Airport runway foreign matter detection method based on YOLO11 optimization network

Abstract

The invention discloses an airport runway foreign matter detection method based on a YOLO11 optimization network, and belongs to the technical field of target detection in computer vision and deep learning. The problems that detection challenges of multi-scale foreign matter features and complex background textures exist when a traditional detection model is used for detecting airfield runway foreign matter, the multi-scale features of millimeter-level bolts and centimeter-level gravel coexist easily cause missing detection, and the texture similarity of metal pieces and runway asphalt further reduces the detection precision are solved. According to the airport runway foreign matter detection network structure based on the YOLO11 optimization network and the detection method, the real-time image of a vehicle-mounted line-scan digital camera is obtained, and the FOD appearing in the image is detected.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of object detection in computer vision and deep learning, and particularly relates to an airport runway foreign object debris (FOD) detection method based on a YOLO11 optimized network. BACKGROUND

[0002] As the core area of aircraft take-off and landing, the surface state of the airport runway is directly related to flight safety, and timely detection of runway foreign objects is a key link in the aviation safety guarantee system. With the rapid development of global air transportation, the airport throughput continues to rise, and the traditional manual inspection and regular cleaning mode has been difficult to meet the safety supervision needs under the high-density flight operation, and there is an urgent need for intelligent and efficient foreign object detection technology innovation.

[0003] Traditional airport runway foreign object detection mainly relies on manual visual inspection and periodic cleaning operations, which not only consumes a lot of manpower and resources, but also has low detection efficiency and the risk of missed detection, and is difficult to deal with emergency situations such as sudden foreign object intrusion. Although technologies such as millimeter wave radar and infrared thermal imaging can achieve non-contact detection, they have high deployment costs and poor environmental adaptability, and are prone to false positives and false negatives in complex weather conditions, and are difficult to accurately identify small foreign objects. In contrast, computer vision-based detection technology has great potential in the field of airport foreign object detection, with low cost and flexible deployment, and can obtain runway images in real time through ordinary surveillance cameras.

[0004] Although deep learning has made significant progress in object detection and is widely used in traffic monitoring, security and other scenarios, there is still a large gap in research on airport runway foreign object detection. The types of runway foreign objects are diverse, with large size differences, and in complex environments such as light changes and shadow occlusion, the appearance features of foreign objects are prone to significant changes, making it difficult for general object detection models to meet the high precision and real-time requirements of aviation safety, and it is necessary to conduct targeted technology optimization combined with the characteristics of the airport runway scene. SUMMARY

[0005] To solve the problem that the traditional detection model has detection challenges of multi-scale foreign object features and complex background textures when detecting airport runway foreign objects: the multi-scale characteristics of millimeter-level bolts and centimeter-level gravel coexisting easily lead to missed detection, and the texture similarity between metal parts and pavement asphalt further reduces the detection accuracy. The method proposes an airport runway foreign object detection network structure and detection method based on a YOLO11 optimized network, which detects FOD in the image by obtaining real-time images of a vehicle-mounted linear array camera.

[0006] The overall implementation process of the application is as follows Figure 1As shown, the technical scheme of the present application is an airport runway foreign matter detection method based on a YOLO11 optimized network, which comprises the following steps:

[0007] S1: determining a foreign matter detection model;

[0008] The foreign matter detection model comprises a backbone network, a neck network and a head network;

[0009] The backbone network comprises modules 0, 1, 2, …, 10 connected in series, wherein the modules 0, 1, 3, 5 and 7 are CBS modules, the modules 2, 4, 6 and 8 are CDEB modules, the module 9 is an SPPF module, and the module 10 is a C2PSA module; the input image is input to the module 0, and the outputs of the modules 4, 6 and 10 are further input to the neck network; the CBS module is a basic convolution unit composed of a standard convolution layer, a batch normalization and a SiLU activation function connected in series, and is used for extracting local features of an image; the SPPF module is a fast spatial pyramid pooling module; the C2PSA module is a convolution block with an attention mechanism; and the CDEB module is a composite detail enhancement and fusion module, which realizes deep enhancement and fusion of image features through multi-stage feature processing, and provides more discriminative feature representation for subsequent networks;

[0010] The neck network comprises modules 11, 12, 13, …, 22 connected in series; the outputs of the module 10 are input to the modules 11 and 21 respectively, the output of the module 6 is input to the module 12, the output of the module 4 is input to the module 15, and the outputs of the modules 16, 19 and 22 are input to the head network; the modules 11 and 14 are Upsample modules, the modules 12, 15, 18 and 21 are Concat modules, the modules 13, 16, 19 and 22 are C3K2 modules, and the module 20 is a CBS module; the Upsample module is an up-sampling module, the Concat module is a feature concatenation module, and the C3K2 module is a module responsible for feature processing tasks in the structure;

[0011] The head network comprises three GNDE-Detect modules, and the input data of the modules correspond to the output data of the modules 16, 19 and 22 respectively;

[0012] S2: training the foreign matter detection model using a public data set FOD-CV and a self-built line array camera data set FOD-Line;

[0013] S3: detecting images using the deployed foreign matter detection model.

[0014] Further, the CDEB module comprises, in sequence: a conv1 module, a split module, n CSP-DSMEnbanceblock modules, a conact1 module, a conv2 module;

[0015] The input data of the CSP-DSMEnbanceblock module is divided into two paths, one path comprises, in sequence: a conv3 module, n CSP-DSMEnbance modules, and the other path is a conv4 module, the outputs of the two paths are jointly input into a conact2 module, and the output of the conact2 module after passing through the conv4 module serves as the output of the CSP-DSMEnbanceblock module;

[0016] The CSP-DSMEnbance module comprises: an Input layer, and then the output of the Input layer is divided into five branches; the first to fourth branches have the same structure and comprise, in sequence: an AdaptiveAvgpool module, a first conv module, a second conv module, an EdgeEnhancer module, and an Upsample module; the fifth branch is a conv module; the outputs of the five branches are jointly input into a Concat module, then pass through a DSM module and a conv module in sequence, and finally output.

[0017] Further, the GNDE-Detect module comprises, in sequence: a Conv_GN module, a first DEConv module, and a second DEConv module, the output of the second DEConv module is divided into two paths, one path is input into a Conv_Reg module, and the other path is input into a Conv_Cls module, and the output of the Conv_Reg module is Scale;

[0018] The normalization processing mode formula of the Conv_GN module is as follows:

[0019]

[0020] wherein, GroupNorm(x ijkl ) represents the normalized output, x ijkl represents a feature value of an i-th sample, a j-th channel group, a k-th row, and an l-th column in an input feature map, μ i,g represents a mean value of a g-th channel group feature of the i-th sample, represents a variance of the g-th channel group feature of the i-th sample, ∈ represents a minimum constant, γ g represents a learnable parameter of the g-th channel group, β g represents a lower table parameter of the g-th channel group.

[0021] The DEConv module represents detail enhancement convolution, which fuses basic features and edge details by parallel ordinary convolution and differential convolution, improves the extraction ability of small targets and texture features, the Conv_Reg module represents a target position regression module, which is used to predict the coordinate offset of the target bounding box, realizes accurate positioning, the Conv_Cls module represents a target classification module, which outputs the probability distribution of the target belonging to each preset category, Scale represents scale adjustment of the regression result, which adapts to the detection needs of different size targets and enhances multi-scale adaptability.

[0022] Further, the loss function for training the foreign matter detection model in S2 is:

[0023]

[0024] wherein, represents the total loss output, and a represents a smoothing coefficient, represents the loss value of the ith sample in the tth iteration, w i (t) represents the weight of the ith sample in the tth iteration, represents the exponential moving average loss of the ith sample in the tth iteration, L total represents the total loss, which is obtained by weighted summation of the sample loss and the category loss, represents the weight of the cth category in the tth iteration, represents the exponential average loss of the cth category in the tth iteration, represents the loss value of the cth category in the tth iteration.

[0025] Further, the input of the DEConv module is divided into two paths, one path passes through the VC module, and the other path passes through the Difference conv module, and then the outputs of the two paths are fused and output; the VC module represents ordinary 3x3 convolution, and the Difference conv module represents central difference convolution (CDC), angle difference convolution (ADC), horizontal difference convolution (HDC), and vertical difference convolution (VDC); finally, the outputs of the VC module and the Difference conv module are spliced and fused by feature fusion.

[0026] Further, the structure of the DSM module is divided into two paths for the input of the DSM module, one path passes through the spatial selection module (SSM), and the other path passes through the frequency selection module (FSM), and then the two paths are fused and output; the SSM module means that the input feature map is first subjected to global average pooling and global maximum pooling, the results of the two are spliced, then subjected to 3x3 convolution, and then subjected to 5x5 and 7x7 depth separable convolution, followed by a tiling (Tile) operation, and finally subjected to element-level multiplication and addition operation; the FSM module means that the input feature map is first subjected to global average pooling, then subtracted from the input feature map, then subjected to element-level multiplication, and finally subjected to residual connection with the input feature map; finally, the outputs of the SSM module and the FSM module are integrated and output by feature fusion.

[0027] Further, the input of the EdgeEnhancer module passes through the average pooling (Avgpool) module, the Conv module, the Subtract module and the Add module; the Avgpool module means that the input feature map is subjected to average pooling operation, the resolution is reduced and the low-frequency feature is extracted; the Conv module means that the pooled feature map is subjected to convolution using 3x3 convolution, and the basic feature is strengthened; the Subtract module means that the original input feature map is subtracted from the pooled and convolved feature map, and the edge high-frequency feature is highlighted; the Add module means that the original input feature map is added to the subtraction result, and the original information is retained through residual connection and the edge is enhanced. Finally, the enhanced feature map is output.

[0028] Based on the YOLO11 optimized GNDE-Detect module combined with the CDEB feature enhancement module, the mAP@0.5 on the FOD-CV dataset reaches 0.963, which is 2.5% higher than the original model, and the parameter amount is reduced to 8.889M through Conv_GN and shared convolution design, and the inference speed is kept at 122FPS, meeting the real-time demand of vehicle-mounted. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The overall implementation process flowchart of the present application.

[0030] Figure 2 The network structure diagram of the improved YOLO11 model.

[0031] Figure 3 The edge enhancement module diagram.

[0032] Figure 4 The CSP-DSMEnhance module structure diagram.

[0033] Figure 5 The improved CDEB module structure diagram.

[0034] Figure 6 Fig. 8 is a schematic diagram of a GNDE-Detect network structure.

[0035] Figure 7 Fig. 9 is a schematic diagram of a DSM structure.

[0036] Figure 8 Fig. 10 is a schematic diagram of a DEConv convolution structure.

[0037] Figure 9 Fig. 11 is a comparison diagram of loss changes of model setting in different ablation experiments in the training process.

[0038] Figure 10 Fig. 12 is a comparison of multi-scale target heat maps of the original model and the improved model; (a) is the IOU and heat map of the improved model; (b) is the IOU and heat map of the original benchmark model.

[0039] Figure 11 Fig. 13 is a hardware deployment device of the present application.

[0040] Figure 12 Fig. 14 is a schematic diagram of a detection result display. DETAILED DESCRIPTION

[0041] The implementation process of the present application is as shown in Figure 1 The following will be described in combination with the drawings and implementation steps

[0042] S1. Model optimization:

[0043] The present application optimizes the feature processing module, detection head structure and loss function of the YOLO11 network model, so that the model can more accurately identify the airport runway FOD with lower computational overhead. The network structure of the improved YOLO11 model is as shown in Figure 2 The steps can be divided into three parts:

[0044] S11. Optimize the C3k2 feature processing module

[0045] The C3K2 module is the main module responsible for feature processing tasks in the Backbone structure. The ordinary C3K and Bottleneck model used by it still has the problem of limited information flow when processing features, which will affect the fusion effect of deep features.

[0046] Therefore, the present application introduces CSP-DSMEnhance (CSP-DSM Scale Edge Information Select) to improve the C3K2 module in Backbone. Based on the idea of DSM, it is introduced as part of the attention mechanism for multi-channel feature enhancement in the improvement of YOLO11, and the CSP-DSMEnhance module is proposed as shown in Figure 3As shown.

[0047] Then, the application draws on the idea of CSP (Cross Stage Partial, cross-stage partial connection) and first processes the image feature branch: the features extracted from the image are adaptively mean-pooled at different scales (mean-pooling 3, 6, 9, 12) to extract local information of different sizes, which helps to capture multi-level features of the image, such as Figure 4 As shown.

[0048] Finally, the input features are transformed at different branches and different scales to extract multi-scale information. Then, each branch is equipped with an EdgeEnhancer module to further enhance the edge information of the processed features of the separated branch. Then, the original input features after local convolution are spliced with the enhanced features of each branch. Finally, the input is input to the DSM attention image edge enhancement module to focus on the complex edge and high-frequency signal area in the image, and to adaptively select features with higher task relevance in multi-scale features. The structure is as shown in Figure 5 As shown for replacing the original C3k2 module.

[0049] S12. Optimize the Head structure

[0050] Head is the core component responsible for multi-scale feature fusion in YOLO11, which uses the Anchor-Free mode. This change makes the network structure more flexible and reduces the dependence on preset anchor boxes, but it is difficult to effectively detect small targets, and too much splicing will lead to large computational complexity and reduced inference efficiency.

[0051] Therefore, the application introduces the Conv_GN structure and makes lightweight improvements based on it. The normalization processing mode formula of Conv_GN is as follows:

[0052]

[0053] It divides the channels of the feature map into multiple groups and normalizes the features in each group. If the input feature map shape is (N, C, H, W), N, C, H, and W represent batch size, channel number, height, and width, respectively. GroupNormalization divides the C channels into G groups, each containing C / G channels, and then calculates the mean and variance of the features in each group to complete the normalization.

[0054] On this basis, in order to further improve the model performance and reduce the computing burden, the lightweight idea of shared convolution is introduced for optimization, the same parameter weight is used in the feature map of each scale, the repeated calculation and memory consumption are reduced, and the model is more lightweight. At the same time, by replacing part of the convolution module responsible for fusion in Detect with DEConv to enhance the extraction ability of high frequency and low frequency information of the image, the feature representation and generalization ability of the model are improved. This helps to more accurately extract the edge, contour and other detailed features of foreign matter in the detection head, thereby improving the detection accuracy of small target objects. The structure is as shown in Figure 8 .

[0055] The improved detection head module is as shown in Figure 6 .

[0056] S13. Optimize the loss function

[0057] EMASlideLoss is improved based on the basic SlideLOSS loss function and the exponential moving average (EMA), which can more accurately focus on samples in different intersection over union (IoU) ranges by dynamically adjusting the loss weight. At the same time, the concept of exponential average is introduced, which enables the model to more effectively learn sample features during training, improves the detection accuracy of airport runway foreign objects, and does not need to directly increase the amount of calculation and parameters like adjusting the network structure or adding additional modules to achieve performance improvement at a small cost. The formula is as follows:

[0058] EMA t =α·Value t +(1-α)·EMA t-1

[0059] Wherein, alpha is a smoothing coefficient, control the weight value of historical information and current, from the formula can be seen that EMA is mainly used to smooth the historical loss or weight information, avoid the interference of training fluctuation to the current loss calculation. And Slide sliding window is to calculate the loss distribution or sample difficulty in a fixed time window (such as the last N batch), dynamically adjust the current loss weight. Prevent the model from paying too much attention to short-term fluctuations and enhance the adaptability to long-term trends. EMASlideLoss is the core idea inherited from this, dynamically adjusting the sample weight and class imbalance processing, reflecting its historical difficulty, the formula is as follows:

[0060]

[0061] The final loss is a weighted sum of the sample weights according to the EMA values. The airport runway foreign object (FOD) is a multi-scale single target, and the detection difficulty and importance of different scales are different. Moreover, the number of different types of metal and natural samples is uneven. EMASlideLoss can dynamically adjust the loss weight according to the target IoU. For FOD samples of different scales, the weight is reasonably allocated according to the matching degree with the predicted box, and special attention is paid to samples with high IoU to improve the detection effect. Through EMA, the loss of difficult samples is continuously tracked, and their influence in training is increased, similar to FocalLoss but more dynamic. The learning and generalization ability of the model to various samples is enhanced. In addition, the loss function is reparameterized and lightweight, which does not increase the computational complexity and parameter amount. The performance is improved without changing the complexity of the model structure, and the detection requirements and lightweight are considered.

[0062] S2. Model training:

[0063] The experiment selected the FOD-CV public dataset published by Khanh on Roboflow for training and verification. In terms of improvement indicators, accuracy, precision, and recall were used for quantitative evaluation of accuracy, which can directly reflect the proportion of correct detection results in the total detection results and preliminarily judge the correctness of the model. At the same time, mAP_0.5, mAP_0.5:0.95, and parameter amount (Params) were considered to evaluate the detection accuracy under different confidence levels. mAP_0.5 calculates the average precision with an IoU threshold of 0.5, and mAP_0.5:0.95 calculates within an IoU threshold range of 0.5-0.95, which can more comprehensively evaluate the performance of the model under different IoU thresholds. The above evaluation indicators are combined to comprehensively evaluate the improved model.

[0064] Firstly, four ablation experiments were set up, and their performance during training is shown in Figure 9 During training, the loss values under different improvement settings show different changes. Compared with the original FOD-v2-YOLO11s model, after adding the CSP-DSM module, the positioning loss (box_loss) and DFL loss (dfl_loss) have similar downward trends, but the numerical values are lower, and the loss can reach a relatively lower value in the early training. Further combined with GNDE-Detect improvement, the loss curve becomes smoother, and the stability is better in the middle training, and the loss value is lower than the previous two. Finally, after introducing the EMASlideLoss loss function, the loss value is significantly lower than other settings in the later training, which indicates that this loss function can help the model converge faster to a lower loss value and optimize the training process.

[0065] The performance results of the four ablation experiments on the test set are shown in the following table:

[0066] Table 1. Ablation experiment results of the original model and the improved model on the FOD-CV validation set

[0067]

[0068] From the mAP_0.5 index, with the gradual addition of improved modules such as CSP-DSM, GNDE-Detect, and EMASlide, the mAP_0.5 gradually increased from 0.938 of the YOLO11s baseline model to 0.963, indicating that the optimized model has significantly improved the detection accuracy of common targets. The mAP_0.5:0.95 index improved more prominently, with the final improved model reaching 0.810, an increase of 0.045 compared to the 0.765 of the YOLO11s baseline model, which means that the improved model has significantly enhanced the average detection accuracy at different IoU thresholds, and has shown more excellent performance in the airport runway FOD detection scene. In terms of Precision and Recall, the improved model Precision reached 0.976 and Recall reached 0.909, both of which have improved compared to the baseline model in terms of accurate target identification and avoiding missed detection. In terms of parameter quantity (Params), the parameter quantity of the model gradually decreased from 9.413M to 8.889M, indicating that the improved network structure effectively reduces parameter redundancy and reduces model storage requirements through optimization design. In terms of inference speed (FPS), although additional modules are added, the model still maintains at 121FPS, only slightly lower than the 130FPS of the baseline model, indicating that the improved structure has reasonably controlled the computational load, balancing detection accuracy and running efficiency.

[0069] The heat map comparison of the original model and the final improved model on part of the dataset is shown in Figure Figure 10 .

[0070] After comparing with the original YOLO11, it is found that the original YOLO11 model has light and scattered heat map colors in the multi-scale airport foreign object area, and for small target objects, the heat map range is large, which has problems of insufficient attention to foreign objects and inaccurate positioning. The improved model has more concentrated heat map for small foreign objects on the runway, with obvious high light color in the real foreign object image area, and the heat map color is dark in the area without foreign objects, with significant improvement in false detection. This indicates that the improved model can better distinguish the complex texture of the airport runway and more accurately identify foreign objects, and performs better in target positioning and background differentiation.

[0071] Finally, the pt (PyTorch) file after training is obtained.

[0072] S3. Model deployment:

[0073] S31. Install appropriate CUDA, ONNX software packages, NCNN environment, and related dependency libraries such as PyTorch on the industrial control computer.

[0074] S32. Use the torch.onnx.export() function to export the pt model as ONNX format.

[0075] S33. Using the NCNN conversion script, load the ONNX model generated in the previous step, generate the NCNN model, and save the generated model to the specified directory of the industrial control computer.

[0076] S34. Write C++ inference call code to load and run the NCNN model, providing an external call interface. Obtain the input data, pass it to the model for inference, and output the inference results.

[0077] S4. Hardware Device Connection and Data Processing:

[0078] S41. For example Figure 11 The diagram shows the hardware deployment device of this invention, which mainly includes an industrial control computer, a camera, and an encoder.

[0079] S42. Connect the camera to the industrial control computer, and use the SDK and camera to trigger the encoder to acquire real-time images from the line scan camera. Then, call the model call interface in step S34 to pass the image to the model for processing, and obtain information such as the location box and confidence level of the airport runway FOD in the image.

[0080] S5. Results Display:

[0081] Based on the output, the target is marked using OpenCV and displayed through a visualization interface, such as... Figure 12 As shown.

Claims

1. An airport runway foreign matter detection method based on a YOLO11 optimization network, the method comprising: S1: determining a foreign matter detection model; The foreign matter detection model comprises: a backbone network, a neck network, and a head network; The backbone network comprises modules 0, 1, 2, …, 10 connected in series, the modules 0, 1, 3, 5, and 7 are CBS modules, the modules 2, 4, 6, and 8 are CDEB modules, the module 9 is an SPPF module, and the module 10 is a C2PSA module; the input image is input to the module 0, and the outputs of the modules 4, 6, and 10 are further input to the neck network; the CBS module is a basic convolution unit composed of a standard convolution layer, a batch normalization, and a SiLU activation function connected in series, and is used for extracting local features of an image; the SPPF module is a fast spatial pyramid pooling module; the C2PSA module is a convolution block with an attention mechanism; the CDEB module is a composite detail enhancement and fusion module, which realizes deep enhancement and fusion of image features through multi-stage feature processing, and provides more discriminative feature representation for subsequent networks; The neck network comprises modules 11, 12, 13, …, 22 connected in series; the outputs of the module 10 are input to the modules 11 and 21 respectively, the output of the module 6 is input to the module 12, the output of the module 4 is input to the module 15, and the outputs of the modules 16, 19, and 22 are input to the head network; the modules 11 and 14 are Upsample modules, the modules 12, 15, 18, and 21 are Concat modules, the modules 13, 16, 19, and 22 are C3K2 modules, and the module 20 is a CBS module; the Upsample module is an up-sampling module, the Concat module is a feature concatenation module, and the C3K2 module is a module responsible for feature processing in the structure; The head network comprises three GNDE-Detect modules, and the input data of the modules correspond to the output data of the modules 16, 19, and 22 respectively; S2: training the foreign matter detection model using a public dataset FOD-CV and a self-built line array camera dataset FOD-Line; S3: detecting images using the deployed foreign matter detection model.

2. The airport runway foreign matter detection method based on YOLO11 optimization network according to claim 1, wherein, The CDEB module comprises modules conv1, split, n CSP-DSMEnbanceblock, conact1, and conv2 connected in series; The input data of the CSP-DSMEnbanceblock module are divided into two paths, one path comprises modules conv3 and n CSP-DSMEnbance in series, and the other path is a module conv4; the outputs of the two paths are jointly input to a module conact2, and the output of the module conact2 is input to a module conv4 to serve as the output of the CSP-DSMEnbanceblock module; The CSP-DSMEnbance module comprises an Input layer, and outputs of the Input layer are divided into five branches; the first to fourth branches have the same structure and comprise, in sequence, an AdaptiveAvgpool module, a first conv module, a second conv module, an EdgeEnhancer module and an Upsample module; the fifth branch is a conv module; outputs of the five branches are input into a Concat module, and then pass through a DSM module and a conv module in sequence, and finally output.

3. The airport runway foreign matter detection method based on YOLO11 optimization network according to claim 1, wherein, The GNDE-Detect module comprises a Conv_GN module, a first DEConv module and a second DEConv module connected in sequence, an output of the second DEConv module is divided into two paths, one path is input into a Conv_Reg module, and the other path is input into a Conv_Cls module, the Conv_Reg module outputs Scale; A normalization processing mode formula of the Conv_GN module is as follows: wherein, GroupNorm(x ijkl ) represents a normalized output, x ijkl represents a feature value of an i-th sample, a j-th channel group, a k-th row and a l-th column in an input feature map, μ i,g represents a mean value of a g-th channel group feature of the i-th sample, σ represents a variance of the g-th channel group feature of the i-th sample, ∈ represents a minimum constant, γ g represents a g-th channel group learnable parameter, and β g represents a g-th channel group lower table parameter. The DEConv module represents a detail enhancement convolution, which fuses basic features and edge details through parallel ordinary convolution and differential convolution, and improves the extraction capability of small targets and texture features, the Conv_Reg module represents a target position regression module, which is used for predicting coordinate offset of a target bounding box, and realizes accurate positioning, the Conv_Cls module represents a target classification module, which outputs probability distribution of a target belonging to each preset category, and Scale represents scale adjustment of a regression result, which adapts to detection requirements of different size targets and enhances multi-scale adaptability.

4. The airport runway foreign matter detection method based on YOLO11 optimization network according to claim 1, wherein, A loss function for training the foreign matter detection model in the S2 is as follows: wherein, represents the total loss output, and a represents a smoothing coefficient, represents a loss value of the i-th sample in the t-th iteration, represents a weight of the i-th sample in the t-th iteration, represents an exponential moving average loss of the i-th sample in the t-th iteration, L total represents a total loss obtained by a weighted sum of a sample loss and a class loss, represents a weight of the c-th class in the t-th iteration, represents an exponential average loss of the c-th class in the t-th iteration, represents a loss value of the c-th class in the t-th iteration.

5. The airport runway foreign matter detection method based on YOLO11 optimization network according to claim 2, characterized in that, The structure of the DSM module is that an input of the DSM module is divided into two paths, one path passes through a spatial selection module, and the other path passes through a frequency selection module, then the two paths are fused and output; the SSM module represents that global average pooling and global maximum pooling are first performed on an input feature map, then the two results are spliced, and then the spliced result is subjected to 3*3 convolution, 5*5 and 7*7 depth separable convolution, then tiling operation is performed, and finally element-level multiplication and addition operations are performed; the FSM module represents that global average pooling is first performed on an input feature map, then the input feature map is subtracted, then element-level multiplication is performed, and finally residual connection is performed with the input feature map; finally, the outputs of the SSM module and the FSM module are integrated by feature fusion and output.

6. The airport runway foreign object detection method based on YOLO11 optimization network according to claim 2, characterized in that, The input of the EdgeEnhancer module passes through an average pooling module, a Conv module, a Subtract module and an Add module; the Avgpool module represents that average pooling operation is performed on an input feature map to reduce resolution and extract low-frequency features; The Conv module represents that 3*3 convolution is used to convolve the pooled feature map to strengthen the basic features; the Subtract module represents that the original input feature map is subtracted from the pooled and convolved feature map to highlight edge high-frequency features; the Add module represents that the original input feature map is added to the subtraction result to retain the original information through residual connection and enhance the edge. Finally, the enhanced feature map is output.

7. The airport runway foreign matter detection method based on YOLO11 optimization network according to claim 3, characterized in that, The input of the DEConv module is divided into two paths, one path passes through the VC module, and the other path passes through the Difference conv module, and then the outputs of the two paths are fused and output; the VC module represents a common 3*3 convolution, and the Difference conv module represents central difference convolution, angle difference convolution, horizontal difference convolution and vertical difference convolution respectively; finally, the outputs of the VC module and the Difference conv module are spliced and fused by feature fusion.

Images