A SAR image transmission tower detection method based on invariant structural features
Through the graph aggregation network and graph mapping process, the feature extraction problem of the transmission pole tower detection algorithm in SAR images under rotation and scale transformation is solved, achieving more efficient detection performance and robustness.
Patent Information
- Application Number
- CN202411012207.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-07-26
AI Technical Summary
The SAR image transmission tower detection algorithm based on deep learning has poor generalization capabilities in complex scenarios, and the feature extraction process cannot maintain rotation and scale invariance, resulting in a decrease in detection accuracy.
Using graph aggregation network combined with graph mapping process, object features are extracted by constructing an adjacency matrix and graph structure, and model optimization is performed using smooth L1 loss and intersection ratio losses to achieve rotation and scale invariance feature extraction.
The detection performance of transmission pole tower detection is improved, the translation, rotation and scale invariance of the model are enhanced, and the detection accuracy and robustness are improved.
Smart Images

Figure CN118537343B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of surveying and mapping remote sensing, and more specifically, to a SAR image transmission tower detection method based on invariant structural features. Background Art
[0002] In recent years, my country has initially formed a power grid system with UHV transmission channels as its backbone. The UHV power grid has a long transmission distance, a wide coverage area, and diverse environmental conditions, and is easily affected by various natural and man-made disasters such as heavy snowfall and floods.
[0003] Satellite-borne and airborne synthetic aperture radar (SAR) has the advantage of being applicable under all-weather and all-day conditions, and can achieve rapid assessment of the damage to large-scale power transmission lines.
[0004] Transmission towers are made of metal materials, which have a strong backscattering effect on electromagnetic waves and can form a certain size of radar cross-section (RCS). For example, cat-head-shaped transmission towers usually present a relatively stable shape, such as a triangle, due to their special rigid structure.
[0005] For SAR images, traditional transmission tower detection methods such as the constant false alarm rate (CFAR) algorithm, the generalized likelihood ratio test (GLRT) target detection algorithm, and the detection algorithm based on extended analysis features (EF) have gradually been replaced by target detection algorithms based on deep learning. These deep learning-based methods extract more representative features from the input SAR images and usually achieve good results. However, SAR systems usually have multiple observation modes and beams, which will produce images at different angles and directions. Such methods usually directly apply the methods in optical images to SAR images, ignoring the association with the SAR imaging characteristics of transmission towers, resulting in poor generalization ability in complex scenes.
[0006] For example, in the feature space extracted by region nomination-based algorithms such as Faster-RCNN and regression-based algorithms such as YOLO, the features are arranged in the form of a two-dimensional grid, which means that although these features have good translation invariance, their rotation invariance and scale invariance are poor.
[0007] During the translation, rotation, and scaling processes, it is required to protect the structural features as much as possible. Obviously, the feature extraction process of the convolutional neural network (CNN) cannot do this.
[0008] Therefore, it is necessary to design a graph aggregation network to effectively measure the importance of each node and integrate the feature information between each node. Summary of the invention
[0009] The present invention provides a SAR image transmission tower detection method based on invariant structural features, which solves the technical problem in the related art that in the feature space extracted by algorithms based on region nomination such as Faster-RCNN and algorithms based on regression such as YOLO, the features are arranged in the form of a two-dimensional grid, which means that although these features have good translation invariance, their rotation invariance and scale invariance are poor.
[0010] The present invention provides a SAR image transmission tower detection method based on invariant structural features, comprising the following steps:
[0011] S100: Based on the graph mapping process, the transmission tower features in the high-resolution SAR image are converted into key points in the contour, and an effective adjacency matrix is constructed to help the graph network extract object features more efficiently;
[0012] S200: Based on the graph aggregation network, it can efficiently measure the importance of each node in the graph structure, obtain valuable features, aggregate the feature information between nodes in the graph structure, and enable the features between nodes of the same type to be stably expressed;
[0013] S300: Based on the graph mapping network, a hybrid of the intersection-over-union loss and the chamfer distance loss is used to guide the learning of the edge point set of the transmission tower. In the detection and positioning of the graph aggregation network, the smooth L1 loss is used as the model optimization criterion to extract category-independent region proposals, and the detector is applied to identify the transmission tower.
[0014] Furthermore, in step S100, the following steps are included:
[0015] S101: The graph mapping process includes a graph mapping network and a graph mapping algorithm. The graph mapping network uses the same network architecture as the region candidate network, that is, The feature map of , and are the width, height and number of channels of the feature map respectively;
[0016] S102: The graph mapping network will predict a total point set Mapped to the target corners as regions of interest to construct a refined target representation:
[0017] in, It means that after The width of the feature map after layer processing, It means that after The height of the feature map after layer processing;
[0018] in is the downsampling rate corresponding to each feature pyramid network layer, is the total number of layers of the feature pyramid network, is the number of layers of the feature pyramid;
[0019] S103: The high-quality proposals generated by the graph mapping network are fed into the extractor to obtain the orientation RoI features, where the target corner points are input into the graph mapping algorithm; the graph mapping algorithm is responsible for focusing on the key parts of the transmission tower and further mapping them into a graph structure representation based on the connection strategy between nodes.
[0020] Furthermore, in step S102, the number of points set in each feature bin is fixed, which is expressed as:
[0021] ;
[0022] ;
[0023] in are the predefined corner point coordinates, is the coordinate of the prediction point, and the model strives to minimize ;
[0024] Right now , Respectively expressed in Prediction point coordinates in layer feature pyramid network Relative to predefined corner coordinates The offset of
[0025] Further, in step S103, the graph mapping algorithm applies the graph Modeling the input features, Figure It is expressed as follows:
[0026] ;
[0027] in For nodes, for the edge;
[0028] Using Node To represent the feature map Elements on ,in ,side corresponds to the connection between any two feature map elements, where , constructing the adjacency matrix based on the optional connection strategy Make the graph network efficiently extract target features, where , HW is the size of the input RoI.
[0029] Furthermore, in step S200, an adjacency matrix is input from a graph aggregation layer of the graph aggregation network. and node feature set , expressed as:
[0030] ;
[0031] in , , Respectively represent the 1st, and node feature vectors, where Indicates the number of nodes, is the channel of each node feature, is a dimensional node feature vector;
[0032] Define a learnable shared linear transformation matrix , , is a parameter matrix, where each row corresponds to a dimensional input vector, a linear transformation can be obtained by using the parameter matrix dimensional output vector, used to extract features on each node;
[0033] For node feature set For each node in , according to the derived adjacency matrix , get all its one-hop neighbor nodes The feature update formula for each node is:
[0034] ;
[0035] in, represents the nonlinear activation function, represents the concatenation of two vectors. express and The normalized fusion factor between is used to measure the fusion ratio of each one-hop neighbor node. The calculation method is as follows:
[0036] ;
[0037] in It is also a nonlinear activation function. represents the transpose operation, is a high-dimensional weight vector, The natural constant The exponential function with base , , which is responsible for extracting the fused importance of each one-hop neighbor node feature.
[0038] Further, Using exponential linear unit activation, With LeakyReLU activation, the latter layer includes two branches: classification branch and regression branch. The classification branch adopts a graph fusion layer that follows the averaging mechanism, and the regression branch realizes target positioning through a simple linear transformation.
[0039] Furthermore, in step S300, it is divided into two stages. The first stage is responsible for extracting region proposals that are independent of the category. The chamfer distance loss is used to describe the difference in point set distribution between two contours. The cross entropy loss is used to learn the region proposal subtask that is independent of the category:
[0040] ;
[0041] in , is the intersection-and-join loss, is the cross entropy loss, is the chamfer distance loss, represents the training phase corresponding to the first phase concatenated together, focusing on extracting region proposals;
[0042] The second stage applies a detector to identify transmission towers. The graph aggregation network in this stage uses logarithmic loss and smoothing L 1 Loss to achieve target positioning:
[0043] ;
[0044] in is the parameter factor, , Corresponding to the training phase in series with the second phase, the target is accurately detected and located based on the proposal. represents the logarithmic loss, is a smooth L1 loss used for target localization.
[0045] Furthermore, in the intersection-over-union loss, It is the intersection-over-union ratio, which describes the intersection ratio between two sets of axis-aligned boxes.
[0046] A transmission tower detection system based on invariant structural features, comprising:
[0047] Preprocessing module: SAR image preprocessing, such as denoising, correction, and generating image pyramids to reflect multi-scale features;
[0048] Graph mapping network module: extracts candidate edge points of transmission towers through feature extraction network and constructs node adjacency matrix;
[0049] Graph aggregation network module: encodes node features, implements attention score and attention weight calculation, and then introduces structural similarity loss and invariant constraint loss to achieve node feature aggregation and update;
[0050] Post-processing and prediction module: Combine edge point set, corner point coordinate regression and smooth L1 loss to guide regression, and obtain the final detection result through non-maximum suppression;
[0051] Training and optimization module: used for total loss function combination, back propagation and parameter update, and realizes model saving and loading.
[0052] A storage medium stores non-temporary computer-readable instructions for executing one or more steps in the above-mentioned SAR image transmission tower detection method based on invariant structural features.
[0053] The beneficial effects of the present invention are:
[0054] For high-resolution SAR images, the present invention designs a transmission tower detection model based on invariant structural features, which defines the transmission tower as a rigid structure composed of consistent components. Graph structure modeling and graph network are used for contour feature extraction. Compared with traditional fully connected layers or fully convolutional networks, the graph network under graph structure modeling can achieve higher detection performance improvement with relatively fewer model parameters.
[0055] Specifically, a graph mapping process is designed to complete the construction of the graph structure from the target contour detection. Compared with the traditional region proposal network, the graph mapping network directly learns the corner points of the target, so the region of interest (RoI) can be more finely located; moreover, the graph mapping network has the ability to better couple with the graph aggregation network module, that is, the output refined positioning features are easy to introduce into the target detection stage.
[0056] The graph aggregation network aggregates and extracts features from the obtained graph structure. The graph aggregation network updates the weights of the nodes to achieve efficient end-to-end training of the graph model. Using the graph aggregation network to extract valuable features and guide the model to decode, identify, and locate results is the core of the graph network module. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 It is a flow chart of a SAR image transmission tower detection method based on invariant structural features proposed by the present invention;
[0058] Figure 2 It is a block diagram of the module composition of a transmission tower detection system based on invariant structural features proposed by the present invention;
[0059] Figure 3It is a transmission tower detector using graph structure modeling in a SAR image transmission tower detection method based on invariant structural features proposed by the present invention;
[0060] Figure 4 It is a schematic diagram of the connection structure of connectable strategy I and strategy II in a SAR image transmission tower detection method based on invariant structural features proposed by the present invention.
[0061] In the figure: 101, preprocessing module; 102, graph mapping network module; 103, graph aggregation network module; 104, post-processing and prediction module; 105, training and optimization module. DETAILED DESCRIPTION
[0062] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.
[0063] The present invention proposes a transmission tower detection method using a graph structure to extract contour features (invariant structural features). First, the dimension is The feature map of is used as the input of the graph mapping network to obtain RoI features and proposed contours, combined with the graph mapping algorithm, such as Figure 4 As shown, using strategy I or II for the adjacency matrix The allocated elements are then stacked with a two-layer graph aggregation network, and then the RoI features and adjacency matrix As input, the first layer uses F =8 multi-head mechanism, and the output feature channel C=128 after splicing operation.
[0064] Specific:
[0065] 1. Dimension The feature map is taken as input and the feature representation of the image is extracted in some way (such as CNN backbone).
[0066] 2. Use some kind of region proposal network (such as RPN) to obtain RoI (Region of Interest) features and proposal contours.
[0067] 3. Build a graph mapping based on the proposed profile and generate an adjacency matrix . Figure 2Strategy I or II shown refers to the adjacency matrix Different strategies for assigning values to elements in .
[0068] 4. Combine RoI features and adjacency matrix As input, a two-layer graph convolutional network is connected. The first layer uses a multi-head attention mechanism with F=8 heads.
[0069] 5. The output features of multi-head attention are concatenated in the channel dimension, and the number of output channels is C=128.
[0070] 6. Perform some kind of supplementary operation on the spliced features, which may be some feature enhancement or adjustment.
[0071] Features that only change slightly when the relative position, direction, or scale of the target changes are called invariant structural features.
[0072] In the SAR scenario, the apparent structure of the transmission tower is relatively stable. In order to extract the invariant structural features (such as contour features) that can stably represent the transmission tower, a graphical model is designed. The completion of the model is divided into two stages.
[0073] The first stage is responsible for extracting category-independent region proposals:
[0074] That is, the model obtains relatively stable rigid structural features by learning the outline of the transmission tower, which is used to locate the proposed area. The specific representation of the outline is as follows:
[0075]
[0076] in Represents the outline of the proposed area, and the coordinates of the corner points in the outline are given by definition, Indicates the number of corner points on the contour. In the SAR scene, the cat-head-shaped transmission tower appears as a triangle, but because it is regarded as a rotating target detection, =4.
[0077] The second stage maps the features within the contour area to the feature space represented by the graph structure:
[0078]
[0079] in Represents the nodes of the graph, It can be regarded as the edge of the graph. Then the graph network is applied to complete the information exchange between nodes and further refine the target representation:
[0080]
[0081] in B represents the rotated bounding box in polar coordinates, Indicates the center point coordinates, Indicates the width and height of the anchor box. represents the rotation angle, represents the rotation angle of the bounding box, and finally a two-class classifier is applied to distinguish the category of the region proposal (object or non-object).
[0082] See also Figure 1 As shown, a SAR image transmission tower detection method based on invariant structural features includes the following steps:
[0083] S100: Based on the graph mapping process, the transmission tower features in the high-resolution SAR image are converted into key points in the contour, and an effective adjacency matrix is constructed to help the graph network extract object features more efficiently;
[0084] The specific steps in step S100 are as follows:
[0085] S101: The graph mapping process includes a graph mapping network and a graph mapping algorithm. The graph mapping network uses the same network architecture as the Region Proposal Network (RPN), that is, ( , and The feature map (width, height and number of channels of the feature map respectively) is taken as input;
[0086] S102: The graph mapping network will predict a total point set Mapped to the target corners as regions of interest to construct a refined target representation;
[0087] in, It means that after The width of the feature map after layer processing, It means that after The height of the feature map after layer processing;
[0088] in is the downsampling rate corresponding to each feature pyramid network layer, is the total number of layers of the feature pyramid network, is the number of layers of the feature pyramid;
[0089] In step S102, the number of points set in each feature bin is fixed, which can be expressed as:
[0090]
[0091]
[0092] in are the predefined corner point coordinates, is the coordinate of the prediction point, and the model strives to minimize .
[0093] Right now , Respectively expressed in Prediction point coordinates in layer feature pyramid network Relative to predefined corner coordinates The offset of
[0094] S103: The high-quality proposals generated by the graph mapping network are fed into the extractor to obtain the orientation RoI features, where the target corner points are input into the graph mapping algorithm; the graph mapping algorithm is responsible for focusing on the key parts of the transmission tower and further mapping them into a graph structure representation based on the connection strategy between nodes.
[0095] In step S103, the graph mapping algorithm is applied to the graph Model the input features, where For nodes, For the edge.
[0096] Using Node To represent the feature map Elements on ,in ,side corresponds to the connection between any two feature map elements, where , constructing the adjacency matrix based on the optional connection strategy Make the graph network efficiently extract target features, where , HW is the size of the input RoI.
[0097] S200: Based on the graph aggregation network, it can efficiently measure the importance of each node in the graph structure, obtain valuable features, aggregate the feature information between nodes in the graph structure, and enable the features between nodes of the same type to be stably expressed;
[0098] In step S200, an adjacency matrix is input from a graph aggregation layer of the graph aggregation network. and node feature set , ,in , , Respectively represent the 1st, and node feature vectors, where Indicates the number of nodes, is the channel of each node feature, is a dimensional node feature vector.
[0099] After being processed by the graph aggregation layer, node features with the same number and different channels will be generated. , ,in , , Respectively represent the 1st, and Aggregate node feature vectors, is a -dimensional aggregate node feature vector.
[0100] Define a learnable shared linear transformation matrix , , is a parameter matrix, where each row corresponds to a dimensional input vector, a linear transformation can be obtained by using the parameter matrix dimensional output vector, used to extract features at each node.
[0101] For each node, according to the derived adjacency matrix , get all its one-hop neighbor nodes The feature update formula for each node is:
[0102]
[0103] in, represents a nonlinear activation function; represents the concatenation of two vectors. express and The normalized fusion factor between is used to measure the fusion ratio of each one-hop neighbor node. The calculation method is as follows:
[0104]
[0105] in It is also a nonlinear activation function. represents the transpose operation, is a high-dimensional weight vector, The natural constant The exponential function with base , , which is responsible for extracting the fusion importance of each one-hop neighbor node feature;
[0106] here Using the Exponential Linear Unit (ELU) activation, With LeakyReLU activation, the latter layer has two branches, the classification branch and the regression branch, which are responsible for two types of classification and target positioning respectively;
[0107] Specifically, the classification branch adopts a graph fusion layer that follows the averaging mechanism, and the regression branch achieves target localization through a simple linear transformation.
[0108] S300: Based on the graph mapping network, a hybrid of the intersection-over-union loss and the chamfer distance loss is used to guide the learning of the edge point set of the transmission tower. In the detection and positioning of the graph aggregation network, the smooth L1 loss is used as the model optimization criterion to extract category-independent region proposals, and the detector is applied to identify the transmission tower.
[0109] In step S300, the loss estimation covers two stages. The first stage is responsible for extracting category-independent region proposals. Used to describe the intersection ratio between two sets of axis-aligned boxes. Specifically, is the intersection-over-union ratio, which is used to describe the intersection ratio of two sets of axis-aligned boxes. for loss, chamfer distance (CD) loss It is used to describe the difference in point set distribution between two contours, using cross entropy loss To learn the category-independent region proposal subtask:
[0110]
[0111] in , Denotes the training stage corresponding to the first stage concatenated together, focusing on extracting region proposals.
[0112] The second stage applies a detector to identify transmission towers. The graph aggregation network in this stage uses the logarithmic loss ( ) and smooth L 1 Loss to achieve target positioning:
[0113]
[0114] in is the parameter factor, , Corresponding to the training phase in series with the second phase, the target is accurately detected and located based on the proposal. is a smooth L1 loss used for target localization.
[0115] Transmission towers in existing SAR datasets are extracted, and their visual features such as intensity, size, orientation and contour are quantitatively analyzed. The results show that with the increase in the number of samples, the changes in various visual features gradually tend to be stable. In contrast, the changes between contours are minimal and smooth. Therefore, contour features can better represent the detected target than other features, which are called invariant structural features.
[0116] The effective extraction and utilization of this structural feature will be beneficial to improving positioning accuracy.
[0117] Compared with traditional convolutional networks that process Euclidean space features, graph networks have a natural advantage in modeling structural features.
[0118] First, the graph structure has inherent rotation invariance, which easily solves the problem of rotation feature misalignment;
[0119] Secondly, the graph structure relies on edges to establish the connection status between any nodes, rather than the fixed pixel connection mode under the convolutional neural network (CNN), which is suitable for general rigid structures such as transmission towers.
[0120] Finally, the graph structure can achieve similar expressions of target features at different scales by controlling the number of nodes, making it scale invariant.
[0121] It is worth noting that in SAR scenarios, the intrinsic relationship between the target contour and its components is more regular. The graph model can further improve the discrimination efficiency through information interaction between structural features.
[0122] However, for SAR target detection tasks, introducing graph networks to model structural features still remains problematic.
[0123] One is the mismatch of feature structure, because the output features of the backbone network formed by stacking convolution operations are presented as a set of grid data structures in Euclidean space, which cannot be directly input into the graph network for processing.
[0124] To this end, the present invention adopts the idea of extracting key points containing target features, that is, completing the conversion of mesh data to key points through contour learning through a graph mapping process. Specifically, the mesh data is converted into a structure containing target key point features through a "graph mapping" process to adapt to the input requirements of the graph network.
[0125] The second is the design problem of the graph network layer. How to deal with the feature interactions between nodes in the graph structure and obtain valuable features is the purpose of the graph network.
[0126] Since the SAR target can be regarded as the aggregation of multiple components, the present invention designs a graph aggregation network to effectively measure the importance of each node and integrate the characteristic information between each node. The proposed "graph aggregation network" can effectively measure the importance of each node (key point) and integrate the characteristic information between each node.
[0127] Finally, the present invention instantiates a detection framework, namely a graph mapping aggregation network to complete the transmission tower detection task and outputs the corresponding results.
[0128] In this embodiment, the present invention proposes a SAR image transmission tower detection method based on invariant structural features, which specifically includes the following parts:
[0129] Graph Mapping Network: Use convolutional neural networks to extract features from high-resolution SAR images, map the edge point set of the transmission tower to a set of coordinates of key points, and construct an adjacency matrix to represent the topological relationship between these key points as the input of the graph convolutional network;
[0130] Graph Aggregation Network: A graph convolutional network based on an attention mechanism is used to extract and aggregate features of the input graph structure. Specifically, a node-level attention mechanism is used to dynamically determine the importance weights between each node and its neighboring nodes, and these attention weights are used to perform weighted feature aggregation to obtain updated feature representations of each node. Residual connections are introduced between graph convolutional layers to retain the original structural information.
[0131] Loss function design: On the graph mapping network side, a weighted combination of intersection-over-union loss and chamfer distance loss is used to guide the model to learn the geometric features of the transmission tower edge; on the graph aggregation network side, smooth L1 loss is used as the optimization target of the regression task to minimize the Euclidean distance between the predicted transmission tower corner coordinates and the true coordinates;
[0132] Data enhancement and training: Perform data enhancement such as affine transformation and interpolation on the training data to enhance the model's robustness to changes in scale, rotation, deformation, etc., and perform end-to-end joint training on the graph mapping network and the graph aggregation network to fully explore the relationship between the two.
[0133] This graph convolutional network-based method effectively utilizes the structural prior of transmission towers, adaptively constructs a topological structure graph, and performs feature aggregation based on the attention mechanism. It can better solve the problem of extracting invariant structural features of transmission towers and improve the accuracy of detection.
[0134] In the graph convolutional network, the node attention weight can also be optimized based on the invariant features of the transmission tower. The calculation process is as follows:
[0135] Node feature encoding: For each input node First, a feedforward neural network is used to encode the feature vector hi to obtain a high-dimensional feature representation :
[0136]
[0137] in is a feedforward network function. This encoding can make the node features have better linear separability in a higher-dimensional vector space.
[0138] Attention score calculation: Then, the calculation node To its neighboring nodes Attention score , the commonly used method is based on node and The feature vector of is used to calculate the score:
[0139]
[0140] in is the feature vector concatenation operation, is a learned vector, is a non-linear activation function.
[0141] Attention weight calculation: Normalize the attention score through the softmax function to get the node To all neighbors The attention weight :
[0142]
[0143] in Is a node The set of neighbor nodes.
[0144] Attention weight optimization: For invariant structures such as transmission towers, we hope that the model can automatically learn that nodes with similar structures should be given higher attention weights.
[0145] Therefore, some auxiliary terms can be introduced into the loss function:
[0146] a) Structural similarity loss: minimize the feature differences of similar structural nodes;
[0147] b) Structural invariance constraint: constrains the attention weight distribution of nodes with similar structures to be similar;
[0148] Through these auxiliary loss terms, the parameters of the attention score can be optimized during backpropagation , so that the attention weights can adaptively capture the invariant structural features of transmission towers.
[0149] Feature aggregation and update: Finally, use the normalized attention weights Perform weighted summation on neighbor node features to obtain node The new feature representation of is:
[0150]
[0151] in is a nonlinear activation function, and the new feature representation Contains the unchanged structure information of the transmission tower.
[0152] Through the above optimization process, the graph convolutional network can adaptively allocate attention weights at the node level, thereby efficiently extracting and aggregating the invariant structural features of transmission towers and providing valuable feature representation for the final detection task.
[0153] It is necessary to add that Represents the input node The original eigenvector of It means node The new feature vector of .
[0154] Represents the nodes through the feedforward neural network FFN The original characteristics Encode to obtain high-dimensional feature representation .
[0155] Represents a compute node To its neighboring nodes Attention score .
[0156] in:
[0157] is a weight vector that needs to be learned;
[0158] Indicates that and Splice into a longer vector;
[0159] It is a commonly used nonlinear activation function;
[0160] Indicates that Normalized nodes To Neighbors The attention weight ,in Is a node The set of neighbor nodes.
[0161] The structural similarity loss is an auxiliary loss term that aims to minimize the feature differences between nodes with similar structures.
[0162] The structural invariance constraint represents another auxiliary loss term, which constrains that nodes with similar structures should be assigned similar attention weight distributions.
[0163] Represents the use of normalized attention weights Neighbor node features Perform weighted summation to get the node New feature representation ,in It is a non-linear activation function.
[0164] In one embodiment of the present invention, a transmission tower detection system based on invariant structural features is also proposed, specifically comprising:
[0165] Preprocessing module: SAR image preprocessing, such as denoising, correction, etc., to generate image pyramids to reflect multi-scale features;
[0166] Graph mapping network module: feature extraction network, such as ResNet and VGGNet, extracts candidate edge points of transmission towers and constructs node adjacency matrix;
[0167] Graph aggregation network module: encodes node features, implements attention score and attention weight calculation, and then introduces structural similarity loss and invariant constraint loss to achieve node feature aggregation and update;
[0168] Post-processing and prediction module: Combine edge point set, corner point coordinate regression and smooth L1 loss to guide regression, and obtain the final detection result through non-maximum suppression (NMS);
[0169] Training and optimization module: total loss function combination, such as classification, regression, structural loss, etc., back propagation and parameter update, and model saving and loading;
[0170] In the implementation process, these modules can be further subdivided into different submodules and components. For example, the graph mapping network can use the existing target detection framework; the attention module can be implemented based on attention mechanisms such as Transformer. Through reasonable modular design and combination, a complete end-to-end transmission tower detection system can be built.
[0171] In addition, attention should be paid to data processing, model evaluation, system deployment and other links to ensure the efficiency and availability of the entire system. In general, this modular design has good scalability and flexibility, which facilitates continuous optimization and future functional expansion.
[0172] At least one embodiment disclosed in the present invention provides a storage medium storing non-transitory computer-readable instructions for executing one or more steps in the aforementioned SAR image transmission tower detection method based on invariant structural features.
[0173] The computer program may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be construed as limiting the scope.
[0174] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.
Claims
1. A SAR image transmission tower detection method based on invariant structural features, characterized in that: The following steps are involved: S100: Based on the graph mapping process, the transmission tower features in the high-resolution SAR image are converted into key points in the contour, and an effective adjacency matrix is constructed to help the graph network extract object features more efficiently; In step S100, the following steps are included: S101: The graph mapping process includes a graph mapping network and a graph mapping algorithm. The graph mapping network uses the same network architecture as the region candidate network, that is, The feature map of , and are the width, height and number of channels of the feature map respectively; S102: The graph mapping network will predict a total point set Mapped to the target corners as regions of interest to construct a refined target representation: in, It means that after The width of the feature map after layer processing, It means that after The height of the feature map after layer processing; in is the downsampling rate corresponding to each feature pyramid network layer, is the total number of layers of the feature pyramid network, is the number of layers of the feature pyramid; S103: The high-quality proposals generated by the graph mapping network are fed into the extractor to obtain the features of the azimuth region of interest, where the target corner points are input into the graph mapping algorithm; the graph mapping algorithm is responsible for focusing on the key parts of the transmission tower and further mapping them into a graph structure representation based on the connection strategy between nodes; S200: Based on the graph aggregation network, it can efficiently measure the importance of each node in the graph structure, obtain valuable features, aggregate the feature information between nodes in the graph structure, and enable the features between nodes of the same type to be stably expressed; In step S200, an adjacency matrix is input from a graph aggregation layer of the graph aggregation network. and node feature set , expressed as: ; in , , Respectively represent the 1st, and node feature vectors, where Indicates the number of nodes, is the channel of each node feature, is a dimensional node feature vector; Define a learnable shared linear transformation matrix , , is a parameter matrix, where each row corresponds to a The input vector of dimension 1 is linearly transformed by the parameter matrix to obtain a dimensional output vector, used to extract features on each node; For node feature set For each node in , according to the derived adjacency matrix , get all its one-hop neighbor nodes The feature update formula for each node is: ; in, represents the nonlinear activation function, represents the concatenation of two vectors. express and The normalized fusion factor between is used to measure the fusion ratio of each one-hop neighbor node. The calculation method is as follows: ; in It is also a nonlinear activation function. represents the transpose operation, is a high-dimensional weight vector, The natural constant The exponential function with base , , which is responsible for extracting the fusion importance of each one-hop neighbor node feature; S300: Based on the graph mapping network, a hybrid of the intersection-over-union loss and the chamfer distance loss is used to guide the learning of the edge point set of the transmission tower. In the detection and positioning of the graph aggregation network, the smooth L1 loss is used as the model optimization criterion to extract category-independent region proposals, and the detector is applied to identify the transmission tower; In step S300, it is divided into two stages. The first stage is responsible for extracting region proposals that are independent of the category. The chamfer distance loss is used to describe the difference in point set distribution between two contours. The cross entropy loss is used to learn the region proposal subtask that is independent of the category: ; in , is the intersection-and-join loss, is the cross entropy loss, is the chamfer distance loss, represents the training phase corresponding to the first phase concatenated together, focusing on extracting region proposals; The second stage applies a detector to identify transmission towers. The graph aggregation network in this stage uses logarithmic loss and smoothing L 1 Loss to achieve target positioning: ; in is the parameter factor, , Corresponding to the training phase in series with the second phase, the target is accurately detected and located based on the proposal. represents the logarithmic loss, is a smooth L1 loss used for target localization.
2. The SAR image transmission tower detection method based on invariant structural features according to claim 1 is characterized in that: In step S102, the number of points set in each feature block is fixed, which is expressed as: ; ; in are the predefined corner point coordinates, is the coordinate of the prediction point, and the model strives to minimize ; Right now , Respectively expressed in Prediction point coordinates in layer feature pyramid network Relative to predefined corner coordinates The offset of 3. The SAR image transmission tower detection method based on invariant structural features according to claim 2 is characterized in that: In step S103, the graph mapping algorithm applies the graph Modeling the input features, Figure It is expressed as follows: ; in For nodes, for the edge; Using Node To represent the feature map Elements on ,in ,side corresponds to the connection between any two feature map elements, where , constructing the adjacency matrix based on the optional connection strategy Make the graph network efficiently extract target features, where , HW is the size of the input RoI.
4. The SAR image transmission tower detection method based on invariant structural features according to claim 3 is characterized in that: Using exponential linear unit activation, With LeakyReLU activation, the latter layer includes two branches: classification branch and regression branch. The classification branch adopts a graph fusion layer that follows the averaging mechanism, and the regression branch realizes target positioning through a simple linear transformation.
5. The SAR image transmission tower detection method based on invariant structural features according to claim 4 is characterized in that: In the intersection-over-union loss, It is the intersection-over-union ratio, which describes the intersection ratio between two sets of axis-aligned boxes.
6. A transmission tower detection system based on invariant structural features, which performs detection according to a SAR image transmission tower detection method based on invariant structural features as described in any one of claims 1 to 5, characterized in that: include: Preprocessing module: SAR image preprocessing, generating image pyramid through denoising and correction to reflect multi-scale features; Graph mapping network module: extracts candidate edge points of transmission towers through feature extraction network and constructs node adjacency matrix; Graph aggregation network module: encodes node features, implements attention score and attention weight calculation, and then introduces structural similarity loss and invariant constraint loss to achieve node feature aggregation and update; Post-processing and prediction module: Combine smooth L1 loss to guide regression and obtain the final detection result through non-maximum suppression; Training and optimization module: used for total loss function combination, back propagation and parameter update, and realizes model saving and loading.
7. A storage medium storing non-transitory computer-readable instructions for executing one or more steps in a SAR image transmission tower detection method based on invariant structural features as described in any one of claims 1 to 5.
