A method for airport target damage detection based on satellite images

Abstract

The present invention relates to the field of airport damage detection technology, and in particular to a method for detecting airport target damage based on satellite images, including constructing an airport facility target detection data set; transforming CSPNet; improving training optimization strategies; optimizing the YOLOv5 network using the DoReFa‑Net multi-value quantization method, and quantizing weights, eigenvalues ​​and gradients at the same time; deploying large image reasoning based on the YOLOv5 network; and performing classified damage detection on oil tank damage, hangar damage, aircraft and runway damage. The present invention provides a deep learning network structure based on structural reparameterization and 8-bit quantization, and detects damage to various airport targets in high-precision satellite remote sensing images in steps; for runway damage, the minimum take-off and landing window is used to perform point-by-point convolution with the runway from different angles, and the area of ​​the minimum take-off and landing window is calculated according to the distribution of craters, so as to determine whether the aircraft can take off after the runway is damaged.

Description

Technical Field

[0001] The invention relates to the technical field of airport damage detection, and in particular to an airport target damage detection method based on satellite images. Background Art

[0002] The remote sensing information support provided by satellite systems has become a crucial source of battlefield information perception. However, problems such as limited data sets and difficult model deployment have become bottlenecks for applications in this field, especially for target identification of various types of infrastructure in airports, which is the key to saving interpreters' time and improving combat efficiency.

[0003] In the field of damage assessment, the existing traditional methods usually use change detection to describe target features and extract damage information. The assessment rules are formulated by establishing the "similarity" of target feature vectors and the "quadratic distance" of geometric and texture features. Although this method has achieved relatively good results by using certain simulation methods, for large images of actual complex targets, the recognition effect is poor when the proportion of various types of targets in the captured image is significantly small. At the same time, this type of algorithm also faces problems such as difficulty in image registration, difficulty in implementation, and difficulty in deployment. Especially when the observation scene, application environment, etc. change, the algorithm is almost ineffective.

[0004] The target damage recognition method based on deep learning neural network model can be beneficial to terminal deployment. It can also combine the specific model of the target and more non-image information to make a more accurate judgment on the target damage. Therefore, it is necessary to carry out research on airport target damage detection technology based on neural network model. Summary of the invention

[0005] In view of the shortcomings of existing algorithms, the present invention provides a deep learning network structure based on structural re-parameterization and 8-bit quantization, and a method for detecting damage to airport targets in high-precision satellite remote sensing images in steps; for runway damage, the minimum take-off and landing window is used to perform point-by-point convolution with the runway from different angles, and the area of ​​the minimum take-off and landing window is calculated according to the crater distribution, so as to determine whether the aircraft can take off after the runway is damaged.

[0006] The technical solution adopted by the present invention is: a method for detecting airport target damage based on satellite images comprises the following steps:

[0007] Step 1: Construct an airport facility target detection dataset, including satellite image datasets of aircraft, hangars, hangar craters, and runway craters;

[0008] Step 2: Using the idea of ​​structural reparameterization, combined with the single-channel model, CSPNet is transformed based on the lightweight YOLOv5 network, and the decoupled training and reasoning architecture is adopted; for the recognition of small targets in remote sensing images, the training optimization strategy is improved;

[0009] Furthermore, CSPNet is modified and the architecture of decoupling training and inference is adopted, including:

[0010] Convert Identity to a 1x1 convolution and construct a 1x1 convolution with the unit matrix as the convolution kernel;

[0011] The 1x1 convolution is equivalently converted to a 3x3 convolution and padded with 0, where the parameters of the 3x3 convolution are 4 3x3 matrices and the parameters of the 1x1 convolution are 1 2x2 matrix; all three branches have BN layers, and the parameters include the accumulated mean and standard deviation, the learned scaling factor and bias. The convolution layer and BN layer during inference are equivalently converted to a convolution layer with bias;

[0012] The 3×3 convolution in CSPNet is structurally reparameterized and re-embedded into the YOLOv5 network as the network architecture during training. During inference, the side branches are fused into the 3×3 convolution module.

[0013] Furthermore, improving the training optimization strategy specifically includes: optimizing the loss function and improving the adaptability of the activation function.

[0014] Furthermore, the optimization of the loss function includes: using the S-IOU loss function, respectively calculating the angle cost, distance cost, shape cost and IoU cost, and obtaining the IOU loss function.

[0015] Furthermore, the adaptability improvement of the activation function includes: merging the 3*3 convolution and the ReLU function into the basic operator, and replacing the swish function with the ReLU function.

[0016] Step 3: Use the DoReFa-Net multi-value quantization method to optimize the YOLOv5 network, quantize the weights, eigenvalues, and gradients at the same time, update the full-precision representation of the weights during backpropagation, and use the quantized representation of the weights during inference; use the direct estimation method to avoid zero gradients during the weight quantization process;

[0017] Furthermore, the formula for DoReFa-Net multi-value quantization is:

[0018]

[0019] The quantization of the eigenvalues ​​is directly done by truncation:

[0020]

[0021] Among them, r i is a real number between 0 and 1, r oIt is a k-bit quantized value ranging from 0 to 1, and h(x) is the activation function.

[0022] Step 4: Deployment of large image reasoning based on YOLOv5 network;

[0023] Further, it specifically includes:

[0024] Implement reasoning for a single image based on the model derived in step 3;

[0025] Use opencv-python to read the image, convert the format, and then normalize it;

[0026] Get the position and confidence of each target box, perform non-maximum suppression, and use the OpenCVcv2.dnn.NMSBoxes function to filter according to the confidence threshold and IOU threshold to obtain the single image inference result;

[0027] The super-large image is cut into small-sized images by step-size cutting, and the appropriate step-size is selected to calculate the position of each small-sized image in the super-large image.

[0028] Step 5: Classify and inspect damage to oil tanks, aircraft hangars, aircraft and runway;

[0029] Furthermore, the detection methods of oil tank damage and aircraft hangar damage include:

[0030] The super large image before damage is used to perform inference recognition of oil tanks and aircraft hangar targets to obtain the actual geographical coordinates of the oil tanks and aircraft hangars;

[0031] The large-scale image of the same area after the damage is inferred and recognized, and the prediction box is checked based on the actual geographic coordinates to determine whether the oil tank and aircraft hangar are damaged;

[0032] The predicted frames of oil tanks and aircraft hangars in the two images before and after the damage are counted to obtain the number of damaged oil tanks and aircraft hangar targets in the area.

[0033] Further, the damage detection method of the aircraft and the runway includes:

[0034] The runway is located according to the coordinates of the four corners of the airport runway, and the crater in the runway area is inferred and identified using the super-large image after damage to obtain the geographical coordinates of the crater;

[0035] The minimum take-off and landing window and the runway are used to perform point-by-point convolution from different angles. When there is no crater, the convolution result is the area of ​​the minimum take-off and landing window, and the position of the take-off and landing window is recorded.

[0036] Further, it specifically includes:

[0037] Input the width and height of the minimum take-off and landing window, the coordinates of the four corners of the runway and the coordinates of the crater, calculate the lengths of the two right-angled sides of the runway and find the slope based on the long side; calculate the coordinates of the center of the runway and the height and width of the runway, find the inclination angle of the runway to determine the rotation matrix, straighten the runway, process the coordinates of the crater in the original image into the coordinates of the rectangle after the runway is straightened, and calculate the minimum take-off and landing window;

[0038] Expand with the runway as the center, exhaustively search the angle range and step of the minimum take-off and landing window, rotate the expanded runway counterclockwise, and keep the size of the expanded runway matrix unchanged; convolve the rotated matrix with the minimum take-off and landing window; under the condition of satisfying the minimum take-off and landing window, if the wider window still has no crater, increase the width of the window, and record the horizontal coordinate, vertical coordinate, width, and height of the take-off and landing window; when the width of the take-off and landing window is the same, if the higher window still has no crater, increase the height of the window; finally, output the coordinates according to the positions of the four corners of the take-off and landing window in the original image.

[0039] Beneficial effects of the present invention:

[0040] 1. Based on the deep learning network structure of structural reparameterization and 8-bit quantization, the damage of each airport target in the high-precision satellite remote sensing image is detected step by step. Compared with the previous interpreter search and change detection method, the present invention can greatly improve the reasoning speed and hardware affinity while ensuring accuracy;

[0041] 2. Use remote sensing maps to construct a detection data set for airport facility targets, improve the YOLOv5 network structure from the perspective of structural reparameterization and loss function and activation function, and implement the inference deployment of super-large satellite images based on the YOLOv5 model based on the reasoning process of a single image. In the super-large image, damage detection is performed according to the different damage effects of various airport targets;

[0042] 3. According to the characteristics of the aircraft and the runway, the minimum take-off and landing window and the runway are used to perform point-by-point convolution from different angles to accurately determine the area of ​​the minimum take-off and landing window, thereby determining whether the airport meets the conditions for aircraft take-off. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the satellite image-based airport target damage detection method of the present invention;

[0044] Figure 2 It is a comparison chart of the model training deployment strategy of the present invention;

[0045] Figure 3 Schematic diagram comparing the RepVGG structure and the ResNet structure of the present invention;

[0046] Figure 4It is a schematic diagram of the improved CSPNet structure of the present invention;

[0047] Figure 5 is a schematic diagram of angle cost calculation of the present invention;

[0048] Figure 6 It is a schematic diagram of IOU loss calculation of the present invention;

[0049] Figure 7 is a flow chart of the deployment of very large image reasoning of the present invention;

[0050] Figure 8 It is the target damage detection process of the present invention;

[0051] Fig. 9 It is the runway crater and the minimum take-off and landing window generated by simulation of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0053] like Figure 1 As shown, a method for detecting airport target damage based on satellite images includes the following steps:

[0054] Step 1: Construct an airport facility target detection dataset;

[0055] Since remote sensing satellites are orbit-fixed satellites, the pixel size in satellite images itself is important feature information, and there is no need to consider its robustness in scale transformation; on this basis, aircraft are divided into smaller fighters and larger special transport aircraft according to their size; for hangars, their arrangement, material characteristics, appearance color and other visual characteristics are considered and divided into 8 categories.

[0056] At the same time, in order to carry out damage detection, data sets are also constructed for hangar craters and runway craters. The shape and color of the craters on the hangar are relatively uniform in black. The ellipse with a diameter of 10 pixels in the image can be used to create a training set by software simulation. It is randomly generated according to the known area range of the crater and merged with the surrounding environment at the edge of the crater; after simulation, it can be automatically labeled to ensure accuracy; for craters near aircraft and runways, their types include irregular circles with diameters of 20 pixels, 40 pixels, and 80 pixels. According to the investigation of actual runway craters, only a single crater needs to be detected, and the arrangement of the craters does not need to be considered. It is only necessary to enrich the type, shape, and color of a single crater on a runway of different materials to generate a data set.

[0057] Step 2: Design of target detection network based on YOLOv5 network;

[0058] Build a lightweight YOLOv5 network to improve inference speed and reduce unnecessary energy consumption while ensuring accuracy, further improve its hardware affinity, and be more conducive to terminal deployment;

[0059] By using the idea of ​​structural reparameterization and combining the characteristics of single-channel model with high speed and multi-channel model with high precision, the CSPNet of YOLOv5 network is transformed. First, the traditional model deployment strategy is transformed to adopt the decoupled training and reasoning architecture; according to the different design performances of the training equipment and the reasoning equipment, different forms of networks are used for design, and then they are equivalently converted by reparameterization. The traditional model training deployment strategy is compared with the model training deployment strategy adopted in this invention. Figure 2 As shown;

[0060] The model training and deployment strategy based on structural reparameterization has the following advantages:

[0061] (1) The 3x3 convolution structure has a very fast computing speed with the support of existing hardware architecture and various optimization libraries;

[0062] (2) The speed of a single-channel architecture is far greater than that of a multi-channel structure with complex branches;

[0063] (3) Single-socket architecture saves memory resources;

[0064] (4) The single-channel architecture is more flexible and it is easy to change the width of each layer.

[0065] like Figure 3 As shown in the figure, after the training is completed, the deployed single-channel model is obtained by equivalently converting the multi-branch model during training. According to the additivity of convolution, the three 3x3 convolution kernels are W1, W2, and W3 respectively:

[0066] conv(x, W1)+conv(x, W2)+conv(x, W3)=conv(x, W1+W2+W3) (1)

[0067] Among them, x is the pixel area in the input image that is convolved with the model.

[0068] like Figure 4, the entire conversion process is divided into two steps: (a) convert Identity to 1x1 convolution and construct a 1x1 convolution with the unit matrix as the convolution kernel; (b) convert the 1x1 convolution equivalently to 3x3 convolution and fill it with 0; since the input and output channels are both 2, the parameters of the 3x3 convolution are 4 3x3 matrices, and the parameters of the 1x1 convolution are 1 2x2 matrix. In addition, all three branches have BN (batch normalization) layers, whose parameters include the accumulated mean and standard deviation and the learned scaling factor and bias. The convolution layer during inference and the subsequent BN layer can be equivalently converted to a convolution layer with bias; using this method, the 3×3 convolution in CSPNet is structurally reparameterized and re-embedded into the YOLOv5 network as the network architecture during training. During inference, the side branches are fused into the 3×3 convolution module, and the model can be restored to its original architecture.

[0069] Furthermore, in order to identify small targets in remote sensing images, the training optimization strategy includes two aspects: (a) optimization and improvement of the loss function; (b) adaptability improvement of the activation function;

[0070] The loss function is aimed at the angle information sensitivity of remote sensing images. The present invention introduces the S-IOU loss function and uses the bounding box regression loss function to supervise the learning of the network. The S-IOU loss function redefines the distance loss by introducing the vector angle between the required regressions, effectively reduces the degrees of freedom of regression, accelerates network convergence, and further improves the regression accuracy.

[0071] The angle cost is the most important part of S-IOU. First, make a prediction on the X or Y axis (whichever is closest), and then continue to approach along the relevant axis. If α < π / 4, the convergence process minimizes α, otherwise minimizes β = π / 2-α; Figure 5 As shown, it is the angle cost calculation process;

[0072] Introduce LF components:

[0073]

[0074] Design weights to balance the dominance of distance and angle factors in different situations:

[0075]

[0076] Taking shape factors into account to correct for irregular regressions:

[0077]

[0078] Simplify and improve the IOU loss:

[0079]

[0080] The specific geometric quantities of IOU are as follows: Figure 6 As shown:

[0081] Finally, the expression of the loss function is obtained:

[0082]

[0083] In terms of activation functions, based on the idea of ​​structural reparameterization, the 3*3 convolution and ReLU function are merged into basic operators, and the swish function is replaced by the ReLU function. Although this will inevitably lead to a loss of accuracy, the consistency of the overall structure and the quantization affinity make this structure gain more benefits in terms of speed.

[0084] Step 3: Airport target training recognition and model export based on lightweight model;

[0085] According to the airport target dataset constructed in step 1, convert the data format according to the input requirements of the YOLOv5 network, place images and labels, and adjust related paths;

[0086] The YOLOv5s network has the smallest depth and feature map width in the YOLOv5 series, the lowest speed, and the lowest AP accuracy. The other three networks are continuously deepened and widened on this basis.

[0087] According to the scale of the data set and the complexity of the scene, the present invention uses YOLOv5s for training and optimizes YOLOv5s using the method in step 2.

[0088] In order to further achieve the lightweight of the model, the present invention adopts the DoReFa-Net multi-value quantization method in training, quantizes weights, eigenvalues ​​and gradients at the same time, updates the full-precision representation of weights during back propagation, and uses the quantized representation of weights during inference; in the process of weight quantization, the straight-through estimator (STE) is used to avoid the problem of zero gradient. The straight-through estimator used for quantization is shown in the following formula:

[0089]

[0090] Among them, r i is a real number between 0 and 1, r o is a k-bit quantized value ranging from 0 to 1.

[0091] The actual weight quantization process can be expressed as follows:

[0092]

[0093] The activation functions of the three constrained quantization input ranges are shown as follows:

[0094]

[0095] h(x)=clip(x, 0, 1) (10)

[0096] h(x)=min(1,|x|) (11)

[0097] In actual quantization training, in order to compromise between computing / storage efficiency and accuracy loss, the DoReFa-Net quantization method was modified to change the quantization interval, constrain the upper and lower limits to avoid taking the value of 1, and remove the symmetric value -1 for the symmetry of the value range. The final quantization algorithm is shown in the following formula:

[0098]

[0099] The quantization of the eigenvalues ​​is directly done by truncation:

[0100]

[0101] Before the quantization training starts, the quantization operations of equations (12) and (13) are added to the convolution operation of the YOLOv5s network. During the forward inference process of the training, the saved full-precision weights and the feature map input from the previous layer are quantized, and then the convolution calculation of this layer is started using the quantized data. In order to reduce the computational complexity, the standardized parameters are also truncated to the 16-bit signed fixed-point representation range; in order to avoid the loss of accuracy, the full-precision gradients are used to update the full-precision weights during the back-propagation process. After the network training is completed, the quantization parameters are extracted for subsequent detection.

[0102] After completing the correct model training and target recognition, the model can be exported. The model export framework used in the present invention is ONNXRuntime. The model file ONNX in the ONNX (Open Neural Network Exchange) format that is most compatible with the framework is selected to export the YOLOv5 network training weights. The YOLOv5 network weight best.pt trained in the previous step is directly exported as an ONNX model for the next step of large image reasoning deployment.

[0103] Step 4: Deployment of large image inference based on the YOLOv5 model;

[0104] The model exported in step 3 is used to realize the reasoning of a single image. The process can be divided into four parts: model loading, image reading and preprocessing, prediction, and post-processing. The exported ONNX format model is loaded, and the reasoning device is selected. The present invention chooses to use GPU reasoning.

[0105] Use opencv-python to read and convert the image format, convert the original image data format into the input data format acceptable to ONNXRuntime, match it with the exported model, rearrange the data and increase the dimension, and finally normalize it;

[0106] The converted input data format is input into the model framework for reasoning, and finally the reasoning result is obtained. The model used has only one input and one output.

[0107] The output results of the Yolov5 model are processed to obtain the final prediction results. After the position and confidence of each target box are obtained, NMS (non-maximum suppression) is performed. Through the OpenCV cv2.dnn.NMSBoxes function, the single image inference results are obtained by filtering according to the confidence threshold and IOU threshold.

[0108] Based on the above single image reasoning process, the reasoning of super-large satellite images can be realized. The process can be divided into three parts: image cutting, small-size image reasoning, and result fusion.

[0109] like Figure 7 A flowchart is deployed for large image reasoning. The 6W×6W large image is cut into 800×800 small-size images that can be input into the network by step cutting. The appropriate step size is selected according to the size of the target and the size of the small-size image to ensure that each target appears at least once in the small-size image after cutting. Pre-cutting is performed to calculate the position of each small-size image in the large image.

[0110] The small-size image calculated in the previous step is cut according to its position in the super-large image. Inference is performed once each image is cut (if batch-size is not 1, multiple images need to be cut at one time and then inference is performed), and the targets close to the edge are deleted. Then the target results are restored to the super-large image and stored. Repeat until the cutting is completed.

[0111] The result obtained in the previous step includes repeated targets caused by step cutting. In the present invention, NMS is used, which is similar to the post-processing process of single image reasoning, to filter out repeated targets and finally obtain the reasoning result of the super-large satellite image.

[0112] Step 5: Classification damage detection method for airport facility targets;

[0113] After the correct inference of each airport target in the super large image is achieved through step 4, the recognition results before and after damage are compared according to the actual geographic coordinates of each airport target to determine whether the target is damaged;

[0114] like Figure 8 The present invention is a target damage detection process. According to the different damage effects of the targets in the airport, the present invention divides the damage into three types: oil tank damage, hangar damage, and aircraft and runway damage for damage detection.

[0115] For flammable and explosive targets such as oil tanks, the targets completely lose their original features after damage. The targets can be identified before damage, but cannot be identified after damage. First, the super-large image before damage is used for inference identification of all oil tank targets through step 4, and the actual geographic coordinates of all oil tanks are obtained. Then, the super-large image of the same area after damage is inferred and identified in the same way, and the actual coordinates are checked to see if there is a prediction frame at the location: if so, it is considered that the oil tank at the coordinate is not damaged; if not, it is considered that the oil tank at the coordinate is damaged. At the same time, the prediction frames of the oil tanks in the two images before and after damage are counted to obtain the number of damaged oil tank targets in the area.

[0116] For non-flammable targets such as hangars, they will not lose their original features after being damaged, and only hangar craters will be left on the targets. Such targets have been produced in step one, and the damaged super-large image is used in step four to perform inference recognition of all hangar targets, and the actual geographic coordinates of all hangars are obtained. According to the actual coordinates, it is checked whether there is a hangar crater prediction frame in the hangar prediction frame at the coordinate: if there is, it is considered that the hangar at the coordinate is damaged; if not, it is considered that the hangar at the coordinate is not damaged; at the same time, the hangar crater prediction frames in the hangar prediction frame in the damaged image are counted, and the number of damaged hangar targets in the area can be obtained.

[0117] For mobile targets such as airplanes, it is necessary to determine whether they can take off normally based on the damage to the runway. In step one, the runway crater target has been created. First, it is located according to the four corner coordinates of the airport runway, and the damaged super-large image is used in step four to infer and identify all craters in the runway area, and the coordinates of all craters are obtained.

[0118] A runway damage detection method was designed based on the identified craters. The main idea is to perform point-by-point convolution with the minimum take-off and landing window and the runway from different angles. When there is no crater, the convolution result is the area of ​​the minimum take-off and landing window, and the position of the take-off and landing window is recorded. In order to facilitate judgment, the rotation matrix and translation matrix are used to flatten the slanted rectangle of the runway.

[0119] First, input the width and height of the minimum take-off and landing window, the coordinates of the four corners of the runway (1*8 matrix, the coordinates of the four corners must be input clockwise or counterclockwise) and the coordinates of all craters (n*2 matrix), calculate the lengths of the two right-angled sides of the runway and find the slope based on the long side; calculate the coordinates of the center of the runway and the height and width of the runway, find the inclination angle of the runway to determine the rotation matrix, straighten the runway, process the coordinates of the craters in the original image into coordinates in the straightened rectangle of the runway, take only the horizontal and vertical coordinates and round them off, and calculate the minimum take-off and landing window.

[0120] Expand with the runway as the center, exhaustively search the angle range and step of the minimum take-off and landing window, rotate the expanded runway counterclockwise, and keep the size of the expanded runway matrix unchanged; convolve the rotated matrix with the minimum take-off and landing window, only consider the valid part, and do not fill in 0; under the condition of satisfying the minimum take-off and landing window, if the wider window still has no crater, increase the width of the window, and record the horizontal coordinate, vertical coordinate, width, and height of the take-off and landing window; when the width of the take-off and landing window is the same, if the higher window still has no crater, increase the height of the window, that is, the height of the window + 1, and finally output the coordinates according to the positions of the four corners of the take-off and landing window in the original image.

[0121] Fig. 9 In order to simulate the generated runway craters and the minimum take-off and landing window, this algorithm can be used to detect runway damage in ultra-large images and then determine whether the aircraft can take off after the damage.

[0122] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for detecting airport target damage based on satellite images, characterized in that: The following steps are involved: Step 1: Construct an airport facility target detection dataset, including satellite image datasets of aircraft, hangars, hangar craters, and runway craters; Step 2: Using the idea of ​​structural reparameterization, combined with the single-channel model, CSPNet is transformed based on the lightweight YOLOv5 network, and the decoupled training and reasoning architecture is adopted; for the recognition of small targets in satellite images, the training optimization strategy is improved; Step 3: Use the DoReFa-Net multi-value quantization method to optimize the YOLOv5 network, quantize the weights, eigenvalues, and gradients at the same time, update the full-precision representation of the weights during backpropagation, and use the quantized representation of the weights during inference; use the direct estimation method to avoid zero gradients during the weight quantization process; Step 4: Deployment of large image reasoning based on YOLOv5 network; Step 5: Classify and inspect damage to oil tanks, aircraft hangars, aircraft and runway; Methods for detecting aircraft and runway damage, including: The runway is located according to the coordinates of the four corners of the airport runway, and the crater in the runway area is inferred and identified using the super-large image after damage to obtain the geographical coordinates of the crater; The minimum take-off and landing window and the runway are used to perform point-by-point convolution from different angles. When there is no crater, the convolution result is the area of ​​the minimum take-off and landing window, and the position of the take-off and landing window is recorded. Specifically include: Input the width and height of the minimum take-off and landing window, the coordinates of the four corners of the runway and the coordinates of the crater, calculate the lengths of the two right-angled sides of the runway and find the slope based on the long side; calculate the coordinates of the center of the runway and the height and width of the runway, find the inclination angle of the runway to determine the rotation matrix, straighten the runway, process the coordinates of the crater in the original image into the coordinates of the rectangle after the runway is straightened, and calculate the minimum take-off and landing window; Expand with the runway as the center, exhaustively search the angle range and step of the minimum take-off and landing window, rotate the expanded runway counterclockwise, and keep the size of the expanded runway matrix unchanged; convolve the rotated matrix with the minimum take-off and landing window; under the condition of satisfying the minimum take-off and landing window, if the wider window still has no crater, increase the width of the window, and record the horizontal coordinate, vertical coordinate, width, and height of the take-off and landing window; when the width of the take-off and landing window is the same, if the higher window still has no crater, increase the height of the window; finally, output the coordinates according to the positions of the four corners of the take-off and landing window in the original image.

2. The method for detecting airport target damage based on satellite images according to claim 1 is characterized in that: CSPNet is modified and the architecture of decoupling training and inference is adopted, including: Convert Identity to a 1x1 convolution and construct a 1x1 convolution with the unit matrix as the convolution kernel; The 1x1 convolution is equivalently converted to a 3x3 convolution and padded with 0, where the parameters of the 3x3 convolution are 4 3x3 matrices and the parameters of the 1x1 convolution are 1 2x2 matrix; all three branches have BN layers, and the parameters include the accumulated mean and standard deviation, the learned scaling factor and bias. The convolution layer and BN layer during inference are equivalently converted to a convolution layer with bias; The 3×3 convolution in CSPNet is structurally reparameterized and re-embedded into the YOLOv5 network as the network architecture during training. During inference, the side branches are fused into the 3×3 convolution module.

3. The method for detecting airport target damage based on satellite images according to claim 1 is characterized in that: Improve training optimization strategies, including: optimization of loss functions and adaptive improvements to activation functions.

4. The method for detecting airport target damage based on satellite images according to claim 3 is characterized in that: The optimization of the loss function includes: using the S-IOU loss function, calculating the angle cost, distance cost, shape cost and IoU cost respectively, and obtaining the IOU loss function.

5. The method for detecting airport target damage based on satellite images according to claim 3 is characterized in that: The adaptive improvements of the activation function include: merging the 3*3 convolution and the ReLU function into the basic operator, and replacing the swish function with the ReLU function.

6. The method for detecting airport target damage based on satellite images according to claim 1 is characterized in that: The formula for DoReFa-Net multi-value quantization is: The quantization of eigenvalues ​​adopts truncation method: Among them, r i is a real number between 0 and 1, r o is a k-bit quantized value ranging from 0 to 1, and h(x) is the activation function.

7. The method for detecting airport target damage based on satellite images according to claim 1 is characterized in that: Step 4 specifically includes: Use opencv-python to read the image, convert the format, and then normalize it; Get the position and confidence of each target box, perform non-maximum suppression, and use the OpenCV cv2.dnn.NMSBoxes function to filter according to the confidence threshold and IOU threshold to obtain the single image inference result; The super-large image is cut into small-sized images by using the step-size cutting method, the step-size is selected, and the position of each small-sized image in the super-large image is calculated.

8. The method for detecting airport target damage based on satellite images according to claim 1 is characterized in that: The detection methods for oil tank damage and aircraft hangar damage include: The super large image before damage is used to perform inference recognition of oil tanks and aircraft hangar targets to obtain the actual geographical coordinates of the oil tanks and aircraft hangars; The large-scale image of the same area after the damage is inferred and recognized, and the prediction box is checked based on the actual geographic coordinates to determine whether the oil tank and aircraft hangar are damaged; The predicted frames of oil tanks and aircraft hangars in the two images before and after the damage are counted to obtain the number of damaged oil tanks and aircraft hangar targets in the area.

Images