An aircraft fuel tank drain port recognition and positioning method based on an improved YOLOv7 model

Abstract

The present application relates to a kind of based on improved YOLOv7 model's aircraft fuel tank drain port identification positioning method, belong to artificial intelligence technical field, including obtaining the picture containing aircraft fuel tank drain port;Picture is preprocessed, obtain available aircraft fuel tank drain port picture;To fuel tank drain port data set is labeled, obtain aircraft fuel tank drain port data set;Coordinate Attention structure and YOLOv7 network model are combined, YOLOv7 network model is improved, constructs YOLO_CA network model;Using aircraft fuel tank drain port data set, improved YOLO_CA network model is trained;Using the YOLO_CA network model of well-trained aircraft fuel tank drain port image to be identified is identified and positioned.The present application can be used to identify and position aircraft fuel tank drain port, then utilize inspection robot to complete aircraft fuel tank automatic water release function.

Description

TECHNICAL FIELD

[0001] The application provides an airplane fuel tank drain port recognition and positioning method based on an improved YOLOv7 model, relates to the field of computer vision, and belongs to the field of artificial intelligence. BACKGROUND

[0002] As a fast and convenient means of transportation, an airplane has become a generally accepted mode of transportation. Flight maintenance and inspection work before and after the takeoff of the airplane is an important link for ensuring flight safety and cannot be ignored. Among them, airplane fuel tank drain port maintenance and inspection is one of the work that must be done before flight. At present, the work is mainly completed by ground staff using a traditional method. The work is time-consuming and highly repetitive, and generally takes several hours to complete. Not only a large amount of manpower is consumed, but also long-term maintenance work can cause fatigue of the staff and thus lead to work errors and omissions. Therefore, an intelligent system that can assist human beings in the work has important social value and economic value.

[0003] In recent years, with the continuous improvement of computer data processing capability, artificial intelligence technology has developed rapidly. As a part of the artificial intelligence system, computer vision is gradually integrated into the production and life of various industries. Compared with pure manual maintenance and inspection, computer vision has low cost, high efficiency and stable operation, and can play a good auxiliary role for human beings. Therefore, many airports begin to invest efforts in researching the combination of computer vision technology and airplane maintenance work, so as to liberate human beings from tedious work.

[0004] Target detection is a basic work in the field of computer vision. In the field of deep learning, target detection can be mainly divided into two categories, one is one-stage target detection algorithm, and the other is two-stage target detection algorithm. The main algorithms of the two-stage target detection algorithm include RCNN, Fast RCNN, Faster RCNN and the like. The main feature of the two-stage target detection algorithm is to divide target detection into two stages of region generation and detection, and the accuracy is high, but the real-time performance is general. The main algorithms of the one-stage target detection algorithm include SSD network, YOLO series network and the like. The main feature of the one-stage target detection algorithm is to realize direct detection of a picture, and the real-time performance is better, but the accuracy is not as high as that of the two-stage target detection algorithm.

[0005] The YOLO (You Only Look Once) series algorithm, as the most representative algorithm in the One-stage algorithm, uses a unified network structure to complete the positioning and classification of objects. The YOLOv7 model is the latest improved algorithm in the YOLO series algorithm, and its detection speed and accuracy have exceeded most known object detectors. Since the airport environment where the aircraft is located is complex and variable, and the size of the aircraft water drain is relatively small, the detection effect of directly applying the YOLOv7 model is general. SUMMARY

[0006] To assist human beings in performing aircraft maintenance inspection work, the present application provides an aircraft fuel tank water drain identification and positioning method based on an improved YOLOv7 model, which can be used to identify and position the aircraft fuel tank water drain, and then use a patrol robot to complete the automatic water draining function of the aircraft fuel tank.

[0007] The technical solution of the present application is as follows:

[0008] An aircraft fuel tank water drain identification and positioning method based on an improved YOLOv7 model, comprising the following steps:

[0009] Step S1: Obtain a picture containing an aircraft fuel tank water drain;

[0010] Step S2: Preprocess the picture obtained in step S1 to obtain a usable aircraft fuel tank water drain picture;

[0011] Step S3: Use the labelImg image labeling tool to label the fuel tank water drain dataset, divide the labeled dataset into training, validation and test sets according to the ratio of 7:2:1, and obtain the aircraft fuel tank water drain dataset;

[0012] Step S4: Combine the Coordinate Attention structure with the YOLOv7 network model to improve the YOLOv7 network model and construct the YOLO_CA network model;

[0013] Step S5: Train the improved YOLO_CA network model using the aircraft fuel tank water drain dataset;

[0014] Step S6: Identify and position the aircraft fuel tank water drain image to be identified using the trained YOLO_CA network model.

[0015] Preferably, step S1 is specifically as follows:

[0016] A large number of aircraft fuel tank drain pictures are acquired by using a camera, and the aircraft fuel tank drain is required to be at each position of the picture when the picture is taken, and meanwhile, the light conditions of the background when the picture is taken are required to be rich and different angles are required to be shot.

[0017] Preferably, the preprocessing process of step S2 includes rotating, flipping, cropping and changing the color saturation of the original picture, and finally the number of pictures reaches about 2500.

[0018] Preferably, step S3 is specifically:

[0019] The position and size of the aircraft fuel tank drain in each picture are labeled by using the labelImg image labeling tool, and the label information is saved in a txt file with the same name as the picture. After completing the labeling work of all pictures, all pictures and labeled labels are divided into training set, validation set and test set according to the ratio of 7:2:1, and the aircraft fuel tank drain dataset is obtained.

[0020] Preferably, step S4 is specifically:

[0021] First, add the Coordinate Attention structure code module to the cmmon.py file in the YOLOv7 network model; then add the CA module to the parse_model() function in the yolo.py file in the YOLOv7 network model; finally, add the CA module to the backbone network in the network structure YOLOv7.yaml file in the YOLOv7 network model, and the addition position is after the ELAN module of the YOLOv7 network, which is used to enhance the feature extraction capability of the backbone network.

[0022] Preferably, step S5 is specifically:

[0023] The YOLO_CA network model is trained by using the aircraft fuel tank drain dataset obtained, and the YOLO_CA weight file is obtained.

[0024] Preferably, the CA module is a CA(Coordinate Attention) structure coding attention module, and its process includes two steps of Coordinate information embedding and Coordinate Attention generation:

[0025] The Coordinate information embedding part first decomposes the global pooling according to the following formula, and converts it into a one-to-one dimensional feature coding operation:

[0026]

[0027] In the above formula, Z cThe meaning of is the output related to the c-th channel; H, W represent the size of the pooling kernel respectively, X c represents the given input information;

[0028] Specifically, given the input X, first use the pooling kernel with size (H, 1) or (1, W) to encode each channel along the horizontal and vertical coordinates respectively; therefore, the output of the c-th channel with height h can be expressed as:

[0029]

[0030] Similarly, the output of the c-th channel with width w can be written as:

[0031]

[0032] The above two transformations aggregate features along two spatial directions respectively, obtaining a pair of direction-aware feature maps; these two transformations also allow the attention module to capture long-term relationships along one spatial direction and preserve precise location information along the other spatial direction, which helps the network to locate the target of interest more accurately;

[0033] The Coordinate Attention generation part, through the concatenate operation after the transformation in the Coordinate information embedding, then uses the 1×1 convolution transformation function F1 to transform it:

[0034]

[0035] The formula [] is the concatenate operation along the spatial dimension, is a nonlinear function, f is the intermediate feature mapping of the spatial information encoding in the horizontal and vertical directions, and then f is decomposed into two separate tensors f h ∈R C / r×H and f w ∈R C / r×W , f h and f w are transformed into tensors with the same number of channels by using another two 1×1 convolution transformations F h and F w respectively, to obtain

[0036] g h = σ (F h (f h ))

[0037] g w = σ (F w (f w))

[0038] Here the sigmoid activation function is used, and the reduction ratio r is used to reduce the channel number of f, and then the output g h and g w is expanded as attention weights respectively; finally, the output Y of the Coordinate Attention Block can be written as:

[0039]

[0040] Preferably, the YOLOv7 network model is composed of two parts, namely a Backbone part and a Head part, wherein the Backbone part is responsible for feature extraction, and the Head part is responsible for multi-scale fusion of the extracted features and prediction using a detection head.

[0041] The Backbone part is composed of a CBS module, an ELAN module and an MP module. The CBS module is composed of a convolution layer, a batch normalization layer and an activation function layer, and this module is the most basic module in the YOLOv7 network model. The ELAN module is composed of different CBS modules stacked together. Without changing the width and height of the input feature layer, it strengthens the mutual relationship between each feature layer through expansion, splicing and random combination, and improves the learning and training ability of the model. The MP module is composed of Maxpool and CBS module double path combination. It compresses the width and height of the input feature layer to expand the receptive field, and then performs feature fusion with the feature information processed by normal convolution, thereby improving the generalization ability of the network.

[0042] The Head part is composed of a SPPCSPC module, an UP module, a RepConv module and an ELAN-H module. The SPPCSPC module is based on the SPP module, and learns from the idea of CSPLayer to add a "residual edge". In this "residual edge", multiple MaxPool operations in parallel are added to avoid the problem of feature information loss caused by image detection in the network, and also solve the problem of repeated image feature extraction by CNN. The UP module is composed of a CBS module and an upsampling function, and its purpose is to facilitate feature fusion of feature layers. The ELAN-H module has basically the same effect as the ELAN module, except that the splicing of the output feature layer is different.

[0043] YOLOv7 network model is selected as the basic framework, and CA module is added after the last 1x1 convolution in the ELAN module of the backbone network part of the YOLOv7 network model to obtain the ELAN_CA module. The backbone network part of the original YOLOv7 model is added with an attention mechanism, thereby obtaining the improved YOLOv7 network model YOLO_CA.

[0044] The application adopts the combination of the Coordinate Attention structure and the YOLOv7 network model, and the speed and accuracy of the YOLOv7 network model in the range of 5 frames / second to 160 frames / second have exceeded most of the known target detectors. On a GPU V100, among all the real-time object detectors known at present, the average accuracy of YOLOv7 is the highest among all the real-time object detectors above 30 frames / second. Since the airport environment where the aircraft is located is complex and variable, and the size of the aircraft water drain is relatively small, in the process of identifying and detecting the target, the complex background information will interfere with the target to be identified. The Coordinate Attention attention mechanism structure is a method proposed in CVPR in 2021, which is a lightweight attention mechanism that considers the spatial dimension and the channel dimension as a whole, mainly designed for mobile networks. Compared with SE Net and CBAM, the CA attention mechanism has the following advantages: first, compared with the SE Net module, the CA module not only pays attention to the channel dimension information, but also considers the spatial position information; second, compared with the CBAM module, the CA module solves the long-distance dependence problem that the CBAM module has not solved. Therefore, the combination of the YOLOv7 network model and the Coordinate Attention attention mechanism can further improve the model accuracy on the basis of ensuring the recognition speed.

[0045] The details of the application can be seen from the prior art.

[0046] The beneficial effects of the application are:

[0047] The application provides an aircraft fuel tank water drain identification and positioning method based on an improved YOLOv7 network model, which uses a large amount of data to train the model, can finally achieve higher accuracy and faster speed than traditional methods, uses the Coordinate Attention structure to better extract the effective features of the detected object, reduces the influence of environmental factors such as background, and can obtain higher robustness. For the single-stage YOLOv7 network model, it can also be smoothly deployed on the edge device for operation, which facilitates the work of the inspection robot. BRIEF DESCRIPTION OF DRAWINGS

[0048] The drawings accompanying the specification of this application form a part of this application and serve to further understand the application, the illustrative embodiments of the application and the description thereof serve to explain the application, and do not constitute an improper limitation on the application.

[0049] Figure 1 The Coordinate Attention structure encoding attention process provided for the embodiments of the application.

[0050] Figure 2 (a) and (b) are ELAN module and ELAN_CA module of YOLOv7 model provided by the embodiment of the application respectively.

[0051] Figure 3 The flowchart of the aircraft water drain port recognition and positioning method based on the improved YOLOv7 network model provided by the embodiment of the application.

[0052] Figure 4 The training effect comparison chart of the YOLO_CA model and the YOLOv7 model provided by the embodiment of the application.

[0053] Figure 5 The recognition effect chart of the YOLO_CA model provided by the embodiment of the application, wherein (a), (b) and (c) are the recognition results of three different pictures. DETAILED DESCRIPTION

[0054] In order to enable the personnel in the technical field to better understand the technical solutions in the specification, the technical solutions in the embodiment of the specification will be described clearly and completely in conjunction with the drawings in the embodiment of the specification, but not limited to this, the unexplained in the specification is according to the conventional technology in the field.

[0055] Embodiment 1

[0056] An aircraft water drain port recognition and positioning method based on an improved YOLOv7 model, as shown in FIG. 1, comprises the following steps: Figures 1-5

[0057] Step S1: obtaining pictures containing aircraft water drain ports;

[0058] A large number of aircraft water drain port pictures are obtained by using a camera, and when the pictures are taken, the aircraft water drain ports are required to be at various positions in the pictures, and at the same time, the light conditions of the background when the pictures are taken should be as rich as possible, and the pictures should be taken at different angles as much as possible.

[0059] Step S2: preprocessing the pictures obtained in step S1 to obtain available aircraft water drain port pictures;

[0060] The preprocessing process of step S2 includes rotating, flipping, cropping and changing the color saturation of the original pictures, etc., and finally the number of pictures reaches about 2500.

[0061] Step S3: using the labelImg image labeling tool to label the water tank drain port data set, dividing the labeled data set into training set, verification set and test set according to the ratio of 7:2:1, and obtaining the aircraft water drain port data set, which is specifically:

[0062] ​The label information of the position and size of the aircraft fuel tank drain in each picture is labeled by using the labelImg image labeling tool, and saved in a txt file with the same name as the picture. After completing the labeling work of all pictures, all pictures and labeled labels are divided into training set, validation set and test set according to the ratio of 7:2:1, and the aircraft fuel tank drain dataset is obtained.

[0063] Step S4: combining the Coordinate Attention structure and the YOLOv7 network model to improve the YOLOv7 network model, and constructing the YOLO_CA network model, specifically:

[0064] First, add the Coordinate Attention structure code module to the cmmon.py file in the YOLOv7 network model; then add the CA module to the parse_model() function in the yolo.py file in the YOLOv7 network model; finally, add the CA module to the backbone network in the network structure YOLOv7.yaml file in the YOLOv7 network model, and the addition position is after the YOLOv7 network ELAN module, which is used to enhance the feature extraction capability of the backbone network.

[0065] The YOLOv7 network model is used to recognize and locate the aircraft fuel tank drain; target detection is usually a combination of two tasks, one is image classification, and the other is image positioning. The YOLO series network model is a single-stage target detection network model, which uses the powerful feature extraction capability of convolutional neural network to extract features, and then uses direct regression to infer the category of the predicted target and the specific position of the predicted target in the image; unlike two-stage target detection network model, it is faster, the model is more lightweight, and it is easier to deploy on edge devices; compared with the traditional manually designed feature extractor, it has stronger robustness, so it is feasible to use the YOLOv7 network model as the aircraft fuel tank drain recognition and positioning target detection framework.

[0066] The CA module is a CA(Coordinate Attention) structure coding attention module, as shown in Figure 1 , which encodes the channel relationship and long-term dependency through precise position information, and the specific operation includes two steps of Coordinate information embedding and Coordinate Attention generation;

[0067] The Coordinate information embedding part first decomposes the global pooling according to the following formula, which is converted into a one-to-one dimensional feature encoding operation:

[0068]

[0069] In the above formula, Z c means the output related to the c-th channel; H, W represent the size of the pooling kernel respectively, X c represents the given input information;

[0070] Specifically, given the input X, first use the pooling kernel with size (H, 1) or (1, W) to encode each channel along the horizontal coordinate and the vertical coordinate respectively; therefore, the output of the c-th channel with height h can be expressed as:

[0071]

[0072] Similarly, the output of the c-th channel with width w can be written as:

[0073]

[0074] The above two transformations aggregate the features along two spatial directions respectively, obtaining a pair of direction-aware feature maps; these two transformations also allow the attention module to capture long-term relationships along one spatial direction and preserve precise location information along the other spatial direction, which helps the network to locate the target of interest more accurately;

[0075] The Coordinate Attention generation part concatenates the transformed results in the Coordinate information embedding, and then uses a 1x1 convolutional transformation function F1 to transform it:

[0076]

[0077] The formula [] is the concatenate operation along the spatial dimension, is a nonlinear function, f is the intermediate feature mapping of the spatial information encoding in the horizontal direction and the vertical direction, and then f is decomposed into two separate tensors f h ∈R C / r×H and f w ∈R C / r×W , f h and f w are transformed into tensors with the same number of channels using another two 1x1 convolutional transformations F h and F w respectively to the input X, obtaining

[0078] g h =σ(F h (f h ))

[0079] gw = σ(F w (f w ))

[0080] Here the sigmoid activation function is used, in order to reduce the complexity and computational overhead of the model, here we usually use a proper reduction ratio r to reduce the channel number of f, then expand the output g h and g w as attention weights respectively; finally, the output Y of the Coordinate Attention Block can be written as:

[0081]

[0082] The YOLOv7 network model consists of two parts, namely the Backbone part and the Head part, wherein the Backbone part is mainly responsible for feature extraction, and the Head part is mainly responsible for multi-scale fusion of the extracted features and prediction using the detection head;

[0083] The Backbone part is composed of CBS Convolution Batch Normalization SiLU) module, ELAN module and MP module, the CBS module is composed of convolution layer, batch normalization layer and activation function layer, this module is the most basic module in the YOLOv7 network model; the structure of the ELAN module is as shown in Figure 2 (a), this module is stacked by different CBS modules using a specific structure, which strengthens the mutual relationship between each feature layer by expansion, splicing and random combination without changing the width and height of the input feature layer, and improves the learning and training ability of the model; the MP module is composed of Maxpool and CBS module double path combination, which expands the receptive field by compressing the width and height of the input feature layer, and then performs feature fusion with the feature information processed by normal convolution, thereby improving the generalization ability of the network;

[0084] The Head part is composed of SPPCSPC module, UP module, RepConv module and ELAN-H module, the SPPCSPC module is based on the SPP module, and learns from the idea of CSPLayer, adding a "residual edge", in which parallel multiple MaxPool operations are added, which avoids the problem of feature information loss caused by image in the network detection process, and also solves the problem of repeated image feature extraction by CNN; the UP module is composed of CBS module and up-sampling function, the purpose is to facilitate feature fusion of feature layers; the ELAN-H module has basically the same effect as the ELAN module, except that the splicing of the output feature layer is different;

[0085] YOLOv7 network model is selected as the basic framework, and a CA module is added after the last 1x1 convolution in the ELAN module of the backbone network part of the YOLOv7 network model, to obtain an ELAN_CA module, as shown in Figure 2 (b), so that the backbone network part of the original YOLOv7 model is added with an attention mechanism, thereby obtaining an improved YOLOv7 network model YOLO_CA.

[0086] In the YOLO_CA model, due to the addition of the Coordinate Attention structure, the global receptive field of the network is enhanced, so that the target recognition ability of the YOLOv7 model is stronger, and the recognition accuracy of the network is improved.

[0087] In this embodiment, the YOLOv7 model and the YOLO_CA model are trained by using the aircraft fuel tank drain data set, and the training results are as shown in Figure 4 Figure 4 It can be seen that the target recognition accuracy of the YOLO_CA model is obviously improved compared with the YOLOv7 network model.

[0088] Step S5: training the improved YOLO_CA network model by using the aircraft fuel tank drain data set, to obtain a YOLO_CA weight file;

[0089] Step S6: using the trained YOLO_CA network model to recognize and position the aircraft fuel tank drain image to be recognized, and the recognition effect diagram is as shown in Figure 5 (a), (b) and (c) are the recognition results of three different pictures, and the objects in the boxes are the positions of the aircraft fuel tank drains recognized by this embodiment.

[0090] The framework model YOLOv7- used in this embodiment is established based on the Pytorch deep learning framework, and the system for training and running is Windows 10. This embodiment uses the CUDA (Compute Unified Device Architecture) operation platform 11.8 version developed by Nvidia company to cooperate with CuDnn for GPU training, and the python language is written in Pycharm.

[0091] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.​

Claims

1. An improved YOLOv7 model-based aircraft fuel tank drain port identification and positioning method, characterized in that, The steps include the following: Step S1: Obtain a picture containing an aircraft fuel tank drain; Step S2: Preprocess the picture obtained in step S1 to obtain a usable aircraft fuel tank drain picture; Step S3: Label the fuel tank drain dataset using the labelImg image labeling tool, divide the labeled dataset into training, validation and test sets according to a ratio of 7:2:1, and obtain the aircraft fuel tank drain dataset; Step S4: Combine the Coordinate Attention structure with the YOLOv7 network model to improve the YOLOv7 network model and build the YOLO_CA network model; The construction process of the YOLO_CA network model is as follows: the YOLOv7 network model is selected as the basic framework, the CA module is added after the last 1x1 convolution in the ELAN module of the backbone network part of the YOLOv7 network model, the ELAN_CA module is obtained, and the attention mechanism is added to the original YOLOv7 model backbone network part, thereby obtaining the improved YOLOv7 network model YOLO_CA; Step S5: Train the improved YOLO_CA network model using the aircraft fuel tank drain dataset; Step S6: Use the trained YOLO_CA network model to recognize and locate the aircraft fuel tank drain image to be recognized.

2. The aircraft water drain port recognition and positioning method based on the improved YOLOv7 model according to claim 1, characterized in that, Step S1 is specifically: A large number of aircraft fuel tank drain pictures are obtained using a camera. When taking the picture, the aircraft fuel tank drain is required to be at different positions in the picture, and at the same time, the background lighting conditions during picture taking need to be rich and different angles need to be taken.

3. The method of claim 2, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, The preprocessing process of step S2 includes rotating, flipping, cropping and changing the color saturation of the original picture, and finally the number of pictures reaches 2500.

4. The method of claim 3, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, Step S3 is specifically: The labelImg image labeling tool is used to label the position and size of the aircraft fuel tank drain in each picture, and the label information is saved in a txt file with the same name as the picture. After completing the labeling work of all pictures, all pictures and labeled labels are divided into training, validation and test sets according to a ratio of 7:2:1, and the aircraft fuel tank drain dataset is obtained.

5. The method of claim 4, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, Step S4 is specifically: First, add the Coordinate Attention structure code module to the cmmon.py file in the YOLOv7 network model; then add the CA module to the parse_model() function in the yolo.py file in the YOLOv7 network model; finally, add the CA module to the backbone network in the network structure YOLOv7.yaml file in the YOLOv7 network model, and the addition position is after the YOLOv7 network ELAN module, which is used to enhance the feature extraction capability of the backbone network.

6. The method of claim 5, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, Step S5 is specifically: The modified YOLO_CA network model is trained using the aircraft fuel tank drain dataset obtained, and the YOLO_CA weight file is obtained.

7. The method of claim 6, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, The CA module is a CA structure encoding attention module, and its process includes two steps of Coordinate information embedding and Coordinate Attention generation. The Coordinate information embedding part first decomposes the global pooling according to the following formula, which is converted into a one-dimensional feature encoding operation: In the above formula, Z c means the output related to the cth channel; H, W represent the size of the pooling kernel, respectively, and X c represents the given input information; Specifically, given the input X, first use the pooling kernel with size (H, 1) or (1, W) to encode each channel along the horizontal and vertical coordinates respectively; therefore, the output of the c-th channel with height h is represented as: Similarly, the output of the c-th channel with width w is written as: The above two transformations respectively aggregate the features along the two spatial directions to obtain a pair of direction-aware feature maps. The Coordinate Attention generation part is concatenated after the transformation in the Coordinate information embedding part, and then uses a 1×1 convolution transformation function F1 to transform it: The equation [ , ] represents the concatenate operation along the spatial dimension, and ∂ is a nonlinear function. f An intermediate feature map is used to encode spatial information in the horizontal and vertical directions, and then f is decomposed into two separate tensors along the spatial dimension. and Using two other 1×1 convolution transformations F h and F w Each f h and f w Transform it into a tensor with the same number of channels as the input X, and obtain Here, the sigmoid activation function is used, and the reduction ratio r is used to reduce the number of channels of f, and then the output g h and g w are extended as attention weights, respectively; finally, the output Y of the Coordinate Attention Block is written as: 。 8. The method of claim 7, wherein the improved YOLOv7 model-based aircraft fuel tank drain port recognition and positioning method is characterized by, The YOLOv7 network model is composed of two parts, namely the Backbone part and the Head part, where the Backbone part is responsible for feature extraction, and the Head part is responsible for multi-scale fusion of the extracted features and prediction using the detection head. The Backbone part is composed of CBS modules, ELAN modules and MP modules. The CBS module is composed of convolution layers, batch normalization layers and activation function layers, and this module is the most basic module in the YOLOv7 network model. The ELAN module is stacked by different CBS modules, which strengthens the mutual relationship between each feature layer by expansion, splicing and random combination without changing the width and height of the input feature layer, and improves the learning and training ability of the model. The MP module is composed of Maxpool and CBS modules in double paths, which compresses the width and height of the input feature layer to expand the receptive field, and then performs feature fusion with the feature information processed by normal convolution, thereby improving the generalization ability of the network. The Head part is composed of SPPCSPC modules, UP modules, RepConv modules and ELAN-H modules. The SPPCSPC module is based on the SPP module, and borrows the idea of CSPLayer to add a "residual edge", in which parallel multiple MaxPool operations are added. The UP module is composed of CBS modules and an upsampling function. The ELAN-H module has basically the same effect as the ELAN module, except that the splicing of the output feature layer is different.

Images