A fractal-based attention-guided remote sensing image feature extraction method and system
Patent Information
- Application Number
- CN202611110338.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-10-16
AI Technical Summary
[0004]然而，现有方法大多基于欧氏空间中的规则特征建模，尚未充分利用分形结构对复杂纹理区域进行显式约束与引导，导致模型对复杂边界及非规则结构的感知能力不足
本发明能够自适应建模遥感图像中的纹理自相似结构。传统特征提取方法往往难以有效区分真实目标结构与复杂背景纹理。本发明引入可学习 Mandelbrot 分形迭代过程，通过图像特征生成的门控函数调节不同空间位置的分形迭代强度，使分形结构响应能够根据图像内容自适应生成，从而更好地表征遥感场景中的纹理自相似性和局部结构复杂性。
Smart Images
Figure CN122929221A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image interpretation technology, and more specifically to a method and system for extracting remote sensing image features based on fractal-guided attention. Background Technology
[0002] Remote sensing image interpretation technology, as an important research direction in the field of Earth observation, has wide application value in national defense security, disaster monitoring, urban planning, traffic management, and marine monitoring. With the development of UAV remote sensing platforms, the ability to acquire high-resolution remote sensing images has been continuously improved, making the fine interpretation of targets in complex scenes a current research hotspot. Currently, most mainstream remote sensing image interpretation methods are based on deep neural networks, using backbone networks such as convolutional neural networks (CNNs) or Transformers for feature extraction, and combining them with prediction heads for corresponding tasks to achieve downstream tasks such as target segmentation, semantic segmentation, instance segmentation, and change detection. Among them, CNNs rely on local receptive fields for feature modeling. Although they have strong capabilities in local texture extraction, their ability to model long-distance dependencies is limited, making it difficult to effectively capture global structural information in complex scenes. Transformers, on the other hand, establish global correlations through self-attention mechanisms, which can enhance long-distance feature interactions, but their ability to model local boundary details and fine-grained texture structures is relatively insufficient.
[0003] For UAV remote sensing images, typical targets (such as ships, aircraft, vehicles, and buildings) are usually characterized by significant scale variations, dense spatial distribution, complex texture structures, and arbitrary target orientations. Simultaneously, complex terrain features generally exhibit obvious multi-scale self-similar texture characteristics and irregular boundary structures, making feature aliasing between target and background regions easy. Especially in complex scenes such as ports, airports, and road intersections, target boundaries often exhibit strong irregularity and random orientation, making it difficult for traditional feature modeling methods to effectively characterize the structural complexity of the target region's interior and edge areas. Existing attention-based remote sensing image segmentation methods typically rely solely on feature similarity for weight allocation, lacking the ability to explicitly model the local geometric complexity of the image, self-similar texture distribution, and chaotic boundary structures. This leads to problems such as target omission, incomplete boundary segmentation, and insufficient representation of local details in complex backgrounds, densely populated small target areas, and weak boundary regions. Furthermore, complex texture regions in remote sensing images often exhibit certain fractal characteristics, meaning they possess statistical self-similar structures at different scales.
[0004] However, most existing methods are based on regular features in Euclidean space and have not made full use of fractal structures to explicitly constrain and guide complex texture regions, resulting in insufficient perception of complex boundaries and irregular structures.
[0005] Therefore, how to enhance the model's ability to model complex textures, self-similar structures, and irregular boundary regions, thereby improving the accuracy and robustness of target segmentation in UAV remote sensing images, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0006] In view of the above problems, the present invention is proposed to provide a method and system for remote sensing image feature extraction based on fractal-guided attention that overcomes or at least partially solves the above problems.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] In a first aspect, embodiments of the present invention provide a remote sensing image feature extraction method based on fractal-guided attention, comprising the following steps: Obtain feature maps from remotely sensed images; For the obtained feature map, the position coordinates and feature vectors of the feature map are obtained respectively, and mapped to complex inputs. The complex inputs of position coordinates and feature vectors at the same spatial location are concatenated to obtain the complex inputs of the learnable Mandelbrot fractal dynamic system. Based on the gating function, an image-driven learnable Mandelbrot fractal iterative model is constructed, and the complex input of the learnable Mandelbrot fractal dynamic system is fed into the learnable Mandelbrot fractal iterative model. The learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M; By embedding the fractal structure response map M as an attention bias term into the Transformer attention mechanism, fractal-guided attention is obtained. Based on the calculation results of fractal-guided attention, the deep feature map of the input feature map is obtained.
[0009] Furthermore, in the step of obtaining the feature map of the remote sensing image, the feature map is obtained by performing initial feature extraction on the remote sensing image using a feature extraction module with N deep stacked layers.
[0010] Furthermore, the complex inputs of the learnable Mandelbrot fractal dynamical system are obtained through the following steps: For the input feature map Its scale is The number of channels is C, and the feature map is defined. If the spatial position is p=(u,v), then the horizontal coordinates u∈{0,1,…,W} are obtained. 1}, the vertical coordinate v∈{0,1,…,H} 1}, the total number of spatial locations is N = H × W; Normalize the horizontal and vertical coordinates of any position p to [ [1,1] yields the normalized coordinates ( , Complex inputs for constructing position coordinates based on normalized coordinates:
[0011] in The imaginary unit; Obtain the feature vector corresponding to any position p in the feature map:
[0012] Obtaining complex inputs from feature vectors:
[0013] in, Real part mapping function, Imaginary part mapping function; The complex inputs of the position coordinates and the eigenvectors are added together to obtain the complex inputs of the learnable Mandelbrot fractal dynamical system:
[0014] Furthermore, based on the gating function, an image-driven, learnable Mandelbrot fractal iterative model is constructed, specifically including the following models:
[0015] in, This represents the complex state of position p in the nth iteration. Fractal gating values generated from image features For the complex number input at position p
[0016] Furthermore, the learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M, specifically including the following steps: After completing the maximum number of fractal iterations In the fractal iteration process within the next few iterations, for each spatial location p, it is determined whether the complex state at that spatial location has escaped, where the escape determination condition is:
[0017] in Complex state The model, The escape threshold; Define the escape iteration function for position p:
[0018] in, This represents the number of iterations at which position p first satisfies the escape condition; If within the maximum number of iterations If no escape occurred, then order This leads to the escape time map representing the total number of position iterations. ; Normalizing the escape time plot yields the fractal structure response plot:
[0019] in, Let be the normalized fractal response at position p.
[0020] Furthermore, the fractal structure response map M is embedded as an attention bias term into the Transformer attention mechanism to obtain fractal-guided attention, which specifically includes the following steps: The fractal structure response map M is embedded as an attention bias term into the Transformer attention mechanism to obtain fractal-guided attention, which specifically includes the following steps: Mapping the fractal structure response map M to a fractal attention bias matrix:
[0021] in, It is a mapping matrix; Adding the fractal attention bias matrix to the standard attention score yields fractal-guided attention:
[0022] in Here, Q represents the learnable weights, K represents the key matrix, V represents the value matrix, and d represents the query and key vector dimensions.
[0023] Furthermore, based on the calculation results of fractal-guided attention, a deep feature map of the input feature map is obtained, specifically including: The calculation results of fractal-guided attention are sequentially subjected to residual connection and layer normalization operations with the input feature map to obtain the deep feature map.
[0024] Furthermore, it also includes inputting the obtained deep feature map into the detection head of the downstream task network to obtain the task result. The downstream task network includes a classification network, a localization network, and a segmentation network.
[0025] Secondly, embodiments of the present invention provide a remote sensing image feature extraction system based on fractal-guided attention, comprising: a fractal-guided attention module; the fractal-guided attention module performs the following steps: For the feature map obtained from the remote sensing image, the position coordinates and feature vectors of the feature map are obtained respectively and mapped to complex inputs. The complex inputs of position coordinates and feature vectors at the same spatial location are concatenated to obtain the complex inputs of the learnable Mandelbrot fractal dynamic system. Based on the gating function, an image-driven learnable Mandelbrot fractal iterative model is constructed, and the complex input of the learnable Mandelbrot fractal dynamic system is fed into the learnable Mandelbrot fractal iterative model. The learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M; By embedding the fractal structure response map M as an attention bias term into the Transformer attention mechanism, fractal-guided attention is obtained. The calculation results of fractal-guided attention are sequentially subjected to residual connection and layer normalization operations with the input feature map to obtain the deep feature map.
[0026] Preferably, the complex inputs of the learnable Mandelbrot fractal dynamical system are obtained through the following steps: For the input feature map Its scale is The number of channels is C, and the feature map is defined. If the spatial position is p=(u,v), then the horizontal coordinates u∈{0,1,…,W} are obtained. 1}, the vertical coordinate v∈{0,1,…,H} 1}, the total number of spatial locations is N = H × W; Normalize the horizontal and vertical coordinates of any position p to [ [1,1] yields the normalized coordinates ( , Complex inputs for constructing position coordinates based on normalized coordinates:
[0027] in The imaginary unit; Obtain the feature vector corresponding to any position p in the feature map:
[0028] Obtaining complex inputs from feature vectors:
[0029] in, Real part mapping function, Imaginary part mapping function; The complex inputs of the position coordinates and the eigenvectors are added together to obtain the complex inputs of the learnable Mandelbrot fractal dynamical system:
[0030] Preferably, the learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M, specifically including the following steps: After completion In the fractal iteration process within the last few iterations, for each spatial location p, determine whether the complex state at that spatial location has escaped: The escape determination condition is as follows:
[0031] in Complex state The model, The escape threshold; Define the escape iteration function for position p:
[0032] in, This represents the number of iterations at which position p first satisfies the escape condition; If within the maximum number of iterations If no escape occurred, then order This yields the escape time map. ; Normalizing the escape time plot yields the fractal structure response plot:
[0033] in The normalized fractal response at position p, This is a fractal structure response diagram.
[0034] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a remote sensing image feature extraction method and system based on fractal-guided attention, which has the following beneficial effects: This invention can adaptively model the self-similar texture structure in remote sensing images. Traditional feature extraction methods often struggle to effectively distinguish between realistic target structures and complex background textures. This invention introduces a learnable Mandelbrot fractal iteration process, adjusting the fractal iteration intensity at different spatial locations through a gating function generated from image features. This allows the fractal structure response to be adaptively generated based on the image content, thereby better representing the texture self-similarity and local structural complexity in remote sensing scenes.
[0035] This invention embeds fractal structure responses into the Transformer attention mechanism, enhancing the structure-awareness of attention computation. Ordinary Transformer attention primarily establishes long-distance dependencies based on feature similarity, lacking explicit constraints on target boundaries and local structural abrupt changes. This invention maps fractal structure response maps into attention bias matrices and embeds learnable modulation coefficients into the attention score calculation process. This allows the attention mechanism to focus on complex boundaries, subtle textures, and locations of regional structural abrupt changes while modeling global relationships, thereby improving the feature representation quality in complex remote sensing scenes. Attached Figure Description
[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0037] Figure 1 This is a structural diagram of the fractal-guided attention module provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the feature modeling backbone network structure based on the fractal-guided attention module provided in this embodiment of the invention. Detailed Implementation
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] This invention discloses a remote sensing image feature extraction method based on fractal-guided attention, which is based on building a fractal-guided attention module for processing remote sensing image feature maps.
[0040] The fractal-guided attention module constructed in this invention has the following structure: Figure 1 As shown, the system includes a complex embedding module, a learnable Mandelbrot fractal dynamics system module, an escape time mapping module, a fractal structure response map generation module, a fractal attention bias construction module, and a Transformer attention modulation module. The main purpose of this fractal-guided attention module is to generate a fractal structure response map through an image feature-driven Mandelbrot fractal dynamics system, and embed this response map into the Transformer attention computation process. This enables the model to explicitly perceive complex target boundaries, local texture self-similarity structures, and abrupt changes in regional structures in remote sensing images, thereby improving the boundary integrity and refined representation capabilities of target segmentation.
[0041] The specific data execution process based on the fractal-guided attention module is as follows: First, the input feature map obtained from the remote sensing image Its scale is The number of channels is C. Define the feature map. If the spatial position is p=(u,v), then the horizontal coordinates u∈{0,1,…,W} are obtained. 1}, the vertical coordinate v∈{0,1,…,H} 1}, the total number of spatial locations is N=H×W.
[0042] Then, the feature map coordinates The input is mapped to a complex number, and an initial fractal state is constructed.
[0043] For any position p=(u,v) in the feature map, its corresponding feature vector is represented as:
[0044] First, normalize the horizontal and vertical coordinates of p to [ 1,1] obtain ( , Constructing complex coordinate embeddings based on normalized coordinates:
[0045]
[0046]
[0047] in It is the imaginary unit.
[0048] To ensure that complex inputs include not only spatial coordinate information but also image semantic information, this invention further incorporates features... Obtaining complex feature embeddings at position p:
[0049] in This represents the feature vector at position p; Real part mapping function, Imaginary part mapping function.
[0050] In one specific embodiment, the real part mapping function and the imaginary part mapping function can be implemented by two independent learnable linear mapping functions:
[0051] in, and Let be the learnable weight matrices for the real and imaginary parts, respectively. and For the corresponding bias term, This represents the feature vector that needs to be mapped, specifically the feature vector at position p. Therefore, the complex feature embedding of the feature vector at position p can be further written as:
[0052] Through the above mapping, the original real-valued feature vectors are projected into the complex domain, where the real and imaginary parts correspond to two sets of learnable projective representations of the feature vectors, thus providing feature inputs for subsequent complex domain modeling or phase correlation modeling.
[0053] Then, the complex coordinate embedding and the complex feature embedding are added together to obtain the complex input of the learnable Mandelbrot fractal dynamical system:
[0054] Right now:
[0055] in, The input is the complex number corresponding to position P. The purpose of this step is to map both the image spatial coordinates and the image semantic features to the complex plane, enabling the learnable Mandelbrot fractal dynamics system to simultaneously perceive geometric position and image content. In this way, the fractal iteration process is no longer determined solely by fixed coordinates, but can adaptively adjust according to the target, background, boundary, and texture features in the remote sensing image.
[0056] Then, a fractal modeling module based on Mandelbrot is constructed. The traditional Mandelbrot fractal iteration process is as follows:
[0057] in Let c be the input point in the complex plane, representing the state of the nth iteration. Traditional Mandelbrot iterations use fixed parameters, making it difficult to adaptively adjust to changes in target structure and texture in remotely sensed images. To enable the fractal dynamical system to adapt to images, this invention introduces a feature map... The generated learnable gating function:
[0058] in Learnable gating functions This is an image-driven fractal gating graph. For learnable convolutional mappings, This refers to the Sigmoid function. In this invention, the image-driven fractal gating map is used to adaptively adjust the fractal iteration intensity at different spatial locations based on image features.
[0059] Based on the gating function, an image-driven, learnable Mandelbrot fractal iterative process is constructed:
[0060] in The complex state of position p in the nth iteration. Fractal gating values generated from image features The input is the complex number corresponding to position P. The purpose of this step is to extend the traditional fixed-form Mandelbrot fractal iteration into an image feature-driven, learnable fractal dynamic system. Through a gating function G(p), the fractal iteration intensity at different spatial locations can adaptively vary according to the image content, enabling the system to respond to complex boundaries, texture details, and local structural complexity in UAV remote sensing images.
[0061] Furthermore, escape-time fractal structure mapping is performed, and the maximum number of fractal iterations is achieved. In the fractal iteration process within the last few iterations, it is determined whether the complex state of each spatial location p has escaped.
[0062] The escape determination condition is defined as follows: ; in Complex state The model, This is the escape threshold.
[0063] Define the escape iteration function for position p:
[0064] in, This represents the number of iterations at which position p first satisfies the escape condition. If the maximum number of iterations is reached... If no escape occurred, then order This yields the escape time map. .
[0065] In this embodiment, escape time refers to the number of the first escape iteration in Mandelbrot fractal iteration, and the escape time map is essentially a representation of the number of iterations at all positions.
[0066] Then, a fractal response map is constructed to facilitate subsequent fusion and attention modulation in the neural network, and the escape time map is normalized:
[0067] in The normalized fractal response at position p, This step involves creating a fractal structure response map. The goal is to explicitly transform the implicit, learnable fractal dynamics system into a structure response map M. This response map characterizes the structural complexity of different regions in a UAV remote sensing image, including target boundary complexity, local texture self-similarity, and the degree of abrupt structural changes in the region. This fractal structure response map provides explicit structural priors for subsequent Transformer attention modulation.
[0068] It should be noted that in this invention, the result of a position p is obtained as one element of the fractal structure response diagram. The complete fractal structure response diagram is obtained by calculating the normalized fractal response for all spatial positions p=(u,v) and then rearranging them according to the original spatial coordinates.
[0069] Furthermore, the fractal structure response map M is used to guide Transformer attention modulation. The standard Transformer attention mechanism is defined as follows:
[0070] The matrix includes the Q query matrix, K key matrix, and V value matrix. The d-column dimensions are query and key vectors. The attention score matrix is...
[0071] in Because the fractal structure response diagram is Therefore, M needs to be mapped to an attention bias matrix:
[0072] in, The mapping matrix is used. In one embodiment, the mapping of the fractal structure response map can be performed using a 1×1 Conv or a linear layer mapping. The purpose of the mapping is to adjust the dimension of the matrix M to fit the subsequent attention calculation.
[0073] Furthermore, attention biases can be constructed using fractal similarity (attention bias matrix P(M)). By incorporating these fractal attention biases into the standard attention score, fractal-guided attention can be obtained.
[0074] in These are learnable weights. The purpose of this step is to embed the fractal structure response map as an attention bias term into the Transformer attention mechanism, enabling attention computation to explicitly perceive complex boundaries, fine textures, and regions of abrupt local structural changes. In this way, the model can pay more attention to target edges and textured regions during target segmentation in UAV remote sensing images, improving the integrity and precision of the segmentation boundaries.
[0075] At this point, the calculation based on the fractal-guided attention module is complete.
[0076] Finally, the calculation results of fractal-guided attention are sequentially coupled with the input feature map using residual connection and layer normalization operations to obtain the deep feature map.
[0077] To further describe the process of acquiring deep feature maps, a backbone feature extraction network based on a fractal-guided attention module can be constructed to obtain deep feature maps. The backbone feature extraction network based on the fractal-guided attention module is as follows: Figure 2 As shown, we refer to typical deep learning model structures (such as ViT, VGG, etc.) as the foundation for construction, and introduce a fractal-guided attention module to build a backbone feature extraction network.
[0078] First, for the input remote sensing image G, which has dimensions H×W and 3 channels, initial feature modeling is performed using a deep, stacked N-layer feature extraction module (such as a convolutional module, a Transformer block, etc.) to extract feature map information in matrix form from the input remote sensing image. The calculation process can be formally represented as follows:
[0079] in, This indicates the feature extraction module. Let be the extracted feature map corresponding to the i-th layer.
[0080] Then, a fractal-guided attention module was introduced. The deepest feature map rich in semantic information The deep feature extraction process can be formally represented as follows:
[0081] At this point, the backbone feature extraction network based on the fractal-guided attention module has been completed.
[0082] This invention focuses on the deepest feature map. This processing is merely one example; in reality, for any... All of these can be processed. To ensure the effectiveness of the modeling, this study selects the deepest feature map Fn, which is rich in semantic information, for processing.
[0083] Then, for typical downstream tasks of remote sensing image interpretation, such as classification, localization, and segmentation, corresponding regression functions are designed and used for prediction. For example, a detection head network can be connected to process the feature map and output location information for a detection task; a classification head network can be connected to output category information for a classification task. The calculation process can be formally represented as follows:
[0084]
[0085]
[0086] in, This represents the classification information obtained from the prediction. This indicates the location information obtained from the prediction. Information is segmented using a mask. Represents the classification function, This represents the location regression function. This is the mask regression function.
[0087] This application focuses on feature extraction based on fractal-guided attention modules, without imposing too many constraints on decoupled networks that can adapt to multiple downstream tasks.
[0088] After constructing the backbone feature modeling network incorporating a fractal-guided attention module, the model was tested using a detection task as an example. The detector constructed using the backbone feature modeling network incorporating a fractal-guided attention module... The network is trained for remote sensing image target segmentation using focal loss and intersection-over-union (IoU) loss, and is formally represented as follows:
[0089] in, This represents the focus loss, used in the classification and prediction branch for target segmentation tasks in remote sensing images. This represents the IoU loss function, used to calculate the distance between the predicted value and the ground truth label of the target corresponding to the input remote sensing image. This represents the total loss used in network training. Once the network training continues until the loss no longer decreases, the training has reached a stable state, the training process ends, and the trained target segmentation model combining a backbone feature modeling network based on a fractal-guided attention module is obtained.
[0090] Furthermore, the trained target segmentation model, which incorporates a backbone feature modeling network based on a fractal-guided attention module, is used to test the remote sensing images to be tested. This can be formally represented as:
[0091] in, This represents the trained target segmentation model incorporating the fractal-guided attention module. and These represent the remote sensing image to be tested and its corresponding target segmentation result, respectively.
[0092] Based on the same inventive concept, this embodiment of the invention also provides a remote sensing image feature extraction system based on fractal-guided attention, which is implemented by constructing a remote sensing image feature extraction module based on fractal-guided attention. Since the principle of the multi-granularity subspace fusion module in solving the problem is similar to the aforementioned remote sensing image feature extraction method based on fractal-guided attention, the implementation of this system can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.
[0093] This invention addresses the aforementioned challenges in remote sensing image feature discrimination and decoupling. It designs a remote sensing image feature extraction strategy based on fractal-guided attention, mapping feature representations at different scales and semantic levels to corresponding subspace structures. Furthermore, it models and measures the geometric relationships between these subspaces within a Grassmanifold space, thereby achieving alignment and collaborative fusion of multi-granularity features in a unified geometric space. Simultaneously, by constructing an adaptive weighting mechanism based on subspace similarity measurement, it suppresses redundant information and feature conflicts, enhances the complementarity and consistency between features of different granularities, and utilizes subspace projection and reconstruction methods to achieve structural enhancement and discriminative optimization of the original features.
[0094] It significantly improves the model's ability to represent the features of typical remote sensing targets, provides a more efficient feature modeling scheme for the field of remote sensing image interpretation, and provides strong technical support for Earth observation and monitoring as well as remote sensing applications.
[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0096] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A remote sensing image feature extraction method based on fractal-guided attention, characterized in that, Includes the following steps: Obtain feature maps from remotely sensed images; For the obtained feature map, the position coordinates and feature vectors of the feature map are obtained respectively, and mapped to complex inputs. The complex inputs of position coordinates and feature vectors at the same spatial location are concatenated to obtain the complex inputs of the learnable Mandelbrot fractal dynamic system. Based on the gating function, an image-driven learnable Mandelbrot fractal iterative model is constructed, and the complex input of the learnable Mandelbrot fractal dynamic system is fed into the learnable Mandelbrot fractal iterative model. The learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M; By embedding the fractal structure response map M as an attention bias term into the Transformer attention mechanism, fractal-guided attention is obtained. Based on the calculation results of fractal-guided attention, the deep feature map of the input feature map is obtained.
2. The method as described in claim 1, characterized in that, In the step of obtaining the feature map of the remote sensing image, the feature map is obtained by performing initial feature extraction on the remote sensing image using a feature extraction module with N deep stacked layers.
3. The method as described in claim 1, characterized in that, The complex inputs of the learnable Mandelbrot fractal dynamical system are obtained through the following steps: For the input feature map Its scale is The number of channels is C, and the feature map is defined. The spatial position is p=(u,v), and the total number of spatial positions is N=H×W; Normalize the horizontal and vertical coordinates of any position p to [ 1,1] yields the normalized coordinates ( , Complex inputs for constructing position coordinates based on normalized coordinates: in, The imaginary unit; Based on the feature vector corresponding to position p in the feature map The complex input to obtain the eigenvectors is: in, Real part mapping function, Imaginary part mapping function; By adding the complex inputs of the coordinates at the same location and the complex inputs of the eigenvectors, we obtain the complex inputs of the learnable Mandelbrot fractal dynamical system.
4. The method according to claim 1, characterized in that, Based on the gating function, an image-driven, learnable Mandelbrot fractal iterative model is constructed, specifically including the following models: in, This represents the complex state of position p in the nth iteration. Fractal gating values generated from image features The input is the complex number corresponding to position p.
5. The method according to claim 1, characterized in that, The learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M. The specific steps include: After completing the maximum number of fractal iterations In the fractal iteration process within the next few iterations, for each spatial location p, it is determined whether the complex state at that spatial location has escaped, where the escape determination condition is: in Complex state The model, The escape threshold; Define the escape iteration function for position p: in, This represents the number of iterations at which position p first satisfies the escape condition; If within the maximum number of iterations If no escape occurred, then order This leads to the escape time map representing the total number of position iterations. ; Normalizing the escape time plot yields the fractal structure response plot: in Let be the normalized fractal response at position p.
6. The method according to claim 1, characterized in that, The fractal structure response map M is embedded as an attention bias term into the Transformer attention mechanism to obtain fractal-guided attention, which specifically includes the following steps: Mapping the fractal structure response map M to a fractal attention bias matrix ; Adding the fractal attention bias matrix to the standard attention score yields fractal-guided attention: in Here, Q represents the learnable weights, K represents the key matrix, V represents the value matrix, and d represents the query and key vector dimensions.
7. The method according to claim 1, characterized in that, Based on the calculation results of fractal-guided attention, a deep feature map of the input feature map is obtained, specifically including: The calculation results of fractal-guided attention are sequentially subjected to residual connection and layer normalization operations with the input feature map to obtain the deep feature map.
8. The method according to claim 1, characterized in that, It also includes inputting the obtained deep feature map into the detection head of the downstream task network to obtain the task result. The downstream task network includes a classification network, a localization network, and a segmentation network.
9. A remote sensing image feature extraction system based on fractal-guided attention, characterized in that, include: A fractal-guided attention module; the fractal-guided attention module performs the following steps: For the feature map obtained from the remote sensing image, the position coordinates and feature vectors of the feature map are obtained respectively and mapped to complex inputs. The complex inputs of position coordinates and feature vectors at the same spatial location are concatenated to obtain the complex inputs of the learnable Mandelbrot fractal dynamic system. Based on the gating function, an image-driven learnable Mandelbrot fractal iterative model is constructed, and the complex input of the learnable Mandelbrot fractal dynamic system is fed into the learnable Mandelbrot fractal iterative model. The learnable Mandelbrot fractal iterative model is iterated, and the escape time T is mapped to the fractal structure response map M; By embedding the fractal structure response map M as an attention bias term into the Transformer attention mechanism, fractal-guided attention is obtained. The calculation results of fractal-guided attention are sequentially subjected to residual connection and layer normalization operations with the input feature map to obtain the deep feature map.
10. The system according to claim 9, characterized in that, The complex inputs of the Mandelbrot fractal dynamical system can be learned, specifically through the following steps: For the input feature map Its scale is The number of channels is C, and the feature map is defined. The spatial position is p=(u,v), and the total number of spatial positions is N=H×W; Normalize the horizontal and vertical coordinates of any position p to [ 1,1] yields the normalized coordinates ( , Complex inputs for constructing position coordinates based on normalized coordinates: in, The imaginary unit; Based on the feature vector corresponding to position p in the feature map The complex input to obtain the eigenvectors is: in, Real part mapping function, Imaginary part mapping function; By adding the complex inputs of the coordinates at the same location and the complex inputs of the eigenvectors, we obtain the complex inputs of the learnable Mandelbrot fractal dynamical system.