Airport runway intrusion identification method based on ESNB algorithm

Abstract

The invention relates to the technical field of airport safety monitoring, in particular to an airport runway intrusion identification method based on an ESNB algorithm, and the method comprises the steps: constructing an airport runway intrusion data set; the method comprises the following steps: constructing an improved Inception U-Net network, and obtaining a teacher network based on the improved Inception U-Net network; based on the teacher network, obtaining an optimal student network through a multi-objective evolutionary algorithm; based on the teacher network and the airport runway intrusion data set, performing knowledge distillation on the optimal student network to obtain an ESNB network; according to the scheme, the ESNB network is deployed in an airport and is used for airport runway intrusion recognition, and the problems that existing airport runway intrusion recognition is low in efficiency and high in computing power requirement can be effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of airport safety monitoring, and particularly relates to an airport runway intrusion identification method based on an ESNB algorithm. BACKGROUND

[0002] Airport runway intrusion is defined by the International Civil Aviation Organization (ICAO) as any intrusion involving an aircraft, vehicle, or pedestrian that occurs in error within the protected area used for aircraft takeoff and landing in an airport. The occurrence of such events not only poses a serious threat to aviation safety, but is also a key link in the prevention of airport accidents. At present, in order to improve the monitoring capability of the airport runway and timely warn of intrusion risks, the following types of solutions are mainly used:

[0003] The first type is based on radar and cooperative surveillance technology. This type of technology can achieve all-weather detection of airport surface targets by transmitting and receiving electromagnetic waves through ASDE-X ground surveillance radar systems. Or it uses ADS-B and other cooperative surveillance technologies, which rely on cooperative targets such as aircraft or vehicles to actively transmit their own position information, which is monitored by ground receiving stations. However, this type of technology has a detection blind area, which is prone to missed reports and poses a major safety hazard. In addition, the deployment and maintenance of radar systems are costly and difficult to popularize in small and medium-sized airports.

[0004] The second type is based on traditional video monitoring technology. This type of technology deploys cameras at key locations in the airport to obtain real-time video streams of the airport runway and surrounding areas. Monitoring personnel manually interpret the video images, but manual interpretation is prone to fatigue and oversight, and the response speed is slow, which cannot meet the stringent requirements of airport runway intrusion warning for immediacy, resulting in serious delays in response to warnings.

[0005] The third type is emerging visual recognition technology based on deep learning, such as models based on the YOLO (You Only Look Once) algorithm, which aims to directly identify and locate the bounding box of targets such as aircraft and vehicles from images. Or models based on network architectures such as U-Net or Inception U-Net, which aim to classify images at the pixel level and accurately segment the outlines of different areas such as aircraft, airport runways, and backgrounds. Then, by analyzing the relative positional relationship between the segmented aircraft mask and the airport runway mask, it is determined whether an intrusion event has occurred. Although visual recognition technology based on deep learning has lower deployment and maintenance costs, is easy to popularize, and can avoid the fatigue and oversight risks of manual interpretation based on traditional video monitoring technology, the complex structure and large number of parameters of deep learning models result in high computational costs and slow inference speeds, which can easily cause information processing delays and prevent accidents from occurring.

[0006] Therefore, how to provide an airport runway intrusion identification method capable of overcoming the low identification efficiency and high calculation cost of the existing airport runway intrusion identification method is a problem that needs to be solved by those skilled in the art. SUMMARY

[0007] Therefore, the present application provides an airport runway intrusion identification method based on an ESNB algorithm, which solves the problems involved in the background art, and the technical solution is as follows:

[0008] S1, constructing an airport runway intrusion data set;

[0009] S2, constructing an improved Inception U-Net network, and obtaining a teacher network based on the improved Inception U-Net network;

[0010] S3, obtaining an optimal student network based on the teacher network through a multi-objective evolutionary algorithm;

[0011] S4, performing knowledge distillation on the optimal student network based on the teacher network and the airport runway intrusion data set, and obtaining an ESNB network;

[0012] S5, deploying the ESNB network in an airport for airport runway intrusion identification.

[0013] In the above technical solution, through the dual optimization mechanism of the multi-objective evolutionary algorithm and the knowledge distillation, the core problems of target missed detection, model delay, etc. in the existing airport runway intrusion identification technology are solved, and an intelligent solution with high precision, high efficiency, and strong robustness is provided, which is suitable for real-time deployment requirements of various airport safety monitoring scenes.

[0014] Further, the improved Inception U-Net network specifically comprises:

[0015] an encoder, a bottleneck network, and a decoder;

[0016] The encoder comprises a first Inception attention module, a first max-pooling layer, a second Inception attention module, a second max-pooling layer, a third Inception attention module, a third max-pooling layer, a fourth Inception attention module, and a fourth max-pooling layer connected in sequence;

[0017] The decoder comprises a first reverse convolution concatenation layer, a sixth Inception attention module, a second reverse convolution concatenation layer, a seventh Inception attention module, a third reverse convolution concatenation layer, an eighth Inception attention module, a fourth reverse convolution concatenation layer, a ninth Inception attention module, and a softmax activation layer connected in sequence;

[0018] The bottleneck network comprises a fifth Inception attention module, a first cropping layer, a second cropping layer, a third cropping layer and a fourth cropping layer, an input end of the fifth Inception attention module is connected with an output end of the fourth maximum pooling layer, and an output end of the fifth Inception attention module is connected to an input end of the first reverse convolution layer;

[0019] The output of the first Inception attention module is connected to an input end of the fourth reverse convolution splicing layer after the first cropping layer; the output of the second Inception attention module is connected to an input end of the third reverse convolution splicing layer after the second cropping layer; the output of the third Inception attention module is connected to an input end of the second reverse convolution splicing layer after the third cropping layer; and the output of the fourth Inception attention module is connected to an input end of the first reverse convolution splicing layer after the fourth cropping layer.

[0020] The first Inception attention module, the second Inception attention module, the third Inception attention module, the fourth Inception attention module, the fifth Inception attention module, the sixth Inception attention module, the seventh Inception attention module, the eighth Inception attention module and the ninth Inception attention module all adopt the same module structure, comprising an Inception layer and a CBAM attention layer connected in sequence.

[0021] In the above technical scheme, multi-level features are gradually extracted through the four-level downsampling process in the encoder, and the airport runway and target information from local details to global semantics are effectively captured. The decoder gradually restores the spatial resolution through the reverse convolution and the alignment of the cropping layer, ensuring the accuracy of the segmentation boundary. The introduced Inception layer fuses the features of different receptive fields through a multi-branch parallel structure, enhancing the adaptability to targets with significant scale differences in the airport scene. The CBAM attention mechanism at the end of each module further focuses the model on key channels and spatial regions, effectively suppressing complex background interference and improving the robustness to challenging scenes such as occlusion and small targets. The overall structure significantly improves the feature expression ability and recognition accuracy of the model in the specific task of airport monitoring on the basis of maintaining the excellent segmentation performance of U-Net through modular design and attention enhancement.

[0022] Further, each Inception layer comprises a branch unit, a feature splicing layer, a full connection layer and a dimension control convolution layer, the four branches of the branch unit are respectively:

[0023] The first branch comprises a 1x1 convolution unit;

[0024] The second branch includes a 1×1 convolutional unit and a 3×3 convolutional unit connected in sequence.

[0025] The third branch includes a 1×1 convolutional unit and a 5×5 convolutional unit connected in sequence.

[0026] The fourth branch comprises a 3×3 max pooling unit and a 1×1 convolutional unit connected in sequence.

[0027] The first, second, third, and fourth branches are all connected to the input of the feature splicing layer. The output of the feature splicing layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the dimension-controlled convolutional layer, which is a 1×1 convolutional unit. The output of the dimension-controlled convolutional layer is connected to the input of the CBAM attention layer.

[0028] In the above scheme, the Inception layer achieves an excellent balance between multi-scale feature perception and computational efficiency through its four-branch parallel structure. The four branches use 1×1, 3×3, and 5×5 convolutional kernels and pooling operations respectively, enabling the simultaneous capture of features from different receptive fields at the same network level, thus effectively addressing the problem of varying aircraft target scales in airport scenarios. This design also possesses efficient dimensionality reduction capabilities; using 1×1 convolutions before the 3×3 and 5×5 convolutions compresses the number of channels, significantly reducing model complexity. Finally, the feature maps output from each branch are concatenated and fused with 1×1 convolutions to form a highly information-rich "feature map," providing high-quality input for the subsequent attention mechanism and enhancing the model's feature extraction capability and robustness for key targets in complex environments.

[0029] Furthermore, the CBAM attention layer includes a channel attention network layer and a spatial attention network layer connected in sequence;

[0030] The channel attention network layer includes two parallel first average pooling layers and a fifth max pooling layer. The outputs of the first average pooling layer and the fifth max pooling layer are connected to a weight sharing layer. The output of the weight sharing layer is connected to a feature summing layer. The output of the feature summing layer is connected to a first sigmoid activation layer. The output of the first sigmoid activation layer is connected to a first data weighting layer. The first data weighting layer also receives data input to the channel attention network layer. The output of the first data weighting layer is connected to the input of the spatial attention network layer.

[0031] The spatial attention network layer includes two parallel second average pooling layers and a sixth max pooling layer. The outputs of the second average pooling layer and the sixth max pooling layer are connected to a concat layer. The output of the concat layer is connected to a weight learning convolutional layer. The output of the weight learning convolutional layer is connected to a second sigmoid activation layer. The output of the second sigmoid activation layer is connected to a second data weighting layer. The second data weighting layer also receives data output from the first data weighting layer.

[0032] In the above technical solution, the CBAM attention layer achieves adaptive enhancement of key features through its cascaded channel and spatial attention mechanism. The channel attention network accurately evaluates and strengthens the importance of different feature channels by using max pooling and average pooling in parallel and synthesizing them through a shared weight layer, enabling the model to focus on more discriminative information dimensions. Subsequently, the spatial attention network locates and highlights key regions of the target in space through similar double pooling operations and convolutional layers, effectively suppressing interference from complex backgrounds and irrelevant noise. This sequential "channel-space" attention mechanism works together to enable the model to dynamically and accurately concentrate computational resources on key targets in the image, significantly improving the model's robustness and recognition accuracy in complex airport environments (such as occlusion and lighting changes).

[0033] Furthermore, the teacher network obtained based on the improved Inception U-Net network is specifically as follows:

[0034] The airport runway intrusion dataset X={x i The input is fed into the improved Inception U-Net network to obtain the predicted x. i The probability that each pixel represents an aircraft within the airport. ,based on Calculate the first loss function L1:

[0035] The first loss function L1 includes binary cross-entropy loss and Dice loss, wherein the binary cross-entropy loss L... BCE Represented as:

[0036]

[0037] Where N represents a single x i Total number of pixels, y j Represents x i The label corresponding to the j-th pixel; Data representing airport runway incursions x i The j-th pixel represents the predicted probability of an aircraft within the airport.

[0038] Dice loss L DiceRepresented as:

[0039]

[0040] in, This is the smoothing constant.

[0041] The first loss function L1 is expressed as:

[0042]

[0043] in, Represents the weights of the binary cross-entropy loss and , Represents the Dice loss weights;

[0044] Based on the first loss value L1, the improved Inception U-Net is trained using the backpropagation method. After the training stops, the improved Inception U-Net network at this point is saved and denoted as the teacher network.

[0045] In the aforementioned scheme, addressing the severe class imbalance caused by the extremely small proportion of intrusion target pixels in airport runway intrusion identification, this technical solution achieves significant technical results by designing a composite loss function L1 that integrates Binary Cross-Entropy Loss (BCE Loss) and Dice Loss. This design allows BCE Loss to focus on pixel-level classification accuracy, effectively constraining false positives that misclassify the background as an aircraft; simultaneously, leveraging the insensitivity of Dice Loss to class imbalance, it focuses on optimizing the overlap between the predicted and real regions, strongly suppressing false negatives that miss aircraft detection. The two loss functions are fused through weighted fusion, working synergistically during training to ensure high performance in both pixel-level accuracy and overall contour integrity of the segmentation results, reducing the risk of false positives and false negatives in airport runway intrusion events.

[0046] Furthermore, obtaining the optimal student network through a multi-objective evolutionary algorithm includes:

[0047] S3-1. Evaluate the importance of each branch in the Inception attention module of the teacher network to obtain the importance vector V={v b The importance score for the b-th branch is calculated as follows:

[0048]

[0049] in, The parameter set represents the b-th branch; ||.||1 represents calculating the L1 norm; b=1,...,36.

[0050] S3-2. Based on importance vectors, the optimal student network is obtained using a genetic algorithm.

[0051] Generate m student control vectors s based on the importance vector V. x ={u b Let S be the student control vector set. Based on S, construct a student network set and evaluate the fitness of the student networks in the student network set.

[0052] (1) Input the validation set into the student network. For each student network, the following formula applies:

[0053]

[0054] Where E represents the set of pixels predicted to belong to the airport when the validation set is input into the student network, T represents the set of pixels in the validation set that actually belong to the airport, and Fitness represents fitness.

[0055] (2) Based on fitness, a new population is constructed using a strategy that combines elite retention and tournament selection.

[0056] Sq1. Based on fitness, select the E students with the highest fitness values ​​from the initial population as elite individuals, and the rest are ordinary individuals. The elite individuals are directly put into the new population without subsequent tournament selection, crossover and mutation operations.

[0057] Sq2, based on ordinary individuals in Sq1, selects mE student networks using a tournament selection method based on fitness. It should be noted that the tournament selection method is an existing technology and will not be elaborated here.

[0058] Sq3. Based on the extracted mE student networks, crossover and mutation operations are performed according to the corresponding student control vectors. The crossover and mutation student networks are then merged with elite individuals to obtain a new population.

[0059] (3) Perform (1)-(3) on the new population. When the maximum number of iterations is reached, the student network with the highest fitness in the new population at this time (3) is recorded as the optimal student network.

[0060] In the above scheme, a multi-objective evolutionary algorithm is used to structurally prune the teacher network, achieving a balance between lightweight model and high accuracy. The scheme first evaluates the importance of each branch of the teacher network based on the L1 norm, using this as prior knowledge to initialize the population, significantly improving search efficiency. Then, using segmentation accuracy (IoU) as the core fitness metric, and combining elite retention and tournament selection strategies, the genetic algorithm is guided to efficiently search within a multi-objective space comprising model complexity, inference speed, and recognition accuracy. This method can automatically explore and retain the most critical network branches for airport runway intrusion detection tasks, eliminating redundant structures. The resulting optimal student network significantly reduces the number of parameters and inference latency while still inheriting the high accuracy characteristics of the teacher network, successfully resolving the contradiction between lightweight deployment and performance assurance.

[0061] Furthermore, S4 specifically includes:

[0062] S4-1. Calculate the soft tag loss L soft ;

[0063] When the training data in the training set is input into the teacher network, the following calculations are performed:

[0064]

[0065] Where, p L For the training data, the Lth pixel is the teacher's soft label for an aircraft within the airport, v L The Lth pixel of the training data is input into the teacher network, and the output data is the ninth Inception attention module of the teacher network; T is the temperature coefficient.

[0066] When the training data in the training set is input into the optimal student network, the calculation is as follows:

[0067]

[0068] Where, q L For training data, the Lth pixel is a student soft label for an aircraft within an airport, z L The Lth pixel of the training data is input into the optimal student network, and the output data is the ninth Inception attention module of the optimal student network.

[0069] Soft tag loss L soft Represented as:

[0070]

[0071] Where α represents the weighting coefficient, α=0.5.

[0072] S4-2, Calculate the hard label loss L hard .

[0073]

[0074] in, The predicted label is obtained by inputting the Lth pixel into the optimal student network. Let L be the actual label corresponding to the Lth pixel.

[0075] S4-3, based on L hard and L soft Construct the distillation loss L2 and train the optimal student network to obtain the ESNB network.

[0076]

[0077] The optimal student network is trained using the backpropagation algorithm based on distillation loss L2. Training stops when the training stopping condition is met, and the optimal student network at this time is saved and denoted as the ESNB network.

[0078] In the above scheme, the generalization ability of the teacher network is efficiently transferred to the lightweight optimal student network through knowledge distillation. The scheme uses a temperature coefficient T to soften the output of the teacher network, generating soft labels rich in inter-category relationship information, guiding the student network to learn its smooth probability distribution; simultaneously, it combines the hard label loss of the real labels to ensure the accuracy of basic classification. A distillation loss function is constructed by weighted fusion of soft and hard label losses, enabling the student network to inherit the powerful feature representation ability of the teacher network while effectively compensating for the performance loss that may be caused by structural pruning. The resulting ESNB network achieves a significant reduction in parameter count and inference latency while maintaining high accuracy.

[0079] Compared with existing technologies, this invention has at least the following advantages: This invention extracts images of two scenarios—"aircraft intrusion on the airport runway" and "no aircraft intrusion on the airport runway"—from an airport aerial image database. After annotating the images, it creates an airport runway intrusion dataset. An improved Inception U-Net network is trained using this dataset to obtain a teacher network. Based on the teacher network, a multi-objective evolutionary algorithm is used to prune the network to obtain an optimal student network. Furthermore, knowledge distillation is used to optimize the optimal student network, which is then designated as the ESNB network. Finally, the ESNB network is deployed at an airport for runway intrusion detection. This method prunes the teacher network using a multi-objective evolutionary algorithm, removing some neural network branches to reduce computational load. Since the optimal student network is calculated using a multi-objective evolutionary algorithm, its superior performance in the detection task is guaranteed. Furthermore, the optimization of the optimal student network using knowledge distillation further ensures that it inherits the excellent detection performance of the teacher network with fewer parameters. Attached Figure Description

[0080] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0081] Figure 1 This is a flowchart of an airport runway intrusion detection method based on the ESNB algorithm.

[0082] Figure 2 This is a diagram of the improved Inception U-Net network structure for an airport runway intrusion detection method based on the ESNB algorithm.

[0083] Figure 3 The structure of the Inception layer in the airport runway intrusion detection method based on the ESNB algorithm is shown.

[0084] Figure 4 This is a structural diagram of the channel attention network layer in the airport runway intrusion detection method based on the ESNB algorithm.

[0085] Figure 5 This is a structural diagram of the spatial attention network layer in an airport runway intrusion detection method based on the ESNB algorithm. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0087] See Figure 1 This embodiment provides an airport runway intrusion identification method based on the ESNB algorithm, specifically including S1 to S5:

[0088] S1. Construct an airport runway intrusion dataset.

[0089] Images covering two scenarios—"aircraft incursion" and "no aircraft incursion"—were extracted from an airport aerial image database. For each image, if an aircraft incursion occurred, pixels in the overlapping area were labeled 1, and pixels in other areas were labeled 0. For images without an incursion, all pixels were labeled 0. Twenty percent of the labeled images were used as validation data to create the validation set, and the remaining eighty percent were used as training data to create the airport runway incursion dataset X={x i}, where i=1,...,n.

[0090] S2. Construct an improved Inception U-Net network, and obtain the teacher network based on the improved Inception U-Net network.

[0091] like Figure 2 As shown, the improved Inception U-Net network includes an encoder, a bottleneck network, and a decoder. The encoder includes a first Inception attention module, a first max pooling layer, a second Inception attention module, a second max pooling layer, a third Inception attention module, a third max pooling layer, a fourth Inception attention module, and a fourth max pooling layer connected in sequence.

[0092] The decoder comprises a first inverse convolutional convolutional layer, a sixth Inception attention module, a second inverse convolutional convolutional layer, a seventh Inception attention module, a third inverse convolutional convolutional layer, an eighth Inception attention module, a fourth inverse convolutional convolutional layer, a ninth Inception attention module, and a softmax activation layer, all connected in sequence.

[0093] The bottleneck network includes a fifth Inception attention module, a first pruning layer, a second pruning layer, a third pruning layer, and a fourth pruning layer. The input of the fifth Inception attention module is connected to the output of the fourth max pooling layer, and the output of the fifth Inception attention module is connected to the input of the first inverse convolutional layer.

[0094] The output of the first Inception attention module is connected to the input of the fourth inverse convolutional convolutional layer after passing through the first pruning layer; the output of the second Inception attention module is connected to the input of the third inverse convolutional convolutional layer after passing through the second pruning layer; the output of the third Inception attention module is connected to the input of the second inverse convolutional convolutional layer after passing through the third pruning layer; and the output of the fourth Inception attention module is connected to the input of the first inverse convolutional convolutional layer after passing through the fourth pruning layer.

[0095] The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth Inception attention modules all adopt the same module structure, and each Inception attention module includes an Inception layer and a CBAM attention layer connected in sequence.

[0096] like Figure 3 As shown, each Inception layer includes a branch unit, a feature concatenation layer, a fully connected layer, and a dimension-controlled convolutional layer. The four branches of the branch unit are as follows:

[0097] The first branch includes a 1×1 convolutional unit;

[0098] The second branch includes a 1×1 convolutional unit and a 3×3 convolutional unit connected in sequence.

[0099] The third branch includes a 1×1 convolutional unit and a 5×5 convolutional unit connected in sequence.

[0100] The fourth branch comprises a 3×3 max pooling unit and a 1×1 convolution unit connected in sequence.

[0101] The first, second, third, and fourth branches are all connected to the input of the feature splicing layer. The output of the feature splicing layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the dimension-controlled convolutional layer, which is a 1×1 convolutional unit. The output of the dimension-controlled convolutional layer is connected to the input of the CBAM attention layer.

[0102] The CBAM attention layer includes a channel attention network layer and a spatial attention network layer connected in sequence;

[0103] like Figure 4 As shown, the channel attention network layer includes two parallel fifth max pooling layers and a first average pooling layer. The outputs of the fifth max pooling layer and the first average pooling layer are connected to a weight sharing layer. The output of the weight sharing layer is connected to a feature summing layer. The output of the feature summing layer is connected to a first sigmoid activation layer. The output of the first sigmoid activation layer is connected to a first data weighting layer. The first data weighting layer also receives data input to the channel attention network layer. The output of the first data weighting layer is connected to the input of the spatial attention network layer.

[0104] like Figure 5 As shown, the spatial attention network layer includes two parallel sixth max pooling layers and second average pooling layers. The outputs of the sixth max pooling layer and the second average pooling layer are connected to a concat layer. The output of the concat layer is connected to a weight learning convolutional layer. The output of the weight learning convolutional layer is connected to a second sigmoid activation layer. The output of the second sigmoid activation layer is connected to a second data weighting layer. The second data weighting layer also receives data output from the first data weighting layer.

[0105] S2-2. The teacher network is obtained based on the improved Inception U-Net network.

[0106] The airport runway intrusion dataset X is input into the improved Inception U-Net network, for x i For each pixel, a probability is given that the pixel is predicted to be an aircraft within the airport. ,based on Calculate the first loss function L1. The first loss function L1 includes the binary cross-entropy loss and the Dice loss, where the binary cross-entropy loss L... BCE Represented as:

[0107]

[0108] Where N represents a single x i Total number of pixels, y j Data representing airport runway incursions x i The label of the j-th pixel, Data representing airport runway incursions x i The j-th pixel represents the predicted probability of an aircraft within the airport.

[0109] Dice loss L Dice Represented as:

[0110]

[0111] in, It is a smoothing constant. .

[0112] The first loss function L1 is expressed as:

[0113]

[0114] in, The weights representing the binary cross-entropy loss are... , This represents the Dice loss weights, and , In this embodiment, the optimal and .

[0115] Based on the first loss value L1, the improved Inception U-Net network is trained for 500 rounds using the backpropagation method. The resulting Inception U-Net network is denoted as the teacher network.

[0116] S3. Based on the teacher network, the optimal student network is obtained through a multi-objective evolutionary algorithm.

[0117] S3-1. Evaluate the importance of each branch in the Inception attention module of the teacher network to obtain the importance vector V={v b}

[0118] To calculate the importance of branches within each Inception attention module in the teacher network, it's important to note that the teacher network contains 9 Inception attention modules, each with 4 branches. Therefore, there are a total of 36 branches, from the first branch of the first Inception attention module to the fourth branch of the ninth Inception attention module. The importance scores for the branches of the first Inception attention module are v1, v2, v3, and v4, and the importance score for the branches of the ninth Inception attention module is v... 33 v 34 v 35 v 36 .

[0119] The importance score for the b-th branch is calculated as follows:

[0120]

[0121] in, The parameter set represents the b-th branch; ||.||1 represents calculating the L1 norm; b=1,...,36.

[0122] S3-2. Based on importance vectors, the optimal student network is obtained using a genetic algorithm.

[0123] S3-2-1. Generating student control vectors s based on importance vector V x ={u b The initial population is constructed based on the student control vector set.

[0124] Student control vector s x ={u bThe student control vector is a 36-bit vector consisting of 0s and 1s, where b = 1, ..., 36. Each bit of the student control vector represents the retain or pruning state of the corresponding branch of the student network. A value of 0 in the student control vector indicates that the corresponding branch of the student network is in a pruned state, and a value of 1 indicates that the corresponding branch of the student network is in a retained state. The student control vector is generated as follows:

[0125] Sp1, Calculate the student control vector s x The b-th element u b The probability P that is 1:

[0126]

[0127] in, This represents the maximum importance score in the importance vector V;

[0128] Sp2, Determine u b The possible values ​​of:

[0129] Generate a random number r, where 0 ≤ r < 1, if r < P(u b =1), then u b =1, otherwise u b =0;

[0130] m student control vectors are generated using the Sq1-Sq2 method, denoted as the student control vector set S={s x}, where x=1,...,m.

[0131] The student network set is constructed based on the student control vector set S. For each student control vector in the student control vector set, the following method is used:

[0132] Using the teacher network as the base network, which inherits the network structure and parameters of the teacher network, the student control vectors are traversed. The base network is pruned according to the student control vectors. That is, when the value in the student control vector is 0, the branch at the corresponding position in the base network is set to the pruned state, and when the value in the student control vector is 1, the branch at the corresponding position in the base network is set to the retained state. After the base network is pruned, a student network is obtained. Then, the student network is constructed in the same way until m student control networks are constructed. The set of constructed student networks is denoted as the initial population.

[0133] S3-2-2. Based on the initial population, the optimal student network is obtained using a genetic algorithm.

[0134] (1) Input the validation set into the student network. For each student network, the following formula applies:

[0135]

[0136] Where E represents the set of pixels predicted to belong to the airport when the validation set is input into the student network, T represents the set of pixels in the validation set that actually belong to the airport, and Fitness represents fitness. Since there are m student networks in the initial population, m fitness values ​​will be obtained.

[0137] (2) Based on fitness, a new population is constructed using a strategy that combines elite retention and tournament selection.

[0138] Sq1. Based on fitness, select the E students with the highest fitness values ​​from the initial population as elite individuals, and the rest are ordinary individuals. The elite individuals are directly put into the new population without subsequent tournament selection, crossover and mutation operations.

[0139] Sq2, based on ordinary individuals in Sq1, selects mE student networks using a tournament selection method based on fitness. It should be noted that the tournament selection method is an existing technology and will not be elaborated here.

[0140] Sq3. Based on the extracted mE student networks, crossover and mutation operations are performed according to the corresponding student control vectors. The crossover and mutation student networks are then merged with elite individuals to obtain a new population. It should be noted that the crossover and mutation operations are implemented using existing technology and will not be elaborated here.

[0141] (3) Perform (1)-(3) on the new population. When the maximum number of iterations is reached, the student network with the highest fitness in the new population at this time (3) is recorded as the optimal student network.

[0142] S4. Based on the teacher network and the airport runway intrusion dataset, knowledge distillation is performed on the optimal student network to obtain the ESNB network.

[0143] S4-1. Calculate the soft tag loss L soft .

[0144] The training data from the airport runway intrusion dataset is input into the teacher network, and the calculation is as follows:

[0145]

[0146] Where, p L For the training data, the Lth pixel is the teacher's soft label for an aircraft within the airport, v L The Lth pixel of the training data is input into the teacher network, and the output data is the ninth Inception attention module of the teacher network; T is the temperature coefficient, T=10.

[0147] The training data from the airport runway intrusion dataset is input into the optimal student network, and the calculation is as follows:

[0148]

[0149] Where, q L For training data, the Lth pixel is a student soft label for an aircraft within an airport, z L The Lth pixel of the training data is input into the optimal student network, and the output data of the ninth Inception attention module of the optimal student network is also shown; T represents the temperature coefficient, T=10.

[0150] Soft tag loss L soft Represented as:

[0151]

[0152] Where α represents the weighting coefficient, α=0.5.

[0153] S4-2, Calculate the hard label loss L hard .

[0154]

[0155] in, The predicted label is obtained by inputting the Lth pixel into the optimal student network. Let L be the actual label corresponding to the Lth pixel.

[0156] S4-3, based on L hard and L soft Construct the distillation loss L2 and train the optimal student network to obtain the ESNB network.

[0157]

[0158] The optimal student network is trained using the backpropagation algorithm based on distillation loss L2. The training stops when the number of training rounds reaches 500, and the optimal student network at this time is saved and denoted as the ESNB network.

[0159] S5. Deploy the ESNB network at the airport for runway intrusion detection.

[0160] The ESNB network is converted to a format suitable for airport monitoring system deployment and integrated into the airport's existing monitoring system. It performs airport runway intrusion detection by processing camera video streams in real time. Once an intrusion event is detected, it immediately triggers an audible and visual alarm and notifies the tower controller. At the same time, the event details are recorded for post-event analysis and system optimization.

[0161] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An airport runway intrusion identification method based on the ESNB algorithm, characterized in that, Includes the following steps: S1. Construct an airport runway intrusion dataset; S2. Construct an improved Inception U-Net network, and obtain the teacher network based on the improved Inception U-Net network; S3. Based on the teacher network, the optimal student network is obtained through a multi-objective evolutionary algorithm; S4. Based on the teacher network and the airport runway intrusion dataset, knowledge distillation is performed on the optimal student network to obtain the ESNB network. S5. Deploy the ESNB network at the airport for runway intrusion detection.

2. The airport runway intrusion identification method based on the ESNB algorithm according to claim 1, characterized in that, The improved Inception U-Net network specifically includes: Encoder, bottleneck network, and decoder; The encoder includes a first Inception attention module, a first max pooling layer, a second Inception attention module, a second max pooling layer, a third Inception attention module, a third max pooling layer, a fourth Inception attention module, and a fourth max pooling layer connected in sequence. The decoder includes a first inverse convolutional concatenation layer, a sixth Inception attention module, a second inverse convolutional concatenation layer, a seventh Inception attention module, a third inverse convolutional concatenation layer, an eighth Inception attention module, a fourth inverse convolutional concatenation layer, a ninth Inception attention module, and a softmax activation layer, which are connected in sequence. The bottleneck network includes a fifth Inception attention module, a first pruning layer, a second pruning layer, a third pruning layer, and a fourth pruning layer. The input of the fifth Inception attention module is connected to the output of the fourth max pooling layer, and the output of the fifth Inception attention module is connected to the input of the first inverse convolutional layer. The output of the first Inception attention module is connected to the input of the fourth inverse convolutional convolutional layer after passing through the first pruning layer; the output of the second Inception attention module is connected to the input of the third inverse convolutional convolutional layer after passing through the second pruning layer; the output of the third Inception attention module is connected to the input of the second inverse convolutional convolutional layer after passing through the third pruning layer; and the output of the fourth Inception attention module is connected to the input of the first inverse convolutional convolutional layer after passing through the fourth pruning layer. The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth Inception attention modules all adopt the same module structure, including an Inception layer and a CBAM attention layer connected in sequence.

3. The airport runway intrusion identification method based on the ESNB algorithm according to claim 2, characterized in that, Each Inception layer includes a branching unit, a feature concatenation layer, a fully connected layer, and a dimension-controlled convolutional layer. The four branches of the branching unit are as follows: The first branch includes a 1×1 convolutional unit; The second branch includes a 1×1 convolutional unit and a 3×3 convolutional unit connected in sequence. The third branch includes a 1×1 convolutional unit and a 5×5 convolutional unit connected in sequence. The fourth branch comprises a 3×3 max pooling unit and a 1×1 convolutional unit connected in sequence. The first, second, third, and fourth branches are all connected to the input of the feature splicing layer. The output of the feature splicing layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the dimension-controlled convolutional layer, which is a 1×1 convolutional unit. The output of the dimension-controlled convolutional layer is connected to the input of the CBAM attention layer.

4. The airport runway intrusion identification method based on the ESNB algorithm according to claim 2, characterized in that, The CBAM attention layer includes a channel attention network layer and a spatial attention network layer connected in sequence; The channel attention network layer includes two parallel fifth max pooling layers and a first average pooling layer. The outputs of the fifth max pooling layer and the first average pooling layer are connected to a weight sharing layer. The output of the weight sharing layer is connected to a feature summing layer. The output of the feature summing layer is connected to a first sigmoid activation layer. The output of the first sigmoid activation layer is connected to a first data weighting layer. The first data weighting layer also receives data input to the channel attention network layer. The output of the first data weighting layer is connected to the input of the spatial attention network layer. The spatial attention network layer includes two parallel sixth max pooling layers and second average pooling layers. The outputs of the sixth max pooling layer and the second average pooling layer are connected to a concat layer. The output of the concat layer is connected to a weight learning convolutional layer. The output of the weight learning convolutional layer is connected to a second sigmoid activation layer. The output of the second sigmoid activation layer is connected to a second data weighting layer. The second data weighting layer also receives data output from the first data weighting layer.

5. The airport runway intrusion identification method based on the ESNB algorithm according to claim 4, characterized in that, The teacher network obtained based on the improved Inception U-Net network is as follows: The airport runway intrusion dataset X={x i The input is fed into the improved Inception U-Net network to obtain the predicted x. i The probability that each pixel represents an aircraft within the airport. ,based on Calculate the first loss function L1: The first loss function L1 includes binary cross-entropy loss and Dice loss, wherein the binary cross-entropy loss L... BCE Represented as: Where N represents a single x i Total number of pixels, y j Represents x i The label corresponding to the j-th pixel; Data representing airport runway incursions x i The j-th pixel represents the predicted probability of an aircraft within the airport. Dice loss L Dice Represented as: in, It is a smoothing constant; The first loss function L1 is expressed as: in, Represents the weights of the binary cross-entropy loss and , Represents the Dice loss weights and ; Based on the first loss value L1, the improved Inception U-Net is trained using the backpropagation method. After the training stops, the improved Inception U-Net network at this point is saved and denoted as the teacher network.

6. The airport runway intrusion identification method based on the ESNB algorithm according to claim 5, characterized in that, The process of obtaining the optimal student network through a multi-objective evolutionary algorithm includes: S3-1. Evaluate the importance of each branch in the Inception attention module of the teacher network to obtain the importance vector V={v b The importance score for the b-th branch is calculated as follows: in, Let represent the parameter set of the b-th branch; ||.||1 represents calculating the L1 norm; b = 1,...,36; S3-2. Based on importance vectors, use a genetic algorithm to obtain the optimal student network; m student control vectors s are generated based on the importance vector V. x ={u b Let S be the student control vector set. Based on S, construct a student network set and evaluate the fitness of the student networks in the student network set. (1) Input the validation set into the student network. For each student network, the following formula applies: Where E represents the set of pixels predicted to belong to the airport when the validation set is input into the student network, T represents the set of pixels in the validation set that actually belong to the airport, and Fitness represents fitness. (2) Based on fitness, a new population is constructed using a strategy that combines elite retention and tournament selection; Sq1. Based on fitness, extract the E students with the highest fitness values ​​from the initial population as elite individuals, and the rest are ordinary individuals. The elite individuals are put into the new population. Sq2, based on ordinary individuals in Sq1, selects mE student networks using tournament selection based on fitness; Sq3. Based on the extracted mE student networks, crossover and mutation operations are performed according to the corresponding student control vectors. The crossover and mutation student networks are then merged with elite individuals to obtain a new population. (3) Perform (1)-(3) on the new population. When the maximum number of iterations is reached, the student network with the highest fitness in the new population at this time (3) is recorded as the optimal student network.

7. The airport runway intrusion identification method based on the ESNB algorithm according to claim 1, characterized in that, Specifically, S4 is: S4-1. Calculate the soft tag loss L soft ; When the training data in the training set is input into the teacher network, the following calculations are performed: Where, p L For the training data, the Lth pixel is the teacher's soft label for an aircraft within the airport, v L The Lth pixel of the training data is input into the teacher network, and the output data is the ninth Inception attention module of the teacher network; T is the temperature coefficient. When the training data in the training set is input into the optimal student network, the calculation is as follows: Where, q L For training data, the Lth pixel is a student soft label for an aircraft within an airport, z L The Lth pixel of the training data is input into the optimal student network, and the output data is the ninth Inception attention module of the optimal student network. Soft tag loss L soft Represented as: Where α represents the weighting coefficient, α=0.5; S4-2, Calculate the hard label loss L hard ; in, The predicted label is obtained by inputting the Lth pixel into the optimal student network. This is the true label corresponding to the Lth pixel; S4-3, based on L hard and L soft Construct the distillation loss L2 and train the optimal student network to obtain the ESNB network; The optimal student network is trained using the backpropagation algorithm based on distillation loss L2. Training stops when the training stopping condition is met, and the optimal student network at this time is saved and denoted as the ESNB network.

Images