Airport air control image target detection method based on improved yolov7

Abstract

The application relates to an airport air control image target detection method based on an improved yolov7, which comprises the following steps: collecting an airport air control image, pre-processing the airport air control image, labeling image features, and obtaining a labeled data set; an improved yolov7 network is used to construct an air control image target detection model; wherein the improved yolov7 network comprises an FEAM attention module; the improved air control image target detection model is trained based on the labeled data set, and target detection is performed through the trained air control image target detection model. The application constructs a data set conforming to the characteristics of an air control image based on network video interception of an airport image, increases an FEAM attention module at a position of a backbone network structure based on a yolov7 target detection structure, and solves the problem of lacking data in the research on the characteristics of an air control image.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of target detection in images, in particular to an airport air control image target detection method based on an improved yolov7. BACKGROUND

[0002] Nowadays, the aviation transportation industry is developing continuously, the flight volume is continuously increasing, the pressure on airport safety guarantee is increasing, and the general civil aviation industry is developing rapidly, which brings new challenges to airport safety management. Airport air control is an important part of airport safety management. However, manual processing of airport monitoring images or videos is unstable, complicated, time-consuming and expensive, so it is very meaningful to use computer vision technology to automatically detect key targets such as airplanes, pedestrians and cars in airport monitoring images or videos to assist airport management personnel in decision-making and improve the ability of airport safety management.

[0003] At present, typical target detection technologies mainly include two-stage detection and one-stage detection methods. The two-stage detection method first generates a detection region candidate frame, and then predicts the position frame and the category, such as the R-CNN, Fast R-CNN and Faster R-CNN target detection algorithms. The one-stage detection method generates a predicted position frame and a category in the detection network, such as the SDD and yolo target detection methods.

[0004] The above aspects have the following problems in constructing the air control image target detection: first, the above-mentioned target detection methods are improved and optimized for common data sets such as PASCAL VOC and MS COOC, and there is a lack of data sets for air control images. Second, in the air control image, the two-stage target detection method is slow in detecting the air control image, and the one-stage target detection method has weak detection ability for the air control image. The application provides an airport air control image target detection method based on an improved yolov7, constructs an air control image feature data set by intercepting airport images from network videos, increases an FEAM attention module at different network structure positions based on the yolov7 target detection structure, constructs a target detection method for air control images, and finally verifies the improved target detection method by using the above data set. SUMMARY

[0005] The application aims to provide an airport air control image target detection method based on an improved yolov7, and solve the problem of a lack of data for air control image features.

[0006] To achieve the above-mentioned purpose, the application provides the following scheme:

[0007] The airport air control image target detection method based on the improved yolov7 comprises the following steps:

[0008] collecting an airport air control image, preprocessing the airport air control image, labeling image features, and obtaining a labeled data set;

[0009] An improved yolov7 network is used to construct an air control image target detection model; wherein the improved yolov7 network includes a FEAM attention module.

[0010] The improved air control image target detection model is trained based on the labeled data set, and target detection is performed through the trained air control image target detection model.

[0011] Optionally, preprocessing the airport air control image includes:

[0012] Based on the airport air control image, an image data set is obtained, and the image data set is randomly divided into a training set and a test set.

[0013] Optionally, labeling image features to obtain a labeled data set includes:

[0014] The target position label is added by a labeling program Labelimg, and the aircraft, pedestrians and vehicles in each picture in the image data set are classified and labeled to construct the labeled data set.

[0015] The labeled data set includes the position information of the aircraft, the pedestrians and the vehicles in the image.

[0016] Optionally, the improved yolov7 network is used to construct an air control image target detection model, which includes:

[0017] The yolov7 backbone network in the yolov7 network is improved, and the FEAM attention module is added to the improved yolov7 backbone network to construct the air control image target detection model.

[0018] The air control image target detection model includes an improved yolov7 backbone network, a yolov7 neck network and a yolov7 head network; the improved yolov7 backbone network, the yolov7 neck network and the yolov7 head network are connected in sequence.

[0019] The yolov7 backbone network after adding the FEAM attention module includes a CBS structure, a FEAM attention module, an ELAN structure, an MP1 structure and a SPPCSPC structure; the CBS structure, the FEAM attention module, the ELAN structure, the MP1 structure and the SPPCSPC structure are connected in sequence.

[0020] Optionally, the FEAM attention module comprises an ECA submodule, an FReLU activation function and a SAM module, which are sequentially connected.

[0021] Optionally, the FReLU activation function depends on the spatial context 2D condition of each pixel to complete the visual task.

[0022] The method for completing the visual task comprises the following steps.

[0023] FReLU(x)=max(x,T(x))

[0024] Wherein, FReLU(x) is the name of the activation function algorithm, max(x,T(x)) is the maximum value of input x and T(x), x is the input of the activation function, and T(x) is the spatial context 2D condition.

[0025] Optionally, training the improved air control image target detection model based on the labeled data set comprises training the FEAM attention module.

[0026] Training the FEAM attention module comprises:

[0027] Based on the labeled data set, a first feature map is obtained, the first feature map is input into the ECA submodule, the ECA submodule is used to average pool the main input end to obtain the channel feature value of each channel, the channel feature value is used to realize the channel weight between the feature vectors by using a 1×1 convolution kernel, and the feature map channel weight is obtained by using a sigmoid activation function. The feature map channel weight is multiplied by the first feature map to obtain the output feature of the ECA submodule, the output feature is activated by the FReLU activation function and transmitted to the SAM module, the SAM module is used to compress the activated output feature in the channel dimension by using average pooling and maximum pooling to obtain a plurality of two-dimensional feature maps, the two-dimensional feature maps are spliced by using channel Concat to obtain a second feature map with a preset number of channels.

[0028] Optionally, training the FEAM attention module further comprises:

[0029] The second feature map is convolved by a hidden layer containing a single convolution kernel, and a spatial attention weight is generated by a sigmoid operation. The spatial attention weight is multiplied by the first feature map to obtain a feature map containing channel and spatial attention weighting, and the training of the FEAM attention module is completed.

[0030] Optionally, the method further comprises verifying the improved air control image target detection model.

[0031] Based on Recall, Precision, F1, mAP as the evaluation index of the model, and the threshold value of IOU is set to a preset value, the improved air control image target detection model is verified.

[0032] Wherein, the Recall is recall rate, the Precision is accuracy rate, the F1 is F1 value, the mAP is mean average precision, and the IOU is a set value.

[0033] The beneficial effects of the present application are:

[0034] The present application is based on network video interception airport image to construct the data set conforming to the characteristics of air control image, based on yolov7 target detection structure, increases the FEAM attention module on the position of the backbone network structure, solves the problem of lacking data for air control image feature research, and proposes a target detection algorithm for air control image. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 The flow chart of the airport air control image target detection method based on improved yolov7 of the embodiment of the present application;

[0037] Figure 2 The part of the data set sample collected by the embodiment of the present application;

[0038] Figure 3 The data sample target position label collected by the embodiment of the present application;

[0039] Figure 4 The FEAM module structure diagram of the embodiment of the present application;

[0040] Figure 5 The CAM module network structure of the embodiment of the present application;

[0041] Figure 6 The ECA module network structure of the embodiment of the present application;

[0042] Figure 7 The SAM module network structure of the embodiment of the present application;

[0043] Figure 8For this embodiment of the invention, FReLU is MAX(x,T(x)).

[0044] Figure 9 This is an improved YOLOv7 target detection network structure according to an embodiment of the present invention;

[0045] Figure 10 These are the experimental results of embodiments of the present invention;

[0046] Figure 11 The present invention describes the detection using the official YOLOv7 model, the model retrained using self-built data, and the improved algorithm in this embodiment; wherein, Figure (a) shows the official model, Figure (b) shows the model retrained using self-built data, and Figure (c) shows the improved YOLOv7 object detection algorithm. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] like Figure 1 As shown, the airport air traffic control image target detection method based on the improved YOLOv7 is characterized by the following steps:

[0050] Step 1: Preprocess the acquired images from the network to obtain an image dataset, such as... Figure 2 As shown.

[0051] Step 2: The preprocessing process is as follows: augment the image dataset and randomly divide the augmented image dataset into training and test sets;

[0052] Step 3: Label image features: Classify and label the fixed structures (airplanes, pedestrians, and vehicles) in each image in the dataset to construct a labeled dataset;

[0053] Step 4: Improve the YOLOv7 air traffic control image target detection model by adding a FEAM attention module to the backbone network;

[0054] Step 5: Train the improved YOLOv7 air traffic control image target detection model constructed in Step 4;

[0055] Step 6: The improved yolov7 air control image target detection model trained in step 5 is used to detect the test set, and the target detection evaluation index is verified.

[0056] Step 7: The weight parameters obtained by training are written into detect.py, and the improved model is run using the python framework to verify that the improved yolov7 model is more accurate and meets the requirements of air control image target detection.

[0057] The present application improves part of the network structure in yolov7, and constructs a target detection method for air control images. The present application uses network video to intercept airport images to construct a data set that meets the characteristics of air control images; design a FEAM attention module; in the yolov7 network, the main network is unchanged, and the FEAM attention module is added at the main position of the network structure; use the established air control image data set to train the network model; and verify it through the target detection evaluation index.

[0058] Establish air control data set

[0059] The present application uses common targets in the airport, mainly including airplanes, pedestrians and vehicles, and part of the data images. According to the airport style and target characteristics, a target position label data set is established, and the data set is labeled. The image target position label is constructed by using the Labelimg tool, and the image and the label are added together to form the data set. The position label, the labeled category is: airplane, person and car three categories, the position label format is two-point-determined position xml format label and center point plus height-width txt format label, as shown in the formula: Figure 3 The position information label uses the Labelimg labeling program to label the data set, and the tool is used to make the label. The image and the label together are used for labeling to construct the data set. A total of 6076 pictures are labeled, and the data contains 12122 airplane targets, 1194 pedestrian targets and 2647 vehicle targets.

[0060] Among them, the position information label uses the Labelimg labeling program to label the data set, and the labeling data set includes: the position information of the airplane, the pedestrian and the vehicle in the image.

[0061] The improved yolov7 air control image target detection model is:

[0062] An air traffic control image target detection model is constructed by using an improved yolov7 network, the yolov7 backbone network in the yolov7 network is improved, and a FEAM attention module is added to the improved yolov7 backbone network to construct the air traffic control image target detection model; wherein the air traffic control image target detection model comprises: an improved yolov7 backbone network, a yolov7 neck network and a yolov7 head network; the improved yolov7 backbone network, the yolov7 neck network and the yolov7 head network are connected in sequence; the yolov7 backbone network after adding the FEAM attention module comprises: a CBS structure, a FEAM attention module, an ELAN structure, an MP1 structure and a SPPCSPC structure; the CBS structure, the FEAM attention module, the ELAN structure, the MP1 structure and the SPPCSPC structure are connected in sequence, and specifically:

[0063] FEAM attention module

[0064] The attention network mechanism reduces useless information to increase detection capability by making the model more interested in a certain type of target. The FEAM attention module mainly includes three parts of ECA module, SAM module and FReLU activation function. For air traffic control images, assigning a larger weight to a region means that the region contains more detection targets, and the model will pay more attention to learning the features of the region, thereby improving the accuracy of the air traffic control image detection algorithm and better assisting airport management personnel in decision-making.

[0065] Inspired by the residual attention module and the CBAM attention module, the core idea of FEAM will adopt the residual module structure to obtain different dimension weight matrices along the channel and space dimensions of the input air traffic control image feature map Y. The input end of the FEAM attention module is changed from the main input to the joint action of the main input and the residual input, and the main input end is inspired by the CBAM attention module, replacing the parameter-rich CAM module with a lightweight ECA module, and using a FReLU activation function specially applied to visual tasks instead of a sigmoid activation function. The structural diagram is as follows: Figure 4As shown, first, the ECA module obtains the channel feature value of the main input through average pooling, and the channel feature value uses a 1x1 convolution kernel to realize the channel weight between the feature vectors, and experiences a sigmoid activation function to obtain the channel weight of the feature map, which is multiplied with the original feature and the feature map input into the FEAM module to obtain the output feature of the ECA module. The output feature is activated by a FReLU activation function and transmitted to the SAM module. The SAM module uses average pooling and maximum pooling to compress the feature map Yc in the channel dimension to obtain two two-dimensional feature maps, which are concatenated together based on the channel to obtain a feature map with a channel number of 2. In order to ensure that the final obtained feature is consistent with the input Yc in the spatial dimension, a hidden layer containing a single convolution kernel is used to perform convolution operation on the spliced feature map, and finally a sigmoid operation is performed to generate a spatial attention weight, which is multiplied with the initial input feature map to obtain a feature map containing channel and spatial attention weighting. The feature map is activated by a FReLU activation function, added to the residual input to obtain an output with better multi-scale information representation capability.

[0066]

[0067] wherein, is the output feature map of the FEAM module, and Y is the input feature map of the FEAM module.

[0068] Step 1: ECA channel attention module

[0069] The CAM module pays more attention to the semantic information of the feature map. The input feature map Y into the CAM module is the feature map, and the feature map Y is the input of the CAM module. In the spatial dimension, the average pooling integrates the information, and the maximum pooling collects the detailed information. The two kinds of pooling operations reduce the network calculation while improving the network expression ability. Finally, the channel attention O c is generated by the addition operation and the sigmoid activation operation. Figure 5

[0070] O c = sigmoid(MLP(MaxPool(Y))+MLP(AvgPool(Y))) (2)

[0071] wherein, O c is the output of the CAM module, sigmoid is the activation function, MLP(AvgPool(Y)) is the fully connected neural network, AvgPool is the global average pooling operation on the input, and MLP(MaxPool(Y)) is the global maximum pooling operation on the input.

[0072] Since CAM uses a fully connected operation to map the features, as shown in Figure 5 ​As shown in the network, the network parameter calculation is large when multiple CBAMs are inserted in the network, therefore, the CAM is improved in view of the problem that the large parameter quantity leads to large calculation quantity. The FEAM selects to use the ECA attention module instead of the traditional CAM, and the one-dimensional convolution with a convolution kernel length of 3 integrates the features of three channels in the channel neighborhood, and finally generates the attention weight Os through the sigmoid activation operation As shown in the structure Figure 6 .

[0073]

[0074] Wherein, AvgPool(Y) is a global average pooling operation on the input, and Conv1 is a one-dimensional convolution operation.

[0075] Step2: SAM spatial attention module

[0076] SAM pays more attention to the position information of the features, and focuses on the region with more effective features in the feature map, which is a supplement to the channel attention. The average pooling and the maximum pooling are used to compress the feature map Yc in the channel dimension to obtain two two-dimensional feature maps, which are concatenated based on the channel to obtain a feature map with a channel number of 2. In order to ensure that the final obtained feature is consistent with the input Yc in the spatial dimension, a hidden layer containing a single convolution kernel is used to perform convolution operation on the spliced feature map, and finally the sigmoid operation generates the spatial attention weight Os, as shown in the structure Figure 7 .

[0077] O s = sigmoid(Conv2(AvgPool(Y),MaxPool(Y))) (4)

[0078] Wherein, MaxPool(Y) is a global maximum pooling operation on the input, Conv2(AvgPool(Y) is a two-dimensional convolution operation, AvgPool is a global average pooling operation on the input, and sigmoid is an activation function.

[0079] Step3: FReLU activation function

[0080] In the convolutional neural network, the activation layer provides the network with nonlinearization ability, and the FReLU activation function is a special activation function designed for visual tasks. By using ordinary convolution to realize the function of the activation function, it has better ability to complete the visual image task. At present, the most widely used activation function is still the ReLU activation function, and its function expression is as follows:

[0081] ReLU(x) = max(0, x) (5)

[0082] Wherein, max(0, x) is the maximum function of two, ReLU is an activation function.

[0083] ReLU activation function has high accuracy in many tasks, FReLU inherits the idea of ReLU and expands to space, depends on the spatial context 2D condition T(x) of each pixel, realizes simple and only increases a small amount of calculation, better completes the visual task, and is expressed as follows, and the structure is as shown in Figure 8

[0084] FReLU = max(x, T(x)) (6)

[0085] Wherein, FReLU(x) is the name of the activation function algorithm, max(x, T(x)) is the maximum value of input x and T(x), x is the input of the activation function, and T(x) is the spatial context 2D condition.

[0086] Improved yolov7 network structure

[0087] In the improved yolov7 network, the yolov7 main network is unchanged, only the FEAM attention module is added before the ELAN module in the yolov7 backbone network, forming the F-ELAN structure, and the overall network structure is as shown in Figure 9 After the air tube feature map is extracted by the F-ELAN module with attention, it is distributed to a larger weight representing a region containing more detection targets, and the model will use more attention to learn the characteristics of the region, thereby improving the accuracy of the air tube image detection algorithm and better assisting the airport management personnel in decision-making. The amount of training parameters increased by adding the FEAM attention module in the yolov7 backbone network is reduced by 96% compared with adding the CBAM attention module.

[0088] Comparison table of the number of training parameters increased by adding the CBAM module and the FEAM module in the yolov7 backbone network, as shown in Table 1:

[0089] Table 1

[0090]

[0091] Overall, the airport air control image target detection method based on the improved yolov7 is similar to the yolov5 network structure, mainly replacing the internal components, and the network is composed of an improved yolov7 backbone network, a yolov7 neck network and a yolov7 head network.

[0092] The improved yolov7 backbone network extracts features and transmits them to the yolov7 neck network for feature fusion, and the fused features are transmitted to the yolov7 head network;

[0093] ​The backbone network comprises a CBS structure, a FEAM attention module, an ELAN structure, an MP1 structure and an SPPCSPC structure. The CBS structure is composed of a convolution module, a normalization module and a SiLU activation function. The air control image is input into the FEAM attention module, the ELAN structure and the MP1 structure through four CBS structures, and the output of the FEAM attention module, the ELAN structure and the MP1 structure is input into the FEAM attention module, the ELAN structure and the MP1 structure for three times, and finally the output of the FEAM attention module, the ELAN structure and the MP1 structure is input into the SPPCSPC structure to form the improved yolov7 backbone network. The ELAN structure mainly enhances the learning ability of the network by the shortest and longest gradient path without destroying the original gradient. The MP1 structure is composed of a maximum pooling module and a convolution with a step of 2, and the two downsampling methods are recombined to enhance the learning ability of the network without destroying the structure. The SPPCSPC structure divides the features into two parts, one of which is obtained by four different maximum pooling operations to distinguish large and small targets, and the other part uses the CBS structure to extract features, and finally the two parts are combined to improve the network precision.

[0094] The neck network and the head network comprise an FPN structure, a PAN structure, an ELAN-H structure and a REP structure. The FPN structure is a top-down feature pyramid, which improves the small target detection capability by using the up-sampling method. The PAN is a bottom-up feature pyramid, which transmits information from the lower layer to the upper layer to improve the detection capability of the occluded target. The ELAN-H structure is similar to the ELAN structure and also improves the shortest and longest gradient path to enhance the learning ability of the network. The REP structure is composed of a convolution kernel of 3, a convolution kernel of 1 and a three-branch topology of the kernel itself, and a large number of experiments prove that this structure has better detection precision and speed.

[0095] Based on the improved yolov7 target detection algorithm file and parameter configuration

[0096] The present application verifies the air control image target detection algorithm based on the constructed air control data set and the improved YOLOv7 target detection algorithm, and the specific steps are as follows:

[0097] Step 1: Configure network training configuration file

[0098] 5000 images are selected from the data set as the training set, and 1076 images are selected as the test set. Before training, the yolov7 data and model configuration file need to be changed and improved. In the data file, the number of object categories is changed to 3, and the object category name is modified in the category list names. In the model configuration file, the anchor box size is set to (12, 16), (19, 36), (40, 28), (36, 75), (76, 55), (72, 146), (142, 110), (192, 243), (459, 401). According to the improved network structure, the feature scale and attention module are added. In the common.py file, the FEAM model code is added, and the FEAM structure is introduced in the yolo.py file. Run the yolo.py file to detect the network change.

[0099] Step2: Configure network training environment

[0100] The test environment is the Dawning cloud computing service system. Each node is configured with an x86 processor with 32 cores and a 2.5GHz main frequency and an NVIDIA Tesla V100 accelerator card. Each node is configured with two 16GB DDR4 2666ECC REG memories, and two Dawning Parastor300S parallel storage systems are configured to provide large-capacity data storage. In terms of network communication, the cluster uses full-speed, non-blocking 200Gb HDR Infiniband dedicated computing network, uses pytorch1.9.0 deep learning framework, and the specific configuration information is shown in the table. Table 2 is the network test hardware environment configuration, Table 3 is the network test software part of the environment configuration, and Table 4 is the yolov7 improved model part parameter setting.

[0101] Table 2

[0102]

[0103] Table 3

[0104]

[0105] Table 4

[0106]

[0107] 2.5, based on the target detection evaluation index verification, and the air traffic control image target detection model after training is verified;

[0108] Step1: Target detection evaluation index selection

[0109] All models are trained and tested in the air control image data established by the patent. In the experiment, Recall, Precision, F1, mAP are used as the evaluation indexes of the model, and the threshold value of IOU is set to 0.5. F1 is the harmonic mean of Recall and Precision, which can give a more accurate response to the model. mAP is the average value of the average accuracy of multiple targets under different Recall conditions. Among them, the definitions of Recall, Precision, F1 and mAP are as follows:

[0110]

[0111]

[0112]

[0113]

[0114] Among them, TP represents the number of correctly identified positive samples, FP represents the number of incorrectly identified positive samples, FN represents the number of missed positive samples, m represents the number of recognition categories and the number of target categories, Recall is the recall rate, Precision is, F1 is the name F1 value, it is a formula, mAP is the average precision, AveP(n) is the single average precision, and IOU is the set value.

[0115] Step2: experimental result analysis

[0116] In order to verify the effectiveness of the improved target detection algorithm proposed in the application, the experimental results are as shown in Table 5.

[0117] Table 5

[0118]

[0119] The input image resolution is 640x640x3, and the results after training for 500 epochs are as shown in Figure 10 The experimental results show that the improved yolov7 target detection algorithm proposed in the application has better performance, and the overall mAP0.5 is improved by 1.5%, although the mAP0.5 number of airplane target decreases by 0.2%, but the truck and person targets are improved, especially the person target is improved by 4.2%. As Figure 11 (a), Figure 11 (b) and Figure 11(c) The yolov7 official model failed to detect the target, while the yolov7 retrained model and the improved model both detected the target, and the improved model had a higher accuracy in detecting the target.

[0120] The above-described embodiments are only descriptions of the preferred modes of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. An airport air control image target detection method based on improved yolov7, characterized in that, The method comprises the following steps: Collecting an airport air control image, preprocessing the airport air control image, labeling image features, and obtaining a labeled data set; An improved yolov7 network is used to construct an air control image target detection model; wherein the improved yolov7 network comprises a FEAM attention module; The improved yolov7 network is used to construct an air control image target detection model, which comprises the following steps: The yolov7 backbone network in the yolov7 network is improved, and the FEAM attention module is added to the improved yolov7 backbone network to construct the air control image target detection model; The air control image target detection model comprises an improved yolov7 backbone network, a yolov7 neck network and a yolov7 head network; the improved yolov7 backbone network, the yolov7 neck network and the yolov7 head network are connected in sequence; The yolov7 backbone network after adding the FEAM attention module comprises a CBS structure, a FEAM attention module, an ELAN structure, an MP1 structure and a SPPCSPC structure; the CBS structure, the FEAM attention module, the ELAN structure, the MP1 structure and the SPPCSPC structure are connected in sequence; The FEAM attention module comprises an ECA submodule, an FReLU activation function and a SAM module, which are connected in sequence; The improved air control image target detection model is trained based on the labeled data set, and target detection is performed through the trained air control image target detection model.

2. The improved yolov7-based airport air control image target detection method according to claim 1, characterized in that, The preprocessing of the airport air control image comprises the following steps: Based on the airport air control image, an image data set is obtained, and the image data set is randomly divided into a training set and a test set.

3. The improved yolov7-based airport air control image target detection method according to claim 2, characterized in that, The labeled data set is obtained by labeling image features, which comprises the following steps: An object position label is added through a labeling program Labelimg, and the aircraft, pedestrians and vehicles in each picture in the image data set are classified and labeled to construct the labeled data set; The labeled data set comprises the position information of the aircraft, the pedestrians and the vehicles in the image.

4. The improved yolov7-based airport air control image target detection method according to claim 1, characterized in that, The FReLU activation function depends on the spatial context 2D condition of each pixel to complete a visual task; The method for completing the visual task comprises the following steps: wherein, is the activation function algorithm name, is the input and takes the maximum value, is the activation function input, is the spatial context 2D condition.

5. The improved yolov7-based airport air control image target detection method according to claim 4, characterized in that, Training the improved air control image target detection model based on the labeled data set comprises training the FEAM attention module; The training of the FEAM attention module comprises the following steps: The first feature map is obtained based on the labeled data set, the first feature map is input into the ECA sub-module, the ECA sub-module is used for average pooling on the main input end, and each channel feature value is obtained, the channel feature value is used for realizing channel weight between feature vectors by using a 1*1 convolution kernel, a feature map channel weight is obtained by using a sigmoid activation function, the feature map channel weight is multiplied by the first feature map, output features of the ECA sub-module are obtained, the output features are activated by using the FReLU activation function, and are transmitted to the SAM module, the SAM module is used for compressing the activated output features in a channel dimension by using average pooling and maximum pooling, a plurality of two-dimensional feature maps are obtained, the two-dimensional feature maps are spliced based on channel Concat, and a second feature map with a preset number of channels is obtained.

6. The improved yolov7-based airport air control image target detection method according to claim 5, characterized in that, The training of the FEAM attention module further includes: The second feature map is subjected to a convolution operation by using a hidden layer containing a single convolution kernel, spatial attention weights are generated by using a sigmoid operation, the spatial attention weights are multiplied by the first feature map, a feature map containing channel and spatial attention weights is obtained, and the training of the FEAM attention module is completed.

7. The improved yolov7-based airport air control image target detection method according to claim 6, characterized in that, The method further includes verifying the improved air control image target detection model: Recall, Precision, F1 and mAP are used as evaluation indexes of the model, an IOU threshold is set as a preset value, and the improved air control image target detection model is verified; The Recall is a recall rate, the Precision is an accuracy rate, the F1 is an F1 value, the mAP is a mean average precision, and the IOU is a set value.

Images