Lightweight segmentation method, device, equipment and medium for low-visibility runway of plateau airport

Abstract

The invention discloses a lightweight segmentation method, device and equipment for a low-visibility runway of a plateau airport and a medium, and relates to the field of airport runway image segmentation. The method comprises the following steps: constructing a plateau region airport runway image data set under extreme climate; a YOLOv8n-seg model is used as a reference model, a StarNet network is used as a trunk module in the YOLOv8n-seg model, a gating channel local attention mechanism is fused before down-sampling operation of the YOLOv8n-seg model, a high-frequency feature fusion segmentation head is used as a head network of the YOLOv8n-seg model, and a lightweight Tibetan area airport runway segmentation model is obtained; and training the model by using a plateau region airport runway image data set, and segmenting the airport runway by using the lightweight Tibetan region airport runway segmentation model. According to the method, the adaptability to extreme scenes can be improved, and efficient and high-precision segmentation of the plateau airport runway under the extreme climate is realized.

Description

Technical Field

[0001] The present application relates to the field of airport runway image segmentation, and in particular to a lightweight segmentation method, device, equipment and medium for low-visibility runways at plateau airports. Background Art

[0002] Amidst the rapid development of the global aviation industry, airports, as national strategic hubs, face the combined challenges of extreme climate (fog, snow, and freeze-thaw cycles) and the complex terrain of the plateau. To address the computational resources and response speed demands of plateau airports in this extreme environment, research on lightweight segmentation networks for low-visibility airport runways is underway. This research can improve aircraft landing safety and reduce runway misjudgment caused by weather changes.

[0003] Existing research on airport runway segmentation primarily focuses on model architecture innovation, training strategy optimization, and scenario adaptability. In terms of model architecture, researchers have improved lightweight networks (such as ShuffleNetV2 and MobileNetV3), introduced dilated convolutions and conditional random fields, and combined multi-stage feature decoders to enhance feature fusion capabilities. VALNet addresses scale variations through context enhancement and directional adaptation modules. DCBP (Dual-field-of-view Context and Boundary Perception Network) employs a dual-field-of-view network to improve boundary accuracy, albeit with a large number of parameters. Rotated Mask R-CNN and CNN-Transformer address the challenges of multi-runway intersection and cross-modal segmentation, respectively. In terms of training strategies, researchers combine focal loss, Lovasz-Softmax loss, and cross-modal attention mechanisms to optimize small-sample generalization and boundary sensitivity. For specialized scenarios, they focus on adapting to multi-source data: unsupervised detection methods based on SAR / PolSAR imagery, dehazing networks for haze scenes, unsupervised K-means / ICA contingency plans, and simulation transfer and satellite pre-training to address data scarcity. Li N proposed a feature-guided framework for addressing complex geographic contexts.

[0004] However, existing research still has shortcomings: (1) There is a lack of plateau scene-specific modeling, and most methods have not been optimized for complex meteorological conditions such as low-oxygen haze and strong light reflection that are unique to the plateau; (2) It is difficult to strike a balance between lightweight and precision. The current lightweight network has insufficient segmentation accuracy in low-contrast plateau images, while high-precision models (such as DCBP) are difficult to meet real-time requirements; (3) The model is not robust enough. Traditional methods have poor adaptability to typical plateau scenes such as blurred runway boundaries and snow cover. Summary of the Invention

[0005] The purpose of this application is to provide a lightweight segmentation method, device, equipment and medium for low-visibility runways at plateau airports, which can improve the adaptability to extreme scenarios and achieve efficient and high-precision segmentation of plateau airport runways under extreme climates.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In the first aspect, the present application provides a lightweight segmentation method for low-visibility runways at plateau airports, including: constructing a plateau area airport runway image dataset under extreme climate; establishing a gated channel local attention mechanism; the gated channel local attention mechanism is used to adaptively fuse dynamic channel attention features and local spatial attention features; establishing a high-frequency feature fusion segmentation head based on the Scharr operator; using the YOLOv8n-seg model as a benchmark model, using the StarNet network as the backbone module in the YOLOv8n-seg model, fusing the gated channel local attention mechanism before the downsampling operation of the YOLOv8n-seg model, and using the high-frequency feature fusion segmentation head as the head network of the YOLOv8n-seg model to obtain a lightweight Tibetan airport runway segmentation model; using the plateau area airport runway image dataset to train the lightweight Tibetan airport runway segmentation model to obtain a trained lightweight Tibetan airport runway segmentation model; inputting the image of the plateau area airport runway to be tested under extreme climate into the trained lightweight Tibetan airport runway segmentation model to output the segmented airport runway.

[0008] Optionally, a dataset of airport runway images in plateau areas under extreme climate conditions is constructed, specifically including:

[0009] Fusion of multi-source heterogeneous images of airport runways in plateau areas;

[0010] Using AI visual models, extreme climate scenes are generated in multi-source heterogeneous images, and a dataset of airport runway images in plateau areas under extreme climate conditions is obtained.

[0011] Optionally, the gated channel local attention mechanism includes: a dynamic channel attention path, a local importance attention path and a gating mechanism;

[0012] The dynamic channel attention path is used to perform a global average pooling operation on the input feature map to obtain a feature matrix; the feature matrix is ​​transposed and then subjected to a one-dimensional convolution to obtain a channel description vector; the channel description vector output by the one-dimensional convolution is globally compressed using a Sigmoid activation function to generate cross-channel attention weights; the one-dimensional convolution adopts an adaptive kernel size;

[0013] The local importance attention path is used to convolve the input feature map to obtain convolution features; after using the operator to perform pooling operations on the convolution features, exponential weighted average is performed to obtain spatial features; multi-scale features in the spatial features are further extracted through convolution with a step size of 2 to obtain a multi-scale feature map; the multi-scale feature map is activated pixel by pixel through convolution and Sigmoid activation function, and finally bilinear interpolation is used for upsampling to generate local importance attention weights;

[0014] The gating mechanism is used to extract the first channel feature from the input feature map and generate the gating factor through the Sigmoid activation function;

[0015] The cross-channel attention weights, local importance attention weights, gating factors, and input feature maps are fused to obtain fused features.

[0016] Optionally, the kernel size is calculated as:

[0017]

[0018] Where k is the kernel size, C is the number of input channels, b is the base offset, and β is the scale factor. Indicates rounding down, l k％2＝0 It means adding 1 when k is an even number.

[0019] Optionally, the high-frequency feature fusion segmentation head includes: a feature processing branch, a Proto branch and a detection branch;

[0020] Feature processing branches include ScharrConv convolution, Conv convolution, group convolution, depth-wise separable convolution, and Conv2d convolution;

[0021] ScharrConv convolution is used to obtain edge features using the Scharr convolution kernel based on the fused feature map;

[0022] Conv convolution is used to obtain one-dimensional convolution features based on the fusion feature map;

[0023] After channel concatenation of edge features and one-dimensional convolution features, input group convolution;

[0024] Group convolution is used to obtain the features after dual-branch convolution fusion based on the features after channel splicing;

[0025] The features fused by the dual-branch convolution are passed through depth-wise separable convolution and Conv2d convolution to obtain the mask coefficients;

[0026] The Proto branch is used to generate a prototype mask based on the first layer feature map in the fusion feature map;

[0027] The detection branch is used to output the predicted box and category based on the fused feature map, and together with the mask coefficient and prototype mask, it constitutes the airport runway segmentation result.

[0028] Optionally, the ScharrConv convolution process is expressed as: F scharr =SiLU(BN(DSConv 3×3 (Scharr(Ft)))); where F scharr is the edge feature, SiLU is the SiLU activation function, BN is the BathNorn2d batch normalization function, DSConv represents the depth separable convolution, DSConv 3×3 represents a 3×3 depth-wise separable convolution, Scharr represents the Scharr convolution kernel, and Ft is the fused feature map;

[0029]

[0030] G x =Conv 3×3 (Ft,K x );

[0031] G y =Conv 3×3 (Ft,K y );

[0032]

[0033] Among them, G x is the horizontal gradient, G y is the vertical gradient; ∈ is a constant, ∈=10 -6 ; E is the edge intensity amplitude, K x For the horizontal template, K y It is a vertical template;

[0034] The Conv convolution process is expressed as: F conv =SiLU(BN(Conv 3×3 (Conv 1×1 (Ft)))); where F conv is a one-dimensional convolution feature, Conv 3×3 Represents 3×3 convolution, Conv 1×1 Represents 1×1 convolution;

[0035] The process of the group convolution, the depth-wise separable convolution and the Conv2d convolution is expressed as: Among them, F concat It is the feature after dual-branch convolution fusion, GroupConv 3×3 represents a 3×3 group convolution, and g represents the number of groups.

[0036] Optionally, the lightweight Tibetan airport runway segmentation model comprises a backbone network, a neck network and a head network.

[0037] The backbone network is configured to generate multi-scale preliminary features according to the plateau airport runway image under extreme climate; the neck network is configured to generate a fusion feature map according to the multi-scale preliminary features; and the head network is configured to generate an airport runway segmentation result according to the fusion feature map.

[0038] The backbone network comprises a backbone module and an SPPF module.

[0039] The neck network comprises an up-sampling path and a down-sampling path, and the gate channel local attention mechanism is fused before the down-sampling operation in the down-sampling path.

[0040] In a second aspect, the application provides a lightweight plateau airport runway segmentation device under low visibility, comprising a data set construction module, an attention mechanism establishment module, a segmentation head establishment module, a segmentation model establishment module, a training module and an application module.

[0041] The data set construction module is configured to construct a plateau airport runway image data set under extreme climate.

[0042] The attention mechanism establishment module is configured to establish a gate channel local attention mechanism; and the gate channel local attention mechanism is configured to adaptively fuse dynamic channel attention features and local spatial attention features.

[0043] The segmentation head establishment module is configured to establish a high-frequency feature fusion segmentation head based on a Scharr operator.

[0044] The segmentation model establishment module is configured to take a YOLOv8n-seg model as a benchmark model, adopt a StarNet network as a backbone module in the YOLOv8n-seg model, fuse the gate channel local attention mechanism before a down-sampling operation of the YOLOv8n-seg model, and use the high-frequency feature fusion segmentation head as a head network of the YOLOv8n-seg model, to obtain a lightweight Tibetan airport runway segmentation model.

[0045] The training module is configured to train the lightweight Tibetan airport runway segmentation model using the plateau airport runway image data set under extreme climate, to obtain a trained lightweight Tibetan airport runway segmentation model.

[0046] The application module is configured to input a to-be-tested plateau airport runway image under extreme climate into the trained lightweight Tibetan airport runway segmentation model, and output a segmented airport runway.

[0047] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described lightweight segmentation methods for low-visibility runways at plateau airports.

[0048] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned lightweight segmentation methods for low-visibility runways at plateau airports.

[0049] According to the specific embodiments provided in this application, this application has the following technical effects:

[0050] The present application provides a lightweight segmentation method, apparatus, equipment and medium for low-visibility runways at plateau airports, a gated channel local attention mechanism, and adaptive fusion of dynamic channel attention features and local spatial attention features, which enhance the network's ability to capture runway features in low-contrast and low-visibility scenarios, and make up for the shortcomings of plateau meteorological condition modeling; the YOLOv8n-seg model is used as the benchmark model, which significantly reduces the number of parameters; the high-frequency feature fusion segmentation head extracts high-frequency edge features based on the Scharr operator, which improves boundary segmentation accuracy while maintaining computational efficiency; the boundary segmentation robustness is enhanced by the high-frequency feature fusion segmentation head, and the gated channel local attention mechanism is combined to optimize multi-scale feature fusion, thereby improving the lightweight Tibetan airport runway segmentation model's resistance to fuzzy boundaries and extreme weather interference, and improving its adaptability to extreme scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0052] Figure 1 A flowchart of a lightweight segmentation method for low-visibility runways at plateau airports provided in one embodiment of the present application;

[0053] Figure 2 A schematic diagram of a dual-path collaborative working mode adopted by a gated channel local attention mechanism provided in another embodiment of the present application;

[0054] Figure 3 A schematic diagram of the high-frequency feature fusion segmentation head network structure provided in another embodiment of the present application;

[0055] Figure 4A schematic diagram of the network structure of a lightweight Tibetan airport runway segmentation model provided by another embodiment of the present application;

[0056] Figure 5 A schematic diagram comparing the segmentation results of the ground truth map, Baseline model, and Baseline+ScharrHead fusion model provided in this application;

[0057] Figure 6 Schematic diagram of the segmentation results of Gongga Airport runway using different lightweight segmentation models in cloudy weather provided for this application;

[0058] Figure 7 Schematic diagram of the segmentation results of Linzhi Airport runway using different lightweight segmentation models on a rainy day provided by this application;

[0059] Figure 8 Schematic diagram of the segmentation results of the runway of Alikunsha Airport in snowy weather provided by this application using different lightweight segmentation models;

[0060] Figure 9 A schematic diagram of the functional modules of a lightweight segmentation device for low-visibility runways at plateau airports provided in one embodiment of the present application;

[0061] Figure 10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0063] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0064] In an exemplary embodiment, Figure 1 As shown, a lightweight segmentation method for low-visibility runways at plateau airports is provided, comprising the following steps 101 to 106. In which:

[0065] Step 101: Construct an airport runway image dataset in plateau areas under extreme climate conditions.

[0066] Step 102: Establish a gated channel local attention mechanism; the gated channel local attention mechanism is used to adaptively fuse dynamic channel attention features and local spatial attention features.

[0067] Step 103: Establish a high-frequency feature fusion segmentation head based on the Scharr operator.

[0068] Step 104: Using the YOLOv8n-seg model as the baseline model, the StarNet network is used as the backbone module in the YOLOv8n-seg model, the gated channel local attention mechanism is integrated before the downsampling operation of the YOLOv8n-seg model, and the high-frequency feature fusion segmentation head is used as the head network of the YOLOv8n-seg model to obtain a lightweight Tibetan airport runway segmentation model.

[0069] Step 105: Using the plateau area airport runway image dataset, a lightweight Tibetan area airport runway segmentation model is trained to obtain a trained lightweight Tibetan area airport runway segmentation model.

[0070] Step 106: Input the image of the airport runway in the plateau area to be tested under extreme climate into the trained lightweight Tibetan airport runway segmentation model, and output the segmented airport runway.

[0071] By implementing the above steps 101 to 106, this application mainly addresses the following key technical issues in the task of runway segmentation at plateau airports:

[0072] 1. Lack of specific modeling of plateau scenes

[0073] Existing methods lack targeted optimization for complex meteorological conditions unique to plateau environments, such as low-oxygen haze, strong light reflections, and snow cover. This application proposes Gated Channel Local Attention (GCLA), which integrates and optimizes multi-scale features to enhance the network's ability to capture runway features in low-contrast and low-visibility scenarios, thus addressing the shortcomings of plateau meteorological modeling.

[0074] 2. Imbalance between lightweight and precision

[0075] Current lightweight networks (such as MobileNet and ShuffleNet) lack segmentation accuracy in low-contrast images from the plateau, while high-precision models (such as DCBP) have high computational complexity and are unable to meet real-time requirements. This application reduces the number of parameters by using a layered residual architecture and a depthwise separable convolutional reconstruction feature extraction network. At the same time, it designs a lightweight segmentation head (ScharrHead) to improve boundary segmentation accuracy while maintaining computational efficiency, achieving a balance between efficiency and accuracy.

[0076] 3. Insufficient robustness in extreme environments

[0077] Traditional methods are less adaptable to typical plateau scenarios (such as blurred runway boundaries, snow cover, and freeze-thaw cycles). This application enhances the robustness of boundary segmentation by using a high-frequency feature fusion segmentation head (ScharrHead) and combines GCLA to optimize multi-scale feature fusion, improving the model's resistance to blurred boundaries and extreme weather interference.

[0078] 4. Demand for network architecture innovation

[0079] Different from the existing general architecture of YOLOv8n-seg, this application proposes a completely redesigned lightweight network that integrates targeted modules such as layered residual, depthwise separable convolution, and attention mechanism to achieve full-process optimization from feature extraction to segmentation head to meet the special needs of plateau airports.

[0080] Summary: The core of this application is to solve the difficult problem of "efficiency, accuracy, and robustness" in runway segmentation at plateau airports through lightweight architecture design (reducing computing resource consumption) and scenario-driven algorithm optimization (improving adaptability to plateau environments), providing technical support for aviation safety in extreme climates.

[0081] In another exemplary embodiment of the present application, the above step 101 may be replaced by the following steps 201 to 202:

[0082] Step 201: Fusing multi-source heterogeneous images of airport runways in plateau areas.

[0083] Data source: Fusion of multi-source heterogeneous data such as Google Earth and Cesium.

[0084] Step 202: Using an AI visual model, generate extreme climate scenes in multi-source heterogeneous images to obtain an image dataset of airport runways in plateau areas under extreme climate conditions.

[0085] Data enhancement: AI visual models are used to generate low-visibility weather scenes (such as haze, snowfall, strong light reflection, etc.) to improve the spatiotemporal diversity and environmental complexity of the data set.

[0086] In another exemplary embodiment of the present application, the present application proposes a gated channel local attention mechanism. By integrating this mechanism before the downsampling operation in the Neck feature fusion network of the YOLOv8n-seg model, irrelevant channels can be pre-filtered, detail information with higher activation response can be retained, a local receptive field can be established, and the spatial activation weight of the runway area can be enhanced. The problems of detail loss and positioning error in the downsampling in the feature fusion stage can be solved, the impact of complex background noise such as ice and snow cover, terrain interference, etc., which are common in plateau airports, can be reduced, and the robustness of the model can be improved.

[0087] The gated channel-based local attention mechanism utilizes a dual-path collaborative architecture, consisting of a dynamic channel-based attention path, a local importance attention path, and a gating mechanism. The dynamic channel-based attention path enhances key channel features and improves sensitivity to low-contrast tracks. The local spatial attention path models the correlation of local structures in the spatial dimension, improving adaptability to blurred boundaries. A learnable gating factor adaptively balances the features of the two paths to optimize feature fusion.

[0088] Figure 2 The dual-path collaborative working mode adopted by the gated channel local attention mechanism is shown, which realizes adaptive feature enhancement through the joint optimization of dynamic recalibration in the channel dimension and local structure perception. In the channel dimension, the mechanism captures long-range dependencies through global average pooling and uses the fully connected layer to generate the channel attention weight W c , to achieve dynamic screening and enhancement of key features; in the spatial dimension, a convolution kernel with a restricted receptive field is used to extract local structural features W s , maintaining sensitivity to fine-grained spatial information. The outputs of the two branches are nonlinearly fused through a learnable gating mechanism, and the gating weights are dynamically generated by the input features, ensuring that the model can adaptively balance the contributions of channel and spatial features.

[0089] The dynamic channel attention path is an improved channel attention interaction mechanism using adaptive convolution kernels. For a given input feature Figure X First, perform a global average pooling operation to obtain the feature matrix Y; then transpose Y and perform a one-dimensional convolution to obtain Y'; finally, use the Sigmoid activation function to globally compress the channel description vector output by the one-dimensional convolution to generate the cross-channel attention weight W c , establish long-range channel dependencies and enhance the feature response of the target main area. The above process can be expressed as:

[0090] Y=AvgPool(X)

[0091] Y′=Conv1D(Y T ).

[0092] W c =Sigmoid(Y′)

[0093] Among them, Conv1D adopts an adaptive kernel size design to dynamically adjust the convolution kernel size in the channel dimension, effectively establish long-range dependencies across channels, and capture nonlinear interactions between channels. The kernel size k is calculated as follows:

[0094]

[0095] Where C is the number of input channels, b is the base offset, and β is the scale factor. Indicates rounding down, l k％2＝0 It means that when k is an even number, 1 is added to ensure that the convolution kernel is an odd size.

[0096] The local importance attention path first performs an Figure X Perform a 1×1 convolution to preliminarily adjust the feature dimension to obtain Z; then use the SoftPool operator to perform a pooling operation on Z, and then use the exponential weighted average method to retain the significant spatial features to obtain Z'; then, use a 3×3 convolution with a step size of 2 to further extract multi-scale features to obtain Z"; then use a 3×3 convolution and a Sigmoid activation function to activate the feature map in the local window pixel by pixel to enhance the positioning accuracy of the boundary area; finally, use bilinear interpolation upsampling to obtain the local importance attention path output W s The above process can be expressed by the following formula. ij It is a 7×7 local window centered at (i, j); Bilinear means bilinear interpolation upsampling.

[0097] Z=Conv 1×1 (X)

[0098] Z′=SoftPool(Z)

[0099]

[0100] Z″=Conv 3×3 (Z′)

[0101] W s =Bilinear(Sigmoid(Conv 3×3 (Z″)))

[0102] The gating mechanism is used to further filter and control the feature map. The first channel in the feature map is used as the initial feature map of the depthwise separable convolution. The low-level texture information it encodes (such as edge gradients and local contrast) has a basic representation capability for the runway boundary in low visibility scenes. Therefore, first, the input feature map is Figure X Take out the first channel X :1 , the gating factor gate is generated through the Sigmoid activation function to dynamically adjust the fusion ratio of the dual attention paths. The gate is calculated as follows:

[0103] gate=Sigmoid(X :1 ).

[0104] Finally, GCLA performs computational processing by integrating dynamic channel attention path, local spatial attention path and gating mechanism, forming a complementary relationship in three dimensions: channel enhancement, spatial positioning and scene adaptation, and jointly improving the feature discrimination ability in low visibility scenes. The final output of GCLA is X out :

[0105] X out =gate⊙W c ⊙W s .

[0106] In another exemplary embodiment of the present application, based on Yolov8n-seg, the Scharr operator is introduced to construct a dual-branch feature extraction module, redesign the segmentation head structure, and propose ScharrHead---a lightweight high-frequency feature fusion segmentation head ( Figure 3 ). Assume that the input feature map Where B represents Batch size, is a set of real numbers, C is the number of feature map channels, H represents the height of the feature map, and W represents the width of the feature map. The Scharr convolution kernel consists of two directional templates K x and K y Definition, horizontal gradient G x , vertical gradient G y The edge strength amplitude E is calculated as follows, where ∈ = 10 -6 , which can prevent zero gradient and avoid gradient explosion.

[0107]

[0108] Dual-branch convolution module F dual By Scharr branch F scharr , conventional convolution branch F conv Composition. F dual Define the following formula. Where DSConv stands for depthwise separable convolution, F concat It is the feature after dual-branch convolution fusion; Represents a channel concatenation operation; g is the number of groups.

[0109]

[0110] Among them, F scharr is the edge feature, SiLU is the SiLU activation function, BN is the BathNorn2d batch normalization function, DSConv represents the depth separable convolution, DSConv 3×3 represents a 3×3 depth-wise separable convolution, Scharr represents the Scharr convolution kernel, and Ft is the fused feature map. conv is a one-dimensional convolution feature, Conv3×3 Represents 3×3 convolution, Conv 1×1 Represents 1×1 convolution. GroupConv 3×3 Represents a 3×3 group of convolutions.

[0111] Therefore, for the improved segmentation head ScharrHead, let the input multi-scale feature be L is the number of feature maps, L=3, Ft0 is the first layer feature output by the backbone network, and the original P is generated by the prototype generation module; X is spliced ​​by multi-scale features. Then generate the mask coefficient M c Finally, the detection output Y inherited from the parent class Detect det , generate the output of the improved segmentation head ScharrHead Output.

[0112]

[0113] like Figure 4 As shown in the figure, the high-frequency feature fusion segmentation head includes a feature processing branch, a Proto branch, and a detection branch. The feature processing branch includes ScharrConv convolution, Conv convolution, group convolution, depthwise separable convolution, and Conv2d convolution. ScharrConv convolution is used to obtain edge features based on the fused feature map using the Scharr convolution kernel. Conv convolution is used to obtain one-dimensional convolution features based on the fused feature map. The edge features and one-dimensional convolution features are channel-wise concatenated and then input into group convolution. Group convolution is used to obtain the features after the two-branch convolution fusion based on the channel-wise concatenated features. The features after the two-branch convolution fusion are passed through depthwise separable convolution and Conv2d convolution to obtain mask coefficients. The Proto branch is used to generate a prototype mask based on the first-layer feature map in the fused feature map. The detection branch is used to output the predicted box and category based on the fused feature map, which together with the mask coefficients and prototype mask form the airport runway segmentation result.

[0114] The characteristics of the edge feature guided segmentation head (ScharrHead) are as follows:

[0115] High-frequency feature enhancement: Extract high-frequency features of the runway edge based on the Scharr operator to enhance boundary segmentation capabilities.

[0116] Lightweight design: While reducing the number of parameters and computational complexity of the baseline model, it maintains a slight improvement in segmentation performance.

[0117] In another exemplary embodiment of the present application, the above step 201 is replaced by the following steps 301 to 303:

[0118] XARSegNet (Xizang Airport Runway Segmentation Network) is a lightweight, real-time segmentation network for airport runways in Tibet under low-visibility conditions. It uses the YOLOv8n-seg model as a baseline architecture for lightweighting and performance improvement. It mainly includes three innovations: (1) The StarNet network is used to replace the feature extraction backbone module of the baseline model to achieve model lightweighting; (2) Gated channel local attention (GCLA) is proposed and introduced into the multi-scale feature fusion Neck stage of the baseline model to improve feature discriminability in interference scenarios such as clouds, fog, and snowfall; (3) The segmentation head module is optimized by embedding the Scharr operator in a differentiable form (Scharr convolution) into the feature decoding process, constructing a dual-branch decoding structure, and enhancing the accuracy of runway boundary feature extraction through high-frequency gradient response (edge ​​intensity amplitude E).

[0119] The overall network structure of XARSegNet is as follows Figure 4 As shown in the figure, the baseline model is optimized through a progressive structure of feature extraction, attention enhancement, and edge optimization, achieving improved airport runway segmentation performance in low visibility conditions while maintaining its lightweight characteristics. The network architecture consists of three main components: the backbone network, the neck network, and the head network (ScharrHead). The input image is passed through the backbone network, generating feature maps of three different scales (1 / 8, 1 / 16, and 1 / 32). These feature maps are then fed into the feature fusion network. After fusion, the feature maps are processed by the head network (ScharrHead) to produce the prediction results.

[0120] The backbone network achieves efficient feature extraction through the STEM module, DownSample module, StarBlock module, and SPPF module. The STEM module performs low-level feature extraction; the DownSample module is responsible for gradually reducing the resolution of the feature map and increasing the number of channels to extract high-level semantic information; the StarBlock module uses depthwise separable convolution to reduce computational complexity and combines it with residual connections to avoid the vanishing gradient problem; and the SPPF module fuses features from different receptive fields through multi-scale pooling to form a preliminary multi-scale representation.

[0121] The neck network is based on a feature pyramid structure and combines upsampling (Upsample), downsampling (Conv), feature concatenation (Concat), and GCLA modules to achieve dynamic fusion of multi-level features. In the upsampling path, Upsample (nearest neighbor interpolation) amplifies deep features by 2 times, concatenates them with shallow high-resolution features through Concat, and then optimizes the feature expression through the C2f module. The downsampling path reduces the resolution through Conv (3×3 convolution with a stride of 2) and concatenates them with deep features. The GCLAttention mechanism module dynamically calculates the weights of features at different levels, highlighting important features and suppressing noise, aiming to improve the segmentation ability of small and occluded targets. It outputs a fusion feature map of three scale features: Ft[0] (1 / 8 resolution), Ft[1] (1 / 16), and Ft[2] (1 / 32).

[0122] The ScharrHead network is responsible for the final semantic segmentation task. It inherits the Detect class and expands the segmentation function based on YOLOv8n object detection. It includes a Proto branch, a feature processing branch, and a detection branch. It integrates the three scale features Ft[0], Ft[1], and Ft[2] output by Neck. The Proto branch uses the first-layer feature map Ft[0] of the backbone network to generate the prototype mask P; the feature processing branch consists of ScharrConv convolution, Conv convolution, group convolution GroupConv, depth-separable convolution DSConv, and Conv2d convolution. After processing the feature map, it outputs the mask coefficient mc; the detection branch (inherits the Detect class) outputs the predicted box and category cls, which together with mc and p form the segmentation result. During training, the three separate data are returned. During inference / export, the detection result and the mask coefficient are spliced ​​and output together.

[0123] This application achieves efficient and high-precision segmentation of plateau airport runways through the technical route of multi-source dataset construction → attention mechanism optimization → edge feature enhancement → lightweight model design, meeting the real-time requirements under extreme climate conditions.

[0124] To verify the effectiveness of the proposed method, a comparative experiment was designed to conduct a comparative analysis with existing mainstream lightweight channel-spatial hybrid attention methods (including CPCA and SCSA) on a test dataset of airport runway images in a plateau area under extreme climate conditions. The results are shown in Table 1. Specifically, each attention module is uniformly embedded before the downsampling operation in the feature fusion stage.

[0125] Table 1 Comparative experimental results of lightweight channel-spatial hybrid attention

[0126] Attention P / ％ mIOU / % GFLOPs Params(M) - 96.6 95.10 12.0 3.259 CPCA

[26] 97.3 96.09 12.4 3.302 SCSA

[27] 97.7 96.14 12.1 3.262 MSCA

[28] 97.2 95.59 12.2 3.301 EMA

[29] 97.6 96.10 12.0 3.259 MLCA

[30] 96.6 96.06 24.8 3.259 GCLA 98.2 96.37 12.0 3.325

[0127] While maintaining the same network depth and backbone structure, we analyzed the differential effects of different attention mechanisms on the model's segmentation performance, computational efficiency, and parameter count. Experimental results show that, with the same hyperparameters for both construction and training, introducing the attention mechanism before the downsampling operation in the quantitative feature fusion phase can enhance the model's segmentation performance. Comparing the experimental results of different attention mechanisms, GCLA achieved the best improvement in the two key metrics of P and mIOU. Although the addition of the GCLA module increased the model's parameter increment by approximately 2% compared to the baseline model, P and mIOU improved by 1.6% and 1.27% compared to the baseline model. Given its optimal improvement in segmentation accuracy, GCLA was ultimately chosen, preserving structural improvement space for subsequent model lightweighting.

[0128] ScharrHead Advantages:

[0129] Compare the Baseline (YOLOv8n-seg) model, the Baseline+ScharrHead fusion model, and the true value map. The comparison results are as follows: Figure 5 shown. Figure 5 Part (a) shows the truth graph, Figure 5 Part (b) shows the segmentation result of the Baseline model. Figure 5 Part (c) shows the segmentation results of the Baseline+ScharrHead fusion model. The results show that the Baseline model has obvious over-segmentation, which is manifested as follows: (1) the runway end area is not effectively segmented; (2) due to cloud occlusion, non-runway areas are misidentified as runways. The Baseline model fused with ScharrHead effectively suppresses over-segmentation by enhancing edge features, achieving segmentation and positioning of the runway edge.

[0130] Advantages of XARSegNet:

[0131] In the low-visibility plateau airport runway segmentation task, lightweight semantic segmentation models are key to ensuring safety and feasibility. To verify the comprehensive performance of the XARSegNet model, we compared its performance with lightweight segmentation models commonly used on edge devices. The comparative experimental results are shown in Table 2.

[0132] Table 2 Comparative experimental results of XARSegNet and lightweight segmentation models

[0133] Model P / ％ mIOU / % GFLOPs Params(M) BiseNetV2 95.29 92.27 98.284 3.341 FastSCNN 94.60 91.35 7.412 1.398 ERFNet 93.41 88.77 121.0 3.716 HRNet-v2 95.17 91.39 77.400 3.935 SegNeXt-T 95.37 90.7 50.279 4.225 Baseline 96.60 95.10 12.0 3.259 XARSegNet 96.70 95.25 8.6 2.043

[0134] XARSegNet shows excellent performance advantage in lightweight segmentation model. Compared with the baseline model Baseline(Yolov8n-seg), the GFLOPs and Params of XARSegNet are reduced by 28.33%(12.0→8.6) and 38.85%(3.259M→2.043M) respectively, while the key performance indicators P and mIOU are improved by 0.1%(96.60%→96.70%) and 0.15%(95.10%→95.25%) respectively, realizing the dual optimization of model efficiency and accuracy. Compared with other models, XARSegNet has the highest P and mIOU while keeping the calculation amount of 8.6 GFLOPs. BiSeNetV2 has higher accuracy(P=95.29%, mIOU=92.27%), but its GFLOPs and Params are 11.43 times and 1.64 times of XARSegNet respectively. FastSCNN has the lowest Params and GFLOPs, but its P(94.60%) and mIOU(91.35%) are reduced by 2.22% and 4.27% compared with XARSegNet. In addition, ERFNet, HRNet-v2 and SegNeXt-T not only have high calculation amount, but also do not reach the accuracy of XARSegNet.

[0135] To evaluate the performance visualization difference between lightweight segmentation models and XARSegNet model, different weather conditions(foggy day, rainy day, snowy day) and different geographical locations of plateau city airports(Gongga, Nyingchi, Ali Kunsha) are selected for visualization comparison and analysis. The comparison results are as follows Figure 6-Figure 8 The runway segmentation details are enlarged and displayed in the yellow frame on the side of each comparison figure, which intuitively presents the robustness of the model in complex environments.

[0136] Figure 6 is the segmentation result of Gongga airport runway by different lightweight segmentation models in cloudy weather. Except for XARSegNet, the rest of the models failed to complete the main runway segmentation. BiSeNetV2, FastSCNN, ERFNet, HRNet-v2 and SegNeXt-T completely failed in the cloud-covered area, and Baseline only retained most of the runway structure integrity. XARSegNet still kept the correct segmentation of runway pixels in the cloud area, verifying its anti-occlusion segmentation ability.

[0137] Figure 7The results of Linzhi Airport runway segmentation using different lightweight segmentation models on a rainy day are shown below. BiseNetV2, FastSCNN, ERFNet, HRNet-v2, and SegNeXt-T all failed to effectively segment the target runway area, with their results exhibiting significant loss due to rain interference. In contrast, the Baseline model and XARSegNet achieved segmentation of the main runway segmentation, with XARSegNet performing better in terms of segmentation completeness and edge continuity at the runway ends.

[0138] Figure 8 The results of runway segmentation for Alikunsha Airport using different lightweight segmentation models in snowy weather. BiseNetV2, ERFNet, and HRNet-v2 suffer from incomplete runway segmentation, while SegNeXt-T fails to effectively segment the runway area. The FastSCNN, Baseline, and XARSegNet models fully segment the runway, but FastSCNN and Baseline exhibit mis-segmentation. The XARSegNet model maintains complete segmentation while eliminating mis-segmentation noise, and its segmented edges match the actual runway edges more closely than the Baseline model.

[0139] For the task of airport runway image segmentation in plateau areas, the innovations of this application are as follows:

[0140] ① Multi-source dataset construction (TARD (Tableland Airport Runway Dataset) dataset)

[0141] This paper first integrates heterogeneous data from Google Earth and Cesium to construct a specialized dataset for airport runways in plateau regions. It also uses AI visual models to enhance low-visibility weather scenarios and improve data diversity. This approach fills a gap in existing research on runway segmentation datasets under extreme plateau climates (such as haze, snow, and strong sunlight).

[0142] ②Gated Channel Local Attention (GCLA)

[0143] A dual-path collaborative architecture (dynamic channel attention + local spatial attention) is proposed to adaptively balance feature fusion through learnable gating factors.

[0144] ③Edge feature guided segmentation head (ScharrHead)

[0145] A lightweight segmentation head designed based on the Scharr operator enhances high-frequency edge features and improves runway boundary segmentation accuracy. This reduces computational complexity while still maintaining a slight improvement in segmentation performance, achieving a balance between lightweighting and accuracy.

[0146] ④ Lightweight Plateau Airport Runway Segmentation Model (XARSegNet)

[0147] The depth-wise separable convolution + layered residual structure is used to optimize computational efficiency.

[0148] Based on the above innovations, the advantages of this application are:

[0149] (1) Solve the problem of lack of specific modeling of plateau scenes

[0150] Disadvantages of existing technologies: Existing methods lack targeted optimization for plateau-specific environments (such as low-oxygen haze, strong light reflection, snow cover, etc.), resulting in insufficient feature capture capabilities of the model in low-contrast and low-visibility scenarios.

[0151] Advantages of this application:

[0152] ① Propose gated channel local attention, enhance key channel features through dynamic channel attention path, and improve sensitivity to low-contrast runways.

[0153] ② Combined with the local spatial attention path, the local structural correlation of the spatial dimension is modeled to improve the adaptability to fuzzy boundaries.

[0154] ③ Through the learnable gating factor, the dual-path features are adaptively balanced, the feature fusion is optimized, and the shortcomings of the plateau meteorological condition modeling are compensated.

[0155] (2) Solve the imbalance between lightweight and precision

[0156] Disadvantages of existing technologies: Current lightweight networks (such as MobileNet and ShuffleNet) have insufficient segmentation accuracy in low-contrast images on the plateau, while high-precision models (such as DCBP) have high computational complexity and are difficult to meet real-time requirements.

[0157] Advantages of this application:

[0158] ① Using a layered residual architecture and depth-wise separable convolutional reconstruction feature extraction network to significantly reduce the number of parameters.

[0159] ② Design the ScharrHead lightweight segmentation head, which extracts high-frequency edge features based on the Scharr operator, improving boundary segmentation accuracy while maintaining computational efficiency.

[0160] ③Achieve a balance between efficiency and accuracy to meet the real-time segmentation needs of plateau airports.

[0161] (3) Solve the problem of insufficient robustness in extreme environments

[0162] Disadvantages of existing technologies: Traditional methods have poor adaptability to typical plateau scenes (such as blurred runway boundaries, snow cover, and freeze-thaw cycles), and their segmentation performance is unstable.

[0163] Advantages of this application:

[0164] ① The high-frequency feature fusion segmentation head (ScharrHead) enhances the robustness of boundary segmentation and effectively copes with fuzzy boundaries and extreme weather interference.

[0165] ②The GCLA module optimizes multi-scale feature fusion and improves the model's resistance to complex background noise (such as ice and snow cover and terrain interference).

[0166] ③ Combined with the TARD dataset (including enhanced data of low-visibility meteorological scenes), the generalization ability of the model in extreme environments is improved.

[0167] (4) Solve the problem of network architecture innovation needs

[0168] Disadvantages of existing technologies: Existing methods (such as YOLOv8n-seg) are general architectures and are not optimized for the special needs of plateau airports.

[0169] Advantages of this application:

[0170] ① Propose a completely redesigned lightweight network XARSegNet, which integrates targeted modules such as layered residual, depthwise separable convolution, and attention mechanism.

[0171] ② Achieve full-process optimization from feature extraction to segmentation head, significantly improving segmentation performance in low-visibility scenarios on the plateau.

[0172] ③Through multi-source dataset construction and AI visual enhancement technology, fill the gap in plateau airport runway segmentation datasets.

[0173] This application solves the difficult problem of "efficiency-accuracy-robustness" coordination in plateau airport runway segmentation through lightweight architecture design (reducing computing resource consumption) and scenario-driven algorithm optimization (improving adaptability to plateau environments), providing efficient and reliable technical support for aviation safety in extreme climates.

[0174] Based on the same inventive concept, embodiments of the present application also provide a device for segmenting low-visibility runways at plateau airports, which is used to implement the aforementioned method for segmenting low-visibility runways at plateau airports. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for segmenting low-visibility runways at plateau airports provided below can be found in the aforementioned limitations of the method for segmenting low-visibility runways at plateau airports, and will not be repeated here.

[0175] In an exemplary embodiment, Figure 9As shown, a lightweight segmentation device for low-visibility runways at plateau airports is provided, including: a dataset building module, an attention mechanism building module, a segmentation head building module, a segmentation model building module, a training module and an application module.

[0176] The dataset construction module is used to construct an airport runway image dataset in plateau areas under extreme climate conditions.

[0177] An attention mechanism establishment module is used to establish a gated channel local attention mechanism; the gated channel local attention mechanism is used to adaptively fuse dynamic channel attention features and local spatial attention features.

[0178] The segmentation head establishment module is used to establish a high-frequency feature fusion segmentation head based on the Scharr operator.

[0179] A segmentation model establishment module is used to build a lightweight Tibetan airport runway segmentation model, using the YOLOv8n-seg model as the baseline model and the StarNet network as the backbone module in the YOLOv8n-seg model. The gated channel local attention mechanism is integrated before the downsampling operation of the YOLOv8n-seg model, and the high-frequency feature fusion segmentation head is used as the head network of the YOLOv8n-seg model.

[0180] The training module is used to use the plateau area airport runway image dataset to train a lightweight Tibetan area airport runway segmentation model to obtain a trained lightweight Tibetan area airport runway segmentation model.

[0181] The application module is used to input the images of airport runways in the plateau area to be tested under extreme climate into the trained lightweight Tibetan airport runway segmentation model and output the segmented airport runways.

[0182] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store segmented airport runways. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a lightweight segmentation method for low-visibility runways at plateau airports is implemented.

[0183] Those skilled in the art will understand that Figure 10 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned method embodiments when executing the computer program.

[0184] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0185] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0186] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0187] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0188] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0189] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A lightweight segmentation method for low-visibility runways at plateau airports, characterized in that: include: Construct an image dataset of airport runways in plateau areas under extreme climate conditions; Establish a gated channel local attention mechanism; The gated channel local attention mechanism is used to adaptively fuse dynamic channel attention features and local spatial attention features; Establish a high-frequency feature fusion segmentation head based on the Scharr operator; The YOLOv8n-seg model was used as the baseline model. The StarNet network was adopted as the backbone module in the YOLOv8n-seg model. A gated channel local attention mechanism was integrated before the downsampling operation of the YOLOv8n-seg model. The high-frequency feature fusion segmentation head was used as the head network of the YOLOv8n-seg model to obtain a lightweight Tibetan airport runway segmentation model. Using the plateau airport runway image dataset, training a lightweight Tibetan airport runway segmentation model to obtain a trained lightweight Tibetan airport runway segmentation model; The images of the airport runways in the plateau area to be tested under extreme climate are input into the trained lightweight Tibetan airport runway segmentation model, and the segmented airport runways are output.

2. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 1 is characterized in that: Construct an image dataset of airport runways in plateau areas under extreme climate conditions, specifically including: Fusion of multi-source heterogeneous images of airport runways in plateau areas; Using AI visual models, extreme climate scenes are generated in multi-source heterogeneous images, and a dataset of airport runway images in plateau areas under extreme climate conditions is obtained.

3. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 1 is characterized in that: The gated channel local attention mechanism includes: a dynamic channel attention path, a local importance attention path and a gating mechanism; The dynamic channel attention path is used to perform a global average pooling operation on the input feature map to obtain a feature matrix; the feature matrix is ​​transposed and then subjected to a one-dimensional convolution to obtain a channel description vector; the channel description vector output by the one-dimensional convolution is globally compressed using a Sigmoid activation function to generate cross-channel attention weights; the one-dimensional convolution adopts an adaptive kernel size; The local importance attention path is used to convolve the input feature map to obtain convolution features; after using the operator to perform pooling operations on the convolution features, exponential weighted average is performed to obtain spatial features; multi-scale features in the spatial features are further extracted through convolution with a step size of 2 to obtain a multi-scale feature map; the multi-scale feature map is activated pixel by pixel through convolution and Sigmoid activation function, and finally bilinear interpolation is used for upsampling to generate local importance attention weights; The gating mechanism is used to extract the first channel feature from the input feature map and generate the gating factor through the Sigmoid activation function; The cross-channel attention weights, local importance attention weights, gating factors, and input feature maps are fused to obtain fused features.

4. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 3 is characterized in that: The formula for calculating the kernel size is: Where k is the kernel size, C is the number of input channels, b is the base offset, and β is the scale factor. Indicates rounding down, l k％2＝0 It means adding 1 when k is an even number.

5. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 1 is characterized in that: The high-frequency feature fusion segmentation head includes: a feature processing branch, a Proto branch and a detection branch; Feature processing branches include ScharrConv convolution, Conv convolution, group convolution, depth-wise separable convolution, and Conv2d convolution; ScharrConv convolution is used to obtain edge features using the Scharr convolution kernel based on the fused feature map; Conv convolution is used to obtain one-dimensional convolution features based on the fusion feature map; After channel concatenation of edge features and one-dimensional convolution features, input group convolution; Group convolution is used to obtain the features after dual-branch convolution fusion based on the features after channel splicing; The features after the dual-branch convolution are fused through depth-wise separable convolution and Conv2d convolution to obtain the mask coefficients; The Proto branch is used to generate a prototype mask based on the first layer feature map in the fusion feature map; The detection branch is used to output the predicted box and category based on the fused feature map, and together with the mask coefficient and prototype mask, it constitutes the airport runway segmentation result.

6. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 5 is characterized in that: The ScharrConv convolution process is expressed as: F scharr =SiLU(BN(DSConv 3×3 (Scharr(Ft)))); where F scharr is the edge feature, SiLU is the SiLU activation function, BN is the BathNorn2d batch normalization function, DSConv represents the depth separable convolution, DSConv 3×3 represents a 3×3 depth-wise separable convolution, Scharr represents the Scharr convolution kernel, and Ft is the fused feature map; G x =Conv 3×3 (Ft,K x ); G y =Conv 3×3 (Ft,K y ); Among them, G x is the horizontal gradient, G y is the vertical gradient; ∈ is a constant, ∈=10 -6 ; E is the edge intensity amplitude, K x For the horizontal template, K y It is a vertical template; The Conv convolution process is expressed as: F conv =SiLU(BN(Conv 3×3 (Conv 1×1 (Ft)))); where F conv is a one-dimensional convolution feature, Conv 3×3 Represents 3×3 convolution, Conv 1×1 Represents 1×1 convolution; The process of the group convolution, the depth-wise separable convolution and the Conv2d convolution is expressed as: Among them, F concat It is the feature after dual-branch convolution fusion, GroupConv 3×3 represents a 3×3 group convolution, and g represents the number of groups.

7. The lightweight segmentation method for low-visibility runways at plateau airports according to claim 1 is characterized in that: The lightweight Tibetan airport runway segmentation model includes: a backbone network, a neck network and a head network; The backbone network is used to generate multi-scale preliminary features based on the images of airport runways in plateau areas under extreme climate conditions; the neck network is used to generate fused feature maps based on the multi-scale preliminary features; and the head network is used to generate airport runway segmentation results based on the fused feature maps. The backbone network includes the backbone module and the SPPF module; The neck network consists of an upsampling path and a downsampling path, and a gated channel-local attention mechanism is fused before the downsampling operation in the downsampling path.

8. A lightweight segmentation device for low-visibility runways at plateau airports, characterized in that: include: A dataset construction module is used to construct an image dataset of airport runways in plateau areas under extreme climate conditions; An attention mechanism establishment module is used to establish a gated channel local attention mechanism; the gated channel local attention mechanism is used to adaptively fuse dynamic channel attention features and local spatial attention features; Segmentation head establishment module, used to establish high-frequency feature fusion segmentation head based on Scharr operator; A segmentation model building module is used to build a lightweight Tibetan airport runway segmentation model, using the YOLOv8n-seg model as the baseline model and the StarNet network as the backbone module in the YOLOv8n-seg model. A gated channel local attention mechanism is integrated before the downsampling operation of the YOLOv8n-seg model, and a high-frequency feature fusion segmentation head is used as the head network of the YOLOv8n-seg model. A training module is used to train a lightweight Tibetan airport runway segmentation model using the plateau area airport runway image dataset to obtain a trained lightweight Tibetan airport runway segmentation model; The application module is used to input the images of airport runways in the plateau area to be tested under extreme climate into the trained lightweight Tibetan airport runway segmentation model and output the segmented airport runways.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the lightweight segmentation method for low-visibility runways at plateau airports according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the lightweight segmentation method for low-visibility runways at plateau airports according to any one of claims 1 to 7 is implemented.

Images