A high-resolution remote sensing target detection method that integrates spatial relationships

Abstract

A high-resolution remote sensing target detection method integrating spatial relationships is proposed. First, a deep convolutional neural network is designed based on the remote sensing target extraction task. Samples are created and trained to obtain a road extraction model. Then, a target detection network model is created to train the samples and labels. The target detection network model is trained, and image feature map information is obtained through convolution. During training, the target detection network model is used for prediction, and confidence loss, category loss, and bounding box position loss are calculated respectively. Road information is used to determine the distance of the predicted location from the road, and the loss value is output for regression calculation, enabling the network to learn that the detected target is near the road. In the post-processing stage, the finally trained target detection model is used to detect the target dataset. The predicted output is processed to obtain the bounding boxes of the identified features in the remote sensing image. The center point information of the bounding boxes is extracted, converted into vector information, and written to a vector file. This invention achieves fine extraction of high-resolution remote sensing targets.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing and target detection in computer vision, and proposes a high-resolution remote sensing target detection method that integrates spatial relationships. Technical Background

[0002] In recent years, with the rapid development of remote sensing technology, high-resolution remote sensing image data has become increasingly abundant, greatly promoting applied research in the field of remote sensing. Remote sensing images contain important targets such as aircraft, ships, and bridges. Combining the spatial relationships between targets during the detection process can achieve higher detection accuracy, and has wide application value in military and civilian fields such as airport aircraft management, port vessel management, and road vehicle management.

[0003] Since the advent of deep learning, mainstream object detection algorithms can be broadly categorized into two types: convolutional neural networks (CNNs) for region proposal-based object detection and end-to-end integrated CNNs for object detection. End-to-end object detection methods, aiming for faster detection speeds, use the results of a single detection as the final object detection result. Their performance is less than ideal compared to multi-stage training based on region proposal-based object detection methods. However, the emergence of the YOLO series offers a more direct approach: directly regressing the bounding box position and its category at the output layer, treating object detection as a regression problem. A single CNN structure can directly predict the bounding box and category probability from the input image. While the accuracy is still not as high as region proposal-based object detection methods, the difference is not significant, and the detection speed is significantly faster than the R-CNN series. Current research on remote sensing targets focuses on using deep learning-based object detection methods for identification and detection. Du Zexing et al. used densely connected networks to detect targets in remote sensing images, and this network exhibits good detection performance for multi-scale remote sensing targets. Ju Moran et al. used an improved YOLOv3 method to detect small remote sensing targets, achieving high detection accuracy. While target detection achieves high accuracy on most datasets, its performance is inconsistent when detecting actual ground features in high-resolution remote sensing images due to numerous interfering factors such as neighboring features, imaging conditions, and image differences. Furthermore, most current detection algorithms for man-made ground features do not actively consider the relationships between targets, which are crucial for human interpretation. Summary of the Invention

[0004] This invention aims to overcome the problem of inaccurate and incomplete target extraction in high-resolution remote sensing images due to factors such as the complexity of ground features. It proposes a high-resolution remote sensing target detection method that integrates spatial relationships. The road extraction algorithm extracts road information from the dataset, converts it into a binary map, and inputs it into the target detection network. A buffer is created based on the road binary map, and the detection range of the buffer is determined to reduce the influence of interference factors in the image. Finally, it achieves fine extraction of high-resolution remote sensing targets.

[0005] To achieve the above objectives, the technical solution proposed by this invention is as follows:

[0006] A high-resolution remote sensing target detection method that integrates spatial relationships includes the following steps:

[0007] Step 1: Prepare a sample, which includes:

[0008] Step 1.1: Acquire high-resolution remote sensing imagery: Use optical satellite remote sensing data with visible-near-infrared sensors or aerial remote sensing data with a general optical camera. Depending on the resolution requirements, multispectral imagery or panchromatic imagery can be used directly.

[0009] Step 1.2: Cropping based on remote sensing image size: Select the area where typical targets are located in the production area and crop the remote sensing image to the same pixel size (the cropping range based on the size of the production area should have a certain degree of randomness and broad coverage).

[0010] Step 1.3: Create training samples for the road extraction model.

[0011] Step 1.4: Based on the task requirements, a large number of images and corresponding annotations are generally collected as samples, and the samples are divided into training set and test set according to a certain ratio.

[0012] Step 2: Train the road extraction model, which specifically includes:

[0013] Step 2.1: Design a deep convolutional neural network, select a deep learning network structure suitable for extracting the target, such as D-LinkNet, HED, RCF, etc., and improve it. The network can be replaced or modified accordingly based on the production target.

[0014] Step 2.2: Initialize weights: Use the existing network model as a pre-trained model to initialize the network weights.

[0015] Step 2.3: Set training hyperparameters: Configure specific values ​​after model tuning.

[0016] Step 2.4: Input samples and train the model: Input the training samples into the network and train it according to the hyperparameters described in Step 2.3 to obtain a road extraction model that can extract road information.

[0017] Step 3: Obtain road information as "prior knowledge," specifically including:

[0018] Step 3.1: Input the high-resolution remote sensing image into the road extraction model to predict the probability that each pixel is a road. The output is a grayscale image from 0 to 255. Adjust the threshold to convert the road extraction image into a binary image. The conversion formula is as follows:

[0019]

[0020] Where threshold is a real number in the interval [0,1], which can be set by the user, with an initial default value of 0.5. x and y are the horizontal and vertical coordinates of the image.

[0021] Step 3.2: Using the road as the center, radiate three pixels as a buffer and extract the buffer range.

[0022] Step 4: Create training samples for the object detection network model, specifically including:

[0023] Step 4.1: The remote sensing image is cropped in the same way as the road extraction model.

[0024] Step 4.2: Create training samples and labels for the object detection network model.

[0025] Step 5: Train the object detection model, which includes the following six steps:

[0026] Step 5.1: Set the training parameters for the target detection model.

[0027] Step 5.2: The image is fed into the network for data augmentation. Data slicing is performed in the backbone to improve detection speed. The neck adopts a structure borrowed from CSP2 to enhance the network's feature fusion capabilities.

[0028] Step 5.3: After passing through the preceding network, a feature map is obtained. This feature map is divided into several grid cells. The size of the grid cells is determined by the size of the input image. If the center of an object falls within a grid cell, then that grid cell is responsible for predicting that object. Each network needs to predict three bounding box locations and one confidence score. A bounding box corresponds to four location information points and one confidence score. The confidence score represents both the confidence that the predicted bounding box contains the object and the accuracy of the bounding box prediction.

[0029] Step 5.4: The target center falls within a grid, record various information about the target, and output its tensor.

[0030] Step 5.5: In the prediction process, the road location is used as "prior knowledge". Since the main ground objects to be detected in this example are service areas and toll stations, which are mainly distributed on both sides of the road and have obvious spatial adjacency, the range is delineated according to the buffer zone formed by the road binary map. The filtering is only performed in the area radiating three grids from the road to reduce the prediction of interference factors and improve the detection speed.

[0031] Step 5.6: The YOLO network has now obtained the scale feature map. Each two-dimensional point on the scale feature map corresponds to a grid in the input image. The feature map is associated with the prediction box. End-to-end training is started. Through linear regression calculation, the anchor box is made close to the true label of the target.

[0032] Step 6: The predicted value is subjected to loss, specifically including:

[0033] Step 6.1: During training, an object detection network model is used for prediction, and three losses are calculated: confidence loss, class loss, and box position loss. The overlap between the predicted box and the label box is continuously corrected according to the bounding box regression formula proposed by the authors.

[0034]

[0035] Step 6.2: Then, the road loss of this invention is input together with the road binary map for regression calculation, so that the network learns that the detected target is close to the road.

[0036] Step 7: Traverse the image and input vector information, specifically including:

[0037] Step 7.1: Parametric Boundary Box: Convert the bounding box parameters output by the object detection model into the coordinates of its lower left and upper right vertices. The conversion formula is as follows:

[0038]

[0039] Step 7.2: Convert the original pixel coordinates to spatial geographic coordinates, traverse the image to obtain high-precision prediction boxes. Compared with the prediction boxes of the original method, this invention has a significant accuracy advantage. Write to a vector file. During the writing process, check if the vector file exists. If it exists, write normally; if it does not exist, create the vector file first and then write. Then concatenate the four spatial geographic coordinates to close the loop.

[0040] Unlike existing target detection methods, high-resolution remote sensing target detection methods that integrate spatial relationships model the visual and semantic features of remote sensing images. A convolutional neural network-based visual module extracts features from the remote sensing images to obtain rich visual features such as remote sensing target region features, location features, and target categories. The extracted visual features are then input into a semantic module to map to a higher-level semantic embedding space. Through deep fusion of visual and semantic features, the spatial relationships between remote sensing targets ultimately improve target detection accuracy, achieving a deep perception and understanding of visual and spatial information in remote sensing images. Integrating spatial relationships in target detection is a future trend and has already been applied in some target detection methods. For example, the RSSRD model proposed in remote sensing target spatial relationship detection methods has good spatial relationship applications. It first accurately identifies different remote sensing targets and their locations, and then detects and displays the spatial positional relationships between different targets, such as distance, vertical alignment, and horizontal alignment. In port identification based on deep learning and spatial analysis, to address the problem of the difficulty in automatically identifying ports, a progressive identification model of ship-terminal-port is proposed by combining deep learning and geospatial analysis. The model detects and locates the terminal category and pixel coordinates, converts them into geographic coordinates, performs hotspot analysis, and then uses density clustering to identify and extract the location and extent of the port.

[0041] The main idea of ​​this invention is to use the spatial relationship between two targets as "prior information," and then use this as a constraint to narrow the detection range and improve the accuracy and recall of target detection. Most artificial targets are planned and constructed with clear spatial relationships, and each category has its own unique attributes and limitations. For example, service areas and toll stations are often built next to roads, and roads can serve as "prior information" for service area detection. By utilizing this corresponding spatial relationship and pre-determining the prior conditions for roads, and then detecting targets around the roads, the accuracy is improved. Furthermore, in large-area detection, only the grid around specific roads needs to be predicted, which greatly improves efficiency. This is the main improvement idea of ​​this method: by introducing prior conditions based on the spatial relationship between remote sensing targets, we can fully understand the deeper spatial semantic information in remote sensing images, thereby improving the overall accuracy of high-resolution remote sensing artificial target detection.

[0042] This invention employs a deep learning-based multi-stage framework to achieve refined extraction of targets with road spatial relationships. The framework comprises two stages: road extraction and target detection. By learning prior extraction methods from the field of salient target detection, the road extraction algorithm identifies and extracts road information from the dataset, which is then fed into the target detection network as "prior information." Next, a post-prior method from four prior combination methods is selected, and a buffer is created based on the road binary map to determine the buffer detection range, thereby reducing the influence of interfering factors in the image.

[0043] Due to the adoption of the above technical solution, the present invention has the following advantages and beneficial effects:

[0044] 1. This invention employs a method that combines a target detection model and a road extraction model. The road extraction model ensures the accuracy of the "prior knowledge," while the target detection model, combined with the "prior knowledge," improves the accuracy of recognition.

[0045] 2. Compared with traditional manual identification of remote sensing targets, the present invention has a faster identification speed. For example, during the identification process, due to the influence of prior knowledge, the target detection model will only detect targets near the road and omit targets at a greater distance. In practical applications, this can reduce the influence of many interference factors and improve efficiency. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method of the present invention;

[0047] Figure 2 This is a sample image of the road extraction model in an embodiment of the present invention;

[0048] Figure 3 Example diagram of the prediction results of the road extraction model in an embodiment of the present invention;

[0049] Figure 4 This is a sample image of the target detection model in an embodiment of the present invention;

[0050] Figure 5 This is an example diagram of the prediction results of the target detection model in an embodiment of the present invention;

[0051] Figure 6 This is an example diagram comparing the final prediction box results in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described in the accompanying drawings can generally be arranged and designed in various different configurations.

[0053] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0054] Figure 1This is a schematic diagram of a high-resolution remote sensing target detection method that integrates spatial relationships.

[0055] Reference Figure 1 As shown, a preferred embodiment of the present invention is provided, including the following steps:

[0056] Step 1: Create a sample, which includes the following four steps:

[0057] Step 1.1: Acquire high-resolution remote sensing imagery: Use optical satellite remote sensing data with visible-near-infrared sensors or aerial remote sensing data with a general optical camera. Depending on the resolution requirements, multispectral imagery or panchromatic imagery can be used directly.

[0058] Step 1.2: Cropping according to the size of the remote sensing image: Select the area where typical targets are located in the production area and crop the remote sensing image to a size of 1280*1280 pixels (the cropping range according to the size of the production area should have a certain degree of randomness and broad coverage).

[0059] Step 1.3: Create training samples for the road extraction model: Use software such as ArcGIS to draw training samples for the road extraction model. Based on the requirements of this example, only the road surface information needs to be drawn to obtain the corresponding labels, such as... Figure 2 As shown.

[0060] Step 1.4: Based on the task requirements, generally collect more than 1,000 images and corresponding annotations as samples, and divide the samples into training set and test set in a 4:1 ratio.

[0061] Step 2: Train the road extraction model, which includes the following four steps:

[0062] Step 2.1: Design a deep convolutional neural network: In order to train the road extraction model, the Dlinknet neural network is selected in this invention. The network can be replaced or modified accordingly based on the production target network.

[0063] Step 2.2: Initialize weights: Use the previously trained network model as a pre-trained model to initialize the Dlinknet network weights.

[0064] Step 2.3: Set training hyperparameters: Configure hyperparameters. The specific values ​​after model tuning are shown below:

[0065] DLinkNet training parameter settings: number of iterations = 400, batch_size = 4, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.0002, learning rate update coefficient = 0.2;

[0066] Step 2.4: Input samples and train the model: Input the training samples into the Dlinknet network and train it according to the hyperparameters described in Step 2.3 to obtain a road extraction model that can extract road information.

[0067] Step 3: Obtain road information as "prior knowledge," which includes the following two steps:

[0068] Step 3.1: Input the high-resolution remote sensing image into the road extraction model, predict the probability that each pixel is a road, and output a grayscale image from 0 to 255. Adjust the threshold to convert the road extraction image into a binary image, such as... Figure 3 As shown, the conversion formula is:

[0069]

[0070] Where threshold is a real number in the interval [0,1], which can be set by the user, with an initial default value of 0.5. x and y are the horizontal and vertical coordinates of the image.

[0071] Step 3.2: Using the road as the center, radiate three pixels as a buffer and extract the buffer range.

[0072] Step 4: Create training samples for the object detection network model, including the following two steps:

[0073] Step 4.1: The remote sensing image is cropped in the same way as the road extraction model.

[0074] Step 4.2: Create training samples and labels for the object detection network model. Keeping the remote sensing image dataset unchanged, label the ground objects corresponding to each image. Each row stores information cls, x, y, w, h, representing the object's category, center point coordinates, length, and width, respectively. For example... Figure 4 As shown.

[0075] Step 5: Train the object detection model, which includes the following six steps:

[0076] Step 5.1: YOLO training parameter settings: number of iterations = 300, batch_size = 32, learning rate update strategy = step, learning rate update step size = [3200, 4800, 6400, 8000], initial learning rate = 0.001, learning rate update coefficient = 0.1.

[0077] Step 5.2: Images are fed into the network and uniformly padded to 640*640 pixels at the input. Data augmentation is performed using methods such as Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling. In the Backbone part, Focus structure data slicing improves detection speed, and the CSP structure allows the model to learn more features. The Neck part adopts the CSP2 structure, which is inspired by CSPnet, to enhance the network's feature fusion capabilities.

[0078] Step 5.3: After the preceding network, a feature map is obtained. This feature map is divided into Sh*Sw grid cells. The size of the grid cells is determined by the size of the input image. If the center of an object falls within a grid cell, then that grid cell is responsible for predicting that object. Each network needs to predict three bounding box locations and one confidence score. A bounding box corresponds to four location information points and one confidence score. The confidence score represents both the confidence that the predicted bounding box contains the object and the accuracy of the bounding box prediction.

[0079] Step 5.4: If the target center falls within a grid cell, the first term is set to 1; otherwise, it is set to 0. The second term is the IoU value between the predicted target box and the actual ground truth. Each target box needs to predict 5 values: (x, y, w, h) and confidence. Each grid cell also needs to predict one class, denoted as C classes. Divide the input image into Sh*Sw grid cells. Each grid cell needs to predict B target boxes and C classes. The output is a Sh*Sw*(5*B+C) tensor.

[0080] Step 5.5: Road location is used as "prior knowledge" in the prediction process. Since the main targets detected in this example are service areas and toll booths, primarily located on both sides of the road and exhibiting clear spatial adjacency, the buffer zone formed by the road binary map is used to define the scope. Filtering is only performed within the area radiating three grids from the road, reducing the prediction of interfering factors and improving detection speed. Once a target is detected, its prediction bounding box is marked, such as... Figure 5 As shown.

[0081] Step 5.6: The YOLO network has now obtained the scale feature map. Each two-dimensional point on the scale feature map corresponds to a grid in the input image. The feature map is associated with the prediction box. End-to-end training is started. Through linear regression calculation, the anchor box is made close to the true label of the target.

[0082] Step 6: The predicted value undergoes loss processing, which includes the following two steps:

[0083] Step 6.1: During training, an object detection network model is used for prediction, and three losses are calculated: confidence loss, class loss, and box position loss. The overlap between the predicted box and the label box is continuously corrected according to the bounding box regression formula proposed by the authors.

[0084]

[0085] Step 6.2: Then, input the road loss of this invention along with the road binary map to obtain the coordinate values ​​of the prediction box. The road binary map is converted into a binary matrix, and the values ​​decrease from near to far from the road. Based on the position of the coordinate values ​​in the binary matrix, the distance value is obtained and input into SmoothL1Loss. The output loss value is used for regression calculation so that the network learns that the detected target is close to the road.

[0086] Step 7: Traverse the image and input vector information, including the following two steps:

[0087] Step 7.1: Parametric Boundary Box: Convert the bounding box parameters output by the object detection model into the coordinates of its lower left and upper right vertices. Specifically:

[0088] Determine the outer frame, and convert the xywh coordinates (center point and width and height of the outer frame) output in the program into x1y1x2y2 coordinates (top left and bottom right corner coordinates of the outer frame) using coordinate transformation formulas, i.e., xmin, xmax, ymin, ymax.

[0089]

[0090] Where x, y, w, and h represent the x-coordinate of the center of the rectangular outer frame, the y-coordinate of the center, the width, and the height, respectively.

[0091] Step 7.2: Convert the original pixel coordinates to spatial geographic coordinates, traverse the image, and obtain the following... Figure 6 -A shows the high-precision prediction box, in which Figure 6 -B represents the prediction result from YOLOv5. A comparison of the two reveals a significant accuracy advantage for this invention. The process involves writing to a vector file, checking if the file exists before writing; if it does, the writing proceeds normally; otherwise, the vector file is created first. Finally, the four spatial geographic coordinates are concatenated to form a closed loop.

[0092] This invention uses a road extraction algorithm to extract road information from a dataset, converts it into a binary image, inputs it into a target detection network, creates a buffer based on the road binary image, determines the detection range of the buffer, reduces the influence of interference factors in the image, and ultimately achieves fine extraction of high-resolution remote sensing targets.

[0093] The implementation of the present invention is not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A high-resolution remote sensing target detection method integrating spatial relationships, comprising the following steps: Step 1: making samples, specifically comprising: Step 1.1: obtaining high-resolution remote sensing images: using optical satellite remote sensing data with visible-near infrared sensors or aerial remote sensing data with optical cameras, and directly using multispectral images or fused panchromatic images according to resolution requirements; Step 1.2: cropping according to the size of the remote sensing image: uniformly cropping the remote sensing image to the same pixel size in the range where the typical target is located in the production area; Step 1.3: making a road extraction model training sample; Step 1.4: collecting a large number of images and corresponding labels according to task requirements as samples, and dividing the samples into a training set and a test set according to the proportion; Step 2: training a road extraction model, specifically comprising: Step 2.1: designing a deep convolutional neural network, selecting a deep learning network structure suitable for extracting targets, and improving it, and the network can be replaced or modified accordingly according to the production target network; Step 2.2: initializing the weight: using an existing network model as a pre-trained model to initialize the network weight; Step 2.3: setting training hyperparameters: configuring the specific values after model tuning; Step 2.4: inputting samples and training the model: inputting the training samples into the network, training according to the hyperparameters described in step 2.3, and obtaining a road extraction model capable of extracting road information; Step 3: obtaining road information as "prior knowledge", specifically comprising: Step 3.1: inputting the high-resolution remote sensing image into the road extraction model to predict the likelihood of each pixel being a road, outputting a gray-scale image from 0 to 255, and adjusting the threshold to convert the road extraction image into a binary image, with the conversion formula as follows: where threshold is a real number in the interval [0, 1], which can be set by the user; x, y are the horizontal and vertical coordinates of the image; Step 3.2: taking the road as the center and extracting a buffer zone with three pixel points as the buffer zone; Step 4: making a target detection network model training sample, specifically comprising: Step 4.1: the remote sensing image is consistent with the one cropped by the road extraction model; Step 4.2: making a target detection network model training sample and label; Step 5: training the target detection model, specifically comprising: Step 5.1: setting the training parameters for training the target detection model; Step 5.2: inputting the image into the network, performing data augmentation at the input end; slicing the data in the Backbone part to improve the detection speed; using the CSP2 structure in the Neck part to enhance the network feature fusion capability; Step 5.3: Through the previous network, a feature map can be obtained, which is divided into several grid cells, the size of which is determined by the size of the input image. If the center of a target falls in this grid, the grid is responsible for predicting the target; each network needs to predict the position information and confidence information of three target boxes, one target box corresponds to four position information and one confidence information; the confidence represents the confidence that the predicted target box contains the target and the accuracy of the target box prediction; Step 5.4: The center of the target falls in a grid, and the information of the target is recorded, and the tensor of the target is output; Step 5.5: In the prediction process, the road position is used as "prior knowledge". Since the main detected ground targets are service areas and toll stations, which are mainly distributed on both sides of the road and have obvious spatial adjacent relationship, the buffer area is determined according to the road binary graph, and only the area within three grids of the road is screened, reducing the prediction of interference factors and improving the detection speed; Step 5.6: Obtain the scale feature map through the YOLO network. Each two-dimensional point on the scale feature map corresponds to a grid in the input image. Associate the feature map with the prediction box, start end-to-end training, and calculate through linear regression to make the anchor box close to the true label of the target; Step 6: The prediction value is subjected to loss, which specifically includes: Step 6.1: In the training process, the target detection network model is used for prediction, and three losses are calculated respectively: confidence loss, class loss, and box position loss; according to the boundary box regression formula, the coincidence degree of the prediction box and the label box is continuously corrected; Step 6.2: Then, regression calculation is performed with the road binary graph and the input road loss, so that the network learns to detect targets near the road; Step 7: Traverse the image and input the vector information, which specifically includes: Step 7.1: Parameterize the bounding box: convert the rectangular bounding box parameters output by the target detection model into the left bottom vertex coordinates and the right top vertex coordinates; the conversion formula is as follows: Where x, y, w, and h represent the center horizontal coordinate, center vertical coordinate, width, and height of the rectangular bounding box, respectively; Step 7.2: Convert the original pixel coordinates to spatial geographic coordinates, traverse the image to obtain high-precision prediction boxes; write to the vector file, and in the writing process, judge whether the vector file exists. If it exists, normally write, if it does not exist, create a vector file and then write; then connect the four spatial geographic coordinates to make them closed.

2. The method of claim 1, wherein the method further comprises: Step 2: Train the hyperparameters of the road extraction model, which specifically includes: The training parameters of DLinkNet are set as follows: iteration times = 400, batch_size = 4, learning rate update strategy = step, learning rate update step = [100, 200, 300, 400], initial learning rate = 0.0002, and learning rate update coefficient = 0.2; Step 5: Train the hyperparameters of the target detection model, which specifically includes: The training parameter settings of Yolo are as follows: iteration number = 300, batch size = 32, learning rate update strategy = step, learning rate update step = [3200, 4800, 6400, 8000], initial learning rate = 0.001, and learning rate update coefficient = 0.

1.

3. The method of claim 1, wherein the method further comprises: In the sample preparation process of step 1, choose ground objects with clear spatial relationships. The greater the role of spatial relationships, the greater the training effect on the network, and the better the target detection effect.

4. The method of claim 1, wherein the method further comprises: Step 1: Cut the remote sensing image into 1280*1280 pixels; Step 5: Fill the remote sensing image into 640*640 pixels.

5. The method of claim 1, wherein the method further comprises: In step 5, the prediction process divides the input image into Sh*Sw grids, and each grid needs to predict B target boxes and C categories. The output is a tensor of Sh*Sw*(5*B+C).

Images