Remote sensing image change detection method and equipment based on structure perception and frequency domain fusion

Abstract

The invention discloses a remote sensing image change detection method and equipment based on structure perception and frequency domain fusion, belongs to the technical field of image change detection, and realizes space alignment and lightweight feature extraction of dual-time-phase features through a dual-branch multi-layer feature extraction encoder sharing weight. A complementary enhancement mechanism is constructed between a frequency domain and a spatial domain by using a structure-perceived frequency domain change module and a multi-level frequency domain fusion decoding module. And false change interference can be effectively suppressed, and the recognition capability of building boundaries and detail areas is enhanced. The whole structure gives consideration to detection precision, generalization performance and calculation efficiency, has good engineering applicability, and can be popularized and applied to various remote sensing scenes such as urban building dynamic monitoring, land utilization change analysis and post-disaster emergency evaluation.

Description

Technical Field

[0001] This invention belongs to the field of image change detection technology, specifically relating to a method and device for detecting changes in remote sensing images based on structure perception and frequency domain fusion. Background Technology

[0002] Change detection, as an important research direction in the field of remote sensing image analysis, has profound and multifaceted significance for the identification, monitoring, and assessment of dynamic changes in the Earth's surface. By conducting differential analysis on remote sensing images from different periods, change detection can accurately capture the increase, decrease, transfer, and morphological changes of land features, providing crucial data support for dynamic land use monitoring, urban expansion analysis, ecological environment assessment, disaster damage detection, and land resource management.

[0003] In recent years, with the rapid development of deep learning technology, end-to-end change detection methods based on structures such as convolutional neural networks and Transformers have gradually become mainstream. These methods typically employ a Siamese network structure, using a dual-branch encoder with shared weights to extract features from images from different time points, and then fusing and reconstructing these features during the decoding stage to achieve accurate identification of changed regions. This type of method effectively improves the automation and accuracy of change detection.

[0004] However, existing deep learning change detection models mainly focus on the extraction and fusion of spatial domain features, neglecting the potential value of frequency domain information in change recognition. Remote sensing images contain rich texture details, boundary features, and structural change information across different frequency components. Relying solely on spatial domain features often fails to fully represent the complexity of ground feature changes, especially under conditions of illumination differences, noise interference, or similar ground feature morphologies, leading to frequent false positives and false negatives. Furthermore, existing models still have certain shortcomings in multi-scale change representation and high- and low-frequency feature modeling, resulting in insufficient capture of change patterns in complex scenes and limited generalization ability.

[0005] In 2022, Wele Gedara Chaminda Bandara, Vishal M. Patel, and other scholars published a research paper entitled "A Transformer-Based Siamese Network for Change Detection" at the IGARSS international conference, proposing a Transformer-based Siamese network architecture, ChangeFormer. This model combines a hierarchical Transformer encoder with a multilayer perceptron decoder for change detection in remote sensing imagery. However, in building change recognition tasks using optical remote sensing imagery, such methods still have shortcomings in feature modeling: they only focus on spatial domain information and lack the ability to mine information between features of different frequencies.

[0006] To address the aforementioned issues, some studies have attempted to introduce multi-scale feature fusion or attention mechanisms into the model structure to enhance the model's discriminative ability. For example, Henan University, in its patented technology "Remote Sensing Image Change Detection Method Based on Wavelet Transform Frequency Domain Perception" (application number: CN202510366947.2, publication number: CN120318564A), proposed improving the response of spatial domain features to changed areas through a multi-scale channel attention module. However, such methods still mainly rely on the contrast enhancement of spatial features and fail to effectively utilize frequency domain information to supplement the modeling of change patterns, thus exhibiting certain limitations when dealing with complex surface textures or fine-grained change scenes. Summary of the Invention

[0007] The purpose of this invention is to provide a method and device for detecting changes in remote sensing images based on structure perception and frequency domain fusion, so as to solve the technical problem that existing remote sensing image change detection methods lack the use of frequency domain information.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: Multi-scale semantic features are extracted from dual-temporal remote sensing images, and the multi-scale semantic features are differentially processed to obtain multi-scale differential features. Structured frequency domain features are extracted from multi-scale differential features using discrete wavelet transform; By fusing multi-scale difference features and structured frequency domain features, pixel-level change prediction results are output.

[0009] Furthermore, the multi-scale semantic features of the dual-temporal remote sensing images are extracted through a weighted, dual-branch, multi-layer feature extraction encoder.

[0010] Furthermore, the multi-scale semantic features are subjected to differential processing, including: A feature difference module is constructed based on multi-scale semantic features. Element-level absolute difference operations are performed on the feature maps of each layer to obtain the difference features. Perform scale alignment on the difference features.

[0011] Furthermore, structured frequency domain features are extracted from the multi-scale difference features using discrete wavelet transform, including: Multi-scale difference features are concatenated and convolved to obtain fused features; A two-dimensional discrete wavelet transform is performed on the fused features to obtain multiple sub-band frequency domain features; The wavelet fusion feature is obtained by adding the frequency domain features of multiple sub-bands; Upsample the wavelet fusion features and obtain the structure-aware wavelet fusion features through convolutional mapping.

[0012] Furthermore, before adding the frequency domain features of the multiple sub-bands to obtain the wavelet fusion features, differentiated convolution enhancement is first applied to different sub-bands.

[0013] Furthermore, the multi-scale difference features and structured frequency domain features are fused to output pixel-level change prediction results, including: After linearly embedding the multi-scale difference features, they are added to the structured frequency domain features pixel by pixel and fused to obtain multiple sets of fused features; Multiple sets of fused features are concatenated, linearly fused, and classified convolutional mapping is performed to obtain a pixel-level probability map; The pixel-level probability map is restored to the original resolution of the input image using bilinear interpolation to obtain the final change prediction result.

[0014] Furthermore, the formula for the change prediction result is:

[0015] In the formula, For pixel-level prediction results of building changes, For upsampling, P is the pixel-level probability map. .

[0016] Secondly, the present invention provides a remote sensing image change detection system based on structure perception and frequency domain fusion, comprising a processing module, a fusion module, and an output module, wherein: Processing module: used to extract multi-scale semantic features from dual-temporal remote sensing images, perform differential processing on the multi-scale semantic features, and obtain multi-scale differential features; Fusion module: used to extract structured frequency domain features from multi-scale differential features using discrete wavelet transform; Output module: Used to fuse multi-scale difference features and structured frequency domain features to output pixel-level change prediction results.

[0017] Thirdly, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0018] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a remote sensing image change detection method based on structure awareness and frequency domain fusion, aiming to overcome the problems of redundant model parameters, insufficient fusion of frequency domain and spatial features, and limited detail recognition capabilities in existing building change detection technologies. Addressing the complexity and high-precision requirements of building target change detection in high-resolution optical remote sensing images, this method captures significant change information between two temporal images using a feature difference mechanism within a shared-weight dual-branch coding framework. It introduces a structure-aware frequency domain change module and a multi-level frequency domain fusion decoding module to achieve joint modeling of multi-scale features in the frequency and spatial domains, and performs pixel-level fine classification. While effectively reducing the number of model parameters and computational complexity, it achieves high-precision identification of building change areas and enhancement of boundary details, significantly improving the overall performance and stability of change detection.

[0020] This system integrates discrete wavelet transform and multilayer perceptron decoding mechanisms to achieve joint feature modeling in the frequency and spatial domains. Wavelet decomposition divides features into low-frequency structural components and high-frequency texture components, which are then complementaryly fused during reconstruction. This enhances the representation of building edges and detailed textures while maintaining the integrity of global structural information. The introduction of learnable wavelet convolution and linear mapping operations enables the network to adaptively aggregate features and model global associations at different scales, thereby improving the classification accuracy and boundary consistency of changed regions. This structure exhibits higher robustness and generalization ability when handling complex backgrounds or fine-grained building changes, effectively suppressing spurious change responses and significantly improving the overall performance of building change detection.

[0021] A dual-branch, multi-layer feature extraction encoder with shared weights is used as the backbone network for feature extraction, performing multi-scale feature extraction on dual-temporal optical remote sensing images. This design not only ensures the consistency of the feature spaces between the two temporal phases but also effectively reduces the number of parameters and storage overhead, significantly improving the network's operational efficiency and deployment flexibility while maintaining detection accuracy. Compared with traditional non-shared weight networks, this method performs better in terms of feature alignment and change sensitivity, and can more accurately capture the semantic differences between temporal images. Attached Figure Description

[0022] Figure 1 This is a flowchart of a remote sensing image change detection method based on structure perception and frequency domain fusion. Figure 2 This is a flowchart illustrating the remote sensing image change detection method based on structure perception and frequency domain fusion in an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall network structure designed for this invention; Figure 4 Image a in the middle is the first temporal image; Figure 4 Image b in the middle is the second temporal image; Figure 4 Figure C shows manually labeled change tags; Figure 4 The middle d-figure shows the prediction results of ChangeFormer; Figure 4 The figure in Figure e shows the prediction results of the network of this invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] It should be noted that the terms "first" and "second" in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, the remote sensing image change detection method based on structure perception and frequency domain fusion includes the following steps: Step 1: Extract multi-scale semantic features from dual-temporal remote sensing images, and perform differential processing on the multi-scale semantic features to obtain multi-scale differential features; We selected the publicly available building change detection dataset LEVIR-CD, which contains dual-temporal remote sensing image samples. and The original image is cropped into several non-overlapping image pairs, and then divided into training, validation, and test sets according to a set ratio to provide a data foundation for subsequent model training and evaluation. Specifically, this includes: We selected the publicly available building change detection dataset LEVIR-CD, which contains remote sensing images of the same area acquired at different time phases. and and its change annotation information; The image data is cropped without overlap to generate dual-temporal image pairs, which are then divided into training, validation, and test sets according to a set ratio for model training and performance evaluation.

[0027] like Figure 3 As shown, the obtained training set is input into a two-branch multilayer feature extraction encoder with shared weights to process the dual-temporal remote sensing images. and Feature extraction is performed; the encoder consists of four consecutive feature extraction blocks, denoted as Block1, Block2, Block3, and Block4 respectively; each feature extraction block contains an Overlap Patch Embedding module and several Transformer feature sub-blocks, which are used to extract multi-scale semantic features layer by layer; in: The overlap patch embedding module in Block 1 uses a 7×7 convolutional kernel with a stride of 4; the overlap patch embedding modules in Blocks 2 through 4 all use a 3×3 convolutional kernel with a stride of 2; the number of Transformer sub-blocks in the four blocks are 2, 2, 2, and 2 respectively; the spatial reduction ratio of the attention mechanism is 8, 4, 2, and 1 respectively, and the number of multi-head attention heads is 1, 2, 5, and 8 respectively; the number of output channels in the four layers is 32, 64, 160, and 256 respectively, corresponding to spatial resolutions of approximately 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input; the output of each layer is normalized and its shape restored before being used as a feature map; Input Image and After the above encoding process, four layers of feature representations are obtained:

[0028] In this context, subscripts indicate levels, and superscripts indicate phases.

[0029] This structure ensures the consistency of feature extraction between the two phases by sharing parameters, which facilitates subsequent difference modeling and multi-scale fusion. Four-layer feature representation of the output A feature difference module is constructed to capture significant change information between two temporal images; Feature Differentiation Module: This module performs feature mapping on each layer separately. and Perform element-wise absolute difference operations to obtain the difference features:

[0030] This operation suppresses the sign uncertainty caused by the difference direction and highlights the regions that may change; For difference features Scale alignment is performed to ensure consistent decoding: deeper features are upsampled to the same spatial size as the first layer using bilinear interpolation, maintaining pixel correspondence. The scale-aligned differential features are used as inputs to the subsequent structure-aware frequency domain change module and multi-level frequency domain fusion decoding module, so as to enhance change cues in both the frequency and spatial domains.

[0031] Step 2: Extract structured frequency domain features from the multi-scale difference features using discrete wavelet transform; A structure-aware frequency domain transformation module is introduced to extract structured frequency domain features through discrete wavelet transform, including the following steps: Will Compared with upsampling , , The features are concatenated along the channel dimension and then compressed and convolved using a 1×1 channel to obtain the fused features. To reduce redundancy and standardize the number of channels: ; Fusion features Performing a two-dimensional discrete wavelet transform (using Haar wavelets) yields four sub-bands:

[0032] in, For low-frequency structural information, and Represents horizontal / vertical texture, high-frequency subband Indicates diagonal details; Differentiated convolutional enhancement strategies are employed for different subbands:

[0033] The above processing enhances the overall outline, directional texture, and sharp details respectively; The wavelet fusion features are obtained by adding the multi-frequency enhancement results:

[0034] It is then upsampled to the same scale as the first layer space, and subsequently mapped to a fixed decoding embedding dimension via a 1×1 embedding convolution, denoted as the structure-aware wavelet fusion feature. :

[0035] This process supplements the high-frequency texture and structural boundary information that is difficult to obtain stably in the spatial domain at the frequency domain level.

[0036] Step 3: Fuse the multi-scale difference features and structured frequency domain features to output pixel-level change prediction results: The design incorporates a multi-level frequency domain fusion decoding module to aggregate multi-scale spatial and frequency domain information and output predictions. Specifically, this includes: Four-layer difference features , , , Perform linear embedding separately, mapping the number of channels to a unified embedding dimension E, and then reconstruct each channel into a two-dimensional feature map: ; The four embedded features are upsampled to the level of the first layer. Same spatial resolution and fused with the obtained structure-aware wavelet features Perform pixel-by-pixel addition to fuse features in the frequency and spatial domains. This operation is defined as follows: ; In the formula, Representing difference features Embedding features, Represents the first layer of difference features height, Represents the first layer of difference features height, Indicates the first A fusion feature.

[0037] Four fusion features The data is concatenated along the channel dimension, linearly fused using 1×1 convolution, and then a Dropout operation (scale 0.1) is applied to obtain the final fused representation: ; In the formula, This indicates the fusion feature.

[0038] By mapping the number of channels to the number of target categories (changed / unchanged) using a 1×1 classification convolution, a pixel-level probability map P is obtained:

[0039] in Use the Sigmoid activation function; The probability map P is restored to the original resolution of the input image using bilinear interpolation. This ensures that the predicted results correspond one-to-one with the input pixels, thus obtaining the final change detection results. :

[0040] Final output This is the pixel-level prediction result of building changes, where the value of each pixel represents the probability that it belongs to the "change" category.

[0041] In one embodiment of the present invention, a method for detecting changes in remote sensing images based on structure perception and frequency domain fusion is provided, comprising the following steps: Step 1. Dataset Acquisition: Select the publicly available building change detection dataset LEVIR-CD, which contains dual-temporal remote sensing image samples. and The original image is cropped into several non-overlapping image pairs, and divided into training, validation, and test sets according to a set ratio to provide a data foundation for subsequent model training and evaluation; Step 2. Network Construction: A Structure-Aware and Frequency-fused Network (SAF-Net) for remote sensing image change detection is constructed. This network employs a shared-weight, dual-branch, multi-layer feature extraction encoder structure to extract deep semantic features at different scales from dual-temporal images. A feature difference module is introduced between the coding features of each layer to enhance the ability to express and align differences in temporal change regions. In the decoding stage, a structure-aware frequency domain change module and a multi-level frequency domain fusion decoding module are introduced. Specifically, the structure-aware frequency domain change module extracts multi-scale frequency domain features through discrete wavelet transform; the multi-level frequency domain fusion module converges and integrates frequency domain change features from different levels at each level across scales, and dynamically fuses them with spatial domain features (differential features) to generate a change representation that combines global semantics and detailed texture, thereby improving the model's ability to identify subtle changes in building textures and structural boundaries. Specifically: The dual-branch multi-layer feature extraction encoder with shared weights, the feature difference module, the structure-aware frequency domain change module, and the multi-level frequency domain fusion decoding module are connected in series to form a remote sensing image change detection network (Structure-Aware and Frequency-fused Network, SAF-Net) based on structure-aware and frequency domain fusion. To facilitate application and evaluation, a threshold τ is set on the probability map P to obtain a binary mask:

[0042] In the formula, This represents the binary mask value obtained at pixel coordinates (x, y). This represents the probability map value at pixel coordinates (x, y), indicating the score of whether the pixel belongs to the foreground.

[0043] The above design, through a chain structure of "parameter-sharing multi-scale extraction - explicit differentiation - wavelet multi-frequency enhancement - spatial and frequency domain aggregation", improves the response capability to high-frequency textures, boundary details and cross-scale structural changes while ensuring controllable computational load, and enhances the overall accuracy and stability of change detection.

[0044] Step 3. Network Training and Validation: The constructed Structure-Aware and Frequency-fused Network (SAF-Net) for remote sensing image change detection is trained end-to-end using the training set obtained in Step 1. The model performance and parameter convergence are dynamically evaluated using the validation set. Finally, a change detection network with excellent detection accuracy and robustness is obtained. Step 4. Change Detection Inference: The trained Structure-Aware and Frequency-fused Network (SAF-Net) for remote sensing image change detection is applied to the test set in Step 1 to perform prediction inference on dual-temporal remote sensing images and output pixel-level change results, thereby realizing the automatic identification and location of changes in building areas. Step 5. Result Evaluation and Performance Analysis: Quantitative analysis is performed on the pixel-level prediction results obtained in Step 4 to verify the effectiveness and stability of the method of the present invention in complex scenarios.

[0045] like Figure 4 The diagram shows the comparison results between the method of this invention and ChangeFormer on the LEVIR-CD test set. Among them, Figure 4 Image a (the first three images in the first column) is the first temporal image. Figure 4 Image b (the third image in the second column) is the second temporal image. Figure 4 The three images in the third column (Figure C) contain manually labeled change tags. Figure 4 The middle d-th graph (the three graphs in the fourth column) shows the prediction results of ChangeFormer. Figure 4The images in the fifth column (three images) show the prediction results of the network of this invention. The comparison shows that this invention exhibits higher accuracy and boundary consistency in detecting building change areas. Taking the first set of samples as an example, ChangeFormer's detection results show obvious noise and some false detection areas, while the results of this invention can more completely restore the building boundary, and the predicted contour is highly consistent with the label map. In the second set of samples, ChangeFormer's results show boundary breaks and false responses, while the change map output by this invention can accurately depict the target shape and effectively suppress false changes. Overall, this invention significantly reduces false detections and missed detections while maintaining contour integrity, fully demonstrating the enhancement effect of the structure-aware frequency domain change module and the multi-level frequency domain fusion decoding module on high-frequency details and boundary features.

[0046] As shown in Table 1, quantitative results on the LEVIR-CD dataset demonstrate that the proposed method outperforms the comparative models in all major metrics. Specifically, precision is improved to 92.90%, recall to 88.30%, F1 score to 90.54%, IoU to 1.83 percentage points higher than ChangeFormer, and overall accuracy to 99.06%. In terms of model complexity, the number of parameters is only 6.13M, a significant reduction compared to ChangeFormer's 41.03M, approximately one-seventh of its complexity. Therefore, the proposed network significantly reduces computational burden while maintaining high-precision detection performance, achieving a comprehensive balance of accuracy, efficiency, and lightweight design in high-resolution optical remote sensing image building change detection tasks, demonstrating clear application advantages.

[0047] Table 1. Comparison of various evaluation metrics between the present invention (bottom) and Changeformer (top) on the same dataset.

[0048] This invention addresses the problems of complex models, difficulty in capturing structured contextual information, and incomplete detection of building change areas in high-resolution optical remote sensing imagery. Existing techniques generally rely on convolutional or Transformer structures with large parameter sets, resulting in high computational overhead and low efficiency in feature extraction and fusion processes, making it difficult to balance detection accuracy with lightweight models. Furthermore, traditional networks have limitations in the interaction of information in the frequency and spatial domains, often neglecting the differences in high-frequency texture and boundary structure features, leading to inaccurate identification of change areas, blurred boundaries, and detection performance that fails to meet the requirements of complex remote sensing scenarios.

[0049] This invention addresses the aforementioned problems in building change detection in optical remote sensing images. It employs a shared-weight dual-branch coding structure as the backbone for feature extraction, achieving semantic alignment and parameter sharing between two-temporal images, significantly reducing the number of model parameters and lowering the dependence on computational resources. Simultaneously, it designs a structure-aware frequency domain change module and a multi-level frequency domain fusion decoding module. By separating and fusing high- and low-frequency features through discrete wavelet transform, it enables full interaction between frequency domain information and spatial features, thereby enhancing the expression of building edges and texture details. This effectively improves the accuracy and robustness of building change detection, better meeting the practical application needs in complex remote sensing scenarios.

[0050] In another embodiment of the present invention, a remote sensing image change detection system based on structure perception and frequency domain fusion is provided, comprising an acquisition module, a processing module, a fusion module, and an output module, wherein: Acquisition module: used to acquire dual-temporal remote sensing images; Processing module: used to extract multi-scale semantic features from dual-temporal remote sensing images and perform differential processing on the multi-scale semantic features; Fusion module: used to extract structured frequency domain features from the fused multi-scale difference features using discrete wavelet transform; Output module: Used to fuse multi-scale difference features and structured frequency domain features to output pixel-level change prediction results.

[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. 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 be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM (optical memory, etc.).

[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or substitutions to the specific implementation of the invention, but these changes, modifications or substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A remote sensing image change detection method based on structural perception and frequency domain fusion, characterized in that, The method comprises the following steps: Multi-scale semantic features of the dual-time-phase remote sensing image are extracted, and the multi-scale semantic features are differentially processed to obtain multi-scale differential features; Structured frequency domain features are extracted from the multi-scale differential features through discrete wavelet transform; The multi-scale differential features and the structured frequency domain features are fused to output pixel-level change prediction results.

2. The structure perception and frequency domain fusion based remote sensing image change detection method according to claim 1, characterized in that, The multi-scale semantic features of the dual-time-phase remote sensing image are extracted by a dual-branch multi-layer feature extraction encoder sharing weights. 3.The structural perception and frequency domain fusion based remote sensing image change detection method according to claim 1, characterized in that, The multi-scale semantic features are differentially processed, including: A feature differential module is constructed based on the multi-scale semantic features, element-level absolute difference operations are respectively performed on feature maps of each layer to obtain differential features; The differential features are scaled and aligned.

4. The structure perception and frequency domain fusion based remote sensing image change detection method of claim 1, wherein, The structured frequency domain features are extracted from the multi-scale differential features through discrete wavelet transform, including: The multi-scale differential features are spliced and convolved to obtain fused features; The fused features are subjected to two-dimensional discrete wavelet transform to obtain a plurality of sub-band frequency domain features; The plurality of sub-band frequency domain features are added to obtain wavelet fused features; The wavelet fused features are up-sampled to obtain structured perception wavelet fused features through convolution mapping.

5. The structural perception and frequency domain fusion based remote sensing image change detection method of claim 4, wherein, Before the plurality of sub-band frequency domain features are added to obtain the wavelet fused features, differential convolution enhancement is performed on different sub-bands.

6. The structural perception and frequency domain fusion based remote sensing image change detection method of claim 1, wherein, The multi-scale differential features and the structured frequency domain features are fused to output pixel-level change prediction results, including: After linear embedding of the multi-scale differential features, the multi-scale differential features are added to the structured frequency domain features pixel by pixel to obtain a plurality of groups of fused features; The plurality of groups of fused features are spliced, linearly fused, and subjected to classification convolution mapping to obtain pixel-level probability maps; The pixel-level probability maps are restored to the original resolution of the input image by bilinear interpolation to obtain final change prediction results.

7. The structural perception and frequency domain fusion based remote sensing image change detection method of claim 6, wherein, The change prediction result formula is: In the formula, is the pixel-level prediction result of the building change, is up-sampling, P is a pixel-level probability map, indicates that the size of the up-sampled .

8. A remote sensing image change detection system based on structural perception and frequency domain fusion, characterized in that, The method comprises a processing module, a fusion module, and an output module, wherein: The processing module is configured to extract multi-scale semantic features of a dual-time-phase remote sensing image, differentially process the multi-scale semantic features, and obtain multi-scale differential features; The fusion module is configured to extract structured frequency domain features from the multi-scale differential features through discrete wavelet transform; The output module is configured to fuse the multi-scale differential features and the structured frequency domain features, and output pixel-level change prediction results.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the computer program when the computer program is executed by the processor to realize the method of claim 1 7. The step of the method of any one.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program, which is executed by a processor, implements the method as claimed in claim 1 7. The step of the method of any one.

Images