A remote sensing image scene classification method and system based on double-filter cooperation
By using a dual-filter collaborative optimization module to process remote sensing image feature maps, the contradiction between noise suppression and edge preservation in remote sensing image scene classification is resolved. This achieves efficient background noise filtering and edge enhancement, thereby improving the accuracy and anti-interference capability of remote sensing image scene classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 耕宇牧星(北京)空间科技有限公司
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-23
AI Technical Summary
Existing remote sensing image scene classification methods tend to lose key edge information when suppressing background noise, resulting in insufficient feature discrimination power and difficulty in distinguishing scene categories with similar textures but different structures.
A dual-filter collaborative approach is adopted, combining Gaussian blur filtering and Gaussian Laplace filtering. The remote sensing image feature map is processed through a dual-channel collaborative optimization module to purify background noise and enhance the edge structure of ground features. The zero-crossing detection characteristic of Gaussian Laplace filtering is used to accurately capture edge features.
It significantly improves the accuracy and anti-interference ability of remote sensing image scene classification, and can retain the edge structure of key ground objects while filtering out redundant background textures, generating a clean and clear scene representation, and improving the classification accuracy of complex scenes.
Smart Images
Figure CN122265858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, remote sensing image processing and artificial intelligence, and in particular to a remote sensing image scene classification method and system based on dual-filter cooperation. Background Technology
[0002] Remote sensing images have broad application prospects in fields such as national defense and security, environmental monitoring, and urban planning. Scene classification technology for remote sensing images, as a core means of understanding the semantic content of high-resolution remote sensing images, has always been a research focus in the field of remote sensing. Current mainstream remote sensing image scene classification methods are mainly based on convolutional neural networks (CNNs) to extract features, obtaining a global feature representation of the image by stacking convolutional layers and pooling layers, and then determining the scene category. However, most existing scene classification methods borrow from natural image processing techniques, ignoring the noise interference introduced by the extremely complex background textures in remote sensing images, and the high dependence of scene classification on the fine edge structure of ground features, resulting in significant bottlenecks in feature extraction under complex scenes.
[0003] In deep learning models, feature extraction often faces an inherent contradiction between noise suppression and edge preservation. On the one hand, remote sensing images often contain a large amount of high-frequency background noise. For example, in a "port" scene, the flickering waves and light reflections on the sea surface create strong interfering textures; in a "mountainous" scene, cloud cover and fine vegetation textures can confuse the viewer. Existing convolutional networks lack targeted noise filtering mechanisms, easily misclassifying these irrelevant background textures as key features, leading to inter-class confusion (e.g., misclassifying wave textures as some kind of ground feature texture). On the other hand, accurate scene classification highly depends on the edge and contour structure of ground features. For example, the difference between an "overpass" and a "highway" lies mainly in the intersection structure of the roads; the difference between an "industrial area" and a "residential area" lies in the arrangement and contour of the buildings. However, to suppress the aforementioned background noise, traditional methods often employ continuous downsampling or global pooling operations, which, while smoothing the noise, inevitably blurs the key edge details of ground features. This simplistic feature aggregation method leads to the loss of edge information, making it difficult for the model to capture the structured semantics of the scene and thus difficult to distinguish scene categories with similar textures but different structures. Therefore, how to accurately remove high-frequency background noise during feature extraction while sharpening and preserving key ground feature edge structures, and resolving the contradiction between noise interference and detail loss, is a key challenge for improving the accuracy of remote sensing image scene classification.
[0004] Given that existing remote sensing image scene classification models tend to lose key edge information when suppressing background noise, resulting in insufficient feature discrimination power, it is necessary to improve remote sensing image classification methods and design a remote sensing image scene classification method that can simultaneously achieve efficient noise reduction and edge enhancement. Summary of the Invention
[0005] This invention addresses the challenges of severe background noise interference and blurred edge structures in remote sensing image scene classification. It proposes a remote sensing image scene classification method and system based on dual-filter collaboration. By designing a dual-filter collaboration optimization module and optimizing the feature extraction network of the scene classification model based on it, remote sensing image scene classification can simultaneously achieve efficient noise reduction and edge enhancement.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides a remote sensing image scene classification method based on dual-filter cooperation, the method comprising the following steps: S1. Construct a dual-filter collaborative optimization module. The dual-filter collaborative optimization module uses parallel dual-channel Gaussian blur filtering and Gaussian Laplace filtering to optimize the feature map and fuses the output of the dual channels. S2. Build a remote sensing image scene classification model based on the dual-filter cooperative optimization module; S3. Train and test the remote sensing scene classification model for remote sensing image scene classification tasks.
[0008] Furthermore, in S1, the constructed dual-filter collaborative optimization module includes: a Gaussian blur filter module, a Gaussian Laplace filter module, a positional encoding mechanism, a state-space model, a gated GELU activation module and a Softmax activation module, a dynamic convolution module, a feature addition and feature concatenation module, and a matrix multiplication operator. The Gaussian blur filter module and the Gaussian Laplace filter module work complementaryly in the frequency domain; the Gaussian blur filter module focuses on purifying background noise, while the Gaussian Laplace filter module focuses on enhancing edge structures.
[0009] Furthermore, the Gaussian blur filtering module uses a Gaussian kernel to perform weighted smoothing on the feature map. This module targets unstructured background noise in remote sensing images by using low-pass filtering to suppress high-frequency interference, purify the feature background, and highlight the main subject area. The Gaussian Laplace filter module performs second-order differential operations on the feature map and utilizes the zero-crossing detection characteristic of the Gaussian Laplace operator to accurately capture abrupt changes in the edges of smoothed features, generating an edge feature map and compensating for the loss of detail caused by the smoothing operation.
[0010] Furthermore, the unstructured background noise includes: water ripples, grass textures, and cloud noise; the abrupt changes in the edges of ground features include: building outlines, road boundaries, and runway lines.
[0011] Furthermore, in S1, the specific process by which the dual-filter collaborative optimization module optimizes the feature map includes: For feature maps First, processing is performed based on Gaussian blur filtering and Gaussian Laplace filtering:
[0012]
[0013] in, This is a Gaussian filter module. This is a Gaussian Laplace filter module. and Feature map information obtained by filtering the corresponding channel; Then, feature addition and dynamic convolution are used to fuse and refine the feature information:
[0014]
[0015] in, To fuse feature maps, This is the feature map refined by dynamic convolution. Indicates a dynamic convolution module; Simultaneously, a state-space model combined with positional encoding is used to analyze the feature map. Encode alignment is performed to optimize the feature map representation of the spatial structure of complex land features:
[0016]
[0017] in, The feature map represents the encoded alignment. This indicates the positional encoding mechanism. This indicates a feature concatenation operation along the channel dimension. Represents a visual state space model. The feature map is processed by the state-space model; Then, a dual activation mechanism is adopted, filtering information through the gated GELU activation module and the Softmax activation module:
[0018]
[0019] in, These are the weight matrices output by gated GELU activation and Softmax activation, respectively; Finally, the weight matrix output by the dual activation mechanism is... and Feature maps after stepwise and dynamic convolution filtering Perform matrix multiplication:
[0020] in, This is the feature map output by the dual-filter collaborative optimization module.
[0021] Secondly, the present invention also provides a remote sensing image scene classification system based on dual-filter cooperation. The system includes a feature extraction module, a dual-filter cooperation optimization module, and a classification module. The above-mentioned method is used to classify remote sensing image scenes.
[0022] Thirdly, the present invention also provides an electronic device, including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to perform the above-described method.
[0023] As can be seen from the above technical solution, compared with the prior art, the beneficial effects of the present invention include: 1. This invention can significantly enhance semantically discernible geometric structural information in a scene while filtering out redundant background textures, generating a clean and clear scene representation. The remote sensing image scene classification model built based on a dual-filter collaborative optimization module effectively solves the contradiction between noise suppression and edge preservation in traditional methods, significantly improving classification accuracy for categories with complex textures and critical structures. This innovation provides a more reliable and efficient solution to the challenge of refined recognition of complex scenes in high-resolution remote sensing imagery, greatly enhancing the model's ability to perceive ground feature details.
[0024] 2. The dual-filter collaborative optimization module proposed in this invention changes the traditional single-dimensional feature aggregation method of convolutional networks. This module effectively smooths non-semantic high-frequency background textures (such as water ripples and fine grass) through a Gaussian blur filter channel, while explicitly capturing and activating the edge structures of key elements within the scene (such as building outlines and road networks) by utilizing the zero-crossing property of the Laplacian of Gaussian (LoG) operator. This mechanism enables the model to reconstruct a scene topology with a clean background and clear structure in the feature space, ensuring that the feature aggregation process is based on the real geometric structure of the ground features, rather than being dominated by disordered background noise, thus laying a precise structural foundation for subsequent global scene understanding.
[0025] 3. Thanks to the dual filtering mechanism's strong protection of edge information and effective removal of background noise, this invention can extract a global representation with high signal-to-noise ratio and significant geometric features. When faced with scenes in remote sensing images that are complex in texture and heavily interfered with (such as "mountainous areas" obscured by clouds and fog, or "ports" disturbed by waves), the model can resist the erosion of environmental noise while fully preserving the subtle structural differences needed to distinguish different scene categories (such as the road network details that distinguish "highway" from "ordinary road"). This ensures that the model can still obtain highly consistent and distinctive semantic responses in complex environments, effectively solving the feature confusion problem caused by edge blurring or background interference in existing technologies, and significantly improving the accuracy and anti-interference ability of remote sensing image scene classification.
[0026] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.
[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0030] Figure 1 This is a schematic diagram of the remote sensing image scene classification method based on dual-filter cooperation provided in an embodiment of the present invention.
[0031] Figure 2 This is a schematic diagram of the dual-filter collaborative optimization module structure provided in an embodiment of the present invention.
[0032] Figure 3 This is a schematic diagram of the remote sensing image scene classification model structure based on dual-filter cooperative optimization provided in an embodiment of the present invention.
[0033] Figure 4 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.
[0035] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0036] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0037] See Figures 1 to 3 As shown, this invention discloses a remote sensing image scene classification method based on dual-filter cooperation, aiming to solve the problem of insufficient capabilities of traditional methods in the feature modeling stage. The method of this invention includes the following steps: Step 1: Build a dual-filter collaborative optimization module; Step 2: Build an improved remote sensing scene classification model that incorporates a dual-filter collaborative optimization module; Step 3: Train and test the improved remote sensing scene classification model, which incorporates the dual-filter collaborative optimization module, for remote sensing image scene classification tasks.
[0038] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Step one above: Build a dual-filter collaborative optimization module; the structure is as follows: Figure 2 As shown, its core modules include a Gaussian blur filter module, a Gaussian-Laplace filter module, a gated GELU activation module and a Softmax activation module, a position encoding mechanism, a state space model module, as well as feature addition and feature concatenation, and a matrix multiplication operator.
[0039] First, in order to eliminate high-frequency noise introduced by background clutter in remote sensing images and enhance the texture details of complex remote sensing scene features, we optimize the feature map by combining Gaussian-Laplace filtering and Gaussian blur filtering.
[0040] Feature maps derived from deep feature extraction Its size is First, processing is performed based on Gaussian-Laplace filtering and Gaussian fuzzy filtering. The calculation process can be formally represented as follows:
[0041]
[0042] in, This is a Gaussian filter module. For Gaussian-Laplace filtering, and The feature map information obtained is filtered for the corresponding branch. Furthermore, feature addition and dynamic convolution are used to fuse and refine the feature information:
[0043]
[0044] in, For dynamic convolution modules, To fuse the feature information after Gaussian-Laplace filtering After Gaussian filtering Feature map, This is the feature map refined by dynamic convolution.
[0045] On the other hand, a state-space model combined with positional encoding is used to process the original feature map. Encoding alignment is performed to clarify the spatial enhancement localization of texture responses and optimize the spatial structural representation of complex terrain features in the feature maps. The computation process can be formally represented as follows:
[0046]
[0047] in, The feature map represents the encoded alignment. Positional encoding mechanisms typically employ relative positional encoding or sinusoidal positional encoding, based on the feature map. The size information is given as a position vector representation along the length and width dimensions. This represents a feature concatenation operation along the channel dimension, used to integrate feature information after positional encoding and Gaussian-Laplace filtering. . It represents a visual state space model, and based on the state space scanning mechanism, it mines fine-grained representations of feature maps in spatial structure. This is the feature map after processing by the state-space model.
[0048] Then, a dual activation mechanism is adopted, combining gated convolution and... Further filtering of information to highlight the highlight features under different activation suppressions can be formally represented as follows:
[0049]
[0050] in, Represents the gating activation mechanism. for activation. This is the weight matrix output by each activation function.
[0051] Furthermore, the weight matrix output by the dual activation mechanism... and Feature maps after stepwise and dynamic convolution filtering Perform matrix multiplication to complete the attention calculation:
[0052] in, This is the feature map output after calculation by the dual-filter collaborative optimization module.
[0053] In summary, the dual-filter collaborative optimization module has completed its calculations.
[0054] Step two above: Build an improved remote sensing scene classification model that incorporates a dual-filter collaborative optimization module, such as... Figure 3 As shown, an improved remote sensing scene classification model combining dual filtering collaboration is constructed.
[0055] First, for the input remote sensing image G, which has dimensions H×W and 3 channels, feature extraction is performed through a series of stacked convolutional blocks or Transformer blocks to obtain a series of feature maps. We define These are shallow, medium, and deep feature maps obtained after feature extraction, each containing feature information.
[0056] The proposed dual-filter cooperative optimization module will then be used to process shallow, medium, and deep feature maps, formally represented as:
[0057]
[0058] in, This indicates the dual-filter collaborative optimization module. This represents a feature extraction unit. Feature extraction units are typically stacked to form a feature extraction network.
[0059] at last, It is used for classification and is formally represented as:
[0060] in, This represents the classification information obtained from the prediction. This represents the classification function.
[0061] At this point, the improved remote sensing scene classification model, combined with the dual-filter collaborative optimization module, has been successfully built.
[0062] Step three above: Training and testing the network for remote sensing image scene classification tasks; after building the scene classification model with the dual-filter collaborative optimization module, the network is trained for remote sensing image scene classification tasks using focal loss, formally represented as:
[0063] in, This represents focus loss, used for classification prediction in remote sensing image scene classification tasks. 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 improved remote sensing scene classification model incorporating the dual-filter collaborative optimization module is obtained.
[0064] Furthermore, the improved remote sensing scene classification model, trained using the dual-filter collaborative optimization module, is tested on the remote sensing images to be tested, formally represented as follows:
[0065] in, This represents the improved remote sensing scene classification model that has been trained and incorporates the dual-filter collaborative optimization module. and These represent the remote sensing image to be tested and its corresponding classification result, respectively.
[0066] As described in the above embodiments, those skilled in the art will understand that the present invention provides a remote sensing image scene classification method based on dual-filter collaboration. This method proposes an innovative solution to the problems of severe background noise interference and blurred edge structures in remote sensing image scene classification. By designing a dual-filter collaboration optimization module and optimizing the feature extraction network of the scene classification model based on it, the feature confusion problem of existing methods when processing complex texture scenes can be solved.
[0067] This module breaks through the limitations of traditional convolutional networks' single-filter processing, innovatively introducing Gaussian filtering and Laplace-Gaussian (LoG) filtering for dual-channel collaborative processing. It leverages the complementary characteristics of these two filters in the frequency domain to decouple and reconstruct features. Specifically, this method includes the following core mechanisms: Gaussian Smoothing Channel: This channel uses a Gaussian kernel to perform weighted smoothing on the feature map. It primarily targets unstructured background noise in remote sensing images (such as water ripples, grass textures, and cloud noise), effectively suppressing high-frequency interference through low-pass filtering, purifying the background, and highlighting the main subject area.
[0068] The LoG Edge Enhancement Channel utilizes the Laplacian of Gaussian operator to perform second-order differential operations on features in parallel. Leveraging the zero-crossing detection property of the LoG operator, this channel accurately captures abrupt changes in smoothed ground edges (such as building outlines, road boundaries, and runway lines), generating high-response edge feature maps to compensate for the loss of detail caused by the smoothing operation.
[0069] Through this dual collaborative mechanism of "Gaussian noise reduction preserving the subject" and "LoG differential edge enhancement," this invention can significantly enhance semantically discernible geometric structural information in a scene while filtering out redundant background textures, generating a pure and clear scene representation. The remote sensing image scene classification model built based on the dual-filter collaborative optimization module effectively solves the contradiction between noise suppression and edge preservation in traditional methods, significantly improving classification accuracy for categories with complex textures and critical structures. This innovation provides a more reliable and efficient solution to the challenge of refined recognition of complex scenes in high-resolution remote sensing imagery, greatly enhancing the model's ability to perceive ground details.
[0070] Furthermore, the present invention also provides a remote sensing image scene classification system based on dual-filter cooperation. The system includes a feature extraction module, a dual-filter cooperation optimization module, and a classification module. The above-described method is used to classify remote sensing image scenes.
[0071] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment, and will not be repeated here.
[0072] Additionally, refer to Figure 4 As shown, this embodiment of the invention also provides an electronic device, which may include a processor, a memory, a communication bus and a communication interface, and may also include a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the above-described method.
[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, software systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0074] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0075] 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 scene classification method based on dual-filter cooperation, characterized in that, The method includes the following steps: S1. Construct a dual-filter collaborative optimization module. The dual-filter collaborative optimization module uses parallel dual-channel Gaussian blur filtering and Gaussian Laplace filtering to optimize the feature map and fuses the output of the dual channels. S2. Build a remote sensing image scene classification model based on the dual-filter cooperative optimization module; S3. Train and test the remote sensing scene classification model for remote sensing image scene classification tasks.
2. The method according to claim 1, characterized in that, In S1, the dual-filter collaborative optimization module includes: a Gaussian blur filter module, a Gaussian Laplace filter module, a position encoding mechanism, a state space model, a gated GELU activation module and a Softmax activation module, a dynamic convolution module, a feature addition and feature concatenation module, and a matrix multiplication operator; wherein: the Gaussian blur filter module and the Gaussian Laplace filter module work complementaryly in the frequency domain, the Gaussian blur filter module focuses on purifying background noise, and the Gaussian Laplace filter module focuses on strengthening edge structures.
3. The method according to claim 2, characterized in that, The Gaussian blur filtering module uses a Gaussian kernel to perform weighted smoothing on the feature map. This module targets unstructured background noise in remote sensing images by using low-pass filtering to suppress high-frequency interference, purify the feature background, and highlight the main area. The Gaussian Laplace filter module performs second-order differential operations on the feature map and utilizes the zero-crossing detection characteristic of the Gaussian Laplace operator to accurately capture abrupt changes in the edges of smoothed features, generating an edge feature map and compensating for the loss of detail caused by the smoothing operation.
4. The method according to claim 3, characterized in that, The unstructured background noise includes: water ripples, grass textures, and cloud noise; the abrupt changes in the edges of ground features include: building outlines, road boundaries, and runway lines.
5. The method according to claim 2, characterized in that, In step S1, the specific process of the dual-filter cooperative optimization module optimizing the feature map includes: For feature maps First, processing is performed based on Gaussian blur filtering and Gaussian Laplace filtering: in, This is a Gaussian filter module. This is a Gaussian Laplace filter module. and Feature map information obtained by filtering the corresponding channel; Then, feature addition and dynamic convolution are used to fuse and refine the feature information: in, To fuse feature maps, This is the feature map refined by dynamic convolution. Indicates a dynamic convolution module; Simultaneously, a state-space model combined with positional encoding is used to analyze the feature map. Encode alignment is performed to optimize the feature map representation of the spatial structure of complex land features: in, The feature map represents the encoded alignment. This indicates the positional encoding mechanism. This indicates a feature concatenation operation along the channel dimension. Represents a visual state space model. The feature map is processed by the state-space model; Then, a dual activation mechanism is adopted, filtering information through the gated GELU activation module and the Softmax activation module: in, These are the weight matrices output by gated GELU activation and Softmax activation, respectively; Finally, the weight matrix output by the dual activation mechanism is... and Feature maps after stepwise and dynamic convolution filtering Perform matrix multiplication: in, This is the feature map output by the dual-filter collaborative optimization module.
6. The method according to claim 1, characterized in that, In step S3, focus loss is used to train the remote sensing image scene classification task.
7. A remote sensing image scene classification system based on dual-filter cooperation, characterized in that, The system includes a feature extraction module, a dual-filter collaborative optimization module, and a classification module, and performs remote sensing image scene classification using the method described in any one of claims 1–6.
8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to perform the method as described in any one of claims 1-6.