A target detection method for airport surface surveillance images based on the YOLO framework

Abstract

This invention discloses a target detection method for airport surface surveillance images based on the YOLO framework, comprising the following steps: acquiring airport surface surveillance images; constructing a dataset conforming to the features of airport surface surveillance images based on the images; labeling the dataset to obtain a labeled dataset; and randomly dividing the labeled dataset into a training set and a test set; constructing an AS-YOLO airport surface surveillance image target detection model; training the AS-YOLO airport surface surveillance image target detection model using the training set; and performing target detection on the test set using the trained AS-YOLO airport surface surveillance image target detection model, and validating the method through target detection evaluation metrics. This invention addresses the problem of insufficient data in airport surface surveillance image feature research and the problem of inaccurate detection of pedestrians and vehicles in airport surface surveillance images due to their small size.

Description

Technical Field

[0001] This invention belongs to the field of image target detection technology, specifically relating to a target detection method for airport surface surveillance images based on the YOLO framework. Background Technology

[0002] With the continuous development of the air transport industry and the sustained growth of flight volume, airport security is facing increasing pressure. The rapid development of the general aviation industry also brings new challenges to airport security management. Manually processing airport surveillance images or videos is unstable, cumbersome, time-consuming, and expensive. Therefore, it is of great significance to develop and design a target detection algorithm for airport surface surveillance images using computer vision technology.

[0003] Currently, typical object detection techniques are mainly divided into two-stage detection and one-stage detection methods. Two-stage detection first generates candidate boxes for detection regions, and then predicts the location boxes and categories, such as object detection algorithms like R-CNN, Fast R-CNN, and Faster R-CNN. One-stage detection generates predicted location boxes and categories in the detection network, such as object detection methods like SDD and YOLO.

[0004] The above-mentioned methods for constructing target detection methods for airport surface surveillance images have the following problems: First, the target detection methods mentioned above are improvements and optimizations of commonly used datasets such as PASCAL VOC and MS COOC, but lack datasets specifically for airport surface surveillance images. Second, the existing YOLO target detection series has low accuracy in detecting small targets. Airport surface surveillance images are characterized by large targets such as aircraft, while targets such as pedestrians and cars are too small. Therefore, due to the presence of large targets such as aircraft, airport surface surveillance image detection cannot be simply reduced to a small target detection problem, while the presence of small targets such as pedestrians and cars affects the accuracy of target detection in airport surface surveillance images. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a target detection method for airport surface surveillance images based on the YOLO framework. A dataset conforming to airport surface image features is constructed using images acquired through airport surface surveillance. Based on the YOLO framework, an AS-YOLO target detection method for airport surface surveillance images is built. Finally, the improved target detection method is validated using the aforementioned dataset.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A target detection method for airport surface surveillance images based on the YOLO framework includes the following steps:

[0008] Collect airport surface surveillance images, construct a dataset that conforms to the characteristics of airport surface surveillance images based on the airport surface surveillance images, annotate the dataset to obtain an annotated dataset, and randomly divide the annotated dataset into a training set and a test set;

[0009] Construct an AS-YOLO target detection model for airport surface surveillance images;

[0010] The AS-YOLO airport surface surveillance image target detection model was trained using the training set.

[0011] The trained AS-YOLO airport surface surveillance image target detection model was used to perform target detection on the test set, and the results were verified by the target detection evaluation index.

[0012] Preferably, the method for labeling the dataset includes:

[0013] The airplanes, pedestrians, and vehicles in each image in the dataset are classified and labeled to construct a labeled dataset.

[0014] Preferably, the method for constructing the AS-YOLO airport surface surveillance image target detection model is as follows:

[0015] The backbone network was constructed using the designed CFEAM, CBF, and MP1 structures.

[0016] Based on the FPN and PAN structures, a neck network is constructed by adding designed F-SPPF and FEAM structures;

[0017] Four head networks are constructed using the CBF structure.

[0018] Preferably, the CBF structure consists of a convolution module, a normalization module, and a FReLU activation function;

[0019] The CFEAM structure is used in combination with multiple FEAM structures;

[0020] The MP1 structure is composed of a max pooling module and a convolution with a stride of 2, which are two downsampling methods combined.

[0021] Preferably, the spatial feature extraction process of the CFEAM structure is as follows:

[0022] The input is manipulated by a CBF with a kernel of 1 and passed to the FEAM structure. This process is repeated four times to generate four different feature maps.

[0023] The first and second feature maps are concatted together and passed to a CBF convolutional layer with a kernel of 3 to obtain the first convolution result.

[0024] The first convolution result is concatted with the third feature map and passed to a CBF convolutional layer with a kernel of 3 to obtain the second convolution result.

[0025] The second convolution result is concatted with the fourth feature map to output the result.

[0026] Preferably, the FPN structure is a top-down feature pyramid;

[0027] The PAN structure is a bottom-up feature pyramid;

[0028] The F-SPPF structure is passed to the CBF structure through a combination of three max pooling operations to extract features;

[0029] The input of the FEAM structure has changed from a main input to a combination of main input and residual input. The main input is based on the CBAM attention module, replacing the parameter-intensive CAM module with a lightweight ECA module, and using the FReLU activation function, which is specifically designed for visual tasks, instead of the sigmoid activation function.

[0030] Preferably, the process of obtaining weight matrices of different dimensions in the FEAM structure is as follows:

[0031] The lightweight ECA module is used to obtain the feature value of each channel by average pooling on the main input edge;

[0032] The channel feature values ​​are implemented using a 1×1 convolution kernel to determine the channel weights between feature vectors.

[0033] The channel weights are processed by a sigmoid activation function to obtain the feature map channel weights;

[0034] The output features of the lightweight ECA module are obtained by multiplying the channel weights of the feature map with the feature map of the initial input FEAM module.

[0035] The output features are activated using the FReLU activation function and then passed to the SAM module;

[0036] The SAM module uses average pooling and max pooling to compress the feature map Yc in the channel dimension, resulting in two two-dimensional feature maps.

[0037] The two two-dimensional feature maps are concatenated based on channel concat to obtain a feature map with 2 channels;

[0038] The concatenated feature maps are convolved using a hidden layer containing a single convolutional kernel.

[0039] Spatial attention weights are generated by performing a sigmoid operation on the convolution result;

[0040] Multiply the spatial attention weights by the feature map of the initial input FEAM module to obtain a feature map that includes channel and spatial attention weights;

[0041] The feature map is activated by the FReLU activation function and added to the residual input to obtain an output with multi-scale information representation capability.

[0042] Preferably, the expression for the output of the multi-scale information representation capability is:

[0043] Where Y is the air traffic control image feature map.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention constructs a dataset that conforms to the characteristics of airport surface surveillance images based on real airport images. The YOLO target detection framework includes a backbone network, a neck network, and a head network. This invention uses the designed CFEAM module and CBF module to construct the backbone network; combining the FPN structure and PAN, the designed F-SPPF module and FEAM module are added to enhance the network feature fusion capability and construct the neck network; the CBF module constructs four head networks, thus constructing a target detection method for airport surface surveillance images. Its advantages are: (1) solving the problem of insufficient data for the study of airport surface surveillance image features; (2) solving the problem of inaccurate detection of pedestrians and vehicles in airport surface surveillance images due to their small size. Attached Figure Description

[0046] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of an airport surface surveillance image target detection method based on the YOLO framework according to the present invention;

[0048] Figure 2 This is a structural diagram of the FEAM module of the present invention;

[0049] Figure 3 This is a schematic diagram of the CAM module network structure of the present invention;

[0050] Figure 4 This is a schematic diagram of the network structure of the ECA module of the present invention;

[0051] Figure 5 This is a network structure diagram of the SAM module of the present invention;

[0052] Figure 6 This is a diagram of the FReLU:MAX(x,T(x)) structure of the present invention;

[0053] Figure 7 This is a schematic diagram of the AS-YOLO target detection network structure of the present invention;

[0054] Figure 8 This is a schematic diagram of the basic CBF convolutional structure of the present invention;

[0055] Figure 9 This is a schematic diagram of the CFEAM structure of the present invention;

[0056] Figure 10 This is a schematic diagram of the MP structure of the present invention;

[0057] Figure 11 This is a schematic diagram of the F-SPPF structure of the present invention. Detailed Implementation

[0058] 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.

[0059] 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.

[0060] Example 1

[0061] like Figure 1 As shown, this invention discloses a target detection method for airport surface surveillance images based on the YOLO framework, comprising the following steps:

[0062] Collect airport surface surveillance images, construct a dataset that conforms to the characteristics of airport surface surveillance images based on the airport surface surveillance images, annotate the dataset to obtain an annotated dataset, and randomly divide the annotated dataset into a training set and a test set;

[0063] Construct an AS-YOLO target detection model for airport surface surveillance images;

[0064] The AS-YOLO airport surface surveillance image target detection model was trained using the training set;

[0065] The trained AS-YOLO airport surface surveillance image target detection model was used to perform target detection on the test set, and the model was validated using target detection evaluation metrics.

[0066] The trained weight parameters were written into detect.py, and a framework was built using Python to run the improved model. The results showed that the AS-YOLO model was more accurate and met the target detection requirements of airport scene monitoring images.

[0067] In this embodiment, the method for labeling the dataset includes:

[0068] The airplanes, pedestrians, and vehicles in each image in the dataset are classified and labeled to construct a labeled dataset.

[0069] The experimental data for this invention comes from real-world airport data. To ensure data diversity, various representative airport environmental images are used, such as those depicting sunny days, late nights, cloudy days, and rainy days. This invention uses common airport targets, primarily including aircraft, pedestrians, and vehicles. A target label dataset is established based on the characteristics of airport targets. Target location labels are categorized into three types: aircraft, pedestrians, and vehicles. The location label format is a .txt file containing the center point plus height and width. The location information labels are annotated using the Labelimg annotation program, with a total of 2000 images annotated. This invention's dataset contains 3466 aircraft targets, 2994 pedestrian targets, and 1911 vehicle targets, totaling 8371 targets. According to the COCO dataset's definition of small targets (targets with pixels smaller than 32×32 are considered absolutely small), small targets account for 39.42% of this invention's dataset.

[0070] In this embodiment, the attention network mechanism increases detection capability by making the model more interested in a certain type of target and reducing useless information. The FEAM attention module mainly consists of three parts: the ECA module, the SAM module, and the FReLU activation function. For air traffic control images, assigning greater weights to areas that contain more detection targets means the model will use more attention to learn the features of that area, thereby improving the accuracy of the air traffic control image detection algorithm and better assisting airport management personnel in making decisions.

[0071] Inspired by residual attention modules and CBAM attention modules, the core idea of ​​FEAM is to adopt a residual module structure, obtaining different dimensional weight matrices along both the channel and spatial dimensions of the input air traffic control image feature map Y. The input of the FEAM attention module changes from a main input to a combination of main and residual inputs. Its main input is borrowed from the CBAM attention module, replacing the parameter-intensive CAM module with a lightweight ECA module, and using the FreLU activation function, specifically designed for visual tasks, instead of the sigmoid activation function. The structure diagram is shown below. Figure 2As shown.

[0072] Specifically, the process of obtaining weight matrices of different dimensions in the FEAM structure is as follows: First, the ECA module obtains the feature value of each channel by average pooling on the main input edge. The channel feature value uses a 1×1 convolution kernel to realize the channel weight between feature vectors, and then undergoes a sigmoid activation function to obtain the channel weight of the feature map. This weight is multiplied with the feature map of the initial input FEAM module to obtain the output feature of the ECA module. The output feature is activated by the FReLU activation function and passed to the SAM module. The SAM module uses average pooling and max pooling to compress the feature map Yc in the channel dimension to obtain two two-dimensional feature maps. These two feature maps are concatenated based on the channels to obtain a feature map with two channels. To ensure that the final feature is consistent with the input Yc in the spatial dimension, a hidden layer containing a single convolution kernel is used to convolve the concatenated feature map. Finally, a sigmoid operation is performed to generate spatial attention weights, which are multiplied with the feature map of the initial input FEAM module to obtain a feature map containing channel and spatial attention weights. This feature map is activated by the FReLU activation function and added to the residual input to obtain an output with better multi-scale information representation capabilities.

[0073]

[0074] Specifically, the CAM module focuses more on the semantic information of the feature map. It uses average pooling to integrate information and max pooling to collect detailed information in the spatial dimension of the input feature map Y. These two pooling operations reduce network computation while improving the network's expressive power. The two pooled one-dimensional vectors are then fed into a fully connected layer for computation. This invention uses a 1×1 convolutional kernel to achieve weight sharing between the feature vectors. Finally, a channel attention Oc is generated through summation and sigmoid activation, as shown in the structure... Figure 3 As shown.

[0075] O c =sigmoid(MLP(MaxPool(Y))+MLP(AvgPool(Y))) (2)

[0076] Because CAM uses fully connected manipulations to map features, such as Figure 3 As shown, inserting multiple CBAMs into the network results in a large computational load on network parameters. Therefore, this invention addresses the problem of high computational load caused by a large number of parameters by improving CAM. FEAM uses an ECA channel attention module instead of the traditional CAM. A one-dimensional convolution with a kernel length of 3 integrates features from three channels within the channel neighborhood, and finally generates attention through sigmoid activation. Structure as Figure 4 As shown.

[0077]

[0078] Specifically, the SAM spatial attention module focuses more on the location information of features, concentrating on regions with more effective features in the feature map, thus complementing the channel attention. Average pooling and max pooling are used to compress the feature map Yc along the channel dimension, resulting in two two-dimensional feature maps. These are then concatenated along the channel dimension to obtain a feature map with two channels. To ensure that the final feature map is spatially consistent with the input Yc, a hidden layer containing a single convolutional kernel is used to convolve the concatenated feature map. Finally, a sigmoid operation is performed to generate the spatial attention weights Os, as shown in the structure below. Figure 5 As shown

[0079] O s =sigmoid(Conv2(AvgPool(Y),MaxPool(Y))) (4)

[0080] Specifically, in convolutional neural networks, activation layers provide the network with non-linearity. The FReLU activation function is specifically designed for visual tasks, implementing its functionality through ordinary convolution, thus offering better capabilities for visual image processing. Currently, the most widely used activation function remains ReLU, whose function expression is as follows:

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

[0082] The ReLU activation function boasts high accuracy across many tasks. FReLU inherits the ideas of ReLU and extends them to spatial context, relying on the 2D condition T(x) of each pixel's spatial context. It is simple to implement with only a small increase in computation, and better suited for visual tasks. Its expression is as follows, and its structure is as follows: Figure 6 As shown:

[0083] FReLU=max(x,T(x)) (6)

[0084] In this embodiment, inspired by self-attention and cross-combination structures, this invention develops a multimodal spatial feature extraction structure, CFEAM, in conjunction with the FEAM module. The CFEAM structure extracts spatial features as follows: the input is manipulated by a CBF layer with a kernel of 1 and passed to the FEAM structure. This process generates four different feature maps four times. The first and second feature maps are concatenated and passed to a CBF convolutional layer with a kernel of 3. The result is then concatenated with the third feature map and passed to another CBF convolutional layer with a kernel of 3. Finally, the result is concatenated with the fourth feature map, and the output is the final result. Using this structure, image fusion at different depths enhances the feature extraction capability of the backbone network. The FEAM attention module reduces environmental interference and also increases the detection capability for small targets in airports.

[0085] In this embodiment, the method for constructing the AS-YOLO airport surface surveillance image target detection model is as follows:

[0086] This invention utilizes real airport surveillance images. First, it constructs an airport surface surveillance image dataset using such images. Based on the YOLO target detection framework, which includes a backbone network, a neck network, and a head network, this invention uses a designed CFEAM module and a CBF module to construct the backbone network. Combining the FPN structure and PAN, it adds a designed F-SPPF module and a FEAM module to enhance the network's feature fusion capability and construct the neck network. The CBF module constructs four head networks, thus building a target detection method for airport surface surveillance images. The established airport surface surveillance image dataset is used to validate the trained network model. Validation is then performed using target detection evaluation metrics.

[0087] In this embodiment, AS-YOLO has the same network framework as YOLO, consisting of a backbone network, a neck network, and a head network. The overall network structure is as follows: Figure 7 As shown.

[0088] The backbone network includes a CBF structure (such as...) Figure 8 As shown), CFEAM structure (as shown) Figure 9 As shown), MP1 structure (as shown) Figure 10(As shown). The CBF structure consists of a convolutional module, a normalization module, and a FReLU activation function. The CFEAM structure is mainly inspired by attention structures and cross-combination structures. By using multiple FEAM structures in combination, it has a good feature learning ability. The MP1 structure consists of a max pooling module and a convolution with a stride of 2. It is composed of two downsampling methods and is recombined to enhance the network's learning ability without breaking the network structure. The SPPCSPC structure divides the features into two parts. One part obtains different receptive fields through four different max pooling operations to distinguish between large and small targets. The other part uses the CBF structure to extract features. Finally, the two parts are merged to improve the network accuracy.

[0089] Neck and head networks include FPN structures, PAN structures, FEAM structures, and F-SPPF structures (such as...). Figure 11 (As shown). The FPN structure is a top-down feature pyramid that uses upsampling to improve small target detection. The PAN structure is a bottom-up feature pyramid that uses information from lower layers to pass to upper layers, improving the detection of occluded targets.

[0090] The FEAM structure is the attention module designed in this invention, which enhances small target detection capabilities by adding a small amount of computational power. The F-SPPF structure combines three max pooling operations and passes them to the CBF structure for feature extraction, reducing computational costs while effectively avoiding image distortion and other problems.

[0091] In the head network, the classic YOLO framework downsamples the original image by 8x, 16x, and 32x before feeding it into the feature detection network, resulting in a 20x20 large target feature map, a 40x40 medium target feature map, and an 80x80 small target feature map. While deep convolutional layers in target detection algorithms are considered to provide rich semantic information, they are less focused on location information and may lose some information about small targets. Considering the characteristics of airport surface surveillance images, which contain numerous small targets such as pedestrians and vehicles, a new scale feature map is added to the original algorithm. The original image is downsampled by 4x, 8x, 16x, and 32x to obtain the same 20x20 large target feature map, 40x40 medium target feature map, 80x80 small target feature map, and 160x160 small target feature map before being fed into the detection network. This improved network structure for airport surface surveillance image target features is beneficial for pedestrian and vehicle target detection, improving detection performance.

[0092] In this embodiment, the airport surface surveillance dataset and the AS-YOLO airport surface surveillance image target detection model constructed in this invention are validated. The specific steps are as follows:

[0093] Step 1: Configure the network training configuration file

[0094] Select 1600 images from the dataset as the training set and 400 images as the test set. Before training, the AS-YOLO data and model configuration files need to be modified. In the data file, change the number of object categories to 3 and modify the names in the category list to include the object category names. In the model configuration file, set the anchor box sizes to (5,9), (6,14), (43,15), (12,16), (19,36), (40,28), (36,75), (76,55), (72,146), (142,110), (192,243), (459,401). Based on the improved network structure, add feature scale and attention modules. Add the CFEAM model code to the common.py file and import the CFEAM structure into the yolo.py file. Run the yolo.py file to check if the network changes are correct.

[0095] Step 2: Configure the network training environment

[0096] This experimental environment uses the Sugon cloud computing service system. Each node is configured with one 32-core x86 processor with a 2.5GHz clock speed and one NVIDIA Tesla V100 accelerator card. Each node also has two 16GB DDR4 2666ECC REG memory modules and two Sugon Parastor300S parallel storage systems to provide large-capacity data storage. For network communication, the cluster uses a full-speed, non-blocking 200Gb HDR Infiniband dedicated computing network and the PyTorch 1.9.0 deep learning framework. Specific configuration information is shown in Table 1.

[0097] Table 1

[0098]

[0099] Network testing software environment configuration

[0100]

[0101] AS-YOLO model parameter settings

[0102]

[0103] Step 3: Selection of Target Detection Evaluation Indicators

[0104] All models were trained and tested using the airport surface surveillance image data established in this invention. Recall, Precision, F1, and mAP were used as evaluation metrics for the models in the experiments, with the IOU threshold set to 0.5. F1 is the harmonic mean of Recall and Precision, which provides a more accurate response for the model. mAP is the average accuracy of multiple targets under different recall conditions. The definitions of Recall, Precision, F1, and mAP are as follows:

[0105]

[0106]

[0107]

[0108]

[0109] Where 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, and m represents the number of identified categories.

[0110] Step 4: Analysis of Experimental Results

[0111] To verify the effectiveness of the improved target detection algorithm proposed in this invention, the experimental results are shown in Table 2 below:

[0112] Table 2

[0113]

[0114]

[0115] The input image resolution was 640×640×3. After 300 epochs of training, the detection accuracy of YOLOv3, YOLOv5l, YOLOv7, and AS-YOLO was compared. Experimental results show that the proposed AS-YOLO object detection algorithm has better performance, improving the mAP0.5 by 5.7 compared to YOLOv3, 1.5% compared to YOLOv5l, and 3.2 compared to YOLOv7. Using YOLOv3, YOLOv5l, YOLOv7, and AS-YOLO object detection algorithms to detect targets in airport surveillance images, the results show that targets missed by the YOLO algorithm can be detected by AS-YOLO, and the accuracy of target detection is higher.

[0116] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for airport scene monitoring image target detection based on a YOLO framework, characterized in that, The method comprises the following steps: Collecting an airport scene monitoring image, constructing a data set conforming to the characteristics of the airport scene monitoring image based on the airport scene monitoring image, labeling the data set, obtaining a labeled data set, and randomly dividing the labeled data set into a training set and a test set; Constructing an AS-YOLO airport scene monitoring image target detection model; Training the AS-YOLO airport scene monitoring image target detection model using the training set; Using the trained AS-YOLO airport scene monitoring image target detection model to perform target detection on the test set and verifying through a target detection evaluation index; The method for constructing the AS-YOLO airport scene monitoring image target detection model comprises: Using a designed CFEAM structure, a CBF structure and an MP1 structure to construct a backbone network; On the basis of combining an FPN structure and a PAN structure, adding a designed F-SPPF structure and a FEAM structure to construct a neck network; Using the CBF structure to construct four head networks.

2. The airport surface surveillance image target detection method based on the YOLO framework according to claim 1, characterized in that, The method for labeling the data set comprises: Classifying and labeling aircraft, pedestrians and vehicles in each picture in the data set to construct a labeled data set.

3. The airport scene monitoring image target detection method based on the YOLO framework according to claim 1, wherein The CBF structure is composed of a convolution module, a normalization module and an FReLU activation function; The CFEAM structure is cross-used by multiple FEAM structures; The MP1 structure is recombined by a maximum pooling module and a convolution with a step of 2, two down-sampling modes.

4. The airport surface surveillance image target detection method based on the YOLO framework according to claim 3, characterized in that, The spatial feature extraction process of the CFEAM structure is as follows: The first feature map and the second feature map are concatenated and transmitted to a CBF convolution layer with a convolution kernel of 3 to obtain a first convolution result; The first convolution result and the third feature map are concatenated and transmitted to a CBF convolution layer with a convolution kernel of 3 to obtain a second convolution result; The second convolution result and the fourth feature map are concatenated and output.

5. The airport scene monitoring image target detection method based on the YOLO framework according to claim 1, wherein The FPN structure is a top-down feature pyramid; The PAN structure is a bottom-up feature pyramid; The F-SPPF structure is combined by three maximum pooling operations and transmitted to the CBF structure to extract features; The input end of the FEAM structure is changed from a main input to a joint action of a main input and a residual input, the main input is a CBAM attention module, a CAM module with more parameters is replaced by a lightweight ECA module, and a FReLU activation function specially applied to a visual task is used instead of a sigmoid activation function. The process of the FEAM structure for obtaining different dimension weight matrices is as follows: 6.The method of claim 5, wherein, The lightweight ECA module is used to obtain feature values of each channel through average pooling on the main input edge. ​ The channel feature values use a 1*1 convolution kernel to realize channel weight between feature vectors; The channel weight is subjected to a sigmoid activation function to obtain feature map channel weight; The feature map channel weight is multiplied with the feature map of the initial input FEAM module to obtain the output feature of the lightweight 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; The two two-dimensional feature maps are spliced based on channel Concat to obtain a feature map with a channel number of 2; A hidden layer containing a single convolution kernel is used to perform convolution operation on the spliced feature map; The convolution result is subjected to a sigmoid operation to generate spatial attention weight; The spatial attention weight is multiplied with the feature map of the initial input FEAM module to obtain a feature map containing channel and spatial attention weighting; The feature map is activated by a FReLU activation function and added to the residual input to obtain the output of the multi-scale information expression ability.

7. The airport surface surveillance image target detection method based on the YOLO framework according to claim 6, characterized in that, The expression of the output of the multi-scale information expression ability is: where Y is an air control image feature map.

Images