Infrared small target detection method for graph structure analysis

Abstract

The invention discloses an infrared small target detection method based on graph structure analysis, and the infrared target detection has important application value in the fields of military and safety air defense, but is often faced with the problem that a real target is difficult to distinguish from various environmental interferences in practical application. The core idea of the method is mainly reflected in three aspects: firstly, by analyzing node features in candidate targets, detail identification of a real target and environmental interference is enhanced; secondly, a node aggregation analysis module is adopted to extract convergence features, and the overall difference between a target and interference is obtained; and finally, on the basis of a graph connection rule, node labels are expanded by a set obtained after node labels of adjacent nodes are sequenced, and the expanded labels are compressed into a reasoning module formed by new labels so as to analyze global association between a target and interference information, so that the recognition precision is further improved. The method can effectively distinguish a target from an interference signal, is suitable for an infrared target detection task in a complex scene, and has relatively high practicability and anti-interference capability.

Description

Technical Field

[0001] The present invention relates to infrared dim small target detection, which is an important research direction of computer vision and image processing, and has wide application value in military, aerospace, security monitoring and other fields. A method for infrared small target detection based on graph structure analysis is provided. Background Art

[0002] In recent years, with the rapid development of infrared technology, infrared imaging technology has been widely used in military, security, medical and other related fields. In particular, infrared target detection plays a huge role in the space-based infrared satellite earth observation system. According to the high-intensity radiation generated by missiles and aircraft during the gliding process, space-based infrared satellites can quickly capture this information using their onboard infrared ground detectors. However, there are still high-intensity radiation false targets close to the target, including small frozen lakes, small stars and high-altitude cirrus clouds, which will interfere with the early warning system.

[0003] To address numerous challenges, such as target capture and false infrared target detection, methods based on tensor decomposition, visual saliency, and deep learning have all made some progress. However, tensor decomposition methods struggle to model the nonlinear relationship between target and background, visual saliency methods struggle to distinguish salient regions of target and false infrared targets when faced with the same high-radiation information, and deep learning relies on extensive data annotation, resulting in poor model interpretability. To overcome these shortcomings, the present invention utilizes graph structure analysis to address the problem of infrared target detection. Graph structures can naturally represent the relationship between target and interfering signal pixels, effectively capturing both global and local features of the target, thereby enabling powerful relationship modeling. The parameters of the graph structure analysis model have clear physical meanings, making them easy to understand and analyze. Therefore, we conducted infrared target detection using graph structure analysis. This method removes false alarms that may interfere with the target from the target detection results, reduces the interference of high-radiation false infrared targets on the infrared warning system, and improves target detection accuracy. This method has important research significance in practical detection algorithms. Summary of the Invention

[0004] The present invention aims to address the high false detection rate in existing infrared small target detection methods, which is caused by the inability to effectively model the nonlinear relationship between the target and the background and to distinguish high-radiation false targets. This method enhances the local information representation of the target and false alarms by leveraging the node features of a graph. It also uses a module that aggregates neighbor node information to extract the graph's converged features, capturing the global characteristics between the target and the interference signal. This approach, combined with the graph's topological properties, further improves the model's recognition accuracy and detection capabilities.

[0005] In order to solve the above technical problems and achieve the above objectives, the technical solutions adopted by the present invention are as follows.

[0006] A method for detecting infrared small targets based on graph structure analysis includes the following steps:

[0007] Step 1: Perform a preliminary screening of suspected targets on the input infrared image to obtain a candidate area image and proceed to step 2;

[0008] Step 2: Build a graph structure using the candidate area image obtained in step 1, including node information, connection links, and system-level topological characteristics, and proceed to step 3;

[0009] Step 3: Embed the constructed graph structure data into the feature extractor to extract node features, aggregate neighbor node features, and topological structure features to achieve broader, multi-level, and multi-dimensional feature extraction. After obtaining the feature extraction information, proceed to step 4;

[0010] Step 4: Input the feature parsing data obtained in step 3 into the classifier, design the loss function, and finally obtain an efficient infrared target detection model through model training.

[0011] In the above technical solution, step 1 includes the following steps:

[0012] Step 1.1: Input an infrared image I of size w×h. Based on the background modeling framework, perform a structural adaptive operation on the input image to remove small bright spots to estimate the background. Then, subtract the estimated background from the input image to obtain the confidence separation result I. separation , as shown in (1), formula (1) is as follows:

[0013]

[0014] in, It is performed on the input image Adaptive operation of structural elements, The structure with r0 as the radius, where the calculation formula of r0 is shown in formula (2), which is as follows:

[0015]

[0016] Step 1.2: Confidence separation result I obtained in step 1.1 separation Then, based on the background perception judgment, θ is used as the threshold to segment the suspicious target to obtain the confidence binary image I bw , the value of θ is determined by bisection, and the termination condition is from I bw The number of connected domains obtained by removing the edge part of the image is between 1 and K.

[0017] Step 1.3: Obtain I through step 1.2 bw Afterwards, It is the inverse operation of the adaptive operation of the structural element to obtain n connected domains Take each connected domain as the center point As the center, in the original image I, the length and width are r c Rectangular area as candidate area r c Calculated by formula (3), formula (3) is as follows:

[0018]

[0019] in, and are the width and height of the minimum circumscribed matrix of the connected region respectively.

[0020] In the above technical solution, step 2 includes the following steps:

[0021] Step 2.1: Convert the result in step 1.3 to the candidate area Calculate the gradient I in the x, y direction x , I y .in,

[0022] Step 2.2: Use the gradients in the x and y directions obtained in step 2.1 to construct the structure tensor M through formula (4), and then use formula (5) to calculate the response value R of the node to screen the key nodes and retain the points where R>threshold. Formulas (4) and (5) are as follows:

[0023]

[0024] R=det(M)-α.trace(M) 2 (5)

[0025] Where w(x, y) is the weight of the Gaussian window, α is the sensitivity coefficient. The smaller the value, the more nodes are detected. α∈(0.04,0.06), threshold=β.R max ,β∈(0.01,0.05),R max is the maximum value of R.

[0026] Step 2.3: Calculate the amplitude in the neighborhood of the key node by screening the key nodes in step 2.2 and direction To obtain the gradient direction histogram around the key node, the peak direction in the histogram is selected as the direction of the key node.

[0027] Step 2.4: Based on the position and direction of the key node calculated in step 2.3, construct an 8×8 pixel area, divide it into 2×2 small blocks, calculate the gradient histogram in 8 directions, and generate a 2×2×8=32-dimensional key node feature vector x=[x1, x1, ..., x d ].

[0028] Step 2.5: The eigenvector obtained in step 2.4 is subjected to explicit nonlinear mapping through equation (9). Equation (6) is as follows:

[0029]

[0030] The expanded dimension is

[0031] In the above technical solution, step 3 includes the following steps:

[0032] Step 3.1: Each candidate area obtained in step 1.3 As a subgraph, each pixel of the subgraph is regarded as a node, the value of the node is the grayscale value of the pixel, and the adjacent pixels are regarded as an edge, and the subgraph structure G can be obtained. Formula (7) is as follows:

[0033] G=(N,E) (7)

[0034] Where N is the set of subgraph nodes and E is the set of subgraph edges.

[0035] Step 3.2: Assign values to each adjacency matrix element based on the correlation between pixels. If node i and node j are connected, then A ij =g ij , otherwise A ij = 0. Each element is calculated by formula (8), which is as follows:

[0036]

[0037] Among them, I i is the grayscale value of node i, I j is the grayscale value of node j, and σ is the parameter that controls weight decay. The set of all elements of the subgraph adjacency matrix A can be obtained as (9), which is as follows:

[0038]

[0039] Step 3.3: For the adjacency matrix A obtained in step 3.2, calculate the degree D of the local unit by counting the number of edges connected to node i ii It can be expressed by formula (10), which is as follows:

[0040] Dii =∑ j A ij (10)

[0041] Step 3.4: In order to enhance the numerical stability, the adjacency matrix is normalized using formula (11), which is as follows:

[0042]

[0043] Among them, I N is the identity matrix, It's A+I N The degree matrix of .

[0044] Step 3.5: For node i, use the normalized adjacency matrix obtained in step 3.4 Aggregate the features of itself and neighboring nodes through formula (12), and then repeat the propagation K times. Formula (12) is as follows:

[0045]

[0046] in, is the neighbor set of node i, X is the node eigenvector matrix.

[0047] Step 3.6: Analyze the topological structure of the subgraph obtained in step 3.1 and assign an initial label to each node v. The node degree used in the initial label is used to obtain the initial feature vector as shown in formula (13). Formula (13) is as follows:

[0048]

[0049] Among them, Am is the initial feature dimension.

[0050] Step 3.7 For each node v, aggregate the labels of the neighbors in t iterations Formula (14) is as follows:

[0051]

[0052] in, is the neighbor set of node v, and AGGREGATE is the aggregation function.

[0053] Step 3.8 combines the node's own label with the aggregated neighbor information from step 3.7 to generate a new label as follows:

[0054]

[0055] In step 3.9, to ensure the collision rate is as low as possible, the labels of the node itself and the aggregated neighbor nodes in step 3.8 are used to obtain the extended label string. The strings are sorted in ascending order and a one-to-one mapping dictionary is generated using formula (17). Formula (16) is as follows:

[0056]

[0057] in, N i is the neighbor node of i, d i , d j is the degree of nodes i and j. This transfer rule refers to the addition of reference parameter W t After T rounds of iteration, the feature vector is updated to:

[0058]

[0059] Among them, when H T =H T-1 When the preset maximum number of iterations T is reached, the node label no longer changes.

[0060] In the above technical solution, step 4 includes the following steps:

[0061] Step 4.1: Through the feature expression of the above steps, the node feature vector X obtained in step 2.5 is converted to poly , the aggregated neighbor node feature vector H extracted in step 3.5 (L) , and the topological structure feature vector C in step 3.9 T Channel connection is performed through formula (19) to obtain the mixed feature layer Z, formula (18) is as follows:

[0062] Z=|X poly ||H (K) ||C T | (18)

[0063] Among them, |*| represents the splicing operation, and the comprehensive feature vector Z is obtained by splicing node features, aggregating neighbor node features and topological structure features.

[0064] Step 4.2: The feature Z fused in step 4.1 is input into the fully connected network through formula (19) for prediction. Formula (20) is as follows:

[0065] Targ et pre =softmax(Re LU(W (l) h (l) +b (l) )) (19)

[0066] Among them, h (0)=Z,W (l) and b (l) represents the weight and bias of the lth layer, ReLU is the activation function, softmax converts the distribution of the function into probability, and cross entropy is used as the loss function as shown in formula (20):

[0067]

[0068] Among them, t i It is the one-hot encoding of the true label. Finally, the loss function is optimized through back propagation and gradient descent method, the network parameters are updated, and the probability prediction of infrared targets is achieved.

[0069] Because the present invention adopts the above technical solution, it has the following beneficial effects:

[0070] 1. The infrared small target detection method based on graph structure analysis proposed in the present invention converts the original infrared data into a graph structure data type, and effectively models the complex relationship between the target and the background, and between the targets according to the graph structure information.

[0071] 2. This method introduces a neighbor node aggregation module and a topology graph reasoning module in the target detection task based on traditional graph structure analysis, which can effectively utilize global context information and improve detection performance.

[0072] 3. The infrared small target detection model based on graph structure analysis proposed by this method can flexibly respond to various detection requirements and be easy to expand by adjusting the graph structure and node features according to different scenarios and detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Design a process for the entire algorithm;

[0074] Figure 2 The relevant images extracted for the candidate targets are the original image, the real target annotation, the candidate area extraction response image, and the segmentation result.

[0075] Figure 3 This is a schematic diagram of the target graph expression. After the graph structured expression, the real target and the false target can be effectively distinguished.

[0076] Figure 4 The topological reasoning process of the graph

[0077] Figure 5 It is a specific network structure diagram, which includes three modules: node feature, aggregation neighbor node feature and topology feature extraction; DETAILED DESCRIPTION

[0078] The present invention will be further described in detail below in conjunction with test examples and specific embodiments. However, this should not be understood as limiting the scope of the present invention to the following embodiments, and all technologies implemented based on the present invention fall within the scope of the present invention.

[0079] The present invention provides a method for infrared small target detection based on graph structure analysis. The technical problem to be solved is to enhance the local information expression of the target and interference signals through the node features of the graph, and then use the node aggregation analysis module to extract the converged features to obtain the overall difference between the target and the interference. Combined with the topological reasoning process of the graph, the recognition accuracy and detection ability of the model are further improved. The entire algorithm design structure framework is as follows Figure 1 As shown, the steps include:

[0080] Step 1.1: Input an infrared image I of size w×h. Based on the background modeling framework, perform a structural adaptive operation on the input image to remove small bright spots to estimate the background. Then, subtract the estimated background from the input image to obtain the confidence separation result I. separation , as shown in (1), formula (1) is as follows:

[0081]

[0082] in, It is performed on the input image Adaptive operation of structural elements, The structure with r0 as the radius, where the calculation formula of r0 is shown in formula (2), which is as follows:

[0083]

[0084] In this example, the image size is 600×400, then It is an opening operation of a structure with a radius of 11.

[0085] Step 1.2: Confidence separation result I obtained in step 1.1 separation Then, based on the background perception judgment, θ is used as the threshold to segment the suspicious target to obtain the confidence binary image I bw , the value of θ is determined by bisection, and the termination condition is from I bw The number of connected domains obtained from removing the edge of the image is between 1 and K. In this example, experiments have shown that 0.3 is a good starting value for θ.

[0086] Step 1.3: Obtain I through step 1.2 bw Afterwards, It is the inverse operation of the adaptive operation of the structural element to obtain n connected domains Take each connected domain as the center point As the center, in Figure Ibw Take the length and width as r c Rectangular area as candidate area r c Calculated by formula (3), formula (3) is as follows:

[0087]

[0088] in, and are the width and height of the minimum circumscribed matrix of the connected region. In this example, The final calculated

[0089] Step 2.1: Convert the result in step 1.3 to the candidate area Calculate the gradient I in the x, y direction x , I y .in,

[0090] Step 2.2: Use the gradients in the x and y directions obtained in step 2.1 to construct the structure tensor M through formula (4), and then use formula (5) to calculate the response value R of the node to screen the key nodes and retain the points where R>threshold. Formulas (4) and (5) are as follows:

[0091]

[0092] R=det(M)-α.trace(M) 2 (5)

[0093] Where w(x, y) is the weight of the Gaussian window, α is the sensitivity coefficient. The smaller the value, the more nodes are detected. α∈(0.04,0.06), threshold=β.R max ,β∈(0.01,0.05),R max is the maximum value of R. In this example, α = 0.02 and β = 0.03.

[0094] Step 2.3: Calculate the amplitude in the neighborhood of the key node by screening the key nodes in step 2.2 and direction To obtain the gradient direction histogram around the key node, the peak direction in the histogram is selected as the direction of the key node.

[0095] Step 2.4: Based on the position and direction of the key node calculated in step 2.3, construct an 8×8 pixel area, divide it into 2×2 small blocks, calculate the gradient histogram in 8 directions, and generate a 2×2×8=32-dimensional key node feature vector X=[x1, x1, ..., xd ].

[0096] Step 2.5: The eigenvector obtained in step 2.4 is subjected to explicit nonlinear mapping through equation (9). Equation (6) is as follows:

[0097]

[0098] The expanded dimension is

[0099] Step 3.1: Each candidate area obtained in step 1.3 As a subgraph, each pixel of the subgraph is regarded as a node, the value of the node is the grayscale value of the pixel, and the adjacent pixels are regarded as an edge, and the subgraph structure G can be obtained. Formula (7) is as follows:

[0100] G=(N,E) (7)

[0101] Where N is the set of subgraph nodes and E is the set of subgraph edges.

[0102] Step 3.2: Assign values to each adjacency matrix element based on the correlation between pixels. If node i and node j are connected, then A ij =g ij , otherwise A ij = 0. Each element is calculated by formula (8), which is as follows:

[0103]

[0104] Among them, I i is the grayscale value of node i, I j is the grayscale value of node j, and σ is the parameter that controls weight decay. The set of all elements of the subgraph adjacency matrix A can be obtained as (9), which is as follows:

[0105]

[0106] In this example, the weight decay parameter σ is controlled by 1e -3 .

[0107] Step 3.3: For the adjacency matrix A obtained in step 3.2, calculate the degree D of the local unit by counting the number of edges connected to node i ii It can be expressed by formula (10), which is as follows:

[0108] D ii =∑ j A ij (10)

[0109] Step 3.4: In order to enhance the numerical stability, the adjacency matrix is normalized using formula (11), which is as follows:

[0110]

[0111] Among them, I N is the identity matrix, It's A+I N The degree matrix of .

[0112] Step 3.5: For node i, use the normalized adjacency matrix obtained in step 3.4 Aggregate the features of itself and neighboring nodes through formula (12), and then repeat the propagation K times. Formula (12) is as follows:

[0113]

[0114] in, is the neighbor set of node i, X is the node eigenvector matrix. In this example, K=3 is used to avoid redundant calculations of multiple ANDs.

[0115] Step 3.6: Analyze the topological structure of the subgraph obtained in step 3.1 and assign an initial label to each node v. The node degree used in the initial label is used to obtain the initial feature vector as shown in formula (13). Formula (14) is as follows:

[0116]

[0117] Among them, Am is the initial feature dimension.

[0118] Step 3.7 For each node v, aggregate the labels of the neighbors in t iterations Formula (14) is as follows:

[0119]

[0120] in, is the neighbor set of node v, and AGGREGATE is the aggregation function.

[0121] Step 3.8 combines the node's own label with the aggregated neighbor information from step 3.7 to generate a new label as follows:

[0122]

[0123] In step 3.9, to ensure the collision rate is as low as possible, the labels of the node itself and the aggregated neighbor nodes in step 3.8 are used to obtain the extended label string. The strings are sorted in ascending order and a one-to-one mapping dictionary is generated using formula (17). Formula (16) is as follows:

[0124]

[0125] in, N i is the neighbor node of i, d i , d j is the degree of nodes i and j. This transfer rule refers to the addition of reference parameter W t After T rounds of iteration, the feature vector is updated to:

[0126]

[0127] Among them, when H T =H T-1 When the maximum number of iterations T is reached, the node label no longer changes. In this example, T=3 is used to avoid redundant calculations.

[0128] Step 4.1: Through the feature expression of the above steps, the node feature vector X obtained in step 2.5 is converted to poly , the aggregated neighbor node feature vector H extracted in step 3.5 (L) , and the topological structure feature vector C in step 3.9 T The channels are connected by formula (18) to obtain the mixed feature layer Z, which is as follows:

[0129] Z=|X poly ‖H (K) ||C T | (18)

[0130] Among them, |*| represents the splicing operation, and the comprehensive feature vector Z is obtained by splicing node features, aggregating neighbor node features and topological structure features.

[0131] Step 4.2: The feature Z fused in step 4.1 is input into the fully connected network for prediction through formula (19), which is as follows:

[0132] Target pre =softmax(ReLU(W (l) h (l) +b (l) )) (19)

[0133] Among them, h (0) =Z,W (l) and b(l) represents the weight and bias of the lth layer, ReLU is the activation function, softmax converts the distribution of the function into probability, and cross entropy is used as the loss function as shown in formula (20):

[0134]

[0135] Among them, t i It is the one-hot encoding of the true label. Finally, the loss function is optimized through back propagation and gradient descent method, the network parameters are updated, and the probability prediction of infrared targets is achieved.

[0136] The infrared small target detection method based on graph structure analysis provided by the present invention has the following significant advantages by converting infrared image data into a graph structure and integrating multi-level feature analysis:

[0137] 1. Improved ability to model complex relationships

[0138] By constructing a graph structure to express the pixel correlation between the target and the background, the nonlinear relationship between the target and the interference signal can be effectively modeled. Graph node features can simultaneously capture local details (such as gradient direction and grayscale differences) and global contextual information (such as topological connectivity and regional consistency). This overcomes the false detection problems caused by traditional methods that rely on linear assumptions or single features, significantly improving the ability to distinguish high-radiation false targets.

[0139] 2. Multi-level feature fusion enhances detection robustness

[0140] The Neighborhood Node Feature Extraction Module is introduced to aggregate the association information between the target and surrounding pixels through multiple rounds of feature propagation, enhancing the differential expression of the spatial distribution and radiation characteristics of the target and interference signals. Combined with the Topology Reasoning Module, it analyzes the connectivity characteristics of the graph structure (such as node degree and neighborhood label propagation) to further explore the essential structural differences between the target and the background, thereby achieving more accurate detection in complex scenarios.

[0141] 3. Model interpretability and flexibility optimization

[0142] Graph-based parameter design has clear physical meaning (e.g., adjacency matrix weights reflect pixel similarity, and topological labels represent structural stability), facilitating analysis of the model's decision-making process and adjustment of key parameters. Furthermore, by adjusting the graph construction strategy (e.g., node definition and connection rules) and feature fusion methods, the system can quickly adapt to different imaging conditions or mission requirements. This system is highly scalable and suitable for a variety of infrared applications, including military reconnaissance and satellite monitoring.

[0143] 4. Reduce dependence on labeled data

[0144] Compared with deep learning methods that rely on large amounts of labeled data, this invention fully utilizes the inherent characteristics of the target and background (such as gradient distribution and structural consistency) for feature learning through graph structure modeling and topological reasoning mechanism, reduces dependence on manual labeling, and can still maintain stable detection performance in data-scarce scenarios.

[0145] 5. Significant effect in suppressing false alarms and misdetections

[0146] Through multi-stage candidate area screening (background separation, connected domain analysis) and joint optimization of graph structure features, interference signals with similar target radiation characteristics (such as ice surface reflection and cloud edge) are effectively filtered out, reducing the false alarm rate and improving the reliability of the early warning system.

[0147] In summary, the present invention is superior to traditional methods in detection accuracy, anti-interference ability, model interpretability and applicability, and provides an efficient and reliable solution for infrared small target detection.

Claims

1. A method for detecting small infrared targets based on graph structure analysis, characterized in that: The steps include: Step 1: Perform a preliminary screening of suspected targets on the input infrared image to obtain a candidate area image; Step 2: Build a graph structure using the candidate area image, including node information, connection links, and system-level topological characteristics; Step 3: Embed the constructed graph structure data into the feature extractor to extract node features, aggregate neighbor node features, and topological structure features, achieving broader, multi-level, and multi-dimensional feature extraction to obtain feature extraction information; Step 4: Input the feature extraction information obtained in step 3 into the classifier, design the loss function, and finally obtain an efficient infrared target detection model through model training.

2. The infrared small target detection method based on graph structure analysis according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Input an infrared image I of size w×h. Based on the background modeling framework, perform a structural adaptive operation on the input image to remove small bright spots to estimate the background. Then, subtract the estimated background from the input image to obtain the confidence separation result I. separation , as shown in (1), formula (1) is as follows: in, It is performed on the input image Adaptive operation of structural elements, is a structure with radius r0, where The calculation formula is shown in formula (2), which is as follows: Step 1.2: Confidence separation result I obtained in step 1.1 separation Then, based on the background perception judgment, θ is used as the threshold to segment the suspicious target to obtain the confidence binary image I bw , the value of θ is determined by bisection, and the termination condition is from I bw The number of connected domains obtained by removing the edge portion of the image is between 1 and K; Step 1.3: Obtain I through step 1.2 bw After that, we get n connected domains Take each connected domain as the center point As the center, in the original image I, the length and width are r c Rectangular area as candidate area r c Calculated by formula (3), formula (3) is as follows: in, and are the width and height of the minimum circumscribed matrix of the connected area, respectively. max(·) means taking the maximum value of (·).

3. The infrared small target detection method based on graph structure analysis according to claim 2, characterized in that: The step 2 comprises the following steps: Node characteristics: Step 2.1: Convert the result in step 1.3 to the candidate area Calculate the gradient I in the x, y direction x , I y ,in, Step 2.2: Use the gradients in the x and y directions obtained in step 2.1 to construct the structure tensor M through formula (4), and then use formula (5) to calculate the response value R of the node to screen the key nodes and retain the points where R>threshold. Formulas (4) and (5) are as follows: R=det(M)-α.trace(M) 2 (5) Where w(x, y) is the weight of the Gaussian window, α is the sensitivity coefficient. The smaller the value, the more nodes are detected. α∈(0.04,0.06), threshold=β.R max ,β∈(0.01,0.05),R max is the maximum value of R; Step 2.3: Calculate the amplitude in the neighborhood of the key node by screening the key nodes in step 2.2 and direction To obtain the gradient direction histogram around the key node, select the peak direction in the histogram as the direction of the key node; Step 2.4: Based on the position and direction of the key node calculated in step 2.3, construct an 8×8 pixel area, divide it into 2×2 small blocks, calculate the gradient histogram in 8 directions, and generate a 2×2×8=32-dimensional key node feature vector X=[x1, x1, ..., x d ]; Step 2.5: The eigenvector obtained in step 2.4 is subjected to explicit nonlinear mapping through formula (6), which is as follows: The expanded dimension is Where d represents the dimension of the key node feature vector.

4. The infrared small target detection method based on graph structure analysis according to claim 3 is characterized in that: The step 3 comprises the following steps: Aggregation neighbor node characteristics: Step 3.1: Each candidate area obtained in step 1.3 As a subgraph, each pixel of the subgraph is regarded as a node, the value of the node is the grayscale value of the pixel, and the adjacent pixels are regarded as an edge, and the subgraph structure G can be obtained. Formula (7) is as follows: G=(N,E) (7) Where N is the set of subgraph nodes, and E is the set of subgraph edges; Step 3.2: Assign values to each adjacency matrix element based on the correlation between pixels. If node i and node j are connected, then A ij =g ij , otherwise A ij =0, each element is calculated by formula (8), which is as follows: Among them, I i is the grayscale value of node i, I j is the grayscale value of node j, σ is the parameter that controls weight attenuation, and the set formula (9) of all elements of the subgraph adjacency matrix A is as follows: Step 3.3: For the adjacency matrix A obtained in step 3.2, calculate the degree D of the local unit by counting the number of edges connected to node i. ii , formula (10) is as follows: D ii =∑ j A ij (10) Step 3.4: In order to enhance the numerical stability, the adjacency matrix is normalized using formula (11), which is as follows: Among them, I N is the identity matrix, It's A+I N degree matrix of ; Step 3.5: For node i, use the normalized adjacency matrix obtained in step 3.4 Aggregate the features of itself and its neighboring nodes through formula (12) and repeat the propagation K times, (13) is as follows: in, X is the node eigenvector matrix H (k) Represents the state of the node feature vector after the kth operation; Topological structure characteristics: Step 3.6: Perform topological structure analysis on the subgraph obtained in step 3.1 and assign an initial label to each node v The node degree used in the initial label is used to obtain the initial feature vector as shown in formula (14). Formula (13) is as follows: Among them, Am is the initial feature dimension. Step 3.7 For each node v, in each iteration of round t, its own label and its neighbors' labels Perform aggregation, formula (14) is as follows: in, is the neighbor set of node v, AGGREGATE is the aggregation function; Step 3.8 combines the node's own label with the aggregated neighbor information from step 3.7 to generate a new label as follows: Step 3.9 uses the labels of the node itself and the aggregated neighbor nodes in step 3.8 to obtain the extended label string, sorts the string in ascending order, and generates a one-to-one mapping dictionary using formula (16). Formula (17) is as follows: in, N i is the neighbor node of i, d i , d j is the degree of nodes i and j. This transfer rule refers to the addition of reference parameter W t After the hash function changes, after T rounds of iteration, the feature vector is updated to: Among them, when H T =H T-1 When the preset maximum number of iterations T is reached, the node label no longer changes.

5. The infrared small target detection method based on graph structure analysis according to claim 4 is characterized in that: The step 4 comprises the following steps: Step 4.1: Through the feature expression of the above steps, the node feature vector X obtained in step 2.5 is converted to poly , the aggregated neighbor node feature vector H extracted in step 3.5 (L) , and the topological structure feature vector C in step 3.9 T Channel connection is performed through formula (19) to obtain the mixed feature layer Z, formula (18) is as follows: Z=|X poly ||H (K) ||C T | (18) Among them, |*| represents the splicing operation, which obtains the comprehensive feature vector Z by splicing node features, aggregating neighbor node features and topological structure features; Step 4.2: The feature Z fused in step 4.1 is input into the fully connected network for prediction through formula (19), which is as follows: <h2 style=";text-align:left;direction:ltr">Targ et<h2 style=";text-align:left;direction:ltr"> pre <h2 style=";text-align:left;direction:ltr"> =softmax(Re LU(W<h2 style=";text-align:left;direction:ltr"> (l) <h2 style=";text-align:left;direction:ltr"> h<h2 style=";text-align:left;direction:ltr"> (l) <h2 style=";text-align:left;direction:ltr"> +b<h2 style=";text-align:left;direction:ltr"> (l) <h2 style=";text-align:left;direction:ltr"> )) (19) Among them, h (0) =Z,W (l) and b (l) Represents the weight and bias of the lth layer, ReLU is the activation function, softmax converts the distribution of the function into probability, and cross entropy is used as the loss function as shown in formula (20): Among them, t i It is the one-hot encoding of the true label. Finally, the loss function is optimized through back propagation and gradient descent method, the network parameters are updated, and the probability prediction of infrared targets is achieved.

Images