High-resolution remote sensing image building change detection network and method

Abstract

The invention provides a high-resolution remote sensing image building change detection network and method, and belongs to the field of image processing. The key problems of task-independent interference, sparse distribution of change areas, fuzzy boundary and the like existing in the existing remote sensing image change detection technology are solved; comprising an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided differential enhancement module, a mixed attention module, a decoder module and a prediction head, the interactive high-frequency enhancement module is used for carrying out feature interaction and enhancement on multi-scale double-time-phase features output by the encoder module, and after the interactive high-frequency enhancement module, the fuzzy-guided differential enhancement module is used for carrying out fuzzy-guided differential enhancement on the multi-scale double-time-phase features. A network adopts a double-path structure, a fuzzy-guided difference enhancement module is arranged on a difference branch in double paths, refined modeling of multi-scale feature difference is realized by fusing standard difference and fuzzy difference in combination with a dynamic weight fusion strategy, and a mixed attention module is arranged on a channel cascade branch in the double paths; the method is applied to remote sensing image building change detection.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing image change detection, and in particular to a network and method for detecting building changes in high-resolution remote sensing images. Background Art

[0002] With continued population growth and urban expansion, the impact of human activities on the Earth's surface environment has reached a historic high. Remote sensing building change detection technology aims to accurately identify changes in the number, density, distribution, and morphology of building complexes by analyzing and comparing spatial information from multi-temporal images of the same geographic area. It has become a key technical tool supporting modern urban governance.

[0003] Traditional change detection methods primarily rely on spectral analysis and manual intervention, and their technical frameworks can be categorized into two typical models: image processing and manual interpretation. Image processing-based approaches typically use methods such as interpolation and change vector analysis to generate difference maps, demarcating change regions by setting thresholds. While simple to implement, these methods rely too heavily on spectral features, leading to non-semantic changes caused by environmental factors such as seasonal changes and varying lighting levels often being misidentified as real changes. Another approach, combining manual interpretation with machine learning, requires professionals to visually interpret the images, manually annotate building outlines as training samples, and then use shallow models such as support vector machines, K-nearest neighbors, and random forests for feature learning. While this approach partially improves efficiency, its generalization capability remains limited due to its reliance on manual feature extraction. Furthermore, model training requires a large amount of labeled data, which is labor-intensive, prone to subjective errors, and involves tedious parameter adjustments, limiting its practicality.

[0004] In recent years, revolutionary advances in deep learning technology have provided a new paradigm for remote sensing building change detection. Compared to traditional approaches that rely on manual feature design, deep learning extracts high-level features from remote sensing imagery through an end-to-end autonomous learning process, thereby improving detection accuracy and stability. Convolutional neural networks (CNNs), with their superior feature extraction capabilities, enable autonomous feature learning from pixel-level textures to object-level semantics. By introducing methods such as attention mechanisms and multi-scale feature fusion, the model's ability to capture details such as building edges and textures is enhanced.

[0005] In recent years, deep learning technology has gained widespread attention and application in the task of detecting building changes in remote sensing imagery due to its excellent feature representation capabilities. However, existing remote sensing imagery change detection techniques still face many challenges and limitations, such as task-irrelevant interference, sparse distribution of changed areas, and blurred boundaries. Therefore, further research and improvement are needed. Summary of the Invention

[0006] In order to solve the above technical problems, this application proposes a high-resolution remote sensing image building change detection network and method.

[0007] The technical solution adopted in this application is: a high-resolution remote sensing image building change detection network, including an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided differential enhancement module, a hybrid attention module, a decoder module and a prediction head. The encoder module adopts a twin encoder for multi-scale feature extraction of the input original dual-phase high-resolution remote sensing image, and the interactive high-frequency enhancement module is used to perform feature interaction and enhancement on the multi-scale dual-phase features output by the encoder module. After the interactive high-frequency enhancement module, the network adopts a dual-path structure, in which the fuzzy-guided differential enhancement module is set in the dual path. On the differential branch in the algorithm, the fuzzy-guided differential enhancement module achieves refined modeling of multi-scale feature differences by fusing standard differentials with fuzzy differentials and combining them with a dynamic weight fusion strategy. The hybrid attention module is set on the channel cascade branch in the dual path, and the channel attention module and the spatial attention module are connected in series to complete the feature enhancement of the channel and spatial dimensions in sequence. The dual-phase features output by the dual-path structure are then upsampled and fused through the decoder module to output a dual-phase feature map of the same size as the original dual-phase high-resolution remote sensing image. The dual-phase feature map is input into the prediction head and the final output is a change map.

[0008] Furthermore, the twin encoder of the encoder module uses a pre-trained CNN as the backbone network to extract multi-level features of dual-temporal high-resolution remote sensing images.

[0009] Furthermore, the interactive high-frequency enhancement module maps the features of the bi-phase image extracted by the encoder module through a 1×1 convolution, and reshapes them to obtain the query, key and value, and then performs the cross-attention mechanism operation to obtain the output after the cross-attention calculation;

[0010] Furthermore, an adaptive Sobel operator is introduced into the interactive high-frequency enhancement module to extract the high-frequency components of the feature map and obtain edge-related information.

[0011] Furthermore, an isotropic Sobel operator with eight directions and a minimum kernel size of 3 is used to extract the high-frequency components of each channel in the input feature map in eight different directions. In this process, a Top-k direction enhancement strategy is introduced to dynamically select the three directions with the most intense response in each channel and perform weighted fusion. The feature map enhanced with high-frequency information generated by the adaptive Sobel operator is then residually connected with the original input feature map. Finally, a 1×1 convolution is performed to generate interactive high-frequency enhancement features consistent with the number of input channels.

[0012] Furthermore, the standard difference in the fuzzy-guided difference enhancement module adopts a spatial normalization strategy to enhance local contrast, performs maximum normalization along the channel dimension, constrains the eigenvalue to the [0,1] interval, and enhances the contrast of the local area;

[0013] Furthermore, the fuzzy difference in the fuzzy-guided difference enhancement module is expressed with the constructed basic differential features as the basic features. Then, based on the fuzzy set theory, channel-adaptive learnable parameters are introduced to construct the nonlinear transformation S-type membership function to obtain the fuzzy differential features. The basic differential features and the fuzzy differential features are spliced through the channels and input into the dynamic fusion subnetwork. The spatial adaptive fusion weight is predicted by the weight generator constructed by depthwise separable convolution and 1×1 convolution, and finally the enhanced hybrid differential features are obtained through weighted fusion.

[0014] Furthermore, the channel attention module in the hybrid attention module adopts the frequency domain channel attention network FCANet, and introduces the two-dimensional discrete cosine transform into the frequency domain channel attention network FCANet. The two-dimensional discrete cosine transform realizes the diversity modeling of channel attention weights by decomposing and selectively fusing the feature map into multi-band components, and finally converts the input features into multi-general attention.

[0015] Furthermore, a combination strategy of binary cross entropy loss and Dice loss is used to train the network and optimize its parameters.

[0016] A method for detecting building changes in high-resolution remote sensing images comprises the following steps:

[0017] S1: Preprocessing of high-resolution remote sensing image change detection dataset;

[0018] S2: Train the model constructed using the high-resolution remote sensing image building change detection network to obtain the optimal model;

[0019] S3: Use the trained model to detect changes in buildings in high-resolution remote sensing images and output a binary image to represent the changes.

[0020] A computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0021] A computer-readable storage medium stores a computer program / instruction thereon, which implements the steps of the method when executed by a processor.

[0022] The beneficial effects of this application compared to the prior art are:

[0023] (1) The IHFM module is introduced so that the change-related areas can be dynamically focused, thereby reducing the interference of irrelevant information and maximizing the accuracy of boundary detection.

[0024] (2) The introduction of the FDEM module can effectively alleviate the problems of noise interference and fuzzy boundary detection in changing areas.

[0025] (3) The HAM module is introduced to capture cross-channel frequency features through frequency domain multi-spectral channel attention, and to focus on key areas with the spatial attention mechanism, forming a channel-space collaborative feature optimization mechanism, thereby improving the model's ability to perceive complex features and effectively improving the accuracy and robustness of change detection.

[0026] (4) The combination strategy of binary cross entropy loss and Dice loss can effectively balance the contribution of changed and unchanged samples to model training, thereby improving the detection accuracy of building change edges while avoiding overfitting or over-refinement problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present application will be further described below with reference to the accompanying drawings:

[0028] Figure 1 A schematic diagram of the structure of the improved network provided in an embodiment of the present application;

[0029] Figure 2 A schematic diagram of the structure of an interactive high-frequency enhancement module provided in an embodiment of the present application;

[0030] Figure 3 Schematic diagram of the isotropic Sobel operator in eight directions provided in the embodiment of the present application;

[0031] Figure 4 A schematic diagram of the structure of a fuzzy-guided differential feature enhancement module provided in an embodiment of the present application;

[0032] Figure 5 A schematic diagram of the structure of the hybrid attention module provided in an embodiment of the present application;

[0033] Figure 6 A schematic diagram of the FcaNet network structure provided in an embodiment of the present application;

[0034] Figure 7 A schematic diagram of the structure of spatial attention provided in an embodiment of the present application;

[0035] Figure 8 A comparison chart of the change detection results of different models provided in the embodiments of this application. DETAILED DESCRIPTION

[0036] like Figures 1 to 8As shown in the figure, this application provides a high-resolution remote sensing image building change detection network, which is an interactive high-frequency enhanced dual-path differential feature extraction network (Interactive High-Frequency Enhanced Dual-Path Differential Network, IHF-DPDNet), and its network structure is as follows Figure 1 As shown in the figure, IHF-DPDNet is designed as a twin architecture, using a pre-trained CNN (ResNet18) as the backbone to extract multi-level features from bi-temporal remote sensing imagery. Specifically, multi-layer feature maps with spatial sizes of 64×64, 32×32, and 16×16 are used in the first three stages of ResNet18. Furthermore, an interactive high-frequency enhancement module establishes dynamic correlations between bi-temporal features. A cross-attention mechanism is used to achieve feature interaction, and an adaptive Sobel operator is combined to enhance high-frequency boundary information, effectively improving the model's ability to perceive subtle changes. Next, the network adopts a dual-path parallel structure, in which a fuzzy-guided difference enhancement module is introduced in the difference branch. By fusing standard difference and fuzzy difference features, a dynamic weighting strategy is used to achieve multi-scale difference modeling, significantly alleviating noise interference and boundary blur. To further enhance the expressiveness of differential features, a hybrid attention module is designed in the channel cascade branch. Multi-spectral channel attention is used in the frequency domain to capture cross-channel feature dependencies. Combined with the spatial attention mechanism, this module focuses on key regions, forming a channel-spatial collaborative optimization mechanism to enhance the model's perception of complex features. Finally, the difference features of different scales are fused through a progressive strategy to generate the final change detection result.

[0037] The specific implementation plan is as follows:

[0038] (1) An interactive high frequency enhancement module (IHFM) is proposed, which has the following structure: Figure 2 As shown in the figure, before capturing the true difference information, interactive feature learning is performed between the two-phase images, and a lightweight edge detection method is used to enhance the high-frequency difference information. Using IHFM, regions related to changes can be dynamically focused, thereby reducing the interference of irrelevant information and maximizing the accuracy of boundary detection.

[0039] The features F1 and F2 of the bi-temporal image extracted in the previous stage are mapped through a 1×1 convolution respectively, and the query Q, key K and value V are obtained by reshaping, and then the cross attention mechanism operation is performed. Its expressions are shown in Equations (1) and (2):

[0040]

[0041] Ci =F i +Cross-Attention(Q i ,K j ,V j )(2);

[0042] Among them, i and j represent the serial numbers of the feature maps at different phases, d i Represents vector Q i Dimensions, represents the cross attention probability distribution, F i and C i Represent the original input feature map and the output after cross attention calculation, F i The additional information is learned through skip connections. The discriminative information obtained through interactive learning of cross-temporal images significantly improves the model's ability to perceive surface changes. This mechanism not only effectively suppresses interference factors irrelevant to the task, but also provides richer contextual information to help the model accurately identify changed areas.

[0043] In order to accurately depict the boundary information of the changing area, especially in high-resolution remote sensing images, details such as building boundaries are easily affected by blur and background interference. Therefore, we designed an adaptive Sobel high-frequency enhancement part in IHFM. By introducing the Sobel operator that is conducive to feature extraction, we can extract the high-frequency components of the feature map with a small amount of computing resources and obtain edge-related information at the same time, thereby improving the model's perception of the boundaries of the changing area and ensuring that the detection results are clearer and more stable. However, since buildings in high-resolution remote sensing images usually have different shapes, the isotropic Sobel operator shows greater advantages due to its robustness to shape changes. The isotropic Sobel operator structures in eight directions are as follows: Figure 3 shown.

[0044] Through the Sobel operator with a minimum kernel size of 3, the high-frequency components of the input feature map in eight different directions can be effectively extracted, so that the high-frequency features of different geometric forms and microstructures can be significantly enhanced. The Sobel kernels in these directions are stored as fixed parameters, so there is no need to learn during the training phase, and they are calculated through group convolution (GroupConvolution). This feature can effectively control the amount of network parameters and reduce the structural risk of the model while ensuring computational efficiency. Since the edge characteristics of different objects are different, the Sobel calculation results in certain directions can better highlight the edges of the target objects, while other directions may be affected by noise. To solve this problem, we introduce the Top-k direction enhancement strategy, dynamically select the three most important directions in each channel, and perform weighted fusion. Specifically, we perform a differential operation between C1 and C2 to obtain the differential feature Cdiff , then the differential feature C diff The interactive features C1 and C2 are spliced in the channel dimension respectively. After the adaptive Sobel operator edge detection, the feature map of high-frequency information enhancement is generated and connected with the original input residual. Finally, the interactive high-frequency enhancement feature with the same number of input channels is generated through 1×1 convolution. The calculation process is shown in equations (3) to (5):

[0045] C diff =C1-C2(3);

[0046]

[0047] in: Indicates the spliced C i and C diff feature;

[0048] i∈{1,2} represents two different phases;

[0049] Conv sobel (·) represents the Sobel operator edge detection.

[0050] IHFM effectively alleviates the problem of partial information loss caused by the increase in the number of neural network layers during training. It maximizes the use of interactive differential discriminant feature information between dual-phase images, while adaptively enhancing high-frequency information, strengthening the model's sensitivity to changing areas and boundary accuracy, and ultimately improving the overall accuracy and robustness of the model.

[0051] (2) Traditional differential methods mainly rely on directly calculating the pixel-level difference between two feature maps. However, this method is sensitive to noise and is easily affected by factors such as illumination and sensor noise. In addition, in remote sensing images, many changing areas often appear as fuzzy transitions, and traditional differential methods are difficult to effectively characterize these areas, which may lead to missed detections or false detections. In order to describe these areas more accurately, we propose a fuzzy-guided differential enhancement module (FDEM), which has a structure as follows: Figure 4 As shown in the figure, this module achieves refined modeling of multi-scale feature differences by fusing standard difference and fuzzy difference and combining it with a dynamic weight fusion strategy.

[0052] The standard difference in FDEM uses a spatial normalization strategy to enhance local contrast, performs maximum normalization along the channel dimension, constrains the eigenvalue to the [0, 1] interval, enhances the local area contrast, and retains the original spatial distribution characteristics. The calculation process is shown in formula (6):

[0053]

[0054] Among them, D norm represents the basic standard difference feature, Represents the feature map after feature interaction enhancement processing, represents a set of real numbers, B represents the batch size, C represents the number of channels, H represents the feature map height, W represents the feature map width, max(·) represents the maximum value operation along the channel dimension, which extracts the maximum difference value of each spatial position across all channels as the local contrast benchmark of the spatial dimension, and the parameter ε is a very small constant (the default value is 10 -6 ), which is used to avoid the numerical instability problem when the denominator is zero, and ensures the robustness of the calculation without affecting the significance of the normalized results.

[0055] In fuzzy difference, first construct the basic standard difference feature D norm As the basic feature expression, based on fuzzy set theory, channel-adaptive learnable parameters are introduced to construct the nonlinear transformation S-type membership function. The calculation process is shown in formula (7):

[0056]

[0057] Among them, D fuzzy represents the fuzzy differential feature, θ c is the learnable threshold parameter of the channel dimension, which determines the activation critical value of the fuzzy differential feature. c , the model can adaptively distinguish significant differences between different channels and background noise, and suppress the interference of low-intensity differences. c is the learnable scaling factor of the channel dimension, which controls the steepness of the fuzzy gradient. The larger its value is, the steeper the fuzzy differential feature is at the threshold θ. c The more sensitive the response is, the more dramatic the gradient change is; conversely, the response is gentler, which is suitable for modeling gradient differences. ⊙ represents the element-by-element multiplication operation (broadcast element in the channel dimension), and the absolute difference feature is mapped to the (0,1) interval through the Sigmoid function. Two differential features D norm With D fuzzy After channel splicing, the dynamic fusion sub-network is input, and the weight generator constructed by depth-separable convolution and 1×1 convolution predicts the spatial adaptive fusion weight W', and finally the enhanced hybrid differential feature D is obtained through weighted fusion. hybrid , the calculation process is shown in equations (8) and (9):

[0058] W'=σ(Conv 1×1 (DSConv 3×3 ([D norm ;D fuzzy ]))) (8);

[0059] D hybrid =W⊙D norm +(1-W)⊙D fuzzy (9);

[0060] Among them, [·;·] represents the channel splicing operation, DSConv(·) represents the 3×3 depth-wise separable convolution, σ represents the Sigmoid function activation function, and ⊙ represents the element-wise multiplication operation.

[0061] (3) For the characteristics of channel cascade branches, inspired by CBAM, this application designs a hybrid attention module (HAM), the structure of which is as follows Figure 5 As shown. The channel attention module and the spatial attention module are connected in series. The channel attention module uses the frequency domain channel attention network FCANet, and then connects to the original spatial attention module to complete the feature enhancement of the channel and spatial dimensions in turn. Input feature F input These are the dual-temporal interactive high-frequency enhancement features H1 and H2 after splicing and 3×3 convolution operations.

[0062] The traditional implementation of the channel attention module mainly relies on scalar calculations for each channel. A typical example is the global average pooling (GAP) strategy adopted by SENet. Although this strategy has the advantages of simplicity and high computational efficiency, it has certain feature representation bottlenecks. Because the global average pooling strategy is too simplified, it is difficult to effectively capture the detailed information in complex inputs, especially the homogenization of feature distribution caused by the mean operation, which will destroy the diverse expression between channels. Therefore, this application introduces the FCANet framework in the channel cascade branch, extends the channel attention to the frequency domain, and reconstructs the channel feature compression into a frequency domain decomposition problem. Based on the perspective of frequency domain analysis, it is proved that the traditional GAP method can actually be regarded as a special case of frequency domain feature decomposition. Thus, while retaining the overall representation capability of the channel, the adaptive selection and compression encoding of multi-band features are realized, significantly improving the discriminability of attention weights.

[0063] In the FCANet framework, the two-dimensional discrete cosine transform (DCT) is applied to the frequency domain channel attention mechanism. By decomposing and selectively fusing the multi-band components of the feature map, it breaks through the representation limitations of the traditional GAP method while maintaining computational efficiency and realizes the diversity modeling of channel attention weights. Specifically, the two-dimensional discrete cosine transform calculation process is shown in Equation (10):

[0064]

[0065] Where H and W are the height and width of the input feature map respectively. is the value of the pixel at (i, j) in the feature map, The 2D DCT spectrum can be viewed as the weighted sum of each pixel in the feature map. The cosine part is equivalent to the weight. The 2D discrete cosine transform can achieve the conversion from the spatial domain to the frequency domain. h is the index in the height direction, and w is the index in the width direction.

[0066] FCANet network structure see Figure 6 , its core component is the multi-spectral channel attention within the rectangular box, which is implemented as follows:

[0067] First, the input feature X is divided into several parts along the channel dimension [X0, X1, ..., X n-1 ],in:

[0068]

[0069] The number of channels C is required to be divisible by the number of partitioned regions n. C′ is the number of channels in the partitioned region. Each partitioned region is assigned a specific two-dimensional DCT frequency component. For each partitioned region, a corresponding two-dimensional DCT frequency component is assigned. The corresponding frequency domain representation is obtained through the transformation process shown in formula (11):

[0070]

[0071] in, Indicates that the two-dimensional DCT transform is in (u i ,v i ) on the weight, represents the value of the input signal at position (h,w), Indicates the DCT basis function at frequency (u i ,v i ) and the value at position (h,w), [u i ,v i ] is with X i The corresponding two-dimensional index of the frequency component is compressed to make Freq i Transformed into a vector with channel dimension C′, and finally the compressed Freq i All vectors are concatenated to obtain multispectral vectors The calculation process is shown in formula (12):

[0072] Freq=compress(X)=cat([Freq 0 ,Freq 1 ,…,Freq n-1 ]) (12);

[0073] In summary, the process of converting the input feature X into multi-spectral attention can be expressed as formula (13):

[0074] FcaAtt = σ(FC(Freq)) (13);

[0075] Among them, FC represents full connection and σ represents Sigmoid activation function. Finally, FCANet weights the obtained multi-spectral channel attention to the input feature X by element-wise multiplication to obtain the output feature F′ out , so as to make full use of the compressed channel information.

[0076] The spatial attention mechanism dynamically adjusts the spatial weight distribution of the feature map to achieve explicit modeling of the relationship between the local area and the global context. Its structure is shown in Figure 7 The mechanism first performs global maximum pooling and average pooling operations on the input features to extract the saliency distribution features and mean statistical features of the spatial dimension, respectively; then, the feature concatenation and convolution compression operations of the channel dimension are applied to construct the spatial weight matrix, which is normalized by the Sigmoid function to form the spatial attention mask. The final output feature is obtained by pixel-by-pixel multiplication of the original feature and the attention mask. The calculation process is shown in Equations (14) and (15):

[0077] M s =σ(Conv 7×7 ([F max ; F avg ])) (14);

[0078] F″ out =M s ×F′ out (15);

[0079] Among them, F max is the feature map processed by maximum pooling, F avg Represents the feature map after average pooling, [;] represents the channel splicing operation, Conv 7×7 (·) represents a two-dimensional convolution operation with a convolution kernel size of 7×7, σ is the Sigmoid activation function, and M s Represents the spatial attention weight of the feature map. To maintain the stability of the gradient flow and retain the original feature information, a learnable residual connection mechanism is introduced. After adjusting the dimension through 1×1 convolution, it is fused with the attention feature. The final output can be expressed as:

[0080] F out =F″ out +Conv 1×1 (F input ) (16).

[0081] (4) Using binary cross entropy loss L BCE With Dice loss L DiceThe complementary characteristics of the two loss functions have been shown to have a synergistic effect in the multi-class imbalanced sample processing scenario, which can effectively balance the contribution of changed and unchanged samples to model training, thereby improving the detection accuracy of building change edges. The total loss function is shown in formula (17):

[0082] L=L BCE +L Dice (17).

[0083] The pixel-level binary cross-entropy (BCE) loss function can effectively measure the difference between the model's prediction results and the true label, promoting the model's accurate detection of the changed area, and is therefore widely used in binary classification problems. In the embodiment of this application, BCE Loss is selected as part of the loss function. The calculation method of BCE Loss is shown in formula (18):

[0084]

[0085] Among them, y i represents the change of pixel i in the ground truth label image (y i =1 indicates a change, y i =0 means no change); p i represents the probability of pixel i changing; N represents the total number of pixels.

[0086] Due to the significant difference in the number of pixels between unchanged and changed areas in dual-temporal remote sensing images, the sample distribution exhibits a highly unbalanced characteristic. When relying solely on binary cross-entropy loss, the model's focus on unchanged pixels is overly magnified, resulting in limited ability to identify change features. To this end, a loss function based on the Dice coefficient is introduced. By measuring the spatial overlap between the predicted area and the true label, it effectively alleviates the impact of category imbalance on the training process and promotes convergence optimization of the network in change-sensitive areas. The calculation method of Dice Loss is shown in formula (19):

[0087]

[0088] Among them, y i Represents the label value of pixel i; Represents the predicted value of pixel i; N represents the total number of pixels.

[0089] This application also proposes a method for detecting building changes in high-resolution remote sensing images using the above network model, including the following steps:

[0090] S1: Preprocess the high-resolution remote sensing image change detection dataset, including geometric registration, radiation correction, image enhancement and other operations, and divide it into training set, validation set and test set;

[0091] S2: Train the network model for building change detection based on high-resolution remote sensing images to obtain the optimal model;

[0092] S3: Use the trained model to detect changes in buildings in high-resolution remote sensing images and output a binary image to represent the changes.

[0093] The network model of this application adopts a training-while-evaluation method during the training process. After each epoch of training, the model is saved and evaluated, and the optimal model is saved. After the training is completed, the optimal model obtained by training is used to predict the test set. In this embodiment, five indicators are used to evaluate the reasoning performance, including precision (Precision, Pre), recall (Recall, Rec), F1 score (F1-score), intersection over union (IoU) and overall accuracy (OA). Specifically, the indicators are defined as follows:

[0094]

[0095] Among them, TP represents the number of pixels that are predicted to change and actually change, TN represents the number of pixels that are predicted to remain unchanged and actually remain unchanged, FP represents the number of pixels that are predicted to change but actually remain unchanged, and FN represents the number of pixels that are predicted to remain unchanged but actually change.

[0096] To objectively and comprehensively compare the performance of the proposed IHF-DPDNet network, this example selected mainstream and advanced change detection methods for comparison. The model comparison methods include classic CNN-based change detection algorithms: FC-EF, FC-Siam-conc, and FC-Siam-diff; a Unet++-based method: SNUNet; an attention-based method: IFNet; and a Transformer-based method: BIT. Table 1 shows the quantitative evaluation results of different methods on the LEVIR-CD dataset.

[0097] Table 1 Quantitative evaluation results of different methods on the LEVIR-CD dataset.

[0098]

[0099] Table 1 shows the comparative results of seven networks for change detection on the LEVIR-CD dataset. Our application achieved the best results in the four core metrics of Recall, F1-score, IoU, and OA, with scores of 89.15%, 90.57%, 82.93%, and 99.06%, respectively. Although IFN performed best in Precision, at 92.68%, its Recall value was relatively low, indicating that it had missed detections during the detection process. In contrast, our application achieved a significant improvement in Recall while maintaining the suboptimal Precision value of 91.61%.

[0100] Comparison of change detection results of different models is shown in the figure below: Figure 8 As shown in the first, second and fourth rows of the figure, the FC series models have obvious problems of false detection and missed detection, especially for small change targets when the categories are unbalanced. In addition, for large continuous buildings, there are problems with blurred boundaries and poor integrity of the change area. In comparison, the IHF-DPDNet of the present application effectively highlights key edge information and captures the continuity of the building outline through an interactive high-frequency enhancement module, while significantly reducing false detection and missed detection in small change areas. As shown in the third row of the figure, thanks to the nonlinear modeling capabilities of the dual-branch mechanism and the fuzzy difference enhancement module for the gradient area, the IHF-DPDNet of the present application suppresses the block false alarms caused by noise, and at the same time, for areas with dense buildings, effectively reduces the adhesion between buildings, so that the model can still maintain stable detection performance in complex scenarios.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A high-resolution remote sensing image building change detection network, characterized by: It includes an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided differential enhancement module, a hybrid attention module, a decoder module and a prediction head. The encoder module adopts a twin encoder to perform multi-scale feature extraction on the input original dual-phase high-resolution remote sensing image. The interactive high-frequency enhancement module is used to perform feature interaction and enhancement on the multi-scale dual-phase features output by the encoder module. After the interactive high-frequency enhancement module, the network adopts a dual-path structure, in which the fuzzy-guided differential enhancement module is set on the differential branch in the dual path. The fuzzy-guided differential enhancement module realizes the refined modeling of multi-scale feature differences by fusing standard differences and fuzzy differences and combining dynamic weight fusion strategies. The hybrid attention module is set on the channel cascade branch in the dual path, and the channel attention module and the spatial attention module are connected in series to complete the feature enhancement of the channel and spatial dimensions in sequence. The bi-phase features output by the dual-path structure are then upsampled and fused through the decoder module to output a bi-phase feature map of the same size as the original bi-phase high-resolution remote sensing image. The bi-phase feature map is input into the prediction head and finally the change map is output.

2. The high-resolution remote sensing image building change detection network according to claim 1, characterized in that: The twin encoder of the encoder module uses a pre-trained CNN as the backbone network to extract multi-level features of dual-temporal high-resolution remote sensing images.

3. The high-resolution remote sensing image building change detection network according to claim 1, characterized in that: The interactive high-frequency enhancement module maps the features of the bi-phase image extracted by the encoder module through a 1×1 convolution, and reshapes them to obtain the query, key and value, and then performs the cross-attention mechanism operation to obtain the output after the cross-attention calculation; The interactive high-frequency enhancement module also introduces an adaptive Sobel operator to extract the high-frequency components of the feature map and obtain edge-related information.

4. The high-resolution remote sensing image building change detection network according to claim 3, characterized in that: The isotropic Sobel operator in eight directions with a minimum kernel size of 3 is used to extract the high-frequency components of each channel in the input feature map in eight different directions. In this process, a Top-k direction enhancement strategy is introduced to dynamically select the three directions with the most intense response in each channel and perform weighted fusion. The feature map enhanced with high-frequency information generated by the adaptive Sobel operator is then residually connected with the original input feature map. Finally, a 1×1 convolution is used to generate interactive high-frequency enhancement features consistent with the number of input channels.

5. The high-resolution remote sensing image building change detection network according to claim 1, characterized in that: The standard difference in the fuzzy-guided difference enhancement module adopts a spatial normalization strategy to enhance local contrast, performs maximum normalization along the channel dimension, constrains the eigenvalue to the [0, 1] interval, and enhances the contrast of the local area. The fuzzy difference in the fuzzy-guided difference enhancement module uses the constructed basic differential features as the basic feature expression. Then, based on the fuzzy set theory, channel-adaptive learnable parameters are introduced to construct the nonlinear transformation S-type membership function to obtain the fuzzy differential features. The basic differential features and the fuzzy differential features are spliced through the channels and input into the dynamic fusion subnetwork. The spatial adaptive fusion weight is predicted by the weight generator constructed by depthwise separable convolution and 1×1 convolution, and finally the enhanced hybrid differential features are obtained through weighted fusion.

6. The high-resolution remote sensing image building change detection network according to claim 1, characterized in that: The channel attention module in the hybrid attention module adopts the frequency domain channel attention network FCANet, and introduces the two-dimensional discrete cosine transform into the frequency domain channel attention network FCANet. The two-dimensional discrete cosine transform realizes the diversity modeling of channel attention weights by decomposing and selectively fusing the feature map into multi-band components, and finally converts the input features into multi-general attention.

7. A high-resolution remote sensing image building change detection network according to any one of claims 1 to 6, characterized in that: A combination strategy of binary cross entropy loss and Dice loss is used to train the network and optimize its parameters.

8. A method for detecting building changes in high-resolution remote sensing images, characterized by: The following steps are involved: S1: Preprocessing of high-resolution remote sensing image change detection dataset; S2: training a model constructed using the high-resolution remote sensing image building change detection network according to any one of claims 1 to 7 to obtain an optimal model; S3: Use the trained model to detect changes in buildings in high-resolution remote sensing images and output a binary image to represent the changes.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 8.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 8 are implemented.

Images