness and contrast of the coating appearance defect image database, random translation, scaling, rotation, random cropping, mirror flipping, random noise addition, random blurring, and other processing to expand the size of the coating appearance defect image database.

[0035] Furthermore, in step 1, the coating appearance defects include flow, dirty particles, scratches, pits and bubbles that appear after the aerospace parts are coated. Each type of coating appearance defect is manually classified and located and labeled to establish an image database of coating appearance defects for aerospace parts.

[0036] The present invention has the following advantages: 1) The coating appearance defect detection system for aerospace parts has high accuracy, fast detection speed, and low hardware resource consumption such as graphics memory for model calculation; 2) The deep neural network model for coating appearance defect recognition has the ability to detect targets with angle rotation, reducing interference from redundant background information and extracting effective defect targets; 3) The channel attention module, spatial attention module, and feature pyramid nonlocal block embedded in the deep neural network model have the ability to extract defect features that highlight the deep neural network model, thereby improving the defect recognition and detection accuracy of the deep neural network model; 4) The intelligent recognition algorithm platform for coating appearance defects of aerospace parts has the function of flexibly adding and deleting recognition functions for various types of coating defects; 5) The coating appearance defect recognition results are fed back to the data management system in real time, the coating defect analysis results are electronically recorded, and the signal prompts are controlled by the network controller to realize real-time alarm prompts.

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and examples. Attached Figure Description

[0038] Figure 1 This is a structural diagram of the aerospace component coating appearance defect detection system of the present invention;

[0039] Figure 2 This is a flowchart of the intelligent identification algorithm for coating appearance defects of aerospace parts according to the present invention;

[0040] Figure 3 This is a schematic diagram of the channel attention module of the present invention;

[0041] Figure 4 This is a schematic diagram of the spatial attention module of the present invention;

[0042] Figure 5 This is a schematic diagram of the non-local block of the feature pyramid of the present invention;

[0043] Figure 6 This is a schematic diagram of the pyramid pooling FPP module of the present invention. Detailed Implementation

[0044] It should be understood that the specific examples described herein are merely illustrative of the invention and are not intended to limit the invention. This invention has been technically verified and comprises a high-definition industrial camera, an auxiliary light source, an image analysis server, a data management system, a network controller, and a signal indicator.

[0045] Implementation: A high-definition industrial camera can acquire real-time images of the coating appearance of aerospace components. These images are transmitted to an image analysis server via a wired network. The image analysis server is equipped with an intelligent recognition algorithm platform for coating appearance defects. This platform includes an image preprocessing module, a data augmentation module, and a deep neural network model for coating appearance defect recognition. The image preprocessing module scales training images to a uniform size and normalizes them. The image data augmentation module expands the coating appearance defect image database, improving the recognition accuracy of the deep neural network model. The deep neural network model for coating appearance defect recognition includes a YOLOv5-OBB convolutional neural network architecture, a channel attention module, a spatial attention module, and non-local feature pyramid blocks. This model performs multi-scale target classification and localization of defects in the real-time acquired coating appearance images, obtaining the final defect detection results and meeting the accuracy requirements for on-site coating appearance image defect detection. The image analysis server performs real-time defect analysis on the coating appearance images acquired by the industrial camera and feeds back defective coating appearance images to the data management system in real time. The data management system stores, manages, and displays the analyzed image data and sends coating defect signals to a signal indicator via a network controller. Figure 1 As shown.

[0046] See Figure 2 A smart algorithm platform for identifying surface defects in coatings of aerospace components, including:

[0047] Step 1: Use a high-definition industrial camera to capture multiple images of the coating appearance of aerospace parts, ensuring that the captured images cover the entire surface of the coating. Import all images into an image analysis server. Manually screen the captured coating appearance images to identify coating defect images. Coating appearance defects include flow, dirty particles, scratches, pits, and bubbles that occur after the aerospace parts are coated. Manually classify and locate each type of coating defect in each coating appearance defect image, generating a label document information for the corresponding defect image. Record the coordinate position of the defect in the image, the defect type, and the defect size. Establish an aerospace part coating appearance defect image database consisting of coating appearance defect images and corresponding label documents.

[0048] Step 2: Construct a deep neural network model for coating appearance defect recognition based on the coating appearance defect image database. Use the coating appearance defect image database as input data for the deep neural network model and train the model. The deep neural network model is built based on the YOLOv5-OBB architecture for target recognition. Design channel attention module, spatial attention module, and feature pyramid nonlocal block and embed them into the deep neural network model YOLOv5-OBB to improve the feature representation ability of the deep neural network model for the target area of ​​coating appearance defects. This enables the deep neural network model to extract defect features containing multi-scale and effective target information, thereby improving the detection capability of different types of defects.

[0049] Figure 3 The design process of the channel attention module is described. To aggregate the spatial information of each feature channel, mean pooling and max pooling are employed. These two pooling methods collect different important features about the target. Then, a fractional weight network is used to further integrate the two pooled features. Given a feature map F∈R C×H×W As input, mean pooling features are obtained simultaneously. With max pooling features To compute channel attention features, average pooling is performed on the features. With max pooling features Cascade generation of an effective feature expressor This serves as the input tensor to the score weighting network. The score weighting network consists of multiple convolutional layers used to generate channel attention-scaled features S. channel ∈R C×1×1 To reduce computation, the sizes of the two convolutional kernels are set to W1∈R. C / r×C×2×1 and W2∈R C×C / r×1×1 Where r is the reduction ratio, the channel attention scaling feature S is then calculated using the sigmoid activation function. channel The channel attention scaling feature is calculated as follows:

[0050]

[0051] In the formula, SWN represents the score-weighted network, Concat represents the cascade function, AvgPool represents average pooling, MaxPool represents maximum pooling, and σ is the sigmoid activation function. Represents convolution operation;

[0052] Figure 4 The design process of the spatial attention module is described. The spatial attention mechanism is used to highlight the effective spatial location information of the target. Inspired by the encoder-decoder network architecture, a spatial attention module is constructed to compute the spatial attention score map S. spatial ∈R1×H×W This is used to highlight the effective spatial location information of the target. Given input features F′∈R C ×H×W To reduce computation, average pooling and max pooling operations are performed along the channel axis to obtain the average pooling feature. and max pooling features Then average pooling features and max pooling features Cascade to obtain input features Instead of directly inputting the feature F′ into the spatial attention module, the spatial attention module consists of an encoder and a decoder. The encoder is constructed using 3×3 convolutional operations and pooling, while the decoder is constructed using upsampling, 2×2 convolutional operations, and a sigmoid activation function to generate the spatial attention score map S. spatial ∈R 1×H×W .

[0053] The nonlocal feature pyramid is composed of a feature pyramid pooling FPP module and feature matrix operations.

[0054] Figure 5 The design process of nonlocal blocks in a feature pyramid is described. Given an input feature map F∈R 2C×H×W As input, it is first fed into the convolutional layer to generate three new feature maps in parallel. Feature F θ and F g Input the feature pyramid pooling FPP module to generate matrix M θ M g ∈R C×135 The calculation is as follows:

[0055] M θ =FPP(F θ ),M g =FPP(F g )

[0056] Feature F φ Dimension deformation and transpose into a matrix Where N = H × W represents the number of pixels, and the similarity matrix M s ∈R N×135 It can be obtained through matrix multiplication and softmax operations, calculated as follows:

[0057] M S =Softmax(M φ T ×M θ )

[0058] Matrix M gand similarity matrix M S Perform matrix multiplication between the transposes of the matrix to produce the output matrix M. g , s ∈R C×N The calculation is as follows:

[0059] M g,S =M g ×M S T

[0060] Where M g,S Size deformation F g,s ∈R C×H×W Then feature F g,S The input convolutional layer W is used to perform convolution to obtain the input features F∈R. 2C×H×W Features F1 have the same output dimension. Finally, feature F1 is multiplied by the scaling parameter α and summed element-wise with the input feature F to obtain the final output feature F as follows. O ∈R 2C×H×W The calculation is as follows:

[0061]

[0062] in Represents matrix multiplication. This represents element-wise summation, where the parameter α is initialized to zero and changes as the deep neural network is trained.

[0063] Figure 6 The design process of the Feature Pyramid Pooling (FPP) module is described. n represents the width or height of the output feature size of the pooling layer. In the Feature Pyramid Pooling (FPP) module, the four output feature sizes are 1×1, 3×3, 5×5, and 10×10, respectively. These four output feature maps are then flattened and concatenated for subsequent matrix multiplication.

[0064] Step 3: Train the deep neural network model using a database of coating appearance defects. The multi-task training optimization objective loss function is expressed as:

[0065] loss=λ1L cls +λ2L obj +λ3L loc +λ4L ang

[0066] In the formula, L cls L obj L loc L angThe loss values ​​are classification loss, target loss, localization loss, and angle classification loss, respectively. λ1, λ2, λ3, and λ4 are hyperparameters used to balance these four components. The final coating appearance image defect recognition model is obtained when the training loss values ​​of the deep neural network model converge and stabilize. Image preprocessing and data augmentation modules are added during training. The image preprocessing module includes scaling the acquired raw images to 640×640 pixels and performing normalization. The data augmentation module includes adjusting the brightness and contrast of the coating appearance defect image database, random translation, scaling, rotation, random cropping, mirroring, random noise addition, and random blurring, etc., to expand the size and diversity of the coating appearance defect image database, enabling the coating appearance defect recognition model to achieve better generalization.

[0067] Step 4: Capture images of the coating appearance of aircraft parts in real time, use a coating appearance image defect recognition model to identify coating defects in the images, and feed the recognition results, including the defect type and coordinates, back to the data management system in real time for storage, management and display. The coating defect recognition result signal is also sent to the network controller via UDP communication protocol to control the operation of the signal indicator.

[0068] The inventors conducted verification of the above method, and the verification process is described below:

[0069] The coating appearance defect image database consists of 1500 images with an original resolution of 4032×3024. Coating appearance defects include flow, dirt particles, scratches, pits, and bubbles that occur after the coating of aerospace parts. The distribution of the number of each type of defect is shown in Table 1.

[0070] Table 1. Distribution of the Number of Appearance Defects in Various Coatings

[0071] Types of defects flowing Dirty particles scratches pit bubble Statistical quantity 1210 2650 1008 930 613

[0072] The deep neural network model, abbreviated as YO-CF, is trained using a database of coating appearance defect images. Training details include:

[0073] The stochastic gradient descent (SGD) optimizer was used to train the deep neural network model for 300 iterations. The minimum batch size was 30, the initial learning rate was set to 0.01, the momentum to 0.9, and the weight decay to 0.0005. The mean average precision (mAP) was calculated to evaluate the performance of the deep neural network model.

[0074] Table 2 Performance metrics of deep neural network models

[0075]

[0076] As can be seen from Table 2, compared with the other five algorithm models, the deep neural network model YO-CF proposed in this embodiment has an mAP value of 71.6%, which can achieve the highest detection accuracy. It also achieves the highest detection accuracy for the average accuracy, i.e., AP value, for each type of defect.

[0077] This invention provides a general framework for an aerospace component coating appearance defect detection system, not limited to specific detection methods. The advantages of this framework are that the intelligent recognition algorithm platform for aerospace component coating appearance defects includes an image preprocessing module, a data augmentation module, and a deep neural network model for coating appearance image defect recognition. The image preprocessing module is used to scale training images to a uniform size and normalize the images. The image data augmentation module is used to expand the size of the coating appearance defect image database and improve the recognition accuracy of the deep neural network model. The deep neural network model for coating appearance image defect recognition includes a convolutional neural network architecture YOLOv5-OBB, a channel attention module, a spatial attention module, and a feature pyramid nonlocal block, used to perform multi-scale target classification and localization of defects in real-time acquired coating appearance images, obtaining the final defect detection result and meeting the accuracy requirements for on-site coating appearance image defect detection.

[0078] Finally, it should be noted that those skilled in the art should understand that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A system for detecting surface defects in the coating of aerospace components, comprising a high-definition industrial camera, an auxiliary light source, an image analysis server, a data management system, a network controller, and a signal indicator, characterized in that, The high-definition industrial camera can acquire real-time images of the coating appearance of aerospace components and transmit these images to an image analysis server via a wired network. The image analysis server is equipped with an intelligent recognition algorithm platform for coating appearance defects of aerospace components. This platform includes an image preprocessing module, a data augmentation module, and a deep neural network model for coating appearance image defect recognition. The image preprocessing module scales training images to a uniform size and normalizes them. The image data augmentation module expands the database of coating appearance defect images, improving the recognition accuracy of the deep neural network model. A high-resolution neural network model, including the YOLOv5-OBB convolutional neural network architecture, channel attention module, spatial attention module, and feature pyramid nonlocal blocks, is used to perform multi-scale target classification and localization of defects in real-time acquired coating appearance images, obtaining the final defect detection results and meeting the accuracy requirements for on-site coating appearance image defect detection. An image analysis server performs real-time defect analysis on coating appearance images acquired by an industrial camera and feeds back defective coating appearance images to the data management system in real time. The data management system stores, manages, and displays the analyzed image data information and sends coating defect signals to a signal indicator via a network controller. The process includes the following steps: Step 1: Use a high-definition industrial camera to capture multiple images of the coating appearance of aerospace components, ensuring the images cover the entire surface of the coating. Import all images into an image analysis server. Manually screen the captured coating appearance images to identify coating defect images. Classify and locate the coating defects in each image to establish an aerospace component coating appearance defect image database. When classifying and locating coating defects in each image, a corresponding label document is generated to record the coordinate position and type of the defect in the image. The aerospace component coating appearance defect image database consists of coating appearance defect images and corresponding label documents. Step 2: Construct a deep neural network model for coating appearance defect recognition based on a database of coating appearance defect images. When constructing this model, the database of coating appearance defect images serves as the input data. The deep neural network model is built using a YOLOv5-OBB architecture based on target recognition. Channel attention modules, spatial attention modules, and feature pyramid nonlocal blocks are designed and embedded into the YOLOv5-OBB deep neural network model. This enhances the deep neural network model's ability to represent the features of target regions of coating appearance defects, enabling it to extract defect features containing multi-scale and effective target information. This improves the detection capability for different types of defects. The design process of the channel attention module includes the following steps: To aggregate the spatial information of each feature channel, mean pooling and max pooling are employed. These two pooling methods collect different important features about the target. Then, a fractional weighted network is used to further integrate the two pooled features, given a feature map F∈R. C×H×W As input, mean pooling features are obtained simultaneously. With max pooling features To compute channel attention features, average pooling is performed on the features. With max pooling features Cascade generation of an effective feature expressor This serves as the input tensor to the score weighting network, which consists of multiple convolutional layers used to generate channel attention scaling features S. channel ∈R C×1×1 To reduce computation, the sizes of the two convolutional kernels are set to W1∈R. C / r×C×2×1 and W2∈R C×c / r×1×1 Where r is the reduction ratio, the channel attention scaling feature S is then calculated using the sigmoid activation function. channel The channel attention scaling feature is calculated as follows: In the formula, SWN represents the score-weighted network, Concat represents the cascade function, AvgPool represents average pooling, MaxPool represents maximum pooling, and σ is the sigmoid activation function. Represents convolution operation; Step 3: Train the deep neural network model using the coating appearance defect image database until the training loss value of the deep neural network model converges to obtain the final coating appearance image defect recognition model. During training, add image preprocessing module and data augmentation module to improve the diversity of coating appearance defect image database and enable the coating appearance defect recognition model to achieve better generalization. Step 4: Capture images of the coating appearance of aircraft parts in real time, use a coating appearance image defect recognition model to identify coating defects in the images, and feed back the recognition results, including the defect type and coordinates, to the data management system in real time. The network controller will then send the defect signal to the signal prompter.

2. The system for detecting surface defects in the coating of aerospace components according to claim 1, characterized in that... The design process of the spatial attention module includes the following steps: Inspired by the encoder-decoder network architecture, a spatial attention module was constructed to compute the spatial attention score map S. spatial ∈R 1×H×W This is used to highlight the effective spatial location information of the target, given the input feature F′∈R. C×H×W To reduce computation, average pooling and max pooling operations are performed along the channel axis to obtain the average pooling feature. and max pooling features Then average pooling features and max pooling features Cascade to obtain input features Instead of directly inputting the feature F′ into the spatial attention module, which consists of an encoder and a decoder, the encoder is constructed using 3×3 convolutional operations and pooling operations. The decoder is constructed using upsampling operations, 2×2 convolutional operations, and a sigmoid activation function to generate the spatial attention score map S. spatial ∈R 1×H×W .

3. The system for detecting surface defects in the coating of aerospace components according to claim 1, characterized in that... The nonlocal feature pyramid block is composed of a feature pyramid pooling FPP module and feature matrix operations, and the process is as follows: Given an input feature map F∈R 2C×H×W As input, it is first fed into the convolutional layer to generate three new feature maps in parallel. Feature F θ and F g Input the feature pyramid pooling FPP module to generate matrix M θ M g ∈R C ×135 The calculation is as follows: M θ =FPP(F θ ),M g =FPP(F g ) feature Dimension deformation and transpose into a matrix Where N = H × W represents the number of pixels, and the similarity matrix M s ∈R N×135 It is obtained through matrix multiplication and softmax operations, and is calculated as follows: Matrix M g and similarity matrix M s Perform matrix multiplication between the transposes of the matrix to produce the output matrix M. g,s ∈R C×N The calculation is as follows: Where M g,s Size deformation F g,s ∈R C×H×W Then feature F g,s The input convolutional layer W is used to perform convolution to obtain the input features F∈R. 2C×H×W Features F1 with the same output dimension are then multiplied by the scaling parameter α and summed element-wise with the input features F to obtain the final output features F0∈R. 2C×H×W The calculation is as follows: in Represents matrix multiplication. This represents element-wise summation, where the parameter α is initialized to zero and changes as the deep neural network is trained.

4. The system for detecting surface defects in the coating of aerospace components according to claim 3, characterized in that... The design process of the feature pyramid pooling FPP module includes the following steps: n represents the width or height of the pooling layer output feature size. In the Feature Pyramid Pooling (FPP) module, the four output feature sizes are 1×1, 3×3, 5×5 and 10×10, respectively. Then, the four output feature maps are flattened and concatenated for subsequent matrix multiplication.

5. A coating appearance defect detection system for aerospace components according to any one of claims 1 to 4, characterized in that: In step 3, when training the deep neural network model using the coating appearance defect image database, loss=λ1L cls +λ2L obj +λ3L loc +λ4L ang In the formula, L cls L obj L loc L ang The loss values ​​are classification loss, target loss, localization loss, and angle classification loss, respectively. λ1, λ2, λ3, and λ4 are hyperparameters used to balance these four loss values. During training, an image preprocessing module and a data augmentation module are added. The image preprocessing module includes scaling the acquired raw images to 640×640 pixels and normalizing them. The data augmentation module includes adjusting the brightness and contrast of the coating appearance defect image database, random translation, scaling, rotation, random cropping, mirroring, random noise addition, and random blurring, thereby expanding the size of the coating appearance defect image database.

6. The system for detecting surface defects in the coating of aerospace components according to claim 1, characterized in that, In step 1, the coating appearance defects include flow, dirty particles, scratches, pits and bubbles that appear after the aerospace parts are sprayed. Each type of coating appearance defect is classified and located manually to establish an image database of coating appearance defects for aerospace parts.

Images