Multi-level semantic information aggregation method and device, equipment and medium

Abstract

The invention provides a multi-level semantic information aggregation method and device, equipment and a medium, and the method comprises the steps: extracting image high-level features of different scales, combining the image high-level features of different scales, and employing a self-attention mechanism to construct a high-precision DFG mapping graph; the DFG mapping graph is embedded into a guide aggregation module in the convolutional neural network for feature fusion, and a fusion feature graph of multi-level semantic information is obtained; and multi-scale fine-grained features are extracted from the fused feature map by adopting convolution operations of different scales, feature fusion is performed on the multi-scale fine-grained features, and a final feature map for remote sensing image change detection is generated. According to the method, the deep features of different scales and a self-attention mechanism are combined, high-precision depth feature guide mapping is constructed, the problems of semantic difference, detail loss and noise interference existing in a traditional method are effectively solved, and the precision and robustness of remote sensing image change detection are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and remote sensing image processing, and in particular to a multi-level semantic information aggregation method, device, equipment and medium. Background Art

[0002] Remote sensing image change detection technology has important applications in land use monitoring, disaster assessment, and ecological and environmental protection. Its primary task is to detect surface changes by analyzing differences between multi-temporal remote sensing images. However, due to the complexity of remote sensing images, including multispectral and multi-scale characteristics, as well as noise interference, traditional change detection methods still have significant room for improvement in terms of accuracy and robustness.

[0003] Currently, multi-level semantic information aggregation methods used for change detection in remote sensing images are mainly divided into two categories: those based on traditional image processing techniques and those based on deep learning techniques. Traditional methods generally rely on hand-designed feature extraction techniques, such as grayscale difference and change vector analysis. These methods have low computational complexity but are poorly adaptable to scene changes and have difficulty dealing with complex landform categories or environmental noise. Deep learning-based methods extract multi-level features through models such as convolutional neural networks, effectively improving the performance of change detection. However, existing deep learning methods often have difficulty efficiently representing spatiotemporal features when processing multi-temporal images, and they ignore the differences in semantic information when aggregating multi-level features. This can lead to problems such as pseudo-change noise and insufficient integrity of changed objects in change detection results.

[0004] Furthermore, in deep learning-driven multi-level semantic information aggregation methods, how to fully utilize the semantic information of multi-level features remains one of the key difficulties. Shallow features typically contain rich spatial detail information, while deep features focus more on global semantic information. However, there is a significant semantic gap between the two, and direct fusion may lead to information loss or noise enhancement. At the same time, traditional convolution operations have limited perception and are unable to capture the global contextual information of the changed area, resulting in insufficient representation of the changed area. Therefore, how to design a method that can both efficiently aggregate multi-level feature semantics and effectively suppress pseudo-change noise is of great significance for improving the accuracy and robustness of change detection. Summary of the Invention

[0005] In view of this, the embodiments of the present invention provide a multi-level semantic information aggregation method, device, equipment and medium, which combine deep features and self-attention mechanism to effectively improve the utilization and robustness of multi-level feature information, while reducing the risk of noise interference and loss of detail information, significantly improving the accuracy and stability of change feature decoding, and achieving the purpose of improving the accuracy and robustness of remote sensing image change detection.

[0006] The embodiment of the present invention provides the following technical solution: a multi-level semantic information aggregation method for remote sensing image change detection, comprising:

[0007] Extract high-level features of images at different scales, combine them, and use a self-attention mechanism to construct a high-precision DFG map.

[0008] The DFG map is embedded into the guided aggregation module of the convolutional neural network to perform feature fusion to obtain a fused feature map of multi-level semantic information;

[0009] Convolution operations of different scales are used to extract multi-scale fine-grained features from the fused feature map, and feature fusion is performed on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

[0010] According to one embodiment of the present invention, combining the high-level image features of different scales and using a self-attention mechanism to construct a high-precision DFG map includes:

[0011] Performing spatial alignment processing and channel dimension compression processing on the high-level image features of different scales so that the high-level image features of different scales are consistent in both spatial dimension and channel dimension;

[0012] The processed features are fused to obtain an overall feature map, and then a convolution operation is performed on the overall feature map to obtain an enhanced feature map;

[0013] The enhanced feature map is processed using a self-attention mechanism to construct a high-precision DFG mapping map.

[0014] According to one embodiment of the present invention, a self-attention mechanism is used to process the enhanced feature map to construct a high-precision DFG map, further comprising:

[0015] The enhanced features processed by the self-attention mechanism are compressed into a single-channel representation through a convolution operation, and the high-precision DFG map is generated through a Sigmoid activation function.

[0016] According to one embodiment of the present invention, the DFG map is embedded in a guided aggregation module in a convolutional neural network to perform feature fusion to obtain a fused feature map of multi-level semantic information, including:

[0017] Performing spatial scale alignment processing on input features of two different levels in the convolutional neural network;

[0018] Perform a feature concatenation operation on the two aligned input features along the channel dimension to generate a preliminary fusion feature map;

[0019] The DFG mapping graph is combined with the preliminary fusion feature graph to obtain the fusion feature graph of the multi-level semantic information.

[0020] According to one embodiment of the present invention, the method further includes:

[0021] The DFG mapping map is combined with the preliminary fusion feature map through element-level multiplication operation, and a convolution operation is performed on the fusion feature map combined with the DFG mapping to further compress the channel dimension of the feature and obtain the fusion feature map of the multi-level semantic information.

[0022] According to one embodiment of the present invention, convolution operations of different scales are used to extract multi-scale fine-grained features from the fused feature map, and the multi-scale fine-grained features are subjected to feature fusion to generate a final feature map for remote sensing image change detection, including:

[0023] Using convolution kernels of different scales to extract fine-grained information of the changed area in the fused feature map to obtain the multi-scale fine-grained features of the changed area;

[0024] The obtained multi-scale fine-grained features are fused by summing and averaging, and then the fused multi-scale fine-grained features are nonlinearly transformed using the ReLU activation function. Finally, a convolution operation is performed to generate the final feature map for remote sensing image change detection.

[0025] According to one embodiment of the present invention, three convolution kernels of different scales, 3×3, 3×1 and 1×3, are used to extract the fine-grained information of the horizontal, vertical and local areas in the fusion feature map to obtain the multi-scale fine-grained features of the changed area.

[0026] The present invention also provides a multi-level semantic information aggregation device for remote sensing image change detection, comprising:

[0027] The DFG mapping construction module is used to extract high-level image features at different scales, combine the high-level image features at different scales, and use the self-attention mechanism to construct a high-precision DFG mapping graph;

[0028] A first feature fusion module is used to embed the DFG map into the guided aggregation module in the convolutional neural network to perform feature fusion, thereby obtaining a fused feature map of multi-level semantic information;

[0029] The second feature fusion module is used to extract multi-scale fine-grained features from the fused feature map by using convolution operations of different scales, and perform feature fusion on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

[0030] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the multi-level semantic information aggregation method for remote sensing image change detection is implemented.

[0031] The present invention also provides a computer-readable storage medium storing a computer program for executing the above-mentioned multi-level semantic information aggregation method for remote sensing image change detection.

[0032] Compared with the prior art, the beneficial effects that can be achieved by at least one of the above-mentioned technical solutions adopted in the embodiments of this specification include at least the following: The embodiments of the present invention propose a multi-level semantic information aggregation method for remote sensing image change detection. By combining deep features and self-attention mechanisms, the utilization and robustness of feature information are significantly improved, the risk of noise interference and detail loss is reduced, and the accuracy and stability of change feature decoding are improved. Compared with the traditional FPN decoding architecture, the present invention effectively solves the problem of semantic differences between feature levels, and greatly improves the effect of change detection through fine feature alignment and fusion, as well as the introduction of self-attention mechanisms. The use of multi-scale convolution operations to extract fine-grained change features enhances the spatial detail representation capability of the change area, which significantly improves the accuracy and robustness of image change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 1 is a flow chart of a multi-level semantic information aggregation method for remote sensing image change detection according to an embodiment of the present invention;

[0035] Figure 2 is a schematic structural diagram of a DFG mapping diagram in an embodiment of the present invention;

[0036] Figure 3 is a schematic structural diagram of a guidance aggregation module in an embodiment of the present invention;

[0037] Figure 4 2 is a schematic diagram of the structure of a multi-level semantic information aggregation device for remote sensing image change detection according to an embodiment of the present invention;

[0038] Figure 5 It is a structural schematic diagram of the computer device of the present invention. DETAILED DESCRIPTION

[0039] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0040] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0041] To overcome the drawback of the traditional Feature Pyramid Network (FPN) decoding architecture in aggregating multi-level semantic information, which is that it does not fully consider the semantic differences between features at different levels and simply superimposes multi-level features, resulting in a decrease in information representation capability. In addition, due to the semantic gaps between feature levels and the loss of detail information, relying solely on convolution operations makes it difficult to capture long-range global information, which in turn affects the accuracy of the generated guidance map. Therefore, an embodiment of the present invention provides a multi-level semantic information aggregation method for remote sensing image change detection.

[0042] like Figure 1 As shown, an embodiment of the present invention provides a multi-level semantic information aggregation method for remote sensing image change detection, including:

[0043] S101. Extract high-level image features at different scales, combine the high-level image features at different scales, and construct a high-precision DFG map using a self-attention mechanism;

[0044] S102. Embedding the DFG map into the guided aggregation module (GA module) in the convolutional neural network for feature fusion to obtain a fusion feature map of multi-level semantic information;

[0045] S103. Use convolution operations of different scales to extract multi-scale fine-grained features from the fused feature map, and perform feature fusion on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

[0046] The embodiment of the present invention proposes a multi-level semantic information aggregation method combining advanced features and self-attention mechanism, using and The advantages of combining two different scales of deep high-level spatiotemporal feature fusion (STFF) features with the self-attention mechanism to construct DFG mapping are:

[0047] 1. Two high-level features at different scales provide richer variation information and detailed features;

[0048] 2. The self-attention mechanism can capture the global information representation of changing features and effectively reduce the interference of noise;

[0049] 3. Avoid the influence of noise such as pseudo changes in shallow features on change decoding and improve the robustness of the results.

[0050] During the change feature decoding phase, the proposed DFG map is embedded into the GA module within the convolutional neural network to enhance the aggregation of multi-level semantic information and further reduce pseudo-changes and noise interference. At the same time, convolution operations at different scales are used to extract fine-grained change features, enhancing the ability to represent spatial details of change information, thereby improving the integrity of the changed objects and the reconstruction accuracy.

[0051] According to some embodiments of the present invention, in S101, the high-level image features of different scales are combined to construct a high-precision DFG map using a self-attention mechanism. Specifically, the following steps are performed:

[0052] S1.1: Spatially align high-level features at different scales and compress the channel dimension to maintain consistency.

[0053] S1.2: The aligned features are fused to extract deeper features to enhance the representation capability.

[0054] S1.3: To overcome the limited receptive field of traditional convolutional models, a self-attention mechanism (SA) is introduced after feature extraction. This mechanism captures the global context of the features to enhance the model's global perception of the input image. The self-attention mechanism enables the model to focus on key task-related features while suppressing irrelevant detail noise.

[0055] S1.4: Compress the fused features into a single channel and generate a DFG map to provide guidance signals for subsequent semantic information aggregation.

[0056] In specific implementation, the specific implementation method of S101 includes the following steps:

[0057] (1) For lower-resolution features, upsampling is performed to make them consistent with higher-resolution features in spatial dimensions;

[0058] (2) Compress the channel dimension of the feature through 1×1 convolution operation to ensure the consistency of features of different scales in the channel dimension, which is convenient for subsequent fusion processing;

[0059] (3) Based on the fused feature map, 3×3 convolution is used to further extract deeper features to enhance the expressiveness and discrimination ability of the features;

[0060] (4) The self-attention mechanism is introduced to process the fused feature maps, enhancing the model's global perception of the input image, enabling the model to focus on key features related to the task while suppressing detail noise that is irrelevant to change detection;

[0061] (5) Using 1×1 convolution, the fused features processed by the self-attention mechanism are compressed into a single-channel representation to further reduce the redundant information of the feature dimension;

[0062] (6) A high-precision DFG map is generated through the Sigmoid activation function, providing important guidance signals for the efficient aggregation of subsequent semantic information.

[0063] In some embodiments of the present invention, in S102, the DFG map is embedded into the GA module in the convolutional neural network to perform feature fusion, so as to reduce noise interference and achieve efficient aggregation of multi-level semantic information, thereby obtaining a fused feature map of multi-level semantic information. Specifically, the following steps are included:

[0064] S2.1: Perform spatial scale alignment on the input features of two different levels in the convolutional neural network to ensure their consistency in the spatial dimension, and then perform feature concatenation along the channel dimension to form a preliminary fusion feature map.

[0065] S2.2: Map the generated DFG to F DFG The DFG map is combined with the preliminary fused feature map through element-wise multiplication as prior information embedded in the feature fusion process. This process can correct the position information of the changing objects while reducing the interference of pseudo-changing noise.

[0066] S2.3: Compress the number of channels of the fused features to reduce the computational complexity of the model and integrate the high-dimensional feature space information into a compact form for subsequent processing.

[0067] In specific implementation, the specific implementation method of S102 includes the following steps:

[0068] (1) For high-level features, upsampling operations are performed to keep them consistent with low-level features in terms of spatial dimensions, ensuring that feature mismatches will not occur due to differences in spatial dimensions during subsequent fusion.

[0069] (2) Connect the aligned high-level features and low-level features along the channel dimension to generate a preliminary fusion feature map to enhance the expressiveness of feature information;

[0070] (2) The DFG map is combined with the preliminary fusion feature map through element-wise multiplication operation to correct the position information of the changed object and suppress the interference of pseudo-change noise on the features;

[0071] (3) Apply a 1×1 convolution operation to the fused feature map combined with the DFG map to further compress the channel dimension of the feature.

[0072] In some embodiments of the present invention, in S103, convolution operations of different scales are used to extract multi-scale fine-grained features from the fused feature map, and the multi-scale fine-grained features are fused to generate a final feature map for remote sensing image change detection. The use of convolution operations of different scales to extract fine-grained change features enhances the spatial details of the change information representation and promotes the complete reconstruction of the changed object. Specifically, it includes:

[0073] S3.1: Use convolution kernels of different scales to extract multi-scale fine-grained features from the fused features, process the horizontal, vertical and local area change information respectively, and fuse the multi-scale features by summing and averaging to enhance the integrity of the reconstruction of spatial details.

[0074] S3.2: Use ReLU activation function to perform nonlinear transformation on multi-scale features to improve the model's fitting ability. Then use 1×1 convolution to further optimize the feature representation and generate the final GA feature map F GA .

[0075] S3.3: Through the above steps, we can effectively achieve efficient aggregation of multi-level semantic information, reduce the interference of pseudo-change noise, enhance the ability to extract fine-grained features of change areas, and provide higher-quality semantic feature representation for subsequent tasks.

[0076] In specific implementation, the specific implementation method of S103 includes the following steps:

[0077] (1) Using convolution kernels of three different scales, including 3×3, 3×1, and 1×3, we extract the change features of horizontal, vertical, and local regions respectively to fully capture the fine-grained information of the change region;

[0078] (2) The feature results extracted by convolution kernels of different scales are fused by summing and averaging, effectively integrating information from various directions and scales and enhancing the integrity of spatial details in the changing area;

[0079] (3) Apply the ReLU activation function to the fused multi-scale features for nonlinear transformation to improve the model’s fitting ability;

[0080] (4) The 1×1 convolution operation is used to further compress the feature dimension and optimize the feature representation to generate the final high-precision GA feature map.

[0081] The present invention will be further described below with reference to the accompanying drawings and implementation examples:

[0082] A multi-level semantic information aggregation method for remote sensing image change detection in this embodiment: Figure 1 As shown, the following steps are included:

[0083] S1: Combine high-level features of different scales and use the self-attention mechanism to build a high-precision DFG mapping, such as Figure 2 shown.

[0084] S1.1: First combine high-level features of different scales and Using upsampling operation Perform spatial alignment and compress the channel dimension of the feature through 1×1 convolution to ensure the consistency of the spatial and channel dimensions of the two features of different scales.

[0085] S1.2: The aligned features are then fused into an overall feature map through feature concatenation, and then subjected to a 3×3 convolution operation to extract deeper features and further enhance the representation capability of the features. The convolution calculation formula is as follows:

[0086]

[0087] Among them, Up represents the upsampling operation, Conv 1×1 Represents 1×1 convolution, Conv 3×3 Represents a 3×3 convolution.

[0088] S1.3: To overcome the limited receptive field of traditional convolutional neural networks, a self-attention mechanism (SA) is introduced after feature extraction. This mechanism captures the global context of the features to enhance the model's global perception of the input image. The self-attention mechanism enables the model to focus on key task-related features while suppressing irrelevant detail noise.

[0089] S1.4: After the self-attention mechanism, the fused feature map is compressed into a single-channel representation using a 1×1 convolution, thereby integrating the multi-dimensional feature information into a compact representation, and the DFG map is generated using the Sigmoid activation function. The calculation expression of the map is as follows:

[0090] F DFG =δ(Conv 1×1 (SA(F Conv ))

[0091] Among them, δ represents the Sigmoid activation function and SA represents the self-attention mechanism.

[0092] S1.5: The final generated DFG map serves as a guiding signal for subsequent multi-level semantic information aggregation, which is used to reduce the interference of pseudo-change noise, alleviate the information loss of high-level features in the fusion process, and enhance the global information representation capability, providing effective prior guidance for the GA module.

[0093] S2: Embed the DFG map into the GA module to reduce noise interference and achieve efficient aggregation of multi-level semantic information, such as Figure 3 As shown on the left;

[0094] S2.1: First, the features of two different levels are and Perform spatial scale alignment to ensure their consistency in spatial dimensions, and then perform feature concatenation along the channel dimension to form a preliminary fusion feature map. Its mathematical expression is:

[0095]

[0096] Among them, Up represents the upsampling operation and Cat represents the feature connection.

[0097] S2.2: Then the generated DFG map F DFG It is embedded into the feature fusion process as prior information and combined with the connection features through element-wise multiplication. This process can correct the position information of the changing object while reducing the interference of pseudo-change noise. The calculation formula is:

[0098]

[0099] in, Represents element-wise multiplication.

[0100] S2.3: The fused feature F′ Cat The 1×1 convolution operation is used to compress the number of channels, thereby reducing the computational complexity of the model and integrating the high-dimensional feature space information into a compact form for subsequent processing. The calculation formula is:

[0101]

[0102] S3: Use convolution operations of different scales to extract fine-grained change features to enhance the spatial details of the change information representation and promote the complete reconstruction of the changed object, such as Figure 3 Shown on the right.

[0103] S3.1: Considering the poor spatial integrity of the changing object, we further use convolution kernels of different scales (3×1, 3×3, 1×3) to fused features. Extract multi-scale fine-grained features. Process horizontal, vertical, and local area change information separately, and fuse multi-scale features by summing and averaging to enhance the integrity of spatial details. The formula is as follows:

[0104]

[0105] Among them, Conv 3×3 、Conv 3×1 and Conv 1×3 They represent 3×3 convolution, 3×1 convolution, and 1×3 convolution respectively.

[0106] S3.2: Then, the ReLU activation function is used to perform nonlinear transformation on the multi-scale features to improve the model’s fitting ability. The feature representation is then further optimized through 1×1 convolution to generate the final GA feature map F GA The mathematical formula is:

[0107]

[0108] Among them, ReLU represents the ReLU activation function.

[0109] The embodiment of the present invention provides a multi-level semantic information aggregation method for remote sensing image change detection, which aims to improve detection accuracy and robustness. The method combines deep features of different scales with the self-attention mechanism to construct a high-precision deep feature guidance (DFG) mapping to reduce pseudo-change noise and enhance feature fusion capabilities. Multi-scale convolution operations are used to extract fine-grained change features, significantly improving the spatial detail integrity of the changed area and the reconstruction accuracy of the changed object. The embodiment of the present invention effectively overcomes the problems of semantic difference, detail loss and noise interference in traditional methods, greatly improves the performance of remote sensing image change detection, and can be widely used in land use monitoring, disaster assessment, ecological and environmental protection and other fields.

[0110] like Figure 4 As shown, the present invention also provides a multi-level semantic information aggregation device 200 for remote sensing image change detection, comprising:

[0111] DFG mapping construction module 201 is used to extract high-level image features of different scales, combine the high-level image features of different scales, and use a self-attention mechanism to construct a high-precision DFG mapping graph;

[0112] A first feature fusion module 202 is configured to embed the DFG map into a guided aggregation module in a convolutional neural network to perform feature fusion, thereby obtaining a fused feature map of multi-level semantic information;

[0113] The second feature fusion module 203 is used to extract multi-scale fine-grained features from the fused feature map by using convolution operations of different scales, and perform feature fusion on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

[0114] In specific implementation, the DFG mapping construction module 201 is mainly used to: perform spatial alignment processing and channel dimension compression processing on the high-level image features of different scales to ensure that the high-level image features of different scales maintain consistency in both spatial and channel dimensions; fuse the processed features to obtain an overall feature map, and then perform a convolution operation on the overall feature map to obtain an enhanced feature map; and use a self-attention mechanism to process the enhanced feature map to construct a high-precision DFG map. The enhanced features processed by the self-attention mechanism are compressed into a single-channel representation through a convolution operation, and the high-precision DFG map is generated through a Sigmoid activation function.

[0115] In specific implementation, the first feature fusion module 202 is mainly used to: align the input features of two different levels in the convolutional neural network in spatial scale; perform feature connection operation on the two aligned input features along the channel dimension to generate a preliminary fusion feature map; combine the DFG mapping map with the preliminary fusion feature map to obtain the fusion feature map of the multi-level semantic information. The DFG mapping map is combined with the preliminary fusion feature map through element-level multiplication operation, and a convolution operation is performed on the fusion feature map combined with the DFG mapping to further compress the channel dimension of the feature to obtain the fusion feature map of the multi-level semantic information.

[0116] In specific implementation, the second feature fusion module 203 is mainly used to: extract the fine-grained information of the change area in the fused feature map using convolution kernels of different scales to obtain the multi-scale fine-grained features of the change area; fuse the obtained multi-scale fine-grained features by summing and averaging, and then use the ReLU activation function to perform nonlinear transformation on the fused multi-scale fine-grained features, and finally perform convolution operation to generate the final feature map for remote sensing image change detection. Among them, three different scales of convolution kernels of 3×3, 3×1 and 1×3 are used to extract the fine-grained information of the horizontal, vertical and local areas in the fused feature map to obtain the multi-scale fine-grained features of the change area.

[0117] The remote sensing image change detection device based on multi-level semantic information aggregation, described in this embodiment of the present invention, combines deep features at different scales with a self-attention mechanism to construct a high-precision deep feature-guided mapping. This effectively overcomes the semantic discrepancies, detail loss, and noise interference issues inherent in traditional methods, significantly improving the accuracy and robustness of remote sensing image change detection. It is widely applicable in fields such as land use monitoring, disaster assessment, and ecological and environmental protection.

[0118] In one embodiment, a computer device is provided, such as Figure 5 As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multi-level semantic information aggregation method for remote sensing image change detection is implemented.

[0119] Specifically, the computer device may be a computer terminal, a server or a similar computing device.

[0120] In this embodiment, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program for executing the above-mentioned multi-level semantic information aggregation method for remote sensing image change detection.

[0121] Specifically, computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable storage media does not include transitory media such as modulated data signals and carrier waves.

[0122] Obviously, those skilled in the art should understand that the various modules or steps of the above-mentioned embodiments of the present invention can be implemented using a general-purpose computing device, they can be concentrated on a single computing device, or distributed across a network composed of multiple computing devices. Alternatively, they can be implemented using program code executable by the computing device, so that they can be stored in a storage device and executed by the computing device. In some cases, the steps shown or described can be performed in a different order than herein, or they can be made into separate integrated circuit modules, or multiple modules or steps can be made into a single integrated circuit module for implementation. Thus, the embodiments of the present invention are not limited to any specific combination of hardware and software.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A multi-level semantic information aggregation method for remote sensing image change detection, characterized in that: include: Extract high-level features of images at different scales, combine them, and use a self-attention mechanism to construct a high-precision DFG map. The DFG map is embedded into the guided aggregation module of the convolutional neural network to perform feature fusion to obtain a fused feature map of multi-level semantic information; Convolution operations of different scales are used to extract multi-scale fine-grained features from the fused feature map, and feature fusion is performed on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

2. The multi-level semantic information aggregation method for remote sensing image change detection according to claim 1, characterized in that: Combining the high-level image features of different scales, a self-attention mechanism is used to construct a high-precision DFG map, including: Performing spatial alignment processing and channel dimension compression processing on the high-level image features of different scales so that the high-level image features of different scales are consistent in both spatial dimension and channel dimension; The processed features are fused to obtain an overall feature map, and then a convolution operation is performed on the overall feature map to obtain an enhanced feature map; The enhanced features processed by the self-attention mechanism are compressed into a single-channel representation through a convolution operation, and the high-precision DFG map is generated through a Sigmoid activation function.

3. The multi-level semantic information aggregation method for remote sensing image change detection according to claim 1, characterized in that: The DFG map is embedded into the guided aggregation module of the convolutional neural network for feature fusion to obtain a fused feature map of multi-level semantic information, including: Performing spatial scale alignment processing on input features of two different levels in the convolutional neural network; Perform a feature concatenation operation on the two aligned input features along the channel dimension to generate a preliminary fusion feature map; The DFG mapping map is combined with the preliminary fusion feature map through element-level multiplication operation, and a convolution operation is performed on the fusion feature map combined with the DFG mapping to further compress the channel dimension of the feature and obtain the fusion feature map of the multi-level semantic information.

4. The multi-level semantic information aggregation method for remote sensing image change detection according to claim 1, characterized in that: Convolution operations of different scales are used to extract multi-scale fine-grained features from the fused feature map, and the multi-scale fine-grained features are fused to generate a final feature map for remote sensing image change detection, including: Using convolution kernels of different scales to extract fine-grained information of the changed area in the fused feature map to obtain the multi-scale fine-grained features of the changed area; The obtained multi-scale fine-grained features are fused by summing and averaging, and then the fused multi-scale fine-grained features are nonlinearly transformed using the ReLU activation function. Finally, a convolution operation is performed to generate the final feature map for remote sensing image change detection.

5. The multi-level semantic information aggregation method for remote sensing image change detection according to claim 4, characterized in that: Convolution kernels of three different scales, 3×3, 3×1 and 1×3, are used to extract the fine-grained information of the horizontal, vertical and local areas in the fusion feature map, respectively, to obtain the multi-scale fine-grained features of the changed area.

6. A multi-level semantic information aggregation device for remote sensing image change detection, characterized in that: include: The DFG mapping construction module is used to extract high-level image features at different scales, combine the high-level image features at different scales, and use the self-attention mechanism to construct a high-precision DFG mapping graph; A first feature fusion module is used to embed the DFG map into the guided aggregation module in the convolutional neural network to perform feature fusion, thereby obtaining a fused feature map of multi-level semantic information; The second feature fusion module is used to extract multi-scale fine-grained features from the fused feature map by using convolution operations of different scales, and perform feature fusion on the multi-scale fine-grained features to generate a final feature map for remote sensing image change detection.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the multi-level semantic information aggregation method for remote sensing image change detection according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program for executing the multi-level semantic information aggregation method for remote sensing image change detection according to any one of claims 1 to 5.

Images