CN117788471A - Method for detecting and classifying aircraft skin defects based on YOLOv5 - Google Patents
Method for detecting and classifying aircraft skin defects based on YOLOv5 Download PDFInfo
- Publication number
- CN117788471A CN117788471A CN202410211173.1A CN202410211173A CN117788471A CN 117788471 A CN117788471 A CN 117788471A CN 202410211173 A CN202410211173 A CN 202410211173A CN 117788471 A CN117788471 A CN 117788471A
- Authority
- CN
- China
- Prior art keywords
- model
- yolov5
- aircraft skin
- interpolation
- original
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 230000007547 defect Effects 0.000 title claims abstract description 18
- 238000001514 detection method Methods 0.000 claims abstract description 43
- 238000012549 training Methods 0.000 claims abstract description 33
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 25
- 230000007246 mechanism Effects 0.000 claims abstract description 18
- 230000006378 damage Effects 0.000 claims abstract description 15
- 230000037380 skin damage Effects 0.000 claims abstract description 14
- 238000012360 testing method Methods 0.000 claims abstract description 11
- 238000010586 diagram Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 6
- 230000007797 corrosion Effects 0.000 claims description 4
- 238000005260 corrosion Methods 0.000 claims description 4
- 230000003213 activating effect Effects 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 6
- 230000004927 fusion Effects 0.000 description 4
- 230000003993 interaction Effects 0.000 description 4
- 238000012423 maintenance Methods 0.000 description 3
- 230000001629 suppression Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000005336 cracking Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a method for detecting aircraft skin defects and classifying based on YOLOv5, which comprises the steps of collecting aircraft skin damage pictures, marking anchor frames on the pictures, classifying the obtained pictures according to different damage, dividing the pictures into a training set and a testing set, carrying out information interpolation by using a bilinear interpolation algorithm, so that the training set data are reinforced, constructing a YOLOv5 model and optimizing the model, inputting the acquired reinforced training set data into the optimized YOLOv5 model to train the model, obtaining trained weight files, storing the optimal model, applying the trained weight files to the picture detection of the testing set to detect the data set, and visualizing the detection result through a Pyside 6. The invention introduces a new attention mechanism in the backbone network backbone layer, can encode the transverse and longitudinal spatial position information into the channel attention, improves the capability of the model in detecting small objects and can accurately position the detected objects when the objects are too dense.
Description
Technical Field
The invention relates to the technical field of aircraft skin defects, in particular to a method for detecting and classifying aircraft skin defects based on YOLOv 5.
Background
The aircraft skin is a covering layer positioned outside the aircraft structural frame, not only forms the appearance of the aircraft and protects internal components, but also has key effects on the aspects of aerodynamic performance, protection, stealth performance, maintainability, weight, appearance quality and the like of the aircraft. Its design and material selection requires balancing a number of factors to ensure that the aircraft performs well under a variety of operating conditions.
When the aircraft skin is damaged, the strength of relevant parts is reduced, so that the aerodynamic performance of the aircraft is affected, and finally the flight safety of the aircraft is endangered. Aircraft skin damage is of several types, namely cracking, corrosion and impact. When damage occurs, the aircraft damage can be checked through the unmanned aerial vehicle or the 3D scanner, and the damage is transmitted to maintenance personnel through video, and the maintenance personnel can judge which type the aircraft skin damage belongs to through video. In order to reduce misjudgment of maintenance personnel, improve detection accuracy and guide workers to repair damaged areas in real time, a target detection algorithm is provided.
The main target detection algorithms can be divided into two categories: one-stage and Two-stage detection algorithms. The Two-stage detection algorithm comprises a fast-r cnn and a Mask-r cnn, so that accurate classification can be achieved, the detection accuracy is high, and the detection speed is low. The One-stage detection algorithm comprises a YOLO algorithm and an SSD algorithm, and has the advantages of high detection speed and lower precision. The YOLOv5 algorithm is the fifth generation of the YOLOv algorithm, which has not only a fast detection speed but also a high detection accuracy, but its effect is poor when the objects are too dense.
Therefore, it is necessary to design an optimized YOLOv5 algorithm model to improve the ability of the model to detect small objects and to accurately locate the detected objects when the objects are too dense.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a method for detecting the defects and classifying the skin of an airplane based on YOLOv5, which solves the problems of poor capability of detecting small objects and poor effect when the objects are too dense in the traditional method; according to the interpolation method based on the depth information, proper consideration of short-distance and long-distance objects in the interpolation process is ensured, so that a more accurate and real multi-view interpolation image is generated, a new attention mechanism is introduced into a backbone network back plane layer through optimizing a YOLOv5 model, transverse and longitudinal space position information can be encoded into channel attention, a mobile network can pay attention to remote space interaction with accurate position information in a large range of position information acquisition without excessive calculation amount, global average pooling is decomposed, detection precision of a small model and attention of the network to channels are improved, the capability of the model in detecting the small object is improved, the detected object can be accurately positioned when the object is too dense, and different damage of an aircraft skin can be identified in real time when the model is applied to aircraft skin damage detection.
In order to solve the technical problems, the invention provides the following technical scheme: a method for detecting and classifying skin defects of an aircraft based on YOLOv5, comprising the steps of:
s1, acquiring an original image of aircraft skin damage, classifying the acquired image according to different damage, and dividing the classified image into a training set and a testing set;
s2, performing information interpolation on the collected damaged pictures of the training sets with different angles by using a bilinear interpolation algorithm, and further obtaining an optimized image so as to enhance the training set data;
s3, optimizing the original YOLOv5 model to obtain an optimized YOLOv5 model;
s4, inputting the acquired enhanced training set data into an optimized YOLOv5 model to obtain a trained weight file;
s5, applying the trained weight file to picture detection of the test set, and visualizing a detection result through a Pyside 6.
Further, the step S1 specifically includes: classifying the original images of the aircraft skin damage into damage images of corrosion, impact and crack, marking the anchor frames, recording the sequence of the categories by using a TXT file, and marking the anchor frames by using a Labelimg target detection marking tool.
Further, in step S2, the specific process includes the following steps:
s21, performing interpolation twice in the x direction to obtain pixel values R1 and R2 after the interpolation twice, and performing interpolation once in the y direction to obtain pixel values in the y direction; the formula for the two interpolations in the x direction is:
;
;
wherein Q is 11 (x 1 ,y 1 ),Q 12 (x 1 ,y 2 ),Q 21 (x 2 ,y 1 ),Q 22 (x 2 ,y 2 ) Four points in an x-y plane rectangular coordinate system;
s22, performing interpolation once in the y direction to obtain a pixel value in the y direction, wherein an interpolation formula in the y direction is as follows:
;
s23, carrying the two formulas in the S21 into the formulas in the S22 can obtain a high-quality and high-resolution multi-view interpolation image:
。
further, in step S3, the optimizing the original YOLOv5 model includes the following steps:
s31, introducing a new attention mechanism CA at a backbone network layer of the backhaul, wherein the new attention mechanism CA divides an input training set in the x and y directions into two partsIndependent directions, lateral and longitudinal spatial position information is encoded into the channel attention, and lateral and longitudinal features are encoded using pooling kernels 1, w and H,1, so the output of the c-th dimensional feature of height H is:;
similarly, the output of the c-th dimension of the width w is:;
wherein,is the height of the feature map; />Is the width of the feature map;x c is an intermediate feature; />,/>,/>,/>Is a known coordinate.
S32, integrating the features of the two formulas from different directions of transverse and longitudinal directions, and outputting a pair of feature graphs with known directionsFeature map +.>Connected to obtain a connected characteristic diagramF 1 And operates on it using a convolution change function to obtain an intermediate profile comprising lateral and longitudinal spatial information +.>The formula is as follows:
;
wherein,is a reduction factor;
s33, respectively encoding the information in the x and y directions into different attention diagrams along the space dimension, and mapping the intermediate feature diagram containing the transverse and longitudinal space informationPerforming a segmentation operation into tensors f in two directions h And f w The final high and wide attention weights are then obtained by a sigmoid activation functiong h Andg w the expression is as follows:
;
;
wherein,,/>is a convolution kernel of 1*1,>activating a function for sigmoid;
s34, CA attention mechanism formula can be expressed as:。
further, in step S3, the optimizing the original YOLOv5 model further includes: the DIoU-NMS algorithm is used for replacing the NMS algorithm in the original model, the DIoU-NMS not only considers IoU, but also considers the distance between the centers of two frames, if the distance between the two frames is large, the algorithm considers that the two frames are frames of two objects, so that filtering is avoided, and for optimizing the NMS in the model, only IoU in the NMS in the original model is needed to be replaced by the DIoU, and the formula is as follows:
;
wherein,,/>representing the center points of the predicted and real frames respectively,ρrepresenting the calculation of the Euclidean distance between two central points,/->Representing the diagonal distance of the minimum closure area that can contain both the predicted and real frames; the formula of the DIoU-NMS is shown below:
;
wherein s is i Is the classification confidence;Mrepresenting a higher scoring bounding box in the prediction box; epsilon is the threshold of the DIoU-NMS,traversing the superposition condition of each frame with high confidence coefficient; r is R DIoU Is a penalty formula.
Further, in step S3, the optimizing the original YOLOv5 model further includes: changing the loss function in the original model to CIoU, wherein the loss function is increased in loss function of detection scale based on DIoU, and increased in loss function of length and width, and the difference of center distance, width and height of the boundary frame is considered, so that regression of the target frame is more stable, and the formula is as follows:
;
;
;
;
wherein a, B can be seen as two sets;is a weight function; />To measure the similarity of aspect ratios.
Further, the step S4 specifically includes: placing the optimized YOLOv5 model into a configured Python environment, training the optimized YOLOv5 model by using the marked pictures in the training set after data enhancement, and accelerating training the model through a GPU.
By means of the technical scheme, the method for detecting the defects and classifying the aircraft skin based on the YOLOv5 has the following beneficial effects:
the interpolation method based on the depth information ensures proper consideration of the short-distance and long-distance objects in the interpolation process, thereby generating more accurate and more real multi-view interpolation images and enhancing the training set data; the method has the advantages that the acquired enhanced training set data is optimized through the YOLOv5 model, a new attention mechanism is introduced into a backbone network back layer, the transverse and longitudinal spatial position information can be encoded into the attention of a channel, the mobile network can pay attention to the large-range position information to acquire remote space interaction with accurate position information without bringing excessive calculation amount, the global average pooling is decomposed, the detection precision of a small model and the attention of the network to the channel are improved, the small object detection capability of the model is improved, the detected object can be accurately positioned when the object is too dense, different damage of the aircraft skin can be identified in real time when the model is applied to the aircraft skin damage detection, the detection speed can be improved, meanwhile, the method also has higher detection precision, and the problem that the effect of the model is poor when the object is too dense is solved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application and do not constitute an undue limitation to the application. In the drawings:
FIG. 1 is a flowchart of a method for detecting and classifying skin defects of an aircraft based on YOLOv 5.
FIG. 2 is a schematic diagram of a specific algorithm model modified by Yolov 5.
Detailed Description
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description. Therefore, the implementation process of how to apply the technical means to solve the technical problems and achieve the technical effects can be fully understood and implemented.
Those of ordinary skill in the art will appreciate that all or a portion of the steps in a method of implementing an embodiment described above may be implemented by a program to instruct related hardware, and thus the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
Fig. 1 and 2 show a specific implementation manner of the present embodiment, in the present embodiment, firstly, an aircraft skin damage picture is collected, an anchor frame is marked on the picture, the obtained picture is classified into a training set and a testing set according to different damage, information interpolation is performed by using a bilinear interpolation algorithm, so that training set data is reinforced, a YOLOv5 model is constructed and optimized, the collected reinforced training set data is input into the optimized YOLOv5 model to train the model, a trained weight file is obtained and the optimal model is stored, the trained weight file is applied to picture detection of the testing set to detect the data set, and a detection result is visualized through a Pyside 6. The interpolation method based on the depth information ensures proper consideration of short-distance and long-distance objects in the interpolation process, thereby generating more accurate and more real multi-view interpolation images, introducing a new attention mechanism into a backbone network back plane layer by optimizing a YOLOv5 model, coding transverse and longitudinal spatial position information into channel attention, enabling a mobile network to pay attention to long-distance spatial interaction with accurate position information in a large range of position information without bringing excessive calculation amount, decomposing global average pooling, improving detection precision of a small model and attention of the network to channels, improving the capability of the model in detecting the small object and accurately positioning the detected object when the object is too dense, being applied to aircraft skin damage detection, identifying different damage of the aircraft skin in real time.
Referring to fig. 1, the present embodiment provides a method for detecting defects and classification of an aircraft skin based on YOLOv5, which includes the following steps:
s1, acquiring an original image of aircraft skin damage, classifying the acquired image according to different damage, and dividing the classified images into a training set and a testing set;
as a preferred embodiment of step S1, S1 specifically includes: classifying the original images of the aircraft skin damage into damage images of corrosion, impact and crack, marking anchor frames, recording the sequence of the categories by using TXT files, and comparing tensors generated after the damage by using a YOLOv5 model. The anchor frame is a reference frame with different preset sizes and different length-width ratios on the image. The calibration anchor frame is marked by using a Labelimg target detection marking tool.
S2, information interpolation is carried out on the collected damaged pictures of the training sets with different angles by using a bilinear interpolation algorithm, so that an optimized image is obtained, and a multi-view interpolation picture with high quality and high resolution is obtained, so that the training set data is enhanced;
as a preferred embodiment of step S2, the specific procedure of S2 includes: a bilinear interpolation algorithm for depth information in multi-view interpolation uses gray values of adjacent pixels or RGB tristimulus values to generate gray values or RGB tristimulus values of unknown pixels. The aim is to generate a higher resolution image from the original image. Let us assume that the target is the pixel value of the resulting function at point Q (x, y), and that Q is known 11 (x 1 ,y 1 ),Q 12 (x 1 ,y 2 ),Q 21 (x 2 ,y 1 ),Q 22 (x 2 ,y 2 ) And corresponding pixel values, wherein Q 11 (x 1 ,y 1 ),Q 12 (x 1 ,y 2 ),Q 21 (x 2 ,y 1 ),Q 22 (x 2 ,y 2 ) Is four points in an x and y plane rectangular coordinate system.
S21, performing interpolation twice in the x direction to obtain pixel values R1 and R2 after the interpolation twice, and performing interpolation once in the y direction to obtain pixel values in the y direction; the formula for the two interpolations in the x direction is:
;
;
wherein Q is 11 (x 1 ,y 1 ),Q 12 (x 1 ,y 2 ),Q 21 (x 2 ,y 1 ),Q 22 (x 2 ,y 2 ) Four points in an x-y plane rectangular coordinate system;
s22, performing interpolation once in the y direction to obtain a pixel value in the y direction, wherein an interpolation formula in the y direction is as follows:
;
s23, carrying the two formulas in the S21 into the formulas in the S22 can obtain a high-quality and high-resolution multi-view interpolation image:
。
in the embodiment, the interpolation method based on the depth information ensures proper consideration of the short-distance and long-distance objects in the interpolation process, thereby generating more accurate and more real multi-view interpolation images.
S3, optimizing an original YOLOv5 model to obtain an optimized YOLOv5 model;
as a preferred embodiment of step S3, in S3, the optimizing the original YOLOv5 model specifically includes the following steps:
s31, introducing a new attention mechanism CA at a backbone network layer of the backhaul, wherein the new attention mechanism CA divides an input training set in x and y directions into two independent directions, codes transverse and longitudinal spatial position information into channel attention, and codes transverse and longitudinal characteristics by using pooling cores 1, W and H,1, so that the output of a c-th dimension characteristic of a height H is as follows:;
similarly, the output of the c-th dimension of the width w is:;
wherein,is the height of the feature map; />Is the width of the feature map;x c is an intermediate feature; />,/>,/>,/>Is a known coordinate.
S32, integrating the features of the two formulas from different directions of transverse and longitudinal directions, and outputting a pair of feature graphs with known directionsFeature map +.>Connected to obtain a connected characteristic diagramF 1 And operates on it using a convolution change function to obtain an intermediate profile comprising lateral and longitudinal spatial information +.>The formula is as follows:
;
wherein,is a reduction factor;
more specifically, a neg middle layer is added after the backbone network, as shown in fig. 2, the neg middle layer is responsible for carrying out multi-scale feature fusion on the feature map and transmitting the features to a Detect detection and identification image layer, and as the size and the position of an object in the image are uncertain, a mechanism is needed to process targets with different scales and sizes. In the Neck layer, feature graphs of different layers are fused together through upsampling and downsampling operations, the Neck middle layer comprises a top-down part and a bottom-up part, and the top-down part realizes fusion of features of different layers through connection fusion of the upsampling and coarser feature graphs, and mainly comprises the following steps:
s321, up-sampling a final layer of feature map of a backbone network to obtain a finer feature map;
s322, carrying out connection fusion on the up-sampled feature map and a feature map of the upper layer (a pair of feature maps with known directions obtained through a new attention mechanism CA) to obtain richer feature expression; the bottom-up part mainly uses convolution layers to connect and fuse characteristic graphs from different layers, and mainly comprises the following steps:
s323, convolving the bottom-most feature map of the top-down part to obtain richer feature expression;
s324, connecting and fusing the convolved feature map and the feature map of the upper layer to obtain richer feature expression;
s325, finally, the final layer of feature map convolved with the backbone network is connected and fused, so that the output feature map of the backbone network is connected and fused with the feature map of the top-down part and the feature map of the bottom-up part to be output, and a final intermediate feature map is obtainedThe method is used for target detection, has higher detection precision, and optimizes the problem that the detection effect of the model is poor when objects are too dense.
S33, respectively encoding the information in the x and y directions into different attention diagrams along the space dimension, and mapping the intermediate feature diagram containing the transverse and longitudinal space informationPerforming a segmentation operation into tensors f in two directions h And f w The final high and wide attention weights are then obtained by a sigmoid activation functiong h Andg w the expression is as follows:
;
;
wherein,,/>is a convolution kernel of 1*1,>activating a function for sigmoid;
s34, CA attention mechanism formula can be expressed as:。
more specifically, in S3, the optimizing the original YOLOv5 model further includes: the non-maximum suppression (Non Maximum Suppression, NMS) algorithm in the original model is replaced with a Distance-IoU Non Maximum Suppression (DIoU-NMS) algorithm that considers not only IoU but also the Distance between the centers of two boxes, which would be considered the boxes of two objects if the Distance between the two boxes were large, and thus would not be filtered out. For optimization of NMS in the model, only IoU in NMS in the original model needs to be replaced with DIoU, which is expressed as follows:
;
wherein,,/>representing the center points of the predicted and real frames respectively,ρrepresenting the calculation of the Euclidean distance between two central points,/->Representing the diagonal distance of the minimum closure area that can contain both the predicted and real frames; the formula of the DIoU-NMS is shown below:
;
wherein s is i Is the classification confidence;Mrepresenting a higher scoring bounding box in the prediction box; epsilon is the threshold of the DIoU-NMS,traversing the superposition condition of each frame with high confidence coefficient; r is R DIoU Is a penalty formula.
More specifically, in S3, the optimizing the original YOLOv5 model further includes: changing the loss function in the original model to CIoU, wherein the loss function is increased in loss function of detection scale based on DIoU, and increased in loss function of length and width, and the regression of the target frame is more stable by considering more factors such as the difference of the center distance, width and height of the boundary frame, wherein the formula is as follows:
;
;
;
;
wherein a, B can be seen as two sets;is a weight function; />To measure the similarity of aspect ratios.
In this embodiment, by optimizing the YOLOv5 model, introducing a new attention mechanism in the backbone network backbone layer, most of the attention mechanisms of the lightweight network currently adopt a Squeeze-and-Specification (SE) module, only the channel information is considered, but the position information is ignored, although the bottleneck attention module (Bottleneck Attention Module, BAM) and the convolution attention module (Convolutional Block Attention Module, CBAM) later attempt to extract the position attention information through convolution after reducing the number of channels, the convolution only can extract local relations, and lack the extraction capability of long-distance relations, so that the new attention mechanism is introduced in the backbone layer, the transverse and longitudinal spatial position information can be encoded into the channel attention, so that the mobile network can pay attention to a large range of position information to obtain remote spatial interaction with accurate position information without bringing excessive calculation amount, and decompose global averaging, so that the detection accuracy of a small model and the channel of the network are improved, the detection capability of detecting small objects and the aircraft skin can not be applied to accurately detecting skin damage of the aircraft when the aircraft is detected, and the aircraft skin can not be detected in real-time.
S4, inputting the acquired enhanced training set data into an optimized YOLOv5 model to obtain a trained best.
As a preferred embodiment of step S4, S4 specifically includes: placing the optimized YOLOv5 model into a configured Python environment, training the optimized YOLOv5 model by using the marked pictures in the training set after data enhancement, and accelerating training the model through a GPU.
And S5, applying the trained weight file to the picture detection of the test set, visualizing a detection result through the Pyside6, training the test set by using the weight file, installing the Pyside6 in a compiler Pycharm, and outputting the picture after detection and identification. Wherein the Pyside6 is a graphical interface (GUI) library of Python developed from Qt in C++ version. The Pyside is a free and commercially available software.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, the different embodiments or examples described in this specification and the features of the different embodiments or examples may be combined and combined by those skilled in the art without contradiction.
Logic and/or steps represented in the flowcharts or otherwise described herein, e.g., a ordered listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions.
The foregoing embodiments have been presented in a detail description of the invention, and are presented herein with a particular application to the understanding of the principles and embodiments of the invention, the foregoing embodiments being merely intended to facilitate an understanding of the method of the invention and its core concepts; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present invention, the present description should not be construed as limiting the present invention in view of the above.
Claims (7)
1. A method for detecting and classifying skin defects of an aircraft based on YOLOv5, comprising the steps of:
s1, acquiring an original image of aircraft skin damage, classifying the acquired image according to different damage, and dividing the classified image into a training set and a testing set;
s2, performing information interpolation on the collected damaged pictures of the training sets with different angles by using a bilinear interpolation algorithm, and further obtaining an optimized image so as to enhance the training set data;
s3, optimizing the original YOLOv5 model to obtain an optimized YOLOv5 model;
s4, inputting the acquired enhanced training set data into an optimized YOLOv5 model to obtain a trained weight file;
s5, applying the trained weight file to picture detection of the test set, and visualizing a detection result through a Pyside 6.
2. The YOLOv 5-based method of detecting aircraft skin defects and classification of claim 1, wherein: the S1 specifically comprises the following steps: classifying the original images of the aircraft skin damage into damage images of corrosion, impact and crack, marking the anchor frames, recording the sequence of the categories by using a TXT file, and marking the anchor frames by using a Labelimg target detection marking tool.
3. The YOLOv 5-based method of detecting aircraft skin defects and classification of claim 1, wherein: in S2, the specific process includes the following steps:
s21, performing interpolation twice in the x direction to obtain pixel values R1 and R2 after the interpolation twice, and performing interpolation once in the y direction to obtain pixel values in the y direction; the formula for the two interpolations in the x direction is:
;
;
wherein Q is 11 (x 1 ,y 1 ),Q 12 (x 1 ,y 2 ),Q 21 (x 2 ,y 1 ),Q 22 (x 2 ,y 2 ) Four points in an x-y plane rectangular coordinate system;
s22, performing interpolation once in the y direction to obtain a pixel value in the y direction, wherein an interpolation formula in the y direction is as follows:
;
s23, substituting the two formulas in S21 into the formulas in S22 can obtain a multi-view interpolation image:
。
4. the YOLOv 5-based method of detecting aircraft skin defects and classification of claim 1, wherein: in S3, the optimizing the original YOLOv5 model includes the following steps:
s31, introducing a new attention mechanism CA at a backbone network layer of the backhaul, wherein the new attention mechanism CA divides an input training set in x and y directions into two independent directions, codes transverse and longitudinal spatial position information into channel attention, and codes transverse and longitudinal characteristics by using pooling cores 1, W and H,1, so that the output of a c-th dimension characteristic of a height H is as follows:
;
similarly, the output of the c-th dimension of the width w is:;
wherein,is the height of the feature map; />Is the width of the feature map;x c is an intermediate feature; />,/>,/>,/>Is a known coordinate;
s32, integrating the features of the two formulas from different directions of transverse and longitudinal directions, and outputting a pair of feature graphs with known directionsFeature map +.>Connected to obtain a connected characteristic diagramF 1 And operates on it using a convolution change function to obtain an intermediate profile comprising lateral and longitudinal spatial information +.>The formula is as follows:
;
wherein,is a reduction factor;
s33, respectively encoding the information in the x and y directions into different attention diagrams along the space dimension, and mapping the intermediate feature diagram containing the transverse and longitudinal space informationPerforming a segmentation operation into tensors f in two directions h And f w The final high and wide attention weights are then obtained by a sigmoid activation functiong h Andg w the expression is as follows:
;
;
wherein,,/>is a convolution kernel of 1*1,>activating a function for sigmoid;
s34, CA attention mechanism formula can be expressed as:。
5. the YOLOv 5-based method of detecting aircraft skin defects and classification of claim 4, wherein: in S3, the optimizing the original YOLOv5 model further includes: the DIoU-NMS algorithm is used for replacing the NMS algorithm in the original model, the DIoU-NMS not only considers IoU, but also considers the distance between the centers of two frames, if the distance between the two frames is large, the algorithm can consider that the two frames are frames of two objects, so that filtering is avoided, and for optimizing the NMS in the model, only IoU in the NMS in the original model is replaced by DIoU, and the formula is as follows:
;
wherein,,/>representing the center points of the predicted and real frames respectively,ρrepresenting the calculation of the Euclidean distance between two central points,/->Representing the diagonal distance of the minimum closure area that can contain both the predicted and real frames; the formula of the DIoU-NMS is shown below:
;
wherein s is i Is the classification confidence;Mrepresenting a higher scoring bounding box in the prediction box; epsilon is the threshold of the DIoU-NMS,traversing the superposition condition of each frame with high confidence coefficient; r is R DIoU Is a penalty formula.
6. The YOLOv 5-based method of detecting aircraft skin defects and classification of claim 4, wherein: in S3, the optimizing the original YOLOv5 model further includes: changing the loss function in the original model to CIoU, wherein the loss function is increased in loss function of detection scale based on DIoU, and increased in loss function of length and width, and the difference of center distance, width and height of the boundary frame is considered, so that regression of the target frame is more stable, and the formula is as follows:
;
;
;
;
wherein a, B can be seen as two sets;is a weight function; />To measure the similarity of aspect ratios.
7. The YOLOv 5-based method of detecting aircraft skin defects and classification of claim 1, wherein: the step S4 specifically comprises the following steps: placing the optimized YOLOv5 model into a configured Python environment, training the optimized YOLOv5 model by using the marked pictures in the training set after data enhancement, and accelerating training the model through a GPU.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410211173.1A CN117788471B (en) | 2024-02-27 | 2024-02-27 | YOLOv 5-based method for detecting and classifying aircraft skin defects |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410211173.1A CN117788471B (en) | 2024-02-27 | 2024-02-27 | YOLOv 5-based method for detecting and classifying aircraft skin defects |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117788471A true CN117788471A (en) | 2024-03-29 |
CN117788471B CN117788471B (en) | 2024-04-26 |
Family
ID=90393057
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410211173.1A Active CN117788471B (en) | 2024-02-27 | 2024-02-27 | YOLOv 5-based method for detecting and classifying aircraft skin defects |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117788471B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN118052811A (en) * | 2024-04-10 | 2024-05-17 | 南京航空航天大学 | NAM-DSSD model-based aircraft skin defect detection method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114119586A (en) * | 2021-12-01 | 2022-03-01 | 中国计量大学 | Intelligent detection method for aircraft skin defects based on machine vision |
CN115063725A (en) * | 2022-06-23 | 2022-09-16 | 中国民航大学 | Airplane skin defect identification system based on multi-scale self-adaptive SSD algorithm |
US20230351573A1 (en) * | 2021-03-17 | 2023-11-02 | Southeast University | Intelligent detection method and unmanned surface vehicle for multiple type faults of near-water bridges |
CN117252842A (en) * | 2023-09-27 | 2023-12-19 | 中国民航大学 | Aircraft skin defect detection and network model training method |
-
2024
- 2024-02-27 CN CN202410211173.1A patent/CN117788471B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20230351573A1 (en) * | 2021-03-17 | 2023-11-02 | Southeast University | Intelligent detection method and unmanned surface vehicle for multiple type faults of near-water bridges |
CN114119586A (en) * | 2021-12-01 | 2022-03-01 | 中国计量大学 | Intelligent detection method for aircraft skin defects based on machine vision |
CN115063725A (en) * | 2022-06-23 | 2022-09-16 | 中国民航大学 | Airplane skin defect identification system based on multi-scale self-adaptive SSD algorithm |
CN117252842A (en) * | 2023-09-27 | 2023-12-19 | 中国民航大学 | Aircraft skin defect detection and network model training method |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN118052811A (en) * | 2024-04-10 | 2024-05-17 | 南京航空航天大学 | NAM-DSSD model-based aircraft skin defect detection method |
CN118052811B (en) * | 2024-04-10 | 2024-06-11 | 南京航空航天大学 | NAM-DSSD model-based aircraft skin defect detection method |
Also Published As
Publication number | Publication date |
---|---|
CN117788471B (en) | 2024-04-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111862126B (en) | Non-cooperative target relative pose estimation method combining deep learning and geometric algorithm | |
CN117788471B (en) | YOLOv 5-based method for detecting and classifying aircraft skin defects | |
Biasutti et al. | Lu-net: An efficient network for 3d lidar point cloud semantic segmentation based on end-to-end-learned 3d features and u-net | |
CN110633661A (en) | Semantic segmentation fused remote sensing image target detection method | |
CN111985376A (en) | Remote sensing image ship contour extraction method based on deep learning | |
CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
CN110910437B (en) | Depth prediction method for complex indoor scene | |
CN112347987A (en) | Multimode data fusion three-dimensional target detection method | |
CN114758337B (en) | Semantic instance reconstruction method, device, equipment and medium | |
CN113838064B (en) | Cloud removal method based on branch GAN using multi-temporal remote sensing data | |
US11544898B2 (en) | Method, computer device and storage medium for real-time urban scene reconstruction | |
CN113191204B (en) | Multi-scale blocking pedestrian detection method and system | |
CN115830004A (en) | Surface defect detection method, device, computer equipment and storage medium | |
CN116612468A (en) | Three-dimensional target detection method based on multi-mode fusion and depth attention mechanism | |
CN115409789A (en) | Power transmission line engineering defect detection method based on image semantic segmentation | |
Laupheimer et al. | The importance of radiometric feature quality for semantic mesh segmentation | |
CN109741358B (en) | Superpixel segmentation method based on adaptive hypergraph learning | |
CN112749741B (en) | Hand brake fastening fault identification method based on deep learning | |
CN117115359B (en) | Multi-view power grid three-dimensional space data reconstruction method based on depth map fusion | |
Wang et al. | Detection of plane in remote sensing images using super-resolution | |
CN116129234A (en) | Attention-based 4D millimeter wave radar and vision fusion method | |
CN112232102A (en) | Building target identification method and system based on deep neural network and multitask learning | |
CN106909936A (en) | A kind of vehicle checking method based on double vehicle deformable part models | |
CN115641510B (en) | Remote sensing image ship detection and identification method | |
CN115953589B (en) | Engine cylinder block aperture size measurement method based on depth camera |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |