Multi-scale fractal deep network infrared false alarm source classification method and system

Abstract

The invention discloses a multi-scale fractal deep network infrared false alarm source classification method and system, belongs to the technical field of remote sensing image processing, and solves the problems that in the prior art, false alarm sources of different scenes and different features are difficult to classify by utilizing texture information with effective generality, a set of general algorithm is difficult to design to solve multi-category false alarm source classification, and the classification efficiency is low. And a classification method is poor in adaptability in a complex background environment and relatively high in training data volume requirement. The method comprises the following steps: reading an infrared remote sensing image, extracting edge contour features of the infrared remote sensing image by using an anisotropic differential diffusion filter, and stacking the edge contour features and an original infrared remote sensing image in an image channel number dimension to be input into a Swin Transform network; and depth features output by each Stage layer in the Swin Transformer network are input into the false alarm source classification network to obtain different false alarm sources, the different false alarm sources are classified and labeled by adopting different identification colors, and the result is labeled on the infrared remote sensing image, and the method is used for multi-scale fractal depth network infrared false alarm source classification.

Description

Technical Field

[0001] A multi-scale fractal depth network infrared false alarm source classification method and system is proposed, which is used for multi-scale fractal depth network infrared false alarm source classification and belongs to the field of remote sensing image processing technology. Background Technology

[0002] Existing Earth observation false alarm source detection methods are mostly based on the detection of single categories or scenarios where clouds and snow coexist, such as cirrus clouds, snow-capped mountains, and forest fires. Furthermore, due to difficulties in data acquisition, research on false alarm source detection is relatively limited. With the increasing complexity of target detection scenarios, the detection and identification of false alarm sources in scenarios where multiple types of false alarm sources coexist are becoming increasingly important. To more accurately detect and eliminate multiple different types of false alarm sources, it is necessary to analyze the different characteristics of different false alarm sources, thereby designing a more universal false alarm source detection scheme.

[0003] Current methods for detecting heterogeneous false alarm sources rely on thresholding, texture features, and supervised classification to process edge information and separate the false alarm source from the background. Therefore, existing technologies suffer from the following technical problems:

[0004] 1. The threshold method has limited effectiveness in detecting and classifying false alarm sources because the gray value distribution of false alarm sources varies in different scenarios and features.

[0005] 2. Texture features contain a lot of information, and it is difficult to select common and effective texture features for different false alarm sources, which can easily lead to poor adaptability when the background is complex;

[0006] 3. Supervised learning requires a large amount of labeled data for training, and there is a risk of overfitting when there is insufficient training data. Summary of the Invention

[0007] To address the problems mentioned above, the present invention aims to provide a multi-scale fractal depth network infrared false alarm source classification method and system, which solves the problems in the existing technology that make it difficult to classify false alarm sources with common and effective texture information for different scenes and characteristics, make it difficult to design a general algorithm to solve the classification of multiple categories of false alarm sources, have poor adaptability to complex background environments, and require a large amount of training data.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multi-scale fractal depth network infrared false alarm source classification method includes the following steps:

[0010] Step 1: Read the infrared remote sensing image, extract the edge contour features of the infrared remote sensing image using an anisotropic differential diffusion filter, and stack the infrared remote sensing image with the original infrared remote sensing image in the number of image channels and input it into the Swing Transformer network.

[0011] Step 2: Input the depth features output from each Stage layer in the Swing Transformer network into the false alarm source classification network to obtain different false alarm sources. Classify and label them with different color labels, and then label the results on the infrared remote sensing image.

[0012] The false alarm source classification network includes convolutional layers that sequentially convolve the depth features output by each Stage layer in the Swing Transformer network, and upsampling layers that upsample the output of the convolutional layers to a uniform size.

[0013] The results obtained from all upsampling layers are merged by multi-scale feature arrangement, and the results obtained from the merged layer are then processed by local difference box counting through a sliding window to obtain the response layer of the fractal texture response map.

[0014] The concat layer concatenates the output of the upsampling layer with the output of the response layer; the Conv layer processes the output of the concat layer using a 1×1 convolution kernel; the fully connected layer processes the output of the Conv layer; and the random forest classifier classifies the output of the fully connected layer for different false alarm sources.

[0015] Furthermore, the specific steps of step 1 are as follows:

[0016] Step 1.1: Based on the characteristics of infrared remote sensing images, design a direction-weighted weighting function ω(θ) and a diffusion function D(x,y), with the following formula:

[0017]

[0018] Where x and y represent the position coordinates of pixels in the infrared remote sensing image in the horizontal and vertical directions, respectively. The image gradient at pixel (x,y) represents the rate of change of intensity of the infrared remote sensing image at that point. α is a parameter that adjusts the gradient magnitude, β is a parameter that adjusts the gradient direction, θ represents the diffusion direction angle to control the directionality of the weighting function, and K is a parameter that controls the diffusion intensity.

[0019] The formula for obtaining the final convolutional diffusion kernel based on ω(θ) and D(x,y) is:

[0020] H(x,y)=ω(θ)·D(x,y)

[0021] Where H(x,y) is the convolution kernel function of the final convolution diffusion kernel;

[0022] Step 1.2: Perform P iterations to update the infrared remote sensing image based on the final convolutional diffusion kernel, obtaining the updated grayscale values ​​of each pixel in the infrared remote sensing image, which in turn yields the edge contour features of the infrared remote sensing image after the grayscale values ​​are updated. Each update process is as follows:

[0023]

[0024] Where I(x,y) represents the pixel grayscale value at (x,y), and P is a parameter controlling the number of iterations. In subsequent updates, I(x,y) is used to calculate the pixel grayscale value from the previous update. new (x, y) represents I(x, y), N(x, y) represents the neighborhood of the current pixel (x, y), Δt is the time step, and (x′, y′) represents the coordinates of a pixel within the neighborhood of the current pixel (x, y). new (x,y) represents the gray value of pixel (x,y) after the final convolution diffusion kernel operation, H(x′,y′) is the convolution kernel function of the (x′,y′) points in the neighborhood of the current pixel (x,y), and I(x′,y′) is the gray value of the pixel at the (x′,y′) point in the neighborhood of the current pixel (x,y).

[0025] Step 1.3: Stack the edge contour features of the updated grayscale infrared remote sensing image with the original infrared remote sensing image in terms of the number of image channels. That is, increase the number of channels in the channel dimension of the updated grayscale infrared remote sensing image so that the dimension of the updated grayscale infrared remote sensing image reaches H×W×3, where H is the height of the updated grayscale infrared remote sensing image and W is the width of the updated grayscale infrared remote sensing image. Then input it into the Swing Transformer network.

[0026] Furthermore, the specific implementation of the response layer in step 2, which involves merging the results obtained from all upsampling layers using a multi-scale feature arrangement, and then performing local difference box counting operations on the results obtained from the merged layer through a sliding window to obtain the fractal texture response map, is as follows:

[0027] Step 2.1: Record the feature block T output by the upsampling layer in Step 2. i And based on feature block T i We obtain set T:

[0028] T = [T1; T2; T3; T4]

[0029] T i =[T i1 ;T i2 ;......;T ij ;Tiz-1 ;T iz (i = 1, 2, 3, 4)

[0030] Where T represents the set of feature blocks output from the original order of the Swing Transformer network output through the upsampling layer, T i Let T be the set of feature maps contained in the i-th feature block. ij Let Z represent the j-th feature map in the i-th feature block, and Z be the number of feature maps in each feature block.

[0031] Then, the set T is reordered and merged according to the following formula to finally obtain the feature block U. j And based on feature block U j We obtain set U:

[0032] U j =[T 1j ;T 2j ;T 3j ;T 4j (j = 1, 2, ..., z-1, z)

[0033] U = [U1; U2; ...; U j ;U z-1 ;U z ]

[0034] Where U represents the set of feature blocks after reordering and merging, U j Let j be the set of feature maps in the j-th feature block;

[0035] Step 2.2: Calculate the feature blocks U obtained after reordering and merging in Step 2.1 using a sliding window. j N can then be obtained. s The mapping relationship with s is calculated using the following formula:

[0036] n s (i, j) = [g max (i, j) / h s ]-[g min (i, j) / h s +1

[0037]

[0038] Among them, g max (i, j) represent each feature block U j Maximum internal grayscale value, g min (i, j) represent each feature block U j Minimum internal grayscale value, s is the box size, n s (i,j) represents the coverage of each feature block U. jMinimum number of boxes required for a grayscale surface, N s To cover feature block U j Minimum number of boxes required for surface grayscale values, h s The height of the box;

[0039] Step 2.3, based on N in Step 2.2 s The relationship with s is used to calculate different N values ​​by selecting different s values. s The value is used to fit a straight line log N using the least squares method. s The image slope k is calculated from the -log l / s image. The value of k is used as the value of the center point of the sliding window, and the feature blocks U in the set U are... j After sliding the window, the fractal texture response map is obtained, where the fractal texture response maps obtained at the same k value point are labeled as the same category.

[0040] A multi-scale fractal depth network infrared false alarm source classification system includes:

[0041] The deep feature extraction module reads infrared remote sensing images, extracts edge contour features of the infrared remote sensing images using an anisotropic differential diffusion filter, and stacks them with the original infrared remote sensing images in the number of image channels before inputting them into the SwinTransformer network.

[0042] Labeling module: Input the depth features output from each stage layer of the Swing Transformer network into the false alarm source classification network to obtain different false alarm sources, classify and label them with different color labels, and label the results on the infrared remote sensing image;

[0043] The false alarm source classification network includes convolutional layers that sequentially convolve the depth features output by each Stage layer in the Swing Transformer network, and upsampling layers that upsample the output of the convolutional layers to a uniform size.

[0044] The results obtained from all upsampling layers are merged by multi-scale feature arrangement, and the results obtained from the merged layer are then processed by local difference box counting through a sliding window to obtain the response layer of the fractal texture response map.

[0045] The concat layer concatenates the output of the upsampling layer with the output of the response layer; the Conv layer processes the output of the concat layer using a 1×1 convolution kernel; the fully connected layer processes the output of the Conv layer; and the random forest classifier classifies the output of the fully connected layer for different false alarm sources.

[0046] Furthermore, the specific implementation steps of the deep feature extraction module are as follows:

[0047] Step 1.1: Based on the characteristics of infrared remote sensing images, design a direction-weighted weighting function ω(θ) and a diffusion function D(x,y), with the following formula:

[0048]

[0049] Where x and y represent the position coordinates of pixels in the infrared remote sensing image in the horizontal and vertical directions, respectively. The image gradient at pixel (x,y) represents the rate of change of intensity of the infrared remote sensing image at that point. α is a parameter that adjusts the gradient magnitude, β is a parameter that adjusts the gradient direction, θ represents the diffusion direction angle to control the directionality of the weighting function, and K is a parameter that controls the diffusion intensity.

[0050] The formula for obtaining the final convolutional diffusion kernel based on ω(θ) and D(x,y) is:

[0051] H(x,y)=ω(θ)·D(x,y)

[0052] Where H(x, y) is the convolution kernel function of the final convolution diffusion kernel;

[0053] Step 1.2: Perform P iterations to update the infrared remote sensing image based on the final convolutional diffusion kernel, obtaining the updated grayscale values ​​of each pixel in the infrared remote sensing image, which in turn yields the edge contour features of the infrared remote sensing image after the grayscale values ​​are updated. Each update process is as follows:

[0054]

[0055] Where I(x, y) represents the pixel grayscale value at (x, y), and P is a parameter controlling the number of iterations. In subsequent updates, I(x, y) is used to calculate the pixel grayscale value from the previous update. new (x, y) represents I(x, y), N(x, y) represents the neighborhood of the current pixel (x, y), Δt is the time step, and (x′, y′) represents the coordinates of a pixel within the neighborhood of the current pixel (x, y). new (x, y) represents the gray value of pixel (x, y) after the final convolution diffusion kernel operation, H(x′, y′) is the convolution kernel function of the (x′, y′) points in the neighborhood of the current pixel (x, y), and I(x′, y′) is the gray value of the pixel at the (x′, y′) point in the neighborhood of the current pixel (x, y).

[0056] Step 1.3: Stack the edge contour features of the updated grayscale infrared remote sensing image with the original infrared remote sensing image in terms of the number of image channels. That is, increase the number of channels in the channel dimension of the updated grayscale infrared remote sensing image so that the dimension of the updated grayscale infrared remote sensing image reaches H×W×3, where H is the height of the updated grayscale infrared remote sensing image and w is the width of the updated grayscale infrared remote sensing image. Then input it into the Swing Transformer network.

[0057] Furthermore, the specific implementation of the response layer in the annotation module, which performs multi-scale feature arrangement and merging on the results obtained from all upsampling layers, and then performs local difference box counting operations on the results obtained from the merged layer through a sliding window to obtain the fractal texture response map, is as follows:

[0058] Step 2.1: Record the feature block T output by the upsampling layer in Step 2. i And based on feature block T i We obtain set T:

[0059] T = [T1; T2; T3; T4]

[0060] T i =[T i1 ;T i2 ;......;T ij ;T iz-1 Ti z (i = 1, 2, 3, 4)

[0061] Where T represents the set of feature blocks output from the original order of the Swing Transformer network output through the upsampling layer, T i Let T be the set of feature maps contained in the i-th feature block. ij Let Z represent the j-th feature map in the i-th feature block, and Z be the number of feature maps in each feature block.

[0062] Then, the set T is reordered and merged according to the following formula to finally obtain the feature block U. j And based on feature block U j We obtain set U:

[0063] U j =[T 1j ;T 2j ;T 3j ;T 4j (j = 1, 2, ..., z-1, z)

[0064] U = [U1; U2; ...; U j ;U z-1 ;U z ]

[0065] Where U represents the set of feature blocks after reordering and merging, U j Let j be the set of feature maps in the j-th feature block;

[0066] Step 2.2: Calculate the feature blocks U obtained after reordering and merging in Step 2.1 using a sliding window. j N can then be obtained. s The mapping relationship with s is calculated using the following formula:

[0067] n s (i, j) = [g max (i, j) / h s ]-[g min (i, j) / h s +1

[0068]

[0069] Among them, g max (i, j) represent each feature block U j Maximum internal grayscale value, g min (i,j) represents each feature block U j Minimum internal grayscale value, s is the box size, n s (i,j) represents the coverage of each feature block U. j Minimum number of boxes required for a grayscale surface, N s To cover feature block U j Minimum number of boxes required for surface grayscale values, h s The height of the box;

[0070] Step 2.3, based on N in Step 2.2 s The relationship with s is used to calculate different N values ​​by selecting different s values. s The value is used to fit a straight line log N using the least squares method. s The image slope k is calculated from the -log l / s image. The value of k is used as the value of the center point of the sliding window, and the feature blocks U in the set U are... j After sliding the window, the fractal texture response map is obtained, where the fractal texture response maps obtained at the same k value point are labeled as the same category.

[0071] Compared with the prior art, the beneficial effects of this invention are as follows:

[0072] This invention extracts edge contour features from infrared remote sensing images using an anisotropic differential diffusion filter. These features are then stacked with the original infrared remote sensing image across multiple image channels and uniformly input into a Swing Transformer network. Layer-by-layer feature maps are extracted and standardized in size. After feature reordering and merging, a fractal texture response map is obtained through differential box counting. This map is then concatenated with the output of the upsampling layer and fused via convolution. Finally, a fully connected layer and a random forest classifier are used to obtain the final image. Different categories of false alarm sources are labeled with different colors on the original infrared remote sensing image. The specific effects are as follows:

[0073] I. This invention performs anisotropic differential diffusion filtering on infrared remote sensing images and stacks them with the original infrared remote sensing images. This can effectively solve the problem of texture noise interference in remote sensing images, filter out useless texture information, and thus specifically enhance the fractal texture features of infrared remote sensing images.

[0074] Second, this invention uses a false alarm source classification network to reorder and merge the depth features output by each Stage layer of the Swin Transformer network at multiple scales to calculate the fractal texture response. By improving the utilization efficiency of fractal features, the requirement for infrared remote sensing image data is reduced, and the overfitting of the network can be effectively mitigated. It makes good use of the different fractal dimensions of different false alarm sources and selects effective texture features to achieve efficient false alarm source classification. This solves the problem that the existing threshold method is difficult to adaptively adjust the gray value discrimination threshold due to the different gray value distribution of false alarm sources with different features in different scenes, which limits the detection and classification effect of false alarm sources.

[0075] Third, this invention re-integrates the output of the upsampling layer with the fractal texture response, which can effectively utilize the nonlinear fitting characteristics of the Swin Transformer network and the mathematical interpretability of fractal geometry at the same time. It also solves the problems of poor interpretability of the Swin Transformer network and low computational efficiency and high computational complexity of traditional fractal geometry applied to the image field. In other words, it solves the problem that existing technologies are difficult to select common and effective texture features for different false alarm sources, which easily leads to poor adaptability when the background is complex. Attached Figure Description

[0076] Figure 1 This is a flowchart of the present invention;

[0077] Figure 2 This is a schematic diagram of the architecture for the false alarm source classification network to process the feature blocks output by each Stage layer in the Swing Transformer network.

[0078] Figure 3 This is the infrared remote sensing image to be detected in this invention;

[0079] Figure 4 This invention is aimed at Figure 3 Edge contour features of infrared remote sensing images extracted using an anisotropic differential diffusion filter;

[0080] Figure 5 This invention is aimed at Figure 3 Fractal texture response map after differential box operation;

[0081] Figure 6 This invention is aimed at Figure 3 The labeled result image. Detailed Implementation

[0082] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0083] A multi-scale fractal depth network infrared false alarm source classification method includes the following steps:

[0084] Step 1: Read the infrared remote sensing image, such as... Figure 3 As shown, the edge contour features of the infrared remote sensing image are extracted using an anisotropic differential diffusion filter and then stacked with the original infrared remote sensing image in the number of image channels before being input into the SwingTransformer network.

[0085] The specific steps are as follows:

[0086] Step 1.1: Based on the characteristics of infrared remote sensing images, design a direction-weighted weighting function ω(θ) and a diffusion function D(x,y), with the following formula:

[0087]

[0088] Where x and y represent the position coordinates of pixels in the infrared remote sensing image in the horizontal and vertical directions, respectively. The image gradient at pixel (x,y) represents the rate of change of intensity of the infrared remote sensing image at that point. α is a parameter that adjusts the gradient magnitude, β is a parameter that adjusts the gradient direction, θ represents the diffusion direction angle to control the directionality of the weighting function, and K is a parameter that controls the diffusion intensity.

[0089] The formula for obtaining the final convolutional diffusion kernel based on ω(θ) and D(x,y) is:

[0090] H(x,y)=ω(θ)·D(x,y)

[0091] Where H(x,y) is the convolution kernel function of the final convolution diffusion kernel;

[0092] Step 1.2: Perform P iterations to update the infrared remote sensing image based on the final convolutional diffusion kernel, obtaining the updated grayscale values ​​of each pixel in the infrared remote sensing image, which in turn yields the edge contour features of the infrared remote sensing image after the grayscale values ​​are updated. Each update process is as follows:

[0093]

[0094] Where I(x, y) represents the pixel grayscale value at (x, y), and P is a parameter controlling the number of iterations. In subsequent updates, I(x, y) is used as the value obtained from the previous update. new (x, y) represents I(x, y), N(x, y) represents the neighborhood of the current pixel (x, y), Δt is the time step, and (x′, y′) represents the coordinates of a pixel within the neighborhood of the current pixel (x, y). neW (x,y) represents the gray value of pixel (x,y) after the final convolution diffusion kernel operation, H(x′,y′) is the convolution kernel function of the (x′,y′) points in the neighborhood of the current pixel (x,y), and I(x′,y′) is the gray value of the pixel at the (x′,y′) point in the neighborhood of the current pixel (x,y);

[0095] Step 1.3: Stack the edge contour features of the updated grayscale infrared remote sensing image with the original infrared remote sensing image in terms of the number of image channels. That is, increase the number of channels in the channel dimension of the updated grayscale infrared remote sensing image so that the dimension of the updated grayscale infrared remote sensing image reaches H×W×3, where H is the height of the updated grayscale infrared remote sensing image and W is the width of the updated grayscale infrared remote sensing image. Then input it into the Swing Transformer network.

[0096] Step 2: Input the depth features output from each stage layer of the Swing Transformer network into the false alarm source classification network. Different false alarm sources are classified and labeled using different color identifiers, and the results are labeled on the infrared remote sensing image, such as... Figure 6 As shown;

[0097] The false alarm source classification network includes convolutional layers that sequentially convolve the depth features output from each Stage layer in the Swing Transformer network, and upsampling layers that upsample the output of the convolutional layers to a uniform size. The Swing Transformer network has four sequentially connected Stage layers, and the depth features obtained after the operation of each Stage layer are denoted as M. i ×N i ×Z i (i = 1, 2, 3, 4), where M i Let N be the height of the i-th feature map. iZ is the width of the i-th feature map. i Let M be the number of channels in the i-th feature map. After convolution and upsampling operations, four sets of M×N×Z feature maps are obtained, where M = maxM. i N = maxN i Z represents the preset number of channels.

[0098] The results from all upsampling layers are processed through a multi-scale feature arrangement and merging layer. Then, the results from the merged layer are processed through a sliding window to perform local difference box counting operations to obtain the response layer of the fractal texture response map. The specific implementation is as follows:

[0099] Step 2.1: Record the feature block T output by the upsampling layer in Step 2. i And based on feature block T i We obtain set T:

[0100] T = [T1; T2; T3; T4]

[0101] T i =[T i1 ;T i2 ;......;T ij ;T iz-1 ;T iz (i = 1, 2, 3, 4)

[0102] Where T represents the set of feature blocks output from the original order of the Swing Transformer network output through the upsampling layer, T i Let T be the set of feature maps contained in the i-th feature block. ij Z represents the j-th feature map in the i-th feature block, and Z is the number of feature maps in each feature block.

[0103] Then, the set T is reordered and merged according to the following formula to finally obtain the feature block U. j And based on feature block U j We obtain set U:

[0104] U j =[T 1j ;T 2j ;T 3j ;T 4j (j = 1, 2, ..., z-1, z)

[0105] U = [U1; U2; ...; U j ;U z-1 ;U z ]

[0106] Where U represents the set of feature blocks after reordering and merging, U j Let j be the set of feature maps in the j-th feature block;

[0107] If there are 4 original feature blocks before sorting and merging (a fixed value), and each feature block has 6 feature maps, then the i-th feature block can be represented as T. i1 T i2 ......T i6 After feature sorting and merging, the first feature map T of each feature block is... i1 Take out (4 images in total) and reassemble them into a feature block U1 = [T 11 T 21 T 31 T 41 This process ultimately results in six new feature blocks, each containing four feature maps.

[0108] Step 2.2: Calculate the feature blocks U obtained after reordering and merging in Step 2.1 using a sliding window. j N can then be obtained. s The mapping relationship with s is calculated using the following formula:

[0109] n s (i, j) = [g max (i, j) / h s ]-[g min (i, j) / h s +1

[0110]

[0111] Among them, g max (i,j) represents each feature block U j Maximum internal grayscale value, g min (i, j) represent each feature block U j Minimum internal grayscale value, s is the box size, n s (i, j) represents the coverage of each feature block U. j Minimum number of boxes required for a grayscale surface, N s To cover feature block U j Minimum number of boxes required for surface grayscale values, h s The height of the box;

[0112] Step 2.3, based on N in Step 2.2 s The relationship with s is used to calculate different N values ​​by selecting different s values. s The value is used to fit a straight line log N using the least squares method. s The image slope k is calculated from the -log l / s image. The value of k is used as the value of the center point of the sliding window, and the feature blocks U in the set U are... j After sliding the window, the fractal texture response map is obtained, where the fractal texture response maps obtained at the same k value point are labeled as the same category.

[0113] The concat layer concatenates the output of the upsampling layer with the output of the response layer. The conv layer processes the output of the concat layer with a 1×1 convolution kernel to obtain a feature map with size M×N×(Z+1). The fully connected layer processes the output of the conv layer. The random forest classifier classifies the output of the fully connected layer for different false alarm sources.

[0114] The above are merely representative embodiments among the many specific applications of this invention, and do not constitute any limitation on the scope of protection of this invention. All technical solutions formed by transformation or equivalent substitution fall within the scope of protection of this invention.

Claims

1. A multi-scale fractal depth network infrared false alarm source classification method, characterized in that, Includes the following steps: Step 1: Read the infrared remote sensing image, extract the edge contour features of the infrared remote sensing image using an anisotropic differential diffusion filter, and stack the infrared remote sensing image with the original infrared remote sensing image in the number of image channels and input it into the Swing Transformer network. Step 2: Input the depth features output from each Stage layer in the Swing Transformer network into the false alarm source classification network to obtain different false alarm sources. Classify and label them with different color labels, and then label the results on the infrared remote sensing image. The false alarm source classification network includes convolutional layers that sequentially convolve the depth features output by each Stage layer in the Swing Transformer network, and upsampling layers that upsample the output of the convolutional layers to a uniform size. The results obtained from all upsampling layers are merged by multi-scale feature arrangement, and the results obtained from the merged layer are then processed by local difference box counting through a sliding window to obtain the response layer of the fractal texture response map. The concat layer concatenates the output of the upsampling layer with the output of the response layer; the Conv layer processes the output of the concat layer using a 1×1 convolution kernel; the fully connected layer processes the output of the Conv layer; and the random forest classifier classifies the output of the fully connected layer for different false alarm sources.

2. The multi-scale fractal depth network infrared false alarm source classification method according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 1.1: Based on the characteristics of infrared remote sensing images, design a direction-weighted weighting function ω(θ) and a diffusion function D(x,y), with the following formula: Where x and y represent the position coordinates of pixels in the infrared remote sensing image in the horizontal and vertical directions, respectively. The image gradient at pixel (x,y) represents the rate of change of intensity of the infrared remote sensing image at that point. α is a parameter that adjusts the gradient magnitude, β is a parameter that adjusts the gradient direction, θ represents the diffusion direction angle to control the directionality of the weighting function, and K is a parameter that controls the diffusion intensity. The formula for obtaining the final convolutional diffusion kernel based on ω(θ) and D(x,y) is: H(x, y) = ω(θ)·D(x, y) Where H(x,y) is the convolution kernel function of the final convolution diffusion kernel; Step 1.2: Perform P iterations to update the infrared remote sensing image based on the final convolutional diffusion kernel, obtaining the updated grayscale values ​​of each pixel in the infrared remote sensing image, which in turn yields the edge contour features of the infrared remote sensing image after the grayscale values ​​are updated. Each update process is as follows: Where I(x, y) represents the pixel grayscale value at (x, y), and P is a parameter controlling the number of iterations. In subsequent updates, I(x, y) is used to calculate the pixel grayscale value from the previous update. new (x, y) represents I(x, y), N(x, y) represents the neighborhood of the current pixel (x, y), Δt is the time step, and (x′, y′) represents the coordinates of a pixel within the neighborhood of the current pixel (x, y). new (x, y) represents the gray value of pixel (x, y) after the final convolution diffusion kernel operation, H(x′, y′) is the convolution kernel function of the (x′, y′) points in the neighborhood of the current pixel (x, y), and I(x′, y′) is the gray value of the pixel at the (x′, y′) point in the neighborhood of the current pixel (x, y). Step 1.3: Stack the edge contour features of the updated grayscale infrared remote sensing image with the original infrared remote sensing image in terms of the number of image channels. That is, increase the number of channels in the channel dimension of the updated grayscale infrared remote sensing image so that the dimension of the updated grayscale infrared remote sensing image reaches H×W×3, where H is the height of the updated grayscale infrared remote sensing image and W is the width of the updated grayscale infrared remote sensing image. Then input it into the Swing Transformer network.

3. The multi-scale fractal depth network infrared false alarm source classification method according to claim 2, characterized in that, The specific implementation of the response layer in step 2, which involves merging the results obtained from all upsampling layers using a multi-scale feature arrangement, and then performing local difference box counting operations on the results obtained from the merged layer through a sliding window to obtain the fractal texture response map, is as follows: Step 2.1: Record the feature block T output by the upsampling layer in Step 2. i And based on feature block T i We obtain set T: T = [T1; T2; T3; T4] T i =[T i1 ;T i2 ;......;T ij ;T iz-1 ;T iz ] (i=1,2,3,4) Where T represents the set of feature blocks output from the original order of the Swing Transformer network output through the upsampling layer, T i Let T be the set of feature maps contained in the i-th feature block. ij Z represents the j-th feature map in the i-th feature block, and Z is the number of feature maps in each feature block. Then, the set T is reordered and merged according to the following formula to finally obtain the feature block U. j And based on feature block U j We obtain set U: U j =[T 1j ;T 2j ;T 3j ;T 4j ] (j=1,2,…,z-1,z) U=[U1;U2;……;U j ;U z-1 ;U z ] Where U represents the set of feature blocks after reordering and merging, U j Let j be the set of feature maps in the j-th feature block; Step 2.2: Calculate the feature blocks U obtained after reordering and merging in Step 2.1 using a sliding window. j N can then be obtained. s The mapping relationship with s is calculated using the following formula: n s (i,j)=[g max (i,j) / h s ]-[g min (i,j) / h s ]+1 Among them, g max (i, j) represent each feature block U j Maximum internal grayscale value, g min (i, j) represent each feature block U j Minimum internal grayscale value, s is the box size, n s (i, j) represents the coverage of each feature block U. j Minimum number of boxes required for a grayscale surface, N s To cover feature block U j Minimum number of boxes required for surface grayscale values, h s The height of the box; Step 2.3, based on N in Step 2.2 s The relationship with s is used to calculate different N values ​​by selecting different s values. s The value is used to fit a straight line logN using the least squares method. s From the -logl / s image, the slope k is calculated, and the value of k is used as the value of the center point of the sliding window for each feature block U in the set U. j After sliding the window, the fractal texture response map is obtained, where the fractal texture response maps obtained at the same k value point are labeled as the same category.

4. A multi-scale fractal depth network infrared false alarm source classification system, characterized in that, include: Deep feature extraction module: Reads infrared remote sensing images, extracts edge contour features of infrared remote sensing images using an anisotropic differential diffusion filter, and stacks them with the original infrared remote sensing images in the number of image channels before inputting them into the Swing Transformer network. Labeling module: Input the depth features output from each stage layer of the Swing Transformer network into the false alarm source classification network to obtain different false alarm sources, classify and label them with different color labels, and label the results on the infrared remote sensing image; The false alarm source classification network includes convolutional layers that sequentially convolve the depth features output by each Stage layer in the Swing Transformer network, and upsampling layers that upsample the output of the convolutional layers to a uniform size. The results obtained from all upsampling layers are merged by multi-scale feature arrangement, and the results obtained from the merged layer are then processed by local difference box counting through a sliding window to obtain the response layer of the fractal texture response map. The concat layer concatenates the output of the upsampling layer with the output of the response layer; the Conv layer processes the output of the concat layer using a 1×1 convolution kernel; the fully connected layer processes the output of the Conv layer; and the random forest classifier classifies the output of the fully connected layer for different false alarm sources.

5. The multi-scale fractal depth network infrared false alarm source classification system according to claim 4, characterized in that, The specific implementation steps of the deep feature extraction module are as follows: Step 1.1: Based on the characteristics of infrared remote sensing images, design a direction-weighted weighting function ω(θ) and a diffusion function D(x,y), with the following formula: Where x and y represent the position coordinates of pixels in the infrared remote sensing image in the horizontal and vertical directions, respectively. The image gradient at pixel (x, y) represents the rate of change of intensity of the infrared remote sensing image at that point. α is a parameter that adjusts the gradient magnitude, β is a parameter that adjusts the gradient direction, θ represents the diffusion direction angle to control the directionality of the weighting function, and K is a parameter that controls the diffusion intensity. The formula for obtaining the final convolutional diffusion kernel based on ω(θ) and D(x,y) is: H(x, y) = ω(θ)·D(x, y) Where H(x, y) is the convolution kernel function of the final convolution diffusion kernel; Step 1.2: Perform P iterations to update the infrared remote sensing image based on the final convolutional diffusion kernel, obtaining the updated grayscale values ​​of each pixel in the infrared remote sensing image, which in turn yields the edge contour features of the infrared remote sensing image after the grayscale values ​​are updated. Each update process is as follows: Where I(x, y) represents the pixel grayscale value at (x, y), and P is a parameter controlling the number of iterations. In subsequent updates, I(x, y) is used to calculate the pixel grayscale value from the previous update. new (x, y) represents I(x, y), N(x, y) represents the neighborhood of the current pixel (x, y), Δt is the time step, and (x′, y′) represents the coordinates of a pixel within the neighborhood of the current pixel (x, y). new (x, y) represents the gray value of pixel (x, y) after the final convolution diffusion kernel operation, H(x′, y′) is the convolution kernel function of the (x′, y′) points in the neighborhood of the current pixel (x, y), and I(x′, y′) is the gray value of the pixel at the (x′, y′) point in the neighborhood of the current pixel (x, y). Step 1.3: Stack the edge contour features of the updated grayscale infrared remote sensing image with the original infrared remote sensing image in terms of the number of image channels. That is, increase the number of channels in the channel dimension of the updated grayscale infrared remote sensing image so that the dimension of the updated grayscale infrared remote sensing image reaches H×W×3, where H is the height of the updated grayscale infrared remote sensing image and W is the width of the updated grayscale infrared remote sensing image. Then input it into the Swing Transformer network.

6. The multi-scale fractal depth network infrared false alarm source classification system according to claim 5, characterized in that, The specific implementation of the response layer in the annotation module, which performs multi-scale feature arrangement and merging on the results obtained from all upsampling layers, and then performs local difference box counting operations on the results obtained from the merged layer through a sliding window to obtain the fractal texture response map, is as follows: Step 2.1: Record the feature block T output by the upsampling layer in Step 2. i And based on feature block T i We obtain set T: T = [T1; T2; T3; T4] T i =[T i1 ;T i2 ;......;T ij ;T iz-1 ;T iz ] (i=1,2,3,4) Where T represents the set of feature blocks output from the original order of the Swing Transformer network output through the upsampling layer, T i Let T be the set of feature maps contained in the i-th feature block. ij Let Z represent the j-th feature map in the i-th feature block, and Z be the number of feature maps in each feature block. Then, the set T is reordered and merged according to the following formula to finally obtain the feature block U. j And based on feature block U j We obtain set U: U j =[T 1j ;T 2j ;T 3j ;T 4j ] (j=1,2,…,z-1,z) U=[U1;U2;……;U j ;U z-1 ;U z ] Where U represents the set of feature blocks after reordering and merging, U j Let j be the set of feature maps in the j-th feature block; Step 2.2: Calculate the feature blocks U obtained after reordering and merging in Step 2.1 using a sliding window. j N can then be obtained. s The mapping relationship with s is calculated using the following formula: n s (i,j)=[g max (i,j) / h s ]-[g min (i,j) / h s ]+1 Among them, g max (i, j) represent each feature block U j Maximum internal grayscale value, g min (i, j) represent each feature block U j Minimum internal grayscale value, s is the box size, n s (i, j) represents the coverage of each feature block U. j Minimum number of boxes required for a grayscale surface, N s To cover feature block U j Minimum number of boxes required for surface grayscale values, h s The height of the box; Step 2.3, based on N in Step 2.2 s The relationship with s is used to calculate different N values ​​by selecting different s values. s The value is used to fit a straight line logN using the least squares method. s The image slope k is calculated from the -log1 / s image. The value of k is used as the value of the center point of the sliding window, and the feature blocks U in the set U are... j After sliding the window, the fractal texture response map is obtained, where the fractal texture response maps obtained at the same k value point are labeled as the same category.

Images