Remote sensing image target detection method fused with multi-scene random field

Abstract

The invention relates to a remote sensing image target detection method fused with a multi-scene random field, and the method comprises the steps: building a model through a method combining multi-task learning and a conditional random field, extracting the context information near a suspected small target, assisting the model to judge a real target, integrating the effective results searched by the model, and outputting an image. Aiming at the problems of wide distribution range of remote sensing images and low model test analysis efficiency, a super-pixel segmentation method of the remote sensing images is designed according to the particularity of similar sizes and shapes of similar targets in the remote sensing images, gridding of a remote sensing large-range region is realized in a model reasoning test process, a retrieval region is divided into a plurality of small regions, and a remote sensing image is obtained. The network target searching efficiency is effectively improved; a target image data set containing different backgrounds is constructed, and the obtained effective feature information is input into a target detection network for training, so that the target detection accuracy of the electric power facility in the remote sensing image is realized, and the inspection efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning, image processing and power transmission line inspection, and in particular to a remote sensing image target detection method integrating multi-scene random fields. Background Art

[0002] With the advancement of science and technology and social development, the demand for electricity is increasing, and the scope and scale of power distribution are expanding. Due to the vast territory of my country and the diverse terrain of power transmission lines, line inspection is essential to ensure the stability of power transmission. Today's power transmission line inspection mainly relies on a combination of manual and drone inspections. However, this inspection method is time-consuming and costly in large-scale transmission line inspections because drones conduct inspections in near-air flight. In recent years, with the continuous development of remote sensing technology, the use of remote sensing images for large-area regional inspections has been rapidly applied in various fields, which not only improves inspection efficiency, but also facilitates the formulation of operation and maintenance strategies in the later stage.

[0003] At present, the target detection methods of remote sensing images can be mainly divided into two categories: methods based on manual features and methods based on deep learning:

[0004] (1) The hand-crafted feature-based method mainly uses manually designed feature descriptors to describe specific targets in remote sensing images, extracts manually designed target features after region selection, and then uses classifiers and regressors to classify and locate targets. Commonly used hand-crafted feature-based algorithms can be divided into SIFT (Scale-Invariant Feature Transform) to extract key points and calculate their local image features, LBP (Local Binary Patterns) to compare the brightness of pixels and their neighbors in binary mode to extract target texture features, HOG (Histogram of Oriented Gradients) to extract edge direction distribution features in images, etc. In different visual tasks, appropriate design methods are selected according to target characteristics for target detection. These algorithms usually perform well under specific conditions, especially before deep learning became popular, and they have been widely used in the field of computer vision. However, with the development of deep learning technology, hand-crafted feature-based target detection algorithms have been replaced by deep learning-based algorithms in many tasks.

[0005] (2) Deep learning-based methods automatically learn the internal structure of data hidden in training data through training methods, and convert the pixel data in the input image into higher-order and more abstract hierarchical features. They are mainly divided into single-stage algorithms and two-stage algorithms. The single-stage algorithm directly classifies and locates the target after feature extraction, such as SSD, YOLO series, etc. The two-stage algorithm first generates candidate regions and then identifies the target, such as R-CNN, Faster R-CNN, etc. This method has better performance in target detection tasks involving remote sensing image data with massive information, and is currently a hot research direction.

[0006] With the continuous advancement of technology, remote sensing image target detection algorithms are also constantly updated and optimized. For example, target detection algorithms for specific applications, combined with deep learning algorithms and feature fusion or multi-angle detection technology, can realize the study of deep information of specific targets, thereby improving recognition accuracy and model performance, and injecting new vitality into the research of image data from different sources. Summary of the invention

[0007] In order to solve the above problems, the purpose of the present invention is to provide a remote sensing image target detection method that integrates multi-scene random fields, so as to achieve the target detection accuracy of power facilities in remote sensing images and improve the inspection efficiency.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A remote sensing image target detection method integrating multi-scene random fields comprises the following steps:

[0010] Step S1: Collect remote sensing image data sets of power inspection transmission towers and lines under different background environments and mark them;

[0011] Step S2: Taking the pole tower as a node, preprocessing the input image, extracting the central color and position of the node by clustering, and representing the basic color and spatial coordinates of the target;

[0012] Step S3: construct a multi-scenario random field model, input multiple nodes and their attributes and known relationship labels, and output the characteristic relationship between the transmission tower and the line;

[0013] Step S4: the number of different background style types is represented as N, each image input is represented as x, and the detection output of the tower and line is represented as y;

[0014] Step S5: The output y obtained by the multi-scenario random field model including the connection relationship between nodes and edges is saved in the form of a labeled file, and combined with the features extracted by the remote sensing image recognition model as feature information to supplement the tower recognition logic and assist the model in performing reasoning tests;

[0015] Step S6: constructing a remote sensing image transmission tower recognition model based on the YOLOv8 algorithm;

[0016] Step S7: construct a superpixel segmentation network, perform gridding preprocessing on the test image, retrieve the grids after superpixel segmentation in a predetermined order, process the test grids, repeat step S2, calculate the color and position of each cluster center in the grid, obtain its color space distribution, and record the color distance and space distance of each pixel area;

[0017] Step S8: Compare the result obtained in step S7 with the result obtained in step S2, set the color distance and space distance thresholds, and the grids that are smaller than the thresholds are similar to the grids, and select the grids that meet the similar rules for the next step of reasoning. If they do not meet the rules, skip the grid and search for the next grid;

[0018] Step S9: The preprocessed result is used as input to the remote sensing image transmission tower recognition model. The model performs the next step of detection on the grids that meet the conditions, determines the target position of the tower based on the context information, and outputs the coordinate information of the tower and the annotation box in the image; the output result is saved and restored to the super large image for matching.

[0019] Furthermore, the multi-scenario random field model includes a conditional layer, a shared layer and a branch layer; the conditional layer converts the image into a representation of a node sequence and uses a convolutional neural network to extract the visual features of the nodes; the shared layer uses the nodes and their edge connections obtained by the conditional random field module as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers and fully connected layers, which are used to learn task-specific features and outputs.

[0020] Define the loss function:

[0021] L w =Σ(w i *f i (y,x))-log(Z(x));

[0022] Among them, L w is the conditional loss function, w i is the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions, and Z(x) is the normalization factor;

[0023] L m =λ1*L1(z1,y)+λ2*L2(z2,y)+...+λ N *L N (z N ,y)

[0024] Among them, L m is the task loss function, L1,L2,...,L N are the loss functions for target detection corresponding to N different background styles, y is the true label, λ1, λ2, ..., λ N is the task weight;

[0025] The overall loss function is defined as:

[0026] L=L+L

[0027] L=L w +L m .

[0028] Further, the conditional layer is as follows:

[0029] A set of feature functions are defined for each node and its relationship. The feature functions calculate feature values ​​based on the visual features, location information, and other contextual information of the node.

[0030] Define the parameterized form of the conditional random field module. For the node sequence in the image, define the conditional random field model as:

[0031]

[0032] Among them, P(y|x) is the conditional probability of the relation label sequence y given the input image feature x, Z(x) is the normalization factor, and w i is the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions;

[0033] Estimate the parameters of the model by maximizing the likelihood of the training data and learn weight parameters that are suitable for the data;

[0034] For a given input image feature x, the conditional probability distribution of the relationship tag sequence y is calculated, and the relationship tag sequence with the highest probability is selected as the prediction result, and different distribution results are carried out to the next step in different branch forms.

[0035] Further, the shared layer is as follows:

[0036] h=f(x,θ shared )

[0037] Among them, f is the function of the shared layer, θ sharedare the parameters of the shared layer; the nodes and their edge connections obtained by the conditional random field module are used as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the representation of these intermediate layers is shared among different tasks, and the output transmission after convolution, pooling, and full connection processing enables the model to learn the correlation and common features between multiple tasks; the parameters of the shared layer are updated through back propagation during the training process to minimize the loss function of multiple tasks.

[0038] Furthermore, the branch layer learns the target features expressed differently in multiple contexts and performs specific analysis for specific situations:

[0039] z1,z2,…,z N =g1(h,θ1),g2(h,θ2),…,g N (h,θ N );

[0040] Among them, g1, g2, …, g N is a function of the branching layer, θ1,θ2,…,θ N are the parameters of the branch layer, corresponding to N different background styles;

[0041] The output of the shared layer is used as input and connected to the branch layer representing each scene; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers, and fully connected layers, which are used to learn task-specific features and outputs; the parameters θ are updated through back propagation during the training process. shared ,θ1,θ2,…,θ N , in order to minimize the loss function of each task; in different scenarios, the node distribution distance and rules are different. In different tasks, new parameters are introduced for learning to improve the adaptability of the model.

[0042] Furthermore, a superpixel segmentation network is constructed to perform gridding preprocessing on the test image, as follows:

[0043] Initialize the seed point K, which represents the number of pre-generated superpixels. Different from the traditional superpixel segmentation method of color contour segmentation, the segmentation result is divided into a grid, that is, the image of size A*B is divided into K superpixel blocks, and the range size of each superpixel block contains [(A*B) / K] pixels, which reduces the probability that other buildings in the remote sensing image are assigned to the same label as the transmission tower due to their similar colors;

[0044] Redefine the seed point in the neighborhood of the seed point, calculate the gradient values ​​of all pixels in the neighborhood, and move the seed point to the place with the smallest gradient in the neighborhood; the length and width of each neighborhood are evenly distributed, defined as S:

[0045]

[0046] Traverse the operation, save the coordinates (x, y) of the seed point of each pixel block and its lab value, and add them to the pre-defined set;

[0047] The default seed point of each pixel block is (S / 2, S / 2). The gradient is calculated by differential method to adjust the seed point:

[0048] Use the 8-domain pixel points of the seed point to calculate the pixel point with the minimum gradient value, and use it as the new seed point. The formula for calculating the gradient by difference is:

[0049] Gradient(x,y)=dx(i,j)+dy(i,j);

[0050] dx(i,j)=I(i+1,j)-I(i,j);

[0051] dy(i,j)=I(i,j+1)-I(i,j);

[0052] Traverse the 8-region pixel points of the current seed point, and take the pixel point with the minimum gradient value as the new center point;

[0053] Since each pixel point will be searched by multiple seed points, each pixel point will have a distance to the surrounding seed points, and the seed point corresponding to the minimum value is taken as the cluster center of the pixel point;

[0054] Considering the similarity between pixel values ​​and coordinate values, the pixels are classified using the transformed Euclidean distance formula:

[0055] Gradient(x,y)=dx(i,j)+dy(i,j);

[0056] dx(i,j)=I(i+1,j)-I(i,j);

[0057] dy(i,j)=I(i,j+1)-I(i,j);

[0058] The proportion of the two distances is coordinated by two parameters m and S, wherein the parameter S is the length value of each pixel block calculated in step S3, and the parameter m is the possible maximum value of the distance in the LAB space;

[0059] Only the pixels in the 2S*2S area around the center point of each superpixel block are traversed to accelerate the convergence of the algorithm. The center point of each superpixel block is calculated to which each pixel in the area is closest and divided into the superpixel block. After completing one iteration, the coordinates of the center point of each superpixel block are recalculated and iterated again.

[0060] The present invention has the following beneficial effects:

[0061] 1. The present invention integrates the model of conditional random fields and multi-task learning, analyzes the line connection relationship between transmission towers, and assists the target detection model to improve the accuracy of tower detection. Aiming at the problem of difficulty in reasoning of super-large images, a reasoning idea based on superpixel segmentation is designed, and the non-target area is quickly filtered by the grid retrieval method, which greatly improves the reasoning speed;

[0062] 2. The method of integrating conditional random fields and multi-task learning in the present invention reduces the difficulty of locating tower targets in remote sensing images, extracts context information from the characteristics of the towers themselves and the line connection relationship between them, and supplements the associated feature information of transmission towers and lines for the target detection model to improve the model accuracy;

[0063] 3. The image target recognition method based on superpixel segmentation of the present invention preprocesses the original image, performs data analysis, data update, and continuously trains and solidifies the model, which directly improves the accuracy and recall rate of the recognition results and optimizes the efficiency of the model for large image detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 This is a schematic diagram of the principle structure of a multi-scenario random field in one embodiment of the present invention;

[0066] Figure 3 Schematic diagram of superpixel segmentation principle in one embodiment of the present invention. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0068] refer to Figure 1 In this embodiment, a remote sensing image target detection method integrating multi-scene random fields is provided, comprising the following steps:

[0069] Step S1: Collect remote sensing image data sets of power inspection transmission towers and lines under different background environments and mark them;

[0070] Step S2: Taking the pole tower as a node, preprocessing the input image, extracting the central color and position of the node by clustering, and representing the basic color and spatial coordinates of the target;

[0071] Step S3: construct a multi-scenario random field model, input multiple nodes and their attributes and known relationship labels, and output the characteristic relationship between the transmission tower and the line;

[0072] Step S4: the number of different background style types is represented as N, each image input is represented as x, and the detection output of the tower and line is represented as y;

[0073] Step S5: The output y obtained by the multi-scenario random field model including the connection relationship between nodes and edges is saved in the form of a labeled file, and combined with the features extracted by the remote sensing image recognition model as feature information to supplement the tower recognition logic and assist the model in performing reasoning tests;

[0074] Step S6: constructing a remote sensing image transmission tower recognition model based on the YOLOv8 algorithm;

[0075] Step S7: construct a superpixel segmentation network, perform gridding preprocessing on the test image, retrieve the grids after superpixel segmentation in a predetermined order, process the test grids, repeat step S2, calculate the color and position of each cluster center in the grid, obtain its color space distribution, and record the color distance and space distance of each pixel area;

[0076] Step S8: Compare the result obtained in step S7 with the result obtained in step S2, set the color distance and space distance thresholds, and the grids that are smaller than the thresholds are similar to the grids, and select the grids that meet the similar rules for the next step of reasoning. If they do not meet the rules, skip the grid and search for the next grid;

[0077] Step S9: The preprocessed result is used as input to the remote sensing image transmission tower recognition model. The model performs the next step of detection on the grids that meet the conditions, determines the target position of the tower based on the context information, and outputs the coordinate information of the tower and the annotation box in the image; the output result is saved and restored to the super large image for matching.

[0078] In this embodiment, the multi-scenario random field model includes a conditional layer, a shared layer and a branch layer; the conditional layer converts the image into a representation of a node sequence and uses a convolutional neural network to extract the visual features of the nodes; the shared layer uses the nodes and their edge connections obtained by the conditional random field module as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers and fully connected layers, which are used to learn task-specific features and outputs.

[0079] Define the loss function:

[0080] L w =Σ(w i *f i (y,x))-log(Z(x));

[0081] Among them, L w is the conditional loss function, w iis the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions, and Z(x) is the normalization factor;

[0082] L m =λ1*L1(z1,y)+λ2*L2(z2,y)+...+λ N *L N (z N ,y)

[0083] Among them, L m is the task loss function, L1,L2,...,L N are the loss functions for target detection corresponding to N different background styles, y is the true label, λ1, λ2, ..., λ N is the task weight;

[0084] The overall loss function is defined as:

[0085] L=L+L

[0086] L=L w +L m .

[0087] In this embodiment, the conditional layer is as follows:

[0088] A set of feature functions are defined for each node and its relationship. The feature functions calculate feature values ​​based on the visual features, location information, and other contextual information of the node.

[0089] Define the parameterized form of the conditional random field module. For the node sequence in the image, define the conditional random field model as:

[0090]

[0091] Among them, P(y|x) is the conditional probability of the relation label sequence y given the input image feature x, Z(x) is the normalization factor, and w i is the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions;

[0092] Estimate the parameters of the model by maximizing the likelihood of the training data and learn weight parameters that are suitable for the data;

[0093] For a given input image feature x, the conditional probability distribution of the relationship tag sequence y is calculated, and the relationship tag sequence with the highest probability is selected as the prediction result, and different distribution results are carried out to the next step in different branch forms.

[0094] In this embodiment, the shared layer is as follows:

[0095] h=f(x,θ shared )

[0096] Among them, f is the function of the shared layer, θ shared are the parameters of the shared layer; the nodes and their edge connections obtained by the conditional random field module are used as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the representation of these intermediate layers is shared among different tasks, and the output transmission after convolution, pooling, and full connection processing enables the model to learn the correlation and common features between multiple tasks; the parameters of the shared layer are updated through back propagation during the training process to minimize the loss function of multiple tasks.

[0097] In this embodiment, the branch layer learns the target features expressed differently under multiple backgrounds and performs specific analysis for specific situations:

[0098] z1,z2,...,z N =g1(h,θ1),g2(h,θ2),...,g N (h,θ N );

[0099] Among them, g1, g2, ..., g N is a function of the branching layer, θ1,θ2,...,θ N are the parameters of the branch layer, corresponding to N different background styles;

[0100] The output of the shared layer is used as input and connected to the branch layer representing each scene; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers, and fully connected layers, which are used to learn task-specific features and outputs; the parameters θ are updated through back propagation during the training process. shared ,θ1,θ2,...,θ N , in order to minimize the loss function of each task; in different scenarios, the node distribution distance and rules are different. In different tasks, new parameters are introduced for learning to improve the adaptability of the model.

[0101] In this embodiment, a superpixel segmentation network is constructed and gridding preprocessing is performed on the test image, as follows:

[0102] Initialize the seed point K, which represents the number of pre-generated superpixels. Different from the traditional superpixel segmentation method of color contour segmentation, the segmentation result is divided into a grid, that is, the image of size A*B is divided into K superpixel blocks, and the range size of each superpixel block contains [(A*B) / K] pixels, which reduces the probability that other buildings in the remote sensing image are assigned to the same label as the transmission tower due to their similar colors;

[0103] Redefine the seed point in the neighborhood of the seed point, calculate the gradient values ​​of all pixels in the neighborhood, and move the seed point to the place with the smallest gradient in the neighborhood; the length and width of each neighborhood are evenly distributed, defined as S:

[0104]

[0105] Traverse the operation, save the coordinates (x, y) of the seed point of each pixel block and its lab value, and add them to the pre-defined set;

[0106] The default seed point of each pixel block is (S / 2, S / 2). The gradient is calculated by differential method to adjust the seed point:

[0107] Use the 8-domain pixel points of the seed point to calculate the pixel point with the minimum gradient value, and use it as the new seed point. The formula for calculating the gradient by difference is:

[0108] Gradient(x,y)=dx(i,j)+dy(i,j);

[0109] dx(i,j)=I(i+1,j)-I(i,j);

[0110] dy(i,j)=I(i,j+1)-I(i,j);

[0111] Traverse the 8-region pixel points of the current seed point, and take the pixel point with the minimum gradient value as the new center point;

[0112] Since each pixel point will be searched by multiple seed points, each pixel point will have a distance to the surrounding seed points, and the seed point corresponding to the minimum value is taken as the cluster center of the pixel point;

[0113] Considering the similarity between pixel values ​​and coordinate values, the pixels are classified using the transformed Euclidean distance formula:

[0114] Gradient(x,y)=dx(i,j)+dy(i,j);

[0115] dx(i,j)=I(i+1,j)-I(i,j);

[0116] dy(i,j)=I(i,j+1)-I(i,j);

[0117] The proportion of the two distances is coordinated by two parameters m and S, wherein the parameter S is the length value of each pixel block calculated in step S3, and the parameter m is the possible maximum value of the distance in the LAB space;

[0118] Only the pixels in the 2S*2S area around the center point of each superpixel block are traversed to accelerate the convergence of the algorithm. The center point of each superpixel block is calculated to which each pixel in the area is closest and divided into the superpixel block. After completing one iteration, the coordinates of the center point of each superpixel block are recalculated and iterated again.

[0119] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0121] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0123] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A remote sensing image target detection method integrating multi-scene random fields, characterized in that: The following steps are involved: Step S1: Collect remote sensing image data sets of power inspection transmission towers and lines under different background environments and mark them; Step S2: Taking the pole tower as a node, preprocessing the input image, extracting the central color and position of the node by clustering, and representing the basic color and spatial coordinates of the target; Step S3: construct a multi-scenario random field model, input multiple nodes and their attributes and known relationship labels, and output the characteristic relationship between the transmission tower and the line; Step S4: the number of different background style types is represented as N, each image input is represented as x, and the detection output of the tower and line is represented as y; Step S5: The output y obtained by the multi-scenario random field model including the connection relationship between nodes and edges is saved in the form of a labeled file, and combined with the features extracted by the remote sensing image recognition model as feature information to supplement the tower recognition logic and assist the model in performing reasoning tests; Step S6: constructing a remote sensing image transmission tower recognition model based on the YOLOv8 algorithm; Step S7: construct a superpixel segmentation network, perform gridding preprocessing on the test image, retrieve the grids after superpixel segmentation in a predetermined order, process the test grids, repeat step S2, calculate the color and position of each cluster center in the grid, obtain its color space distribution, and record the color distance and space distance of each pixel area; Step S8: Compare the result obtained in step S7 with the result obtained in step S2, set the color distance and space distance thresholds, and the grids that are smaller than the thresholds are similar to the grids, and select the grids that meet the similar rules for the next step of reasoning. If they do not meet the rules, skip the grid and search for the next grid; Step S9: The preprocessed result is used as input to the remote sensing image transmission tower recognition model. The model performs the next step of detection on the grids that meet the conditions, determines the target position of the tower based on the context information, and outputs the coordinate information of the tower and the annotation box in the image; the output result is saved and restored to the super large image for matching.

2. The method for remote sensing image target detection by integrating multi-scene random fields according to claim 1, characterized in that: The multi-scene random field model includes a conditional layer, a shared layer and a branching layer; the conditional layer converts the image into a representation of a node sequence and uses a convolutional neural network to extract visual features of the nodes; The shared layer uses the nodes and their edge connections obtained by the conditional random field module as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers and fully connected layers, which are used to learn task-specific features and outputs, and define the loss function: L w =∑(w i *f i (y,x))-log(Z(x)); Among them, L w is the conditional loss function, w i is the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions, and Z(x) is the normalization factor; L m =λ1*L1(z1,y)+λ2*L2(z2,y)+...+λ N *L N (z N ,y) Among them, L m is the task loss function, L1,L2,…,L N are the loss functions for target detection corresponding to N different background styles, y is the true label, λ1, λ2,…,λ N is the task weight; The overall loss function is defined as: L=L w +L m .

3. The method for remote sensing image target detection by integrating multi-scene random fields according to claim 2, characterized in that: The conditional layer is specifically as follows: A set of feature functions are defined for each node and its relationship. The feature functions calculate feature values ​​based on the visual features, position information and other contextual information of the node. Define the parameterized form of the conditional random field module. For the node sequence in the image, define the conditional random field model as: Among them, P(y|x) is the conditional probability of the relation label sequence y given the input image feature x, Z(x) is the normalization factor, and w i is the weight parameter of the model, f i is the characteristic function, Σ represents the sum of all characteristic functions; Estimate the parameters of the model by maximizing the likelihood of the training data and learn weight parameters that are suitable for the data; For a given input image feature x, the conditional probability distribution of the relationship tag sequence y is calculated, and the relationship tag sequence with the highest probability is selected as the prediction result, and different distribution results are carried out to the next step in different branch forms.

4. The method for remote sensing image target detection by integrating multi-scene random fields according to claim 2, characterized in that: The shared layer is as follows: h=f(x,θ shared ) Among them, f is the function of the shared layer, θ shared are the parameters of the shared layer; the nodes and their edge connections obtained by the conditional random field module are used as the tower and line information through the forward propagation process of the shared layer to obtain the representation of the intermediate layer; the representation of these intermediate layers is shared among different tasks, and the output transmission after convolution, pooling, and full connection processing enables the model to learn the correlation and common features between multiple tasks; the parameters of the shared layer are updated through back propagation during the training process to minimize the loss function of multiple tasks.

5. The method for remote sensing image target detection by integrating multi-scene random fields according to claim 2, characterized in that: The branch layers learn target features expressed differently under multiple backgrounds and perform specific analysis for specific situations: z1,z2,...,z N =g1(h,θ1),g2(h,θ2),...,g N (h,θ N ); Among them, g1, g2, …, g N is a function of the branching layer, θ1,θ2,...,θ N are the parameters of the branch layer, corresponding to N different background styles; The output of the shared layer is used as input and connected to the branch layer representing each scene; the branch layer contains a task-specific layer structure, including multiple convolutional layers, pooling layers, and fully connected layers, which are used to learn task-specific features and outputs; the parameters θ are updated through back propagation during the training process. shared ,θ1,θ2,…,θ N , in order to minimize the loss function of each task; in different scenarios, the node distribution distance and rules are different. In different tasks, new parameters are introduced for learning to improve the adaptability of the model.

6. The method for remote sensing image target detection by integrating multi-scene random fields according to claim 1, characterized in that: The superpixel segmentation network is constructed to perform gridding preprocessing on the test image, as follows: Initialize the seed point K, which represents the number of pre-generated superpixels. Different from the traditional superpixel segmentation method of color contour segmentation, the segmentation result is divided into a grid, that is, the image of size A*B is divided into K superpixel blocks, and the range size of each superpixel block contains [(A*B) / K] pixels, which reduces the probability that other buildings in the remote sensing image are assigned to the same label as the transmission tower due to their similar colors; Redefine the seed point in the neighborhood of the seed point, calculate the gradient values ​​of all pixels in the neighborhood, and move the seed point to the place with the smallest gradient in the neighborhood; the length and width of each neighborhood are evenly distributed, defined as S: Traverse the operation, save the coordinates (x, y) of the seed point of each pixel block and its lab value, and add them to the pre-defined set; The default seed point of each pixel block is (S / 2, S / 2). The gradient is calculated by differential method to adjust the seed point: Use the 8-domain pixel points of the seed point to calculate the pixel point with the minimum gradient value, and use it as the new seed point. The formula for calculating the gradient by difference is: Gradient(x,y)=dx(i,j)+dy(i,j); dx(i,j)=I(i+1,j)-I(i,j); dy(i,j)=I(i,j+1)-I(i,j); Traverse the 8-region pixel points of the current seed point, and take the pixel point with the minimum gradient value as the new center point; Since each pixel point will be searched by multiple seed points, each pixel point will have a distance to the surrounding seed points, and the seed point corresponding to the minimum value is taken as the cluster center of the pixel point; Considering the similarity between pixel values ​​and coordinate values, the pixels are classified using the transformed Euclidean distance formula: Gradient(x,y)=dx(i,j)+dy(i,j); dx(i,j)=I(i+1,j)-I(i,j); dy(i,j)=I(i,j+1)-I(i,j); The proportion of the two distances is coordinated by two parameters m and S, wherein the parameter S is the length value of each pixel block calculated in step S3, and the parameter m is the possible maximum value of the distance in the LAB space; Only the pixels in the 2S*2S area around the center point of each superpixel block are traversed to accelerate the convergence of the algorithm. The center point of each superpixel block is calculated to which each pixel in the area is closest and divided into the superpixel block. After completing one iteration, the coordinates of the center point of each superpixel block are recalculated and iterated again.

Images