SAR image road intersection center point extraction method based on improved YOLOv8 model and tensor voting algorithm

Abstract

The invention discloses an SAR image road intersection center point extraction method based on an improved YOLOv8 model and a tensor voting algorithm. The invention provides a road intersection center point extraction method based on an improved YOLOv8 model and a tensor voting algorithm in order to solve the problems that in SAR image geometric calibration and correction, a traditional ground control point extraction method depends on manual point pricking, efficiency is low, and subjective errors are likely to be introduced. In the recognition stage, a target detection model suitable for a satellite-borne SAR image is constructed, and the recognition precision and robustness are effectively improved; in the extraction stage, a tensor voting algorithm is adopted to model the intersection area structure, and the center point coordinate is accurately obtained. The method specifically comprises the steps of data preprocessing, target detection model construction and optimization, road intersection automatic identification, geometric constraint modeling, tensor voting technology-based center point quality inspection screening and precision verification. The method has important significance in the aspects of on-orbit data processing, intelligent geometric correction, on-orbit geometric calibration automation and the like of a satellite-borne SAR system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of synthetic aperture radar image control point extraction, and particularly relates to a SAR image road intersection center point extraction method based on an improved YOLOv8 model and a tensor voting algorithm. BACKGROUND

[0002] In the geometric correction and scaling of synthetic aperture radar (SAR) images, accurate extraction of ground control points (GCPs) is crucial. Traditional methods that rely on manual point insertion are labor-intensive, inefficient, and prone to subjective errors. To alleviate the burden of manual work, there is an urgent need to develop automatic extraction methods based on deep learning and other technologies. Therefore, it is of great significance to study automatic SAR image control point extraction methods based on deep learning models.

[0003] Currently, in the field of SAR image ground control point extraction, the commonly used control points are usually artificial corner reflectors or natural features with specific geometric configurations. Road intersections, as a typical geometric feature in SAR images, have the advantages of regular geometric configuration and widespread distribution. The center point can be used as an extraction object, providing stable and reliable ground control points for SAR image geometric correction, geometric scaling, and high-precision reconstruction of three-dimensional scenes. Therefore, the present application focuses on identifying road intersection center points, aiming to fully utilize their advantages, improve the efficiency and accuracy of ground control point extraction, and provide reliable support for SAR image geometric processing and remote sensing applications.

[0004] Existing methods for accurately extracting road intersection center points can be divided into several categories:

[0005] The first type of method involves geometric feature extraction, which determines the center location by extracting straight line features from road edges and calculating their intersections. Wang et al. extracted straight line features from road edges that form intersections and calculated the corresponding straight line intersections to determine the intersection center. Yin Yong et al. obtained road surface data and its corresponding centerline data to construct a hybrid line-surface topological relationship. Based on this, they first extracted dangling nodes on the centerline. Then, based on these dangling nodes, they broke the road line features after dimensionality reduction. Using the established topological relationship, they selected nodes with at least three associated arcs within the centerline topological nodes and identified them as intersection nodes. Next, they extracted the intersection surface by calculating the nearest point, ultimately identifying relevant information about the intersection. Guo Fengcheng et al. applied a morphological bottom-hat operation to SAR images using circular structuring elements to highlight dark regions, which are often associated with intersections. Subsequently, they used a morphological closing operation to eliminate noise and fill gaps, improving the recognition of candidate regions. After binarization, a seed-filling algorithm is used to mark connected regions, and finally, candidate regions are precisely located by calculating center point coordinates. However, in complex terrain or scenes with numerous distractions, this method may have difficulty accurately extracting candidate road intersections. Zhou Weiwei and others used the deep learning FasterR-CNN algorithm to detect road intersections. They first prepared and annotated a remote sensing image dataset containing a variety of scenes and types, then divided and trained the model to identify intersection areas. They also used graph-theory-based image segmentation combined with road target features to screen potential intersection areas. They also enhanced road features through bilateral filtering, morphological filtering, and other operations. Finally, they used tensor voting technology to extract the coordinates of the center points of road intersections.

[0006] The second type of method is trajectory data analysis method, which uses vehicle trajectory data to extract the center point of road intersection. Zheng Xiaoyu et al. use crowd-sourced trajectory data to detect outliers, identify position offset and motion outliers through adaptive density clustering and (Squeeze-and-Excitation Temporal Convolutional Network Auto-Encoder, SE_TCN_AE) model, then analyze the turning behavior in the trajectory data to extract the turning point, and determine the center position and coverage of the intersection based on the clustering pattern, so as to realize the accurate positioning and information extraction of the intersection. Li Yali et al. designed a multi-element integrated method to identify road intersections, which comprehensively considers the spatial distribution and implicit dynamic characteristics of vehicle trajectories, as well as the visual features of remote sensing images. Specifically, seed intersections are extracted through morphological method, clustering method and tensor voting method, etc. In the process of extracting intersections based on vehicle trajectories, both the road center line is first extracted in the grid space and then the intersection is detected, and the trajectory dynamics characteristics are directly used for extraction in the vector space. The trajectory data analysis method can reflect the changes of road network in real time, but may be affected by factors such as data sparsity and noise interference.

[0007] In summary, although the current research on SAR image road intersection extraction has achieved rich results, the method focusing on the accurate positioning of intersection center point still has obvious shortcomings. Specifically, the field faces two key problems that need to be solved: first, how to automatically identify road intersection targets in massive complex SAR images. Second, how to accurately extract the geometric center point coordinates of road intersection. In view of the above two problems, the present application proposes a SAR image road intersection center point extraction method based on the fusion of improved YOLOv8 model and tensor voting algorithm. Since L1 level SAR image is a commonly used downstream application data type, it has completed system geometric correction and radiation correction, has good visualization, stable radiation response and other characteristics, and is widely used in target detection and geometric processing tasks. However, this level of image still has certain geometric distortion and radiation residual, especially in complex terrain or strong scattering background, which easily affects the accurate identification and positioning of feature targets. In view of the above problems, based on YOLOv8 network, a model suitable for automatic identification of spaceborne SAR image road intersection (CoTAttention_ASPP_You Only Look Once, CoT_AS_YOLO) is proposed, which improves the accuracy and robustness of road intersection identification in complex spaceborne SAR image background. On the basis of constructing the automatic identification model of spaceborne SAR image road intersection, the tensor voting algorithm is introduced to model the structure of the road intersection area, and the accurate coordinates of the center point are extracted. SUMMARY

[0008] To overcome the shortcomings of the prior art, the present application proposes a SAR image road intersection center point extraction method based on an improved YOLOv8 model and a tensor voting algorithm. The CoT_AS_YOLO model is constructed based on the YOLOv8 network, and the tensor voting technology is introduced to realize the automatic identification of the road intersection of the spaceborne SAR image and the accurate extraction of the road intersection center point. First, the image enhancement technology is used to expand the original data and construct a high-quality training data set. On this basis, the CoT_AS_YOLO deep learning model is designed based on the YOLOv8 network, which combines the Atrous Spatial Pyramid Pooling (ASPP) and the Contextual Transformer Attention (CoT) mechanism, to extract the coarse positioning coordinates of the road intersection. Then, the ground coordinates are calculated to the image plane by using the RPC (Rational Polynomial Coefficient) model, and the initial image coordinates of the road intersection in the SAR image are obtained, and the initial position of the intersection center point is confirmed by eliminating the gross error through the geometric constraint model. Finally, the tensor voting technology is applied to extract the fine geometric structure based on the planar feature of the road intersection area in the SAR image, and the high-precision positioning of the road intersection center point is realized. The specific process includes SAR image quality optimization, CoT_AS_YOLO model detection, RPC coordinate reverse projection, geometric constraint gross error filtering, tensor voting feature refinement, and stereo verification experiment. The method has the advantages of high automation, strong recognition accuracy, small positioning error, and wide adaptability. It can effectively replace the manual piercing operation, improve the efficiency and accuracy of road intersection identification and center point extraction in SAR images, and is suitable for large-scale and complex background geometric correction and calibration tasks, and can provide key support for high-precision geometric processing of SAR images.

[0009] To achieve the above functions, the present application provides a SAR image road intersection center point extraction method based on an improved YOLOv8 model and a tensor voting algorithm, comprising the following steps:

[0010] S1: data preprocessing:

[0011] (1) Image cropping: crop a fixed size image (640x640 pixels) from the original SAR image, and select the target image.

[0012] (2) Data expansion:

[0013] Data augmentation is achieved using Python and its rich open-source libraries.

[0014] 1) Rotation operation: rotate 90° counterclockwise, enrich road intersection angle features.

[0015] 2) Noise injection: add Gaussian noise to simulate real imaging interference and improve noise resistance.

[0016] 3) Contrast enhancement: highlight the differences between roads and the environment to improve recognition accuracy.

[0017] (3) Dataset creation: use Labelimg tool for sample labeling and divide data into training set / validation set / test set.

[0018] S2: CoT_AS_YOLO target detection model construction and optimization:

[0019] (1) Build YOLOv8 base model architecture.

[0020] (2) Module integration optimization:

[0021] 1) To address the problem of target blur and few visible features caused by background noise, embed ASPP module in the backbone network.

[0022] 2) To address the problem of the model's difficulty in fully perceiving target detail features in complex backgrounds, introduce CoAttention collaborative attention mechanism.

[0023] S3: SAR image road intersection automatic recognition: use CoT_AS_YOLO model for training to recognize road intersections in SAR images.

[0024] S4: Geometric constraint modeling:

[0025] (1) Calculate the image coordinates of the road intersection center point by using the RPC algorithm to calculate the ground road intersection center point coordinates to the image.

[0026] (2) According to the geometric constraint model, remove gross errors and confirm the rough position of the measured road intersection center point.

[0027] S5: Center point quality inspection and screening based on tensor voting technology:

[0028] (1) Data preprocessing

[0029] Crop the detection frame output by the CoT_AS_YOLO model to exclude areas unrelated to road intersections in the image, reduce interference factors, and retain more gray-scale information and details. Convert the cropped image to a binary image.

[0030] (2) Road intersection area recognition

[0031] Crossing area features are extracted using tensor voting technology.

[0032] (3) Center point extraction

[0033] The center of mass in the intersection area is calculated as the intersection center point. After determining the boundary and area of the road intersection, the center of mass of the area is calculated by the image pixel method. The coordinates of the center of mass are taken as the coordinates of the road intersection center point, and the entire extraction process is completed.

[0034] S6: Accuracy verification:

[0035] (1) CoT_AS_YOLO model accuracy evaluation.

[0036] (2) Road intersection center point extraction accuracy evaluation.

[0037] Further, the data preprocessing step S1 mainly includes the following steps:

[0038] (1) Image cropping: Crop fixed-size images (640x640 pixels) from the original SAR image and select them.

[0039] (2) Data augmentation:

[0040] Data augmentation is achieved using Python and its rich open source libraries.

[0041] 1) Rotation operation: Use the OpenCV image processing library to rotate the SAR image dataset. Specifically, rotate 90° counterclockwise to enrich the angle features of the road intersection.

[0042] 2) Noise injection: By adding Gaussian noise, simulate the noise interference that may occur during actual imaging, which is conducive to training the model to accurately detect and recognize road intersections in real-world scenarios with noise.

[0043] 3) Contrast enhancement: Adjust the contrast of the image to highlight key features in the image, allowing the model to better learn the differences between road intersections and their surroundings, improving the recognition accuracy of the model.

[0044] (3) Dataset creation: Use the Labelimg tool for sample labeling and divide the data into training set / verification set / test set.

[0045] Further, the CoT_AS_YOLO target detection model construction and optimization step S2 mainly includes the following steps:

[0046] On the basis of the YOLOv8 network, a CoT_AS_YOLO model for road intersection detection is proposed. The structure of the model includes four stages: input end, backbone network, neck network, and detection head. The role and composition structure of each stage are introduced as follows:

[0047] (1) Input end

[0048] In order to meet the input size requirements of the backbone network, the input image is scaled to a fixed size at the input end.

[0049] (2) Backbone network

[0050] The backbone network, as the basic feature extraction module of the model, plays a key role in the entire CoT_AS_YOLO model. Its structure is composed of multiple functional modules connected in order.

[0051] 1) Convolutional layer (Conv): The convolutional layer uses a convolution kernel to perform sliding convolution on the image. Each convolution operation can capture different local features in the image, from simple edge and texture information to more complex structural features.

[0052] 2) C2f module: The C2f module further connects and fuses features across stages based on the features extracted by the convolutional layer. This cross-stage feature fusion mechanism helps the model better understand the semantic information in the image and is important for accurately detecting road intersections, which have complex spatial structures.

[0053] 3) Atrous Spatial Pyramid Pooling (ASPP) module: At the end of the backbone network, an ASPP module is configured. The ASPP module is composed of multiple parallel and different atrous convolutions. Typically, several different atrous rates are set, such as 1, 3, 6, 9, etc. Different atrous rate convolution operations allow the model to capture multi-scale context information. Through this feature extraction method, the model can better handle complex backgrounds and fuzzy targets, improving the accuracy and robustness of road intersection detection.

[0054] (3) Neck

[0055] The neck network plays a key role in the CoT_AS_YOLO model, mainly responsible for fusing and up-sampling the features extracted by the backbone network to better adapt to the needs of the detection head.

[0056] 1) C2f module: The neck network also contains three C2f modules, which continue to process features, further integrate and optimize feature information, and enhance the model's ability to express and extract road intersection features.

[0057] 2) Concatenation: The concatenation operation integrates multi-scale feature information, making the feature representation richer and more comprehensive. Feature maps at different levels contain information at different scales. Concatenating them provides the model with a wider range of context and details, helping to improve detection accuracy.

[0058] 3) Upsampling: Upsampling enables features at different levels to complement each other, avoiding excessive information loss during the backbone network downsampling process. By restoring resolution, the model can better capture the details of the target, which is crucial for accurate road intersection detection.

[0059] 4) Contextual Transformer Attention (CoT): The CoT module is based on an attention mechanism. Its core function is to assign weights to different regions of a feature map, thereby focusing on features related to road intersections. The CoT module uses complex computational units to perform in-depth analysis of feature maps, assigning higher weights to regions that may contain road intersections while suppressing interference from irrelevant information. This attention mechanism significantly enhances the model's perception and targeting of road intersection features, thereby improving detection accuracy and efficiency.

[0060] (4) Head Network

[0061] The detection head is the key component of the CoT_AS_YOLO model's final object detection output and consists of three convolutional layers. The detection head outputs feature maps of sizes 80×80×45, 40×40×45, and 20×20×45, respectively. Large-scale feature maps (such as 80×80×45) capture more detailed information and are suitable for detecting smaller road intersection targets. Small-scale feature maps (such as 20×20×45) have a wider field of view and can detect larger road intersection targets. Through this multi-scale feature output and detection mechanism, the CoT_AS_YOLO model can achieve accurate detection of road intersections of varying sizes, meeting the diverse needs of practical applications.

[0062] Furthermore, step S3 of automatically identifying road intersections in SAR images mainly includes the following steps:

[0063] Use the CoT_AS_YOLO model for training to identify road intersections in SAR images.

[0064] Further, the step S4 of geometric constraint modeling mainly includes the following steps:

[0065] (1) The ground road intersection center point coordinates are calculated to the image by the RPC algorithm, and the image coordinates of the road intersection center point are calculated.

[0066] The center point of the intersection detection frame recognized by CoT_AS_YOLO is taken as the rough position of the road intersection center point, and the ground measured road intersection center point coordinates are calculated to the image by the RPC algorithm, and the image coordinates of the road intersection center point are calculated

[0067] (2) According to the geometric constraint model, the rough position of the measured road intersection center point is confirmed.

[0068] The geometric constraint relationship of the road intersection center point in the ground and the image in distance, angle, slope and area is constructed, and the rough error elimination of the SAR image road intersection center point is realized through the three geometric constraint relationships of distance, angle and area.

[0069] Further, the step S5 of center point quality inspection and screening based on tensor voting technology mainly includes the following steps:

[0070] (1) Data preprocessing

[0071] The detection frame output by the CoT_AS_YOLO model is cropped to exclude the area in the image irrelevant to the road intersection, reduce the interference factors, and retain more gray information and details. The cropped image is converted into a binary image.

[0072] (2) Road intersection area recognition

[0073] Based on the tensor voting algorithm, the intersection area features are extracted. First, the binary image is converted into a non-directional spherical tensor by a specific mapping function to realize tensor coding and unify the data basis. Secondly, the weight matrix is constructed according to the gray difference and spatial distance between pixel points in the spherical domain to vote for edge information, and the data direction is given and the features are highlighted according to the position and direction consistency in the rod-shaped domain to complete the tensor transmission. Then the transmitted tensor is decomposed by singular value decomposition method, and the intersection area saliency map and feature saliency are analyzed. Finally, the polarity and type are determined according to the voting direction of the image element, and the road intersection is recognized by setting the threshold combined with the saliency features.

[0074] The specific operation steps of the tensor voting algorithm are as follows:

[0075] 1) Tensor coding

[0076] In two-dimensional image data I, the tensor coding form of any image element is as follows:

[0077]

[0078] where, and are the second order derivatives of the image I along the x and y directions, the Laplace operator is used to calculate the second order partial derivatives of the image I:

[0079]

[0080] 2) Tensor Propagation

[0081] After the tensor encoding of all the pixels in the image is completed, each voting unit and the pixels in its neighborhood begin to exchange geometric information. The tensor voting contains two voting calculations in total, which are spherical domain voting and rod domain voting. Through the first spherical domain voting, the edge information of the image can be extracted, and an initial direction is assigned to each pixel; the second rod domain voting is performed according to the direction information of each pixel point, so that each pixel point highlights the significant feature information. The voting results received by each pixel point are counted, and a new tensor is formed through vector superposition to realize the information propagation between pixels.

[0082] 3) Tensor Decomposition

[0083] For each pixel point in the image I, the voting results of the tensor voting are counted, and each pixel will obtain a new tensor T. Singular value and matrix spectral decomposition are performed on the tensor T, as follows:

[0084]

[0085] where λ1, λ2 are eigenvalues, and satisfy λ1≥ λ2≥ 0; are the corresponding eigenvectors; and correspond to the rod component and the spherical component respectively, and λ1and λ1- λ2 correspond to the rod feature and the spherical feature significance index respectively.

[0086] 4) Feature Structure Entity Extraction

[0087] The voting decomposition result can well judge the rod significance and the spherical significance of the pixel point, and the structure attribute of the pixel point can be judged with the help of the polarity. The direction information of the polarity can be expressed by the following formula:

[0088]

[0089] where (x, y) represents the pixel coordinates, u(x, y) represents the voting result of the pixel, represents the significant direction of the pixel. sgn represents the sign function.

[0090] The pixel points inside the curve have low polarity because the two opposite directions of voting cancel each other out. The pixel points at the end points of the curve have high polarity because the voting on one side is more significant. The pixel points at the intersection points of the curve also have low polarity because the voting cancels each other out.

[0091] Considering the characteristics of the saliencies λ1 and λ2, λ1≥λ2 at the internal and end point positions of the curve, and λ2 has a local maximum value at the intersection point position of the curve. By comprehensively considering the saliency features and polarity, the position information of the pixel can be determined.

[0092] (3) Center point extraction

[0093] The center of mass in the intersection area is calculated as the intersection center point. After determining the boundary and area of the road intersection, the center of mass of the area is calculated by the image pixel method. The coordinates of the center of mass are taken as the coordinates of the road intersection center point, and the whole extraction process is completed.

[0094] Further, the precision verification of step S6 mainly includes the following steps:

[0095] (1) CoT_AS_YOLO model precision evaluation.

[0096] In order to evaluate the recognition effect of CoT_AS_YOLO model on SAR image road intersection, four commonly used evaluation indexes, including precision (P), recall (R), F1-Score (F1) and average precision (AP), are selected to quantitatively analyze the model detection results. Through the joint evaluation of the above indexes, the robustness, generalization ability and recognition effect of the model on the road intersection target in complex SAR image in different scenes can be systematically analyzed.

[0097] (2) Road intersection center point extraction precision evaluation.

[0098] Root Mean Square Error (RMSE) is very sensitive to large or small errors in a group of measurements, and can accurately reflect the precision of the coordinate difference of the road intersection center point. Therefore, RMSE is used to measure the extraction precision of the road intersection center point. In the actual measurement process, the number of observations n is often limited, and the true value is difficult to obtain exactly. Here, the average value of the coordinate difference of the n road intersection center points is regarded as the true value. Let the coordinate of the i-th road intersection center point extracted by the tensor voting algorithm be The actual i-th road intersection center point coordinate on the ground is The calculation formula of the i-th point error value is:

[0099]

[0100] The calculation formula for the average value of the point error is:

[0101]

[0102] The point error calculation formula is as follows

[0103]

[0104] Where n represents the number of center points of road intersections; x i Indicates the point error value; f(x i ) represents the average value of the point error value. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] The description of the contents of the present invention will become more apparent and easier to understand when taken in conjunction with the following drawings, in which:

[0106] Figure 1 This is a flow chart of a method for extracting center points of road intersections from SAR images based on an improved YOLOv8 model and a tensor voting algorithm.

[0107] Figure 2 This is a schematic diagram of the CoT_AS_YOLO network structure optimized based on the YOLOv8 model in the SAR image road intersection center point extraction method based on the improved YOLOv8 model and tensor voting algorithm of the present invention;

[0108] Figure 3 This is a CoT_AS_YOLO configuration flow chart for a method for extracting center points of road intersections from SAR images based on an improved YOLOv8 model and a tensor voting algorithm.

[0109] Figure 4 This is a flowchart of the method for extracting the center point of a road intersection from a SAR image based on an improved YOLOv8 model and a tensor voting algorithm in the present invention, which uses a tensor voting algorithm to accurately extract the center point of a road intersection. DETAILED DESCRIPTION

[0110] according to Figure 1 The steps shown in the figure are used to describe in detail a method for extracting the center point of a road intersection in a SAR image based on an improved YOLOv8 model and a tensor voting algorithm.

[0111] Step 1: Data preprocessing. This includes the following specific steps:

[0112] The original SAR image is cropped to select target data to ensure quality. According to the position information and approximate range of the road intersection, a fixed-size image is cropped from the original SAR image, and the target image is selected to ensure that the data used for model training has high availability and reliability. After data cropping, image enhancement techniques are applied to expand the cropped data to improve the generalization ability of the model, and a high-quality training set is prepared.

[0113] Step 2: CoT_AS_YOLO target detection model construction and optimization. Including the following specific steps:

[0114] As shown in Figure 2 and Figure 3 , the CoT_AS_YOLO target detection model is based on the YOLOv8 network framework, integrating the Atrous Spatial Pyramid Pooling (ASPP) and Contextual Transformer Attention (CoT) attention mechanisms for automatic identification of road intersections in SAR images. The model mainly includes four parts: input, backbone network, neck network, and detection head:

[0115] The input is used for image size standardization; the backbone network integrates convolutional layers, C2f modules, and ASPP modules to strengthen multi-scale context perception; the neck network introduces multi-layer feature fusion modules, upsampling, and COT attention mechanisms to enhance the recognition ability of fuzzy road structures; the detection head uses multi-scale output to detect large, medium, and small intersection targets, improving overall detection accuracy and robustness.

[0116] Step 3: Automatic identification of road intersections in SAR images. Including the following specific steps:

[0117] The built CoT_AS_YOLO model is used for training and prediction tasks. Through model inference, the detection box and initial position of the center point of the road intersection target are obtained.

[0118] Step 4: Geometric constraint modeling. Including the following specific steps:

[0119] (1) To improve the geometric accuracy of the recognition results, the RPC model is used to inverse the ground-truth intersection coordinates to the image plane to obtain image coordinates.

[0120] (2) Utilizing the geometric characteristics of the center point in terms of distance, angle, and area, multiple geometric constraint rules are constructed to eliminate gross errors in the recognition results, excluding obvious false positives and non-cooperative targets.

[0121] Step 5: Center point quality inspection and screening based on tensor voting technology. Including the following specific steps:

[0122] As Figure 4 shown, the processing flow of high-precision extraction of road intersection center points using tensor voting algorithm includes the following three stages:

[0123] (1) Data preprocessing

[0124] The intersection detection frame region output by the CoT_AS_YOLO model is image cropped, irrelevant regions are removed to reduce background interference, and the gray scale detail information of the road intersection is maximally preserved. Then the processed image is converted into a binary image to provide a unified image basis for subsequent tensor analysis.

[0125] (2) Road intersection region feature extraction and analysis

[0126] The tensor voting algorithm is used to model and identify the structural features of the intersection region, which mainly includes four stages: tensor encoding, tensor transmission, tensor decomposition, and feature structure extraction.

[0127] 1) Tensor encoding

[0128] First, any pixel in the image is directionless tensor encoded, and its encoding form is as follows:

[0129]

[0130] In the formula, and are the second-order derivatives along the x and y directions of the image I, which are processed using the Laplace operator, and the specific expression is as follows:

[0131]

[0132] 2) Tensor transmission

[0133] After completing the tensor encoding, the tensor information transmission between pixels is carried out in two stages:

[0134] In the spherical domain, a weight matrix is constructed according to the gray difference and spatial distance, and an initial directionless voting is performed to extract edge information;

[0135] In the rod-shaped domain, the second stage of directional voting is performed according to the geometric features such as direction consistency to enhance the saliency of the structural features in the image. Finally, the tensor superposition between pixels is used to complete the tensor transmission process of the overall structure.

[0136] 3) Tensor decomposition

[0137] Each pixel will obtain a new tensor T after completing the tensor transmission, which is processed by singular value decomposition and has the following form:

[0138]

[0139] where λ1, λ2 are eigenvalues and satisfy λ1≥ λ2≥ 0; are the corresponding eigenvectors; and correspond to the rod and spherical components respectively, and λ1 and λ1- λ2 are the significance indicators of the rod and spherical features respectively.

[0140] 4) Feature structure entity extraction

[0141] By combining the tensor feature significance analysis with the polarity function, the geometric structure type of the image element is further determined, and the calculation formula of the polarity direction is as follows:

[0142] where (x, y) represents the image element coordinates, u(x, y) represents the image element voting result, represents the image element significance direction. sgn represents the sign function.

[0143] According to the distribution characteristics of the tensor significance indicator and the polarity direction, the center area of the road intersection can be identified: inside the road line segment or edge, the polarity is low and the significance is stable; at the end of the intersection, the polarity is enhanced; in the center of the intersection area, the polarity appears a local minimum but the significance reaches a local maximum. By judging the combined characteristics of the polarity and the tensor significance, the automatic recognition of the structure area of the road intersection can be realized.

[0144] (3) Center point extraction

[0145] After the structure recognition of the intersection area is completed, the geometric center of the intersection area is extracted as the center point of the road intersection by using the image centroid calculation method, and the specific steps are as follows:

[0146] 1) According to the contour extraction of the intersection area in the binary image;

[0147] 2) The pixel weighted centroid calculation is performed on the area to determine the geometric center point coordinates;

[0148] 3) The calculated centroid is taken as the final output of the center point of the road intersection, and the high-precision automatic extraction is realized.

[0149] Step 6: Precision verification. Including the following specific steps:

[0150] In order to evaluate the performance of the method, the recognition precision of the automatic recognition model of the road intersection and the extraction precision of the center point of the road intersection are quantitatively verified:

[0151] (1) The road intersection recognition effect of the CoT_AS_YOLO model is evaluated by the indicators of precision, recall, F1-Score and average precision (AP), and quantitative analysis and the like;

[0152] (2) The root mean square error is used to measure the deviation between the predicted coordinates and the measured coordinates. The coordinate difference is counted at the pixel level to verify the road intersection center point extraction accuracy of the method.

[0153] The above only describes the best embodiments of the present application and is not used to limit the present application. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for extracting the center point of road intersections from SAR images based on an improved YOLOv8 model and a tensor voting algorithm, characterized in that: The steps include: S1: Dataset preprocessing: (1) Image cropping: cropping fixed-size images (640 × 640 pixels) from the original SAR image and selecting them; (2) Data expansion: Use Python and its rich open source libraries to implement data augmentation: 1) Rotation operation: Rotate 90° counterclockwise to enrich the angle features of the road intersection; 2) Noise injection: Add Gaussian noise to simulate real imaging interference and improve noise immunity; 3) Contrast enhancement: highlighting the difference between the road and the environment to improve recognition accuracy; (3) Dataset preparation: Use Labelimg tool to label samples and divide the data into training set / validation set / test set; S2: CoT_AS_YOLO target detection model construction and optimization: (1) Build the YOLOv8 basic model architecture; (2) Module integration optimization: 1) To address the problem of target blur and few visible features caused by background noise, the ASPP module is embedded in the backbone network; 2) To address the problem that the model has difficulty in fully perceiving the detailed features of the target in complex backgrounds, the CoAttention mechanism is introduced; S3: Automatic recognition of road intersections in SAR images: Use the CoT_AS_YOLO model for training to identify road intersections in SAR images; S4: Geometric Constraint Modeling: (1) The coordinates of the center point of the ground road intersection are calculated back onto the image using the RPC algorithm to obtain the image coordinates of the center point of the road intersection; (2) Eliminate gross errors based on the geometric constraint model and confirm the rough position of the measured road intersection center point; S5: Center point quality inspection and screening based on tensor voting technology: (1) Data preprocessing Crop the detection box output by the CoT_AS_YOLO model to exclude areas in the image that are not related to the road intersection, reduce interference factors, retain more grayscale information and details, and convert the cropped image into a binary image; (2) Road intersection area recognition Use tensor voting technology to extract cross-region features; (3) Center point extraction Calculate the centroid of the intersection area as the center point of the intersection; after determining the boundaries and areas of the road intersection, use the image pixel method to find the centroid of the area; use the coordinates of the centroid as the coordinates of the center point of the road intersection to complete the entire extraction process; S6: Accuracy verification: (1) CoT_AS_YOLO model evaluation indicators; (2) Evaluation of the accuracy of road intersection center point extraction.

2. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1 is characterized in that: The step S1 comprises the following steps: (1) Image cropping: crop a fixed-size image (640 × 640 pixels) from the original SAR image and select the target image; (2) Data expansion: Use Python and its rich open source libraries to implement data augmentation: 1) Rotation operation: By using the OpenCV image processing library, the SAR image dataset is rotated 90° counterclockwise to enrich the angle features of the road intersection; 2) Noise Injection: Gaussian noise is added to simulate the noise interference that may occur in actual imaging. This helps the trained model accurately detect and identify road intersections even when faced with noisy images in real scenes. 3) Contrast enhancement: By adjusting the contrast of the image, key features in the image are highlighted, allowing the model to better learn the differences between the road intersection and the surrounding environment, thereby improving the model's recognition accuracy; (3) Dataset preparation: Use the Labelimg tool to label samples and divide the data into training set / validation set / test set.

3. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1 is characterized in that: The step S2 comprises the following steps: Based on the YOLOv8 network framework, a target detection model CoT_AS_YOLO integrating the context transformer attention mechanism and the atrous spatial pyramid pooling module is constructed for automatic recognition of road intersections in SAR images.

4. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1, characterized in that: The step S3 comprises the following steps: The constructed CoT_AS_YOLO model is used for training and prediction tasks. Through model reasoning, the detection box of the road intersection target and the initial position of its center point are obtained.

5. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1, characterized in that: The step S4 comprises the following steps: (1) The center point of the intersection detection frame identified by CoT_AS_YOLO is used as the rough position of the center point of the road intersection. At the same time, the coordinates of the center point of the road intersection measured on the ground are reversed to the image through the RPC algorithm to obtain the image coordinates of the center point of the road intersection; (2) Construct the geometric constraint relationship between the ground and the center point of the road intersection in the image in terms of distance, angle, slope and area. Through the three-dimensional geometric constraint relationship of distance, angle and area, the gross error of the center point of the road intersection in the SAR image is eliminated.

6. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1, characterized in that: The step S5 comprises the following steps: (1) Data preprocessing Crop the detection box output by the CoT_AS_YOLO model to exclude areas in the image that are not related to the road intersection, reduce interference factors, retain more grayscale information and details, and convert the cropped image into a binary image; (2) Road intersection area recognition Use tensor voting technology to extract cross-region features: First, the binary image is converted into an undirected spherical tensor through a specific mapping function to achieve tensor encoding and unify the data foundation. Secondly, in the spherical domain, a weight matrix is ​​constructed based on the grayscale difference and spatial distance between pixels to obtain edge information. In the rod domain, weighted voting is calculated based on position and direction consistency to give data direction and highlight features, completing tensor transfer. Then, the transferred tensor is decomposed using methods such as singular value decomposition to analyze the intersection area saliency map and feature significance. Finally, the polarity and type are determined according to the pixel voting direction. Combined with the significant features, the road intersection is identified by setting a threshold. The specific operation steps of the tensor voting algorithm are as follows: 1) Tensor Encoding In the two-dimensional image data I, the tensor encoding form of any pixel is as follows: In the formula and is the second-order derivative along the x and y directions in image I, and the Laplace operator is used to calculate the second-order partial derivative of image I: 2) Tensor transfer After completing the tensor encoding of all pixels in the image, each voting unit begins to interact with the pixels in its neighborhood to carry out geometric information exchange. Tensor voting involves two voting calculations: spherical domain voting and rod domain voting. The first spherical domain voting extracts the edge information of the image and assigns an initial direction to each pixel. The second rod domain voting is performed based on the direction information of each pixel, so that each pixel highlights significant feature information. The voting results received by each pixel are counted, and a new tensor is formed by vector superposition to realize information transfer between pixels. 3) Tensor decomposition For each pixel point in the image I, the results of its tensor voting are counted, and each pixel will obtain a new tensor T. The singular value and matrix spectrum decomposition of tensor T are performed as follows: In the formula, λ1 and λ2 are used as eigenvalues ​​and satisfy λ1≥λ2≥0; is the corresponding eigenvector; and They correspond to the rod-like component and the spherical component respectively, and λ1 and λ1-λ2 correspond to the significance indexes of the rod-like feature and the spherical feature respectively; 4) Feature structure entity extraction The voting decomposition result can well determine the rod saliency and ball saliency of the pixel point. At the same time, the polarity can be used to determine the structural attributes of the pixel point. The direction information of the polarity can be expressed as follows: In the formula, (x, y) represents the pixel coordinates, u(x, y) represents the pixel voting result, Represents the direction of pixel significance, sgn represents the sign function, and the polarity of the pixel points inside the curve is lower because of the mutual cancellation of votes in two opposite directions; the pixel points at the endpoints of the curve have a higher polarity because the voting influence on one side is more significant; the polarity of the pixel points at the intersection of the curve is also lower due to the mutual cancellation of votes. Considering the characteristics of significance λ1 and λ2, λ1 ≥ λ2 inside the curve and at the endpoints; and λ2 has a local maximum at the intersection of the curve. By comprehensively considering the significance characteristics and polarity, the position information of the pixel can be determined; (3) Center point extraction Calculate the centroid of the intersection area as the center point of the intersection; after determining the boundary and area of ​​the road intersection, use the image pixel method to find the centroid of the area; use the coordinates of the centroid as the coordinates of the center point of the road intersection to complete the entire extraction process.

7. The method for automatically extracting the center point of a road intersection based on the CoT_AS_YOLO target detection model and the tensor voting algorithm according to claim 1, characterized in that: The step S6 comprises the following steps: (1) CoT_AS_YOLO model accuracy evaluation To evaluate the CoT_AS_YOLO model's recognition performance for road intersections in SAR images, four commonly used evaluation metrics, Precision (P), Recall (R), F1-Score (F1), and Average Precision (AP), were used to quantitatively analyze the model's detection results. By combining these metrics, we can systematically analyze the model's robustness and generalization capabilities in different scenarios, as well as its recognition performance for road intersection targets in complex SAR images. (2) Evaluation of the accuracy of road intersection center point extraction The root mean square error (RMSE) is extremely sensitive to very large or very small errors in a set of measurements and can accurately reflect the precision of the coordinate difference of the center point of the road intersection. Therefore, RMSE is used to measure the extraction accuracy of the center point of the road intersection. In the actual measurement process, the number of observations n is often limited, and the true value is difficult to obtain accurately. Here, the average value of the coordinate difference of the center points of the n road intersections is regarded as the true value. Let the coordinates of the center point of the i-th road intersection extracted by the tensor voting algorithm be The actual coordinates of the center point of the i-th road intersection on the ground are The calculation formula for the i-th point error value is: The calculation formula for the average value of the point error is: The point error calculation formula is as follows Where n represents the number of center points of road intersections; x i Indicates the point error value; f(x i ) represents the average value of the point error.

Images