Infrared small target segmentation method and device based on double dynamic frequency domain filtering

Abstract

The present application relates to an infrared small target segmentation method and device based on double dynamic frequency domain filtering. The segmentation device comprises: an image acquisition unit for acquiring an infrared image to be segmented; a model processing unit with a pre-trained neural network segmentation model, the neural network segmentation model adopts an encoder-decoder structure, and a double dynamic frequency domain filtering module is embedded in at least one scale layer of the encoding path and the decoding path of the neural network segmentation model; the double dynamic frequency domain filtering module is used to perform the frequency domain filtering step in the infrared small target segmentation method; and a result output unit is used to obtain the segmentation result of the infrared small target according to the output of the neural network segmentation model. The technical problem solved by the present application is to provide an infrared small target segmentation method based on double dynamic frequency domain filtering, and the core innovation lies in the introduction of a "double dynamic frequency domain filtering" mechanism. The mechanism realizes the dual dynamic adjustment of "single core dynamic variable" and "multi-core dynamic weighting".

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and specifically relates to an infrared small target segmentation method, device, electronic device and storage medium based on dual dynamic frequency domain filtering. Background Technology

[0002] Infrared small target segmentation is a key technology in remote sensing image processing, widely used in fields such as UAV perception, military guidance, and early warning. Compared with conventional vision tasks, infrared small target segmentation has significant unique characteristics: the target size is extremely small, occupying only a small number of pixels in the image; structural information such as shape, edges, and texture is very scarce; and the signal-to-noise ratio is low, making it susceptible to interference from complex background clutter. Therefore, this task places stringent demands on the network's global feature perception capability, background suppression capability, and fine-grained feature representation capability.

[0003] In recent years, deep learning methods have gradually become the main technical approach in the field of infrared small target segmentation due to their strong end-to-end feature learning capabilities. Most existing methods employ encoder-decoder networks as their basic framework, and introduce strategies such as multi-scale fusion, attention mechanisms, and dynamic convolution to improve segmentation performance. However, these methods still face several technical bottlenecks.

[0004] First, spatial convolution, as a fundamental module of deep learning networks, is simple to implement and has strong local modeling capabilities, but its receptive field is limited by the size of the convolution kernel, mainly relying on local neighborhood information and making it difficult to directly capture global context information.

[0005] Secondly, while attention mechanisms can enhance feature representation through global correlation modeling, their computational complexity is on the order of quadratic, resulting in high computational costs and hindering lightweight and real-time deployment. Furthermore, existing dynamic convolution methods typically only dynamically adjust the weighting coefficients of multiple baseline filter kernels, while the baseline filter kernels themselves remain fixed. They cannot adaptively adjust the frequency response based on input features. Therefore, the expressive power of the dynamic filter kernel obtained by weighting the baseline filter kernels is limited to the subspace spanned by the baseline filter kernels, thus restricting its dynamic adjustment capability.

[0006] From the perspective of frequency domain analysis, frequency domain filtering can perceive global features at a lower computational cost, balancing global perception capability with computational overhead, and providing a feasible technical approach for infrared small target segmentation. However, if the reference filter kernel itself is fixed, its adaptability to different input features is still insufficient, making it difficult to flexibly adjust the frequency response according to changes in the target and background.

[0007] Therefore, existing technologies still suffer from problems such as strong locality of spatial convolution and insufficient global modeling ability; high computational cost of attention mechanism; limited dynamic adjustment ability of existing dynamic convolution methods due to the fixed nature of the reference filter kernel; and lack of fine-grained adaptive adjustment mechanism for frequency response in existing frequency domain filtering methods. Summary of the Invention

[0008] The technical problem to be solved by this invention is: addressing the problems in infrared small target segmentation such as weak targets, complex backgrounds, limited receptive fields of spatial convolution, high computational costs of attention mechanisms, and insufficient frequency response flexibility of existing dynamic convolution methods. This invention proposes an infrared small target segmentation method and device based on dual dynamic frequency domain filtering. It adopts a dual dynamic frequency domain filtering strategy to achieve stronger global feature extraction capabilities and more flexible frequency response adjustment capabilities with lower computational overhead, thereby being more effectively applied to infrared small target segmentation tasks.

[0009] To address the aforementioned technical problems, this invention provides an infrared small target segmentation method based on dual dynamic frequency domain filtering. Its core innovation lies in the introduction of a "dual dynamic frequency domain filtering" mechanism. This mechanism achieves dual dynamic adjustment: "single-core dynamic variability" and "multi-core dynamic weighting." On one hand, through "band-level dynamic adjustment," adjustment coefficients for each frequency band are adaptively generated based on input features, and different frequency bands are modulated group by group. This allows the frequency response of the reference filter kernel itself to dynamically change with the input features, solving the problem of insufficient expressive power of a fixed reference kernel. On the other hand, through "filter kernel-level dynamic weighting," based on the already dynamically variable reference filter kernel, weighting coefficients for multiple reference filter kernels are further adaptively generated according to input features, thus forming a final dynamic filter kernel with stronger expressive power and greater flexibility. This dual dynamic mechanism, while maintaining the advantages of global receptive field and low computational overhead of frequency domain filtering, greatly enhances the expressive flexibility of the filter kernel.

[0010] The technical solution of this invention is the infrared small target segmentation method based on dual dynamic frequency domain filtering, which is characterized by including the following steps: (1) Obtain the infrared image to be segmented; (2) The infrared image is input into a pre-trained neural network segmentation model, which adopts a pyramid encoder-decoder structure and embeds a dual dynamic frequency domain filtering module in at least one scale layer of its encoding and decoding paths. (3) The input feature map is frequency-domain filtered using the dual dynamic frequency domain filtering module to obtain the output feature map. The frequency domain filtering includes: Based on the input feature map, frequency band adjustment coefficients and filter kernel weighting coefficients are adaptively generated; The frequency band components of multiple preset reference filter cores are modulated group by group using the frequency band adjustment coefficient to obtain the modulated reference filter cores; The final dynamic filter kernel is obtained by weighting and summing multiple modulated reference filter kernels using the weighting coefficients of the filter kernel. In the frequency domain, the input feature map is filtered using the dynamic filtering kernel; (4) Based on the output of the neural network segmentation model, the segmentation result of the infrared small target is obtained.

[0011] Preferably, step (3) adaptively generates the frequency band adjustment coefficient and filter kernel weighting coefficient based on the input feature map, specifically including: (3.1) Perform global average pooling and global maximum pooling on the input feature map respectively, and concatenate the pooling results along the channel dimension to form a global feature representation; (3.2) Input the global feature representation into the first fully connected branch and the second fully connected branch in parallel, respectively; (3.3) The frequency band adjustment coefficients are generated through the first fully connected branch and the Sigmoid activation function; (3.4) The filter kernel weighting coefficients are generated through the second fully connected branch and the Softmax activation function.

[0012] Preferably, the pre-set plurality of benchmark filter kernels in step (3) are: k learnable benchmark filter kernels in complex form set for each channel of the input feature map, wherein the size of the benchmark filter kernels is related to the size of the input feature map. Figure 1 Furthermore, each of the aforementioned reference filter cores is pre-divided into n frequency bands (Ω1, Ω2, ..., Ω...) according to its frequency magnitude. n And generate a corresponding mask matrix (M1, M2, ..., M) for each frequency band. n ), used to extract components of each frequency band; The mask matrix M j The corresponding frequency band index is j, and its element belongs to the frequency band Ω in the frequency domain coordinates. j If the value is 1, then the value is 0; otherwise, the value is 0. (1) In the formula, (u,v) are frequency domain coordinates.

[0013] Preferably, the process of obtaining the modulated reference filter kernel and the final dynamic filter kernel specifically includes: Using the aforementioned frequency band adjustment coefficients, the n frequency band components of each reference filter core are weighted and summed to obtain the modulated reference filter core: (2) In the formula, c is the channel index, i is the reference filter kernel index, and j is the frequency band index. This represents the j-th frequency band component of the c-th channel and the i-th reference filter core. Indicates and The corresponding frequency band adjustment coefficient, This represents the c-th channel and the ith reference filter core after dynamic frequency band adjustment. The reference filter core can adjust its frequency response according to the input characteristics, thereby achieving "single-core dynamic variable". Using the weighting coefficients of the filter kernels, the k modulated reference filter kernels in the same channel are weighted and summed to obtain the dynamic filter kernel of that channel: (3) In the formula, Indicates and The corresponding filter kernel weighting coefficients.

[0014] As a preferred option: Step (3) involves filtering the input feature map in the frequency domain using a dynamic filtering kernel, specifically including: For each channel of the input feature map Perform a two-dimensional discrete Fourier transform to obtain the corresponding frequency domain representation. : (4)

[0015] In the formula, Represents a two-dimensional discrete Fourier transform; The frequency domain representation of each channel of the input feature map is multiplied element-wise with the corresponding dynamic filter kernel. The result of the element-wise multiplication is then subjected to a two-dimensional discrete Fourier inverse transform to obtain the filtered spatial domain feature map. (5) In the formula, This represents element-wise multiplication. This represents the two-dimensional discrete Fourier inverse transform.

[0016] Preferably, after obtaining the filtered spatial feature map, the following steps are also included: The filtered spatial feature maps of all channels are stitched together. The concatenated feature maps are then fused using 1×1 convolution and ReLU activation functions to generate the final output feature map of the dual dynamic frequency domain filtering module. (6) In the formula, the output features H represents the height of the input feature map, and W represents the width of the input feature map. Indicates the number of output channels. Indicates the number of input channels. This represents a 1×1 convolution.

[0017] Preferably, the encoding path of the neural network segmentation model includes a max pooling layer for downsampling after the dual dynamic frequency domain filtering module.

[0018] Preferably, in the decoding path of the neural network segmentation model described in step (2), the steps for feature fusion include: (2.1) Upsample the feature map of the current decoding layer; (2.2) The upsampled feature map is concatenated with the feature map of the corresponding scale in the encoding path along the channel dimension to obtain the concatenated feature map; (2.3) The spliced ​​feature map is used as the input of the dual dynamic frequency domain filtering module to extract context information.

[0019] Another technical solution of the present invention is the infrared small target segmentation device based on dual dynamic frequency domain filtering, which is characterized by including: Image acquisition unit, used to acquire infrared images to be segmented; The model processing unit has a pre-trained neural network segmentation model built in. The neural network segmentation model adopts a pyramid encoder-decoder structure and embeds a dual dynamic frequency domain filtering module in at least one scale layer of its encoding and decoding paths. The dual dynamic frequency domain filtering module is used to perform the frequency domain filtering step in the aforementioned infrared small target segmentation method. The result output unit is used to obtain the segmentation result of the infrared small target based on the output of the neural network segmentation model.

[0020] Another technical solution of the present invention is the electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The special feature is that when the processor executes the program, it implements the infrared small target segmentation method based on dual dynamic frequency domain filtering as described above.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention balances global perception capability with computational overhead. It employs frequency domain filtering for feature extraction, which, compared to spatial domain convolution, provides a global receptive field, enabling better perception of global contextual information; compared to attention mechanisms, it has lower computational overhead, making it more suitable for lightweight and real-time deployment.

[0022] (2) This invention greatly improves the dynamic flexibility of the filter kernel. Existing dynamic convolution methods typically only involve a weighted combination of multiple fixed reference convolution kernels. However, this invention further introduces a dynamic bandgap adjustment mechanism, enabling the reference filter kernel itself to dynamically adjust its frequency response according to changes in input features. Through the dual dynamic mechanism of "single-kernel dynamic variable" and "multi-kernel dynamic weighting," the expressive flexibility of the filter kernel is enhanced, which is beneficial for improving the feature discrimination ability of infrared small targets and complex backgrounds.

[0023] (3) This invention has good adaptability and scalability. Dual dynamic frequency domain filtering is not only applicable to infrared small target segmentation networks, but can also be embedded as a general feature extraction module in other vision task frameworks, demonstrating good adaptability and scalability. Attached Figure Description

[0024] Figure 1 This is a diagram of the infrared small target segmentation network structure based on dual dynamic frequency domain filtering of the present invention; Figure 2 This is a diagram of the dual dynamic frequency domain filtering structure of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] The core idea of ​​dual-dynamic frequency domain filtering is to realize a dual dynamic adjustment mechanism for the filter kernel in the frequency domain. Its "dual-dynamic" nature is manifested in: (1) Frequency band-level dynamic adjustment. The reference filter core is divided into multiple frequency bands according to the frequency magnitude, and the adjustment coefficients of each frequency band are adaptively generated according to the input characteristics. Different frequency bands are modulated group by group, so that the frequency response of the reference filter core can change dynamically with the input characteristics.

[0027] (2) Dynamic weighting at the filter kernel level. Based on the dynamic variability of the reference filter kernel itself, weighting coefficients of multiple reference filter kernels are adaptively generated according to the input characteristics, thus forming the final dynamic filter kernel.

[0028] In the specific implementation of dual dynamic frequency domain filtering, such as Figure 2 As shown, for the input feature map, global average pooling and global maximum pooling are first applied, and then concatenated along the channel dimension to form a global feature representation. Then, the frequency band adjustment coefficients and filter kernel weighting coefficients are output through two parallel branches. For each channel of the input feature map, k complex-form baseline filter kernels are set, with the size of the filter kernels varying with the input feature map. Figure 1 The input feature map is then divided into n frequency bands. Through band-level adjustment and filter kernel-level weighting, the final dual dynamic filter kernel is obtained, which is then applied to the input feature map. Frequency domain filtering is performed. Finally, channel information is fused in the spatial domain using a 1×1 convolution to obtain the output feature map.

[0029] Example 1: Infrared Small Target Segmentation Network Based on Dual Dynamic Frequency Domain Filtering This embodiment provides an infrared small target segmentation method based on dual dynamic frequency domain filtering. Its core is to use a pre-trained neural network segmentation model to process the input infrared image and output a high-precision binary classification segmentation result image.

[0030] This neural network segmentation model employs a pyramid encoder-decoder structure. For example... Figure 1 As shown, the left side of the network is the encoding path, which is used to gradually extract semantic features at different levels; the right side of the network is the decoding path, which is used to gradually restore spatial resolution and fuse shallow detail information with deep semantic information.

[0031] The dual dynamic frequency domain filtering module is embedded at various scale positions in both the encoding and decoding ends, serving as the core feature extraction unit. Its feature extraction process can be represented as follows:

[0032] In the formula, and These represent the output feature map and the input feature map, respectively. This indicates a dual dynamic frequency domain filtering operation.

[0033] Input single-channel infrared image For example, at the encoding end, a 2×2 window with a stride of 2 is used for downsampling after the DDFF module. At the decoding end, to fully fuse information, the decoded feature maps at each scale are compressed into a single channel through a 1×1 convolution, then upsampled to the same size as the original input image (256×256), and concatenated along the channel dimension. Finally, a 1×1 convolutional layer and a sigmoid activation function are used to output a pixel-level binary classification segmentation result image. .

[0034] This invention uses the IoU loss function as the optimization objective for network training and employs the Adam optimization algorithm to update network parameters. The IoU loss function... The definition is as follows:

[0035] In the formula, N represents the total number of pixels in the image. and Let represent the predicted probability value and the true labeled value of the i-th pixel, respectively. ε is a very small positive number (e.g., 1e to 7) used to avoid the denominator being zero and to improve numerical stability. This loss function directly optimizes the core evaluation metrics of the segmentation task, which helps the model converge faster and achieve higher accuracy.

[0036] Example 2: Internal structure and working principle of dual dynamic frequency domain filtering: This embodiment details the internal structure and working principle of the core innovative module—dual dynamic frequency domain filtering. This module generates a flexible and variable dynamic frequency domain filtering kernel through a dual dynamic mechanism of "single-core dynamic variable" and "multi-core dynamic weighting," achieving global information perception of input features. Its structure is as follows: Figure 2 As shown.

[0037] Let the input feature map be In the formula, H and W are the height and width, respectively, and C in This is the number of input channels.

[0038] First, for any channel of the input feature map X Perform a two-dimensional discrete Fourier transform (DFT) to transform it from the spatial domain to the frequency domain, obtaining the frequency domain representation. :

[0039] In the formula, This represents the two-dimensional discrete Fourier transform.

[0040] This invention sets a set of k learnable, complex-form reference filter kernels for each channel of the input feature map X, denoted as... Where c is the channel index and i is the reference filter kernel index, these reference filter kernels are the foundation of frequency domain filtering. To address the deficiency of fixed reference kernels in existing technologies, this invention introduces a band-level dynamic adjustment mechanism. Each complete reference filter kernel is:

[0041] The frequency bands are divided into n equal-width bands Ω1, Ω2, ..., Ω according to their frequency magnitude. n The frequency magnitude is measured from the frequency domain coordinates (u,v) to the center point coordinates. The Euclidean distance d(u,v) metric:

[0042] For each frequency band Define a mask matrix Its elements belong to the frequency domain coordinates. If the value is 1, then the value is 0; otherwise, the value is 0.

[0043] Therefore, the component of the i-th reference filter kernel in the j-th frequency band can be obtained by element-wise multiplication:

[0044] In the formula, This indicates element-wise multiplication.

[0045] This completes the preparatory work for splitting the fixed reference core into multiple independent components according to frequency band; The dual dynamic adjustment mechanism is implemented through two parallel branches. Global average pooling is performed on the input feature map X separately. and global maximum pooling The results are then concatenated along the channel dimension. This forms a compact vector Z that can comprehensively represent global information:

[0046] In the formula, and These represent global average pooling and global max pooling, respectively. This indicates a channel splicing operation.

[0047] In the first branch, the global feature representation Z passes through a fully connected layer and a sigmoid activation function to generate frequency band adjustment coefficients. In the formula, the element These are adjustment coefficients for the c-th channel, the i-th reference kernel, and the j-th frequency band. The generation of these coefficients enables the network to adaptively enhance or suppress specific frequency components based on the global characteristics of the current input features.

[0048] In the second branch, the global feature representation Z passes through another fully connected layer and a Softmax activation function to generate filter kernel weighting coefficients. In the formula, the element These are the weighting coefficients for the c-th channel and the i-th reference kernel, and satisfy... This coefficient is responsible for making dynamic trade-offs among multiple modulated reference cores.

[0049] After obtaining two sets of dynamic coefficients, the "single-core dynamic variable" step is executed. This utilizes the frequency band adjustment coefficient. The modulated reference filter core is obtained by weighted summation of the frequency band components of the c-th channel and the i-th reference filter core:

[0050] In the formula, This represents the reference filter core after dynamic frequency band adjustment. Therefore, the reference filter core can adjust its frequency response according to the input characteristics, thereby achieving "single-core dynamic variable".

[0051] After this step, each reference filter core It is no longer fixed, but adaptively adjusts its response intensity in different frequency bands according to the input characteristics, realizing dynamic variability at the "single core" level; Next, the "multi-core dynamic weighting" step is executed. This utilizes the weighting coefficients of the filter kernel. By weighted fusion of k modulated reference filter kernels within the same channel c, a final dynamic filter kernel with extremely strong expressive power for that channel is obtained. :

[0052] Thus, the two mechanisms of "single-core dynamic variable" and "multi-core dynamic weighting" produce a synergistic effect, resulting in a dynamic filter kernel that is generated in the end. It can achieve great flexibility and adaptability in both frequency response and inter-core combination dimensions according to the input content, and its expression space far exceeds that of existing technologies.

[0053] Finally, the generated dynamic filter kernel is used The input feature map is then filtered in the frequency domain. For channel c, the filtering process is as follows:

[0054] In the formula, This represents the two-dimensional discrete Fourier inverse transform, which restores the filtered frequency domain features to a spatial domain feature map. .

[0055] After completing the channel-by-channel frequency domain filtering, in order to promote information exchange between channels, the output feature maps of all channels are... The features are concatenated along the channel dimension and then fused using a 1×1 convolutional layer and a ReLU activation function to generate the final output feature map.

[0056] In the formula, the output features , Indicates the number of output channels. This represents a 1×1 convolution.

[0057] In this way, the present invention closely combines the global perception advantage of frequency domain filtering with a highly flexible dynamic kernel adjustment mechanism, providing an efficient and powerful technical solution for solving the extremely challenging task of infrared small target segmentation.

[0058] Creative discourse: The inventiveness of this invention compared to the prior art is mainly reflected in the following aspects: Dual Dynamic Filtering Mechanism: Existing dynamic convolution typically only weights a fixed reference filter kernel at the linear combination level. This invention, however, proposes for the first time a mechanism combining band-level dynamic adjustment and filter kernel-level dynamic weighting. The frequency response of each reference filter kernel can adaptively change according to the input features, and after weighted combination, the expression space of the filter kernel is significantly expanded, greatly improving the ability to distinguish complex backgrounds and small targets.

[0059] Unified global perception in the frequency domain and lightweight design: The global receptive field is achieved by using frequency domain filtering, which avoids the locality defects of spatial convolution. At the same time, it avoids the secondary computational complexity of the attention mechanism, significantly reducing computational overhead while ensuring global modeling capabilities. It is especially suitable for application scenarios with high real-time requirements, such as infrared small targets.

[0060] The end-to-end embedded system design seamlessly integrates the aforementioned modules into the pyramid encoder-decoder network, organically unifying global perception, detail recovery, and multi-scale fusion, achieving segmentation accuracy and robustness superior to existing methods. These technical features are not simply a superposition of existing technologies, but rather a synergistic overall technical solution with outstanding substantive characteristics and significant progress.

[0061] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be covered by the claims of the present invention.

Claims

1. A method for segmenting small infrared targets based on dual dynamic frequency domain filtering, characterized in that, Includes the following steps: (1) Obtain the infrared image to be segmented; (2) The infrared image is input into a pre-trained neural network segmentation model, which adopts a pyramid encoder-decoder structure and embeds a dual dynamic frequency domain filtering module in at least one scale layer of its encoding and decoding paths. (3) The input feature map is frequency-domain filtered using the dual dynamic frequency domain filtering module to obtain the output feature map. The frequency domain filtering includes: Based on the input feature map, frequency band adjustment coefficients and filter kernel weighting coefficients are adaptively generated; The frequency band components of multiple preset reference filter cores are modulated group by group using the frequency band adjustment coefficient to obtain the modulated reference filter cores; The final dynamic filter kernel is obtained by weighting and summing multiple modulated reference filter kernels using the weighting coefficients of the filter kernel. In the frequency domain, the input feature map is filtered using the dynamic filtering kernel; (4) Based on the output of the neural network segmentation model, the segmentation result of the infrared small target is obtained.

2. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 1, characterized in that, Step (3) adaptively generates the frequency band adjustment coefficients and filter kernel weighting coefficients based on the input feature map, specifically including: (3.1) Perform global average pooling and global maximum pooling on the input feature map respectively, and concatenate the pooling results along the channel dimension to form a global feature representation; (3.2) Input the global feature representation into the first fully connected branch and the second fully connected branch in parallel, respectively; (3.3) The frequency band adjustment coefficients are generated through the first fully connected branch and the Sigmoid activation function; (3.4) The filter kernel weighting coefficients are generated through the second fully connected branch and the Softmax activation function.

3. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 1, characterized in that, Step (3) refers to the pre-defined plurality of reference filter kernels as follows: k learnable reference filter kernels in complex form are set for each channel of the input feature map, the size of the reference filter kernels being consistent with the input feature map; and each reference filter kernel is pre-divided into n frequency bands (Ω1, Ω2, ..., Ω) according to frequency magnitude. n And generate a corresponding mask matrix (M1, M2, ..., M) for each frequency band. n ), used to extract components of each frequency band; The mask matrix M j The corresponding frequency band index is j, and its element belongs to the frequency band Ω in the frequency domain coordinates. j If the value is 1, then the value is 0; otherwise, the value is 0. （1） In the formula, (u,v) are frequency domain coordinates.

4. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 3, characterized in that, The process of obtaining the modulated reference filter kernel and the final dynamic filter kernel specifically includes: Using the aforementioned frequency band adjustment coefficients, the n frequency band components of each reference filter core are weighted and summed to obtain the modulated reference filter core: （2） In the formula, c is the channel index, i is the reference filter kernel index, and j is the frequency band index. This represents the j-th frequency band component of the c-th channel and the i-th reference filter core. Indicates and The corresponding frequency band adjustment coefficient, This represents the c-th channel and the ith reference filter core after dynamic frequency band adjustment. The reference filter core can adjust its frequency response according to the input characteristics, thereby achieving "single-core dynamic variable". Using the weighting coefficients of the filter kernels, the k modulated reference filter kernels in the same channel are weighted and summed to obtain the dynamic filter kernel of that channel: （3） In the formula, Indicates and The corresponding filter kernel weighting coefficients.

5. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 1, characterized in that, Step (3) involves using a dynamic filter kernel to filter the input feature map in the frequency domain, specifically including: For each channel of the input feature map Perform a two-dimensional discrete Fourier transform to obtain the corresponding frequency domain representation. : （4） In the formula, Represents a two-dimensional discrete Fourier transform; The frequency domain representation of each channel of the input feature map is multiplied element-wise with the corresponding dynamic filter kernel. The result of the element-wise multiplication is then subjected to a two-dimensional discrete Fourier inverse transform to obtain the filtered spatial domain feature map. （5） In the formula, This represents element-wise multiplication. This represents the two-dimensional discrete Fourier inverse transform.

6. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 5, characterized in that, After obtaining the filtered spatial feature map, the following steps are also included: The filtered spatial feature maps of all channels are stitched together. The concatenated feature maps are then fused using 1×1 convolution and ReLU activation functions to generate the final output feature map of the dual dynamic frequency domain filtering module. （6） In the formula, the output features H represents the height of the input feature map, and W represents the width of the input feature map. Indicates the number of output channels. Indicates the number of input channels. This represents a 1×1 convolution.

7. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 1, characterized in that, In the encoding path of the neural network segmentation model, after the dual dynamic frequency domain filtering module, there is also a max pooling layer for downsampling.

8. The infrared small target segmentation method based on dual dynamic frequency domain filtering according to claim 1, characterized in that, In step (2), the feature fusion steps in the decoding path of the neural network segmentation model include: (2.1) Upsample the feature map of the current decoding layer; (2.2) The upsampled feature map is concatenated with the feature map of the corresponding scale in the encoding path along the channel dimension to obtain the concatenated feature map; (2.3) The spliced ​​feature map is used as the input of the dual dynamic frequency domain filtering module to extract context information.

9. An infrared small target segmentation device based on dual dynamic frequency domain filtering, characterized in that, include: Image acquisition unit, used to acquire infrared images to be segmented; The model processing unit has a pre-trained neural network segmentation model built in. The neural network segmentation model adopts a pyramid encoder-decoder structure and embeds a dual dynamic frequency domain filtering module in at least one scale layer of its encoding and decoding paths. The dual dynamic frequency domain filtering module is used to perform the frequency domain filtering step in the infrared small target segmentation method as described in any one of claims 1 to 6; The result output unit is used to obtain the segmentation result of the infrared small target based on the output of the neural network segmentation model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the infrared small target segmentation method based on dual dynamic frequency domain filtering as described in any one of claims 1 to 6.

Images