A semantic segmentation method for remote sensing images that combines boundary induction and semantic compensation mechanisms

Abstract

In semantic segmentation of remote sensing images, boundary information of objects tends to be unclear, and semantic information is imbalanced between different layers of the network, resulting in poor segmentation accuracy and robustness.In particular, there is a demand for a highly accurate segmentation method that combines boundary information and semantic information for areas with complex boundary shapes and fine-grained features. [Solution] The present invention provides a semantic segmentation method for remote sensing images that combines boundary guidance and semantic compensation mechanisms, and constructs a boundary-guided semantic compensation network that includes an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module, and an auxiliary boundary monitoring module. The network extracts hierarchical features from the input remote sensing image, then uses a cross-layer semantic compensation module to combine high-level semantic information with detailed spatial information, and then uses a Transformer to integrate global context to generate semantic segmentation results. Additionally, an auxiliary boundary monitoring module is used during the training process to perform auxiliary monitoring of boundary information. This allows for compensation of semantic information while suppressing loss of boundary information, improving the accuracy and robustness of semantic segmentation of remote sensing images.

Description

[Technical Field]

[0001] The present invention relates to the technical field of remote sensing image processing, and more particularly to a method for semantic segmentation of remote sensing images that combines boundary guidance and semantic compensation mechanisms. [Background technology]

[0002] As the resolution of remote sensing images continues to improve, the texture and structure complexity of objects in remote sensing scenes has increased significantly. These images typically contain rich details, diverse categories, and similar textures, accompanied by background noise and interference, posing significant challenges for segmentation tasks. In recent years, the CNN-Transformer hybrid architecture has gradually become the mainstream segmentation model. CNNs can extract local spatial features, but their limited receptive field makes them poor at modeling long-range dependencies. Transformers have global modeling capabilities, but their performance is relatively poor when dealing with fine-grained edges and small targets. While the CNN-Transformer hybrid architecture combines the advantages of both to some extent, it still suffers from the following problems:

[0003] (1) Insufficient cross-layer feature fusion: The semantic hierarchical differences between features at different scales in the CNN encoder are large. Therefore, using simple combination or layer-by-layer upsampling fusion leads to insufficient semantic alignment, preventing effective fusion of the semantics of the upper layer and the texture of the lower layer, which impacts the feature representation of small targets and complex structures.

[0004] (2) Loss of boundary information: The Transformer decoder mainly focuses on global semantic reconstruction and lacks an explicit boundary guidance mechanism, which results in blurred or broken edges and even category confusion during the prediction stage, severely impacting the accuracy of the target boundary.

[0005] In recent years, the increasing resolution of remote sensing images has led to a demand for more advanced semantic segmentation techniques in fields such as land cover classification and urban analysis. Conventional techniques have been proposed that improve segmentation accuracy by incorporating multi-scale feature fusion and attention mechanisms based on an encoder-decoder architecture using convolutional neural networks. For example, Chinese Patent CN120355928B discloses a technique for improving segmentation performance of remote sensing images by extracting and fusing deep features.

[0006] However, in these conventional techniques, boundary information of objects is often processed implicitly, and there is still room for improvement in segmentation accuracy for regions with complex boundary shapes and small targets. It has also been pointed out that imbalances in semantic information between different network layers can lead to local semantic loss and misclassification. Furthermore, a clear mechanism for collaboratively learning and optimizing boundary information and semantic information has not been fully established.

[0007] To address these issues, a new segmentation method is required that explicitly introduces boundary information and compensates for semantic information between layers. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] China Patent Publication CN120355928B Summary of the Invention [Problem to be solved by the invention]

[0009] The objective of the present invention is to provide a semantic segmentation method for remote sensing images that combines boundary guidance and semantic compensation mechanisms, thereby solving the problems of semantic misalignment, blurred boundaries, difficulty in identifying small targets, and reduced computational efficiency in existing remote sensing image segmentation models, and improving the completeness, continuity, and category distinction of feature boundaries in complex scenes. [Means for solving the problem]

[0010] In order to achieve the above object, the present invention provides a semantic segmentation method for remote sensing images that combines boundary guidance and semantic compensation mechanisms, including: Step 1: constructing a boundary-guided semantic compensation network, including an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module, and an auxiliary boundary monitoring module; Step 2: input the preprocessed original remote sensing image into the boundary-guided semantic compensation network for feature extraction; Step 3: performing semantic compensation and feature enhancement using a cross-layer semantic compensation module; Step 4 performs global-local context decoding; Step 5: performing boundary guidance and auxiliary monitoring using an auxiliary boundary monitoring module; Step 6: Generate a segmentation result and perform joint optimization by combining the main segmentation loss and the auxiliary boundary loss; and (7) performing semantic segmentation using the optimized boundary-guided semantic compensation network.

[0011] Preferably, the execution process of step 2 comprises: Step 2.1 inputs the original RGB remote sensing image and performs preprocessing; Step 2.2 performs multi-scale feature extraction using an EfficientNet-B3 based encoder.

[0012] Preferably, in the execution process of step 3, a cross-layer semantic compensation module is used to dynamically inject high-level semantic information into low-level features, and the execution process includes: Step 3.1 aligns high-level semantic features; Step 3.2 generates dual attention weights that integrate the spatial attention gating unit and the channel compression and excitation unit; Step 3.3 optimizes the low-level feature space; and step 3.4, which performs feature fusion and receptive field enhancement.

[0013] Preferably, the process of executing step 4 comprises: Step 4.1, which serializes the features and processes them with a Transformer; Step 4.2 involves incremental upsampling and feature fusion to reconstruct the semantic mapping.

[0014] Preferably, the process of executing step 5 comprises: Step 5.1 fuses multi-scale decoded features; Step 5.2, which performs edge-sensitive feature enhancement; and step 5.3 of performing auxiliary boundary prediction. [Effects of the Invention]

[0015] This invention provides a semantic segmentation method for remote sensing images that combines boundary guidance and semantic compensation mechanisms. The method uses an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module, and an auxiliary boundary monitoring module to perform semantic segmentation. After inputting the original remote sensing image, the encoder extracts hierarchical features with strong semantic information and different spatial resolutions. The cross-layer semantic compensation module then generates fusion features that combine high-level semantics and fine spatial details. The powerful modeling capabilities of the Transformer are then used to integrate global context and gradually reconstruct a high-resolution semantic mapping through fusion with the compensated features, ultimately outputting the segmentation result. During the training process, the boundary-guided semantic compensation network further utilizes an auxiliary boundary monitoring module to reduce the loss of boundary details, thereby improving the accuracy and robustness of the segmentation.

[0016] In order to more clearly describe the technical solutions in the embodiments of the present invention or the prior art, the following briefly describes the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without paying creative labor. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a structural schematic diagram of a boundary-guided semantic compensation network in the remote sensing image semantic segmentation method that combines boundary guidance and semantic compensation mechanisms according to the present invention; [Figure 2] 1 is a structural schematic diagram of a cross-layer semantic compensation module (CLSCM) according to the present invention; [Figure 3] 1 is a structural schematic diagram of an auxiliary boundary monitoring module (ABSM) according to the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0018] The following detailed description of the embodiments of the present invention is provided in the accompanying drawings, in which the same or similar reference numerals throughout the drawings represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are for the purpose of explaining the present invention and should not be construed as limiting the present invention.

[0019] The present invention provides a remote sensing image semantic segmentation method that combines boundary guidance and semantic compensation mechanisms, including: Step 1: constructing a boundary-guided semantic compensation network, including an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module, and an auxiliary boundary monitoring module; Step 2: input the preprocessed original remote sensing image into the boundary-guided semantic compensation network for feature extraction; Step 3: performing semantic compensation and feature enhancement using a cross-layer semantic compensation module; Step 4 performs global-local context decoding; Step 5: performing boundary guidance and auxiliary monitoring using an auxiliary boundary monitoring module; Step 6: Generate a segmentation result and perform joint optimization by combining the main segmentation loss and the auxiliary boundary loss; and (7) performing semantic segmentation using the optimized boundary-guided semantic compensation network.

[0020] The following description will be further explained with reference to the specific structure and execution process. The structure of the boundary-guided semantic compensation network in step 1 is shown in Figure 1. Its core consists of four main components: an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module (CLSCM), and an auxiliary boundary monitoring module (ABSM). The encoder can efficiently extract multi-scale local features, while the CLSCM is embedded in the encoder stage to inject high-level semantics into low-level features. This dynamic semantic compensation enhances the expressive ability for small targets and fine-grained structures. The decoder uses a global-local Transformer block to capture long-range dependencies and aggregate context information. At the same time, the ABSM obtains features from each decoder stage and uses them as input to construct the auxiliary boundary monitoring branch. By introducing auxiliary loss during training, the ABSM explicitly guides the network to focus on boundary regions, thereby improving boundary integrity and category separability.

[0021] The purpose of step 2 is to input the original remote sensing image and extract hierarchical features with strong semantic information and different spatial resolutions through an encoder network. The process is as follows:

[0022] Step 2.1: Input Preprocessing Input an original RGB remote sensing image I with size H × W × 3. Then, perform the standardization process: I norm = (1-μ) / σ, where μ and σ are the mean and standard deviation of the dataset. Data enhancement techniques such as random cropping and inversion are applied to obtain the network input tensor X.

[0023] Step 2.2: Multiscale feature extraction X is input to an EfficientNet-B3 based encoder, which goes through multiple stages of downsampling to generate a set of feature maps E i (i=1,2,3,4). The main output features are defined as follows:

[0024] Deep features: Low resolution, contain a lot of high-level semantic information, and have a wide receptive field.

[0025] Mid-level features: Medium level of resolution and semantics.

[0026] Shallow features: High resolution and preserves spatial details such as rich edges and textures, but low semantics.

[0027] Output of step: a set of multi-scale features E i (i=1,2,3,4).

[0028] Step action: Complete an initial multi-layer understanding of the input image, laying the foundation for subsequent fusion and compensation.

[0029] In step 3, a cross-layer semantic compensation module (CLSCM) is used to dynamically inject high-level semantic information into low-level features to bridge the semantic gap and improve the feature representation ability for small targets and complex structures. The detailed structure of CLSCM is shown in Figure 2.

[0030] Step 3.1: Aligning high-level semantic features Specifically, each encoder stage E i (i=1,2,3), all the deeper features (E i+1 The combined tensors (from E1 to E4) are combined along the channel dimension to integrate richer high-level semantic information. Then, a pixelshuffle upsampling operation is used to process the combined tensor, reallocating some of the channel information to the spatial dimension to improve spatial resolution and detail representation. After that, a 1x1 convolution is used to adjust the number of channels to ensure that the high-level features are perfectly aligned with the low-level features in terms of resolution and dimension. The formula is as follows:

[0031] JPEG0007821458000002.jpg13133where F h denotes the combined high-level features, and Up(F h ) denotes the upsampling operation by pixelshuffle pixel rearrangement, and F h ’ denotes the upsampled and channel-aligned high-level feature map, which serves as the input for the subsequent attention-based fusion.

[0032] Step 3.2: Generating dual attention weights Next, CLSCM performs attention-based weighting enhancement on the original low-level features and the aligned high-level features. For the high-level feature branch, the module integrates a spatial attention gating unit (SAGU) and a channel compression and excitation unit (CSEU) to dynamically adjust semantic responses through a dual attention mechanism in spatial and channel dimensions. Specifically, the input features are first projected by 1x1 convolution to generate three feature groups: Query (Q), Key (K), and Value (V). Here, (Q) and (V) are processed by the SAGU and CSEU, respectively, to enhance spatial saliency and channel selectivity. The enhanced Q and K are then summed and element-wise multiplied with V to obtain semantically weighted high-level features. Finally, convolution and dropout operations are performed to reduce overfitting and generate discriminative high-level semantic feature weights. The relevant formulas are as follows:

[0033] JPEG0007821458000003.jpg49154Here, the circle (with a circle inside it) indicates the Hadamard product, σ indicates a sigmoid activation function, δ indicates a ReLU activation function, and B indicates batch normalization. s and E c correspond to the outputs of SAGU and CSEU, respectively. By stacking them, we can realize joint modeling of spatial and channel attention, thereby obtaining an attention-enhanced feature representation, which is denoted as φ(·), and F h ’ ’ denotes the refined high-level semantic weight.

[0034] Step 3.3: Optimizing the low-level feature space In the low-level branch, the original low-level features are first normalized by layer normalization, and a spatial attention map (Att) is generated using a 3x3 depth-separable convolution. At the same time, a value tensor (V) is generated using a 1x1 convolution. Then, the Hadamard product (Att〇V) is used to weight the key regions. Then, a 1x1 convolution is applied to match the number of channels, thereby obtaining spatially optimized low-level features. The relevant formula is as follows:

[0035] JPEG0007821458000004.jpg32158 Step 3.4: Feature fusion and receptive field enhancement Finally, we multiply the weighted high-level semantic features by the spatially optimized low-level features element-wise and pass the result through one convolutional layer to generate a fused feature F. fused This allows for dynamic and hierarchical integration of fine-grained details and semantic context. However, its limited receptive field limits its ability to model global context. To address this issue, a large-scale selective kernel (LSK) module is introduced into CLSCM, which combines multi-scale deep unit separable convolutions with an attention-based selection mechanism to enhance the ability to sense small objects and fine structures. JPEG0007821458000005.jpg53170This module avoids the sampling defects of dilated convolution, expands the receptive field, and improves robustness to blurred boundaries, small objects, and background noise. It compensates for the limitations of CLSCM in spatial modeling, improving segmentation accuracy and generalization ability in complex scenes. The corresponding formula is defined as follows:

[0036] JPEG0007821458000006.jpg55155Here, σ denotes the sigmoid activation function and DSConv(·) denotes the depthwise separable convolution.

[0037] Output of step: Semantically compensated and enhanced features F out .

[0038] Step action: Fill in the semantic gaps between features at different levels and generate fusion features that combine high-level semantics and fine spatial details.

[0039] In step 4, a Transformer decoder is used to upsample the compensated encoder features and aggregate global and local context information to gradually restore the spatial resolution of the segmentation map. The steps are as follows:

[0040] Step 4.1: Serializing features and processing them with the Transformer The deepest features output from the encoder (or the features after fusion with the CLSCM output) are flattened into a sequence and then input into a Transformer-based decoder, where a global-local attention block (GLB) can simultaneously capture long-range dependencies (global context) and the relevance of local details.

[0041] Step 4.2: Incremental upsampling and feature fusion The decoder upsamples in multiple stages, with each upsampling stage performing skip connections and fusion between the corresponding resolution features in the encoder that have been compensated by CLSCM. This design ensures that the enhanced spatial details in the encoder are still utilized while recovering the resolution.

[0042] Output of the step: A set of decoder features {D1, D2, D1} of increasing resolution, and the final primal partitioning logical feature map.

[0043] Stepwise operation: By utilizing the powerful modeling capabilities of Transformer to integrate global context and fuse it with compensated features, a high-resolution semantic mapping is reconstructed stepwise.

[0044] Step 5 is the boundary guidance and auxiliary monitoring process, where the auxiliary boundary monitoring module (ABSM) explicitly guides the network to focus on the boundary regions of objects, improving the boundary accuracy and continuity of the segmentation results. The detailed structure of ABSM is shown in Figure 3. However, it should be noted that this module is only enabled during the training phase.

[0045] Step 5.1: Fusion of multi-scale decoded features First, the multi-stage decoder features {D1, D2, D1} are processed in two parallel branches: one 3x3 convolution is used to obtain contextual cues, and the other 1x1 convolution is used to extract fine details. The output undergoes normalization (batch normalization) and activation (ReLU6 activation function) processing, followed by fusion and upsampling via bilinear interpolation. Combining these with earlier higher-resolution features builds a hierarchical detail enhancement pathway, which effectively integrates multi-scale information, strengthens boundary representation, and improves segmentation accuracy and robustness. The relevant formula is as follows:

[0046] JPEG0007821458000007.jpg35148where δ represents the ReLU activation function, B represents the batch normalization operation, i represents the decoder output stage index, N represents the number of decoder layers participating in multi-scale fusion, and L represents the fused features after local details are enhanced. The output of the step: a set of decoder features {D1, D2, D1} with increasing resolution, and the final main partitioning logical feature map.

[0047] Step 5.2: Edge-sensitive feature enhancement By subtracting the original features from the pooling operation, the boundary regions are enhanced and an edge attention map is generated. The main purpose of the Edge Sensitive Feature Enhancement Module (EFEM) is to explicitly extract edge information and combine it with the original features to enhance the model's sensitivity to boundary regions and structural details, which helps improve the accuracy and continuity of boundary segmentation in remote sensing images.

[0048] As shown in Figure 3, the processing flow of EFEM is as follows: First, a pooling operation is applied to the input features to extract low-frequency background information, and the pooling result is subtracted element-by-element from the original features to generate an edge-enhanced feature map, which highlights sudden boundary transitions. Next, the map is processed using 1x1 convolution, batch normalization, and a sigmoid activation function to generate an edge attention map for weighting and fusing the original features. Subsequently, an additional attention mechanism is applied to further enhance the responses of salient regions and suppress redundant information, thereby obtaining the initial enhanced feature map. The relevant formula is as follows:

[0049] JPEG0007821458000008.jpg18160, where σ denotes the sigmoid activation function and B denotes batch normalization. To further enhance the model's attention effect on key regions, ABSM calculates the average and maximum responses in the channel dimension, generating two complementary spatial attention maps w avg and w max These attention maps allow for a more comprehensive search of salient regions. Finally, these two attention maps are multiplied element-wise with the previously obtained feature L to perform spatial enhancement, resulting in an edge-enhanced feature map F e The formula is as follows:

[0050] JPEG0007821458000009.jpg26166, where C denotes the number of channels. After undergoing the above enhancement process, the edge feature map is further processed to generate intermediate category-level prediction results suitable for auxiliary monitoring.

[0051] Step 5.3: Auxiliary boundary prediction F e Then, operations such as convolution, pixel relocation upsampling, and pooling are performed on the resulting auxiliary boundary prediction map, which has the same resolution as the primary output. First, 1x1 convolution is applied to adjust the number of feature channels, and then pixel relocation upsampling is performed, which spatially redistributes the semantic cues originally encoded in the channel dimension in the spatial domain. Unlike traditional point-based prediction, this allows each pixel to integrate the semantic information of its surrounding regions, thereby realizing a transformation from single prediction to local region fusion.

[0052] To further emphasize robustness, average pooling is applied to smooth the local fusion features, effectively suppressing noise and reducing redundancy. The combination of pixel relocation upsampling and average pooling not only emphasizes semantic continuity at boundaries, but also significantly improves the stability and clarity of edge prediction, especially in scenes with noisy backgrounds or complex adjacent category situations. Finally, a bilinear interpolation method is used to upsample the edge map, thereby matching the spatial resolution of the primary segmentation output. The corresponding calculation formula is as follows:

[0053] JPEG0007821458000010.jpg20153Here, δ represents the ReLU activation function, B denotes batch normalization, and Interp(·) denotes upsampling by bilinear interpolation.

[0054] The obtained prediction forms an auxiliary supervision branch, whose loss is used in combination with the main decoding loss in the training process. Combined with predefined weights, this auxiliary branch provides complementary gradients to guide the backbone network to focus on boundary regions, capturing fine structures and explicitly modeling the semantics of edges, thereby reducing the loss of boundary details and improving segmentation accuracy and robustness, especially in complex segmentation scenes with blurred edges, dense targets, or adjacent categories. Furthermore, it improves the structural consistency of the prediction mask, highlighting its applicability in practical applications.

[0055] In step 6, the final segmentation result is generated, and the main segmentation loss and auxiliary boundary loss are combined to perform joint optimization of the entire network.

[0056] (1) Main split output: The final output of the decoder goes through a 1x1 convolutional layer to adjust the number of channels to the target number of categories C. Through the Softmax function, the category probability distribution map P main Generate.

[0057] Final prediction map: JPEG0007821458000011.jpg17142(2) Multitask loss calculation: The total loss function is the main split loss L main and auxiliary boundary loss L aux It is constructed by weighted addition of

[0058] Principal Segmentation Loss: By combining cross-entropy loss and Dice loss, we achieve both pixel-level classification accuracy and consistency of region shape. The formulas for the cross-entropy loss function and Dice loss function are as follows:

[0059] JPEG0007821458000012.jpg52170

[0060] The combined formula for the main losses is as follows:

[0061] JPEG0007821458000013.jpg17129The formula for auxiliary boundary loss is as follows:

[0062] JPEG0007821458000014.jpg29165

[0063] The formula for total loss is:

[0064] JPEG0007821458000015.jpg15139 where λ denotes the weight of the auxiliary loss, which is set as 0.4 in the present test.

[0065] Output of the step: the final semantic segmentation prediction map Y, and the total loss L for optimizing the network during the training phase. main .

[0066] Step action: By jointly optimizing two related tasks, semantic segmentation and boundary prediction, using a multi-task learning framework, the model maintains high classification accuracy while producing segmentation results with clear and continuous boundaries. During the inference / testing phase, only the main branch path is used and the ABSM branch is discarded, so no additional computational overhead is incurred.

[0067] In step 7, semantic segmentation is performed using the optimized boundary-guided semantic compensation network.

[0068] Furthermore, the present invention further proposes examples for comparative explanation.

[0069] The encoder uses the EfficientNet-B3 network model, stages 2 to 5, and the decoder uses a Transformer global-local attention structure. A cross-layer semantic compensation module (CLSCM) is inserted between each coding layer. A decoder-driven boundary monitoring module (ABSM) is connected to the output of the decoding stage and optimizes the boundary in combination with an auxiliary loss function.

[0070] To evaluate the effectiveness of the proposed BGSC-Net, we conducted extensive comparison and ablation tests on four public datasets: Potsdam, Vaihingen, LoveDA, and UAVid. The tests were performed on a server running the Ubuntu 18.04 operating system and equipped with an NVIDIA GeForce RTX 4090 graphics processor (24GB graphics memory). The development environment was based on Python 3.8, and PyTorch 2.0.0+cu118 was used as the deep learning framework. The AdamW optimizer was used, and the CosineAnnealingLR scheduler was used to smoothly decay the learning rate. The initial learning rate of the backbone network was 1×10 -3 and the remainder is 9 × 10 -3The weight decay rate is 0.01. All tests are trained with a batch size of 8. For the UAVid dataset, input images were resized to 1024 x 1024, and the model was trained for 40 epochs. For the ISPRS Potsdam and Vaihingen datasets, images were randomly cropped to 512 x 512, and the model was trained for 45 and 105 epochs, respectively. For the LoveDA dataset, small blocks of 512 x 512 were used, and the model was trained for 45 epochs. During the training process, various data enhancement techniques, such as random rotation, flipping, brightness / contrast adjustment, cropping, scaling, and sharpening, were used to emphasize robustness to complex boundaries and diverse scenes. This setting enables the model to effectively learn fine-grained segmentation features while ensuring training efficiency and optimal resource utilization. To comprehensively evaluate the segmentation performance of the model, the present invention uses the overall accuracy rate (OA), mean intersection over union (mIoU), and mean F1 score (mF1) as the main evaluation metrics.

[0071] The comparison methods include MANet, A2-FPN, UNetFormer, and SFANet, and the method of the present invention is referred to as BGSC-Net.

[0072] Test results: (1) Comparative tests are conducted on four datasets: Potsdam, Vaihingen, LoveDA, and UAVid.

[0073] Table 1 Comparison test in Potsdam JPEG0007821458000016.jpg36170

[0074] Table 2 Comparative study in Vaihingen JPEG0007821458000017.jpg35170

[0075] Table 3 Comparative test in LoveDA JPEG0007821458000018.jpg34170

[0076] Table 4 Comparison test on UAVid JPEG0007821458000019.jpg35170

[0077] The results show that the experimental results of the present invention on the four datasets are clearly improved in three indicators.

[0078] (2) We conduct ablation experiments on four datasets: Potsdam, Vaihingen, LoveDA, and UAVid.

[0079] Table 5 Ablation experiments in Potsdam JPEG0007821458000020.jpg37170

[0080] Table 6 Ablation experiments in Vaihingen JPEG0007821458000021.jpg37170

[0081] Table 7 Ablation experiment on LoveDA JPEG0007821458000022.jpg37170

[0082] Table 8 Ablation experiment using UAVid JPEG0007821458000023.jpg37170

[0083] According to the results, the two innovative modules proposed in the method of the present invention improved the segmentation performance of the model.

[0084] What has been disclosed above is only one or more preferred embodiments of the present invention, which of course cannot limit the scope of the claims of the present invention. Those skilled in the art can understand that they can implement all or part of the flow of the above embodiments and make equivalent changes according to the claims of the present invention, and still belong to the scope covered by the invention.

Claims

1. A method for semantic segmentation of remote sensing images that combines boundary induction and semantic compensation mechanisms, comprising: Step 1: constructing a boundary-guided semantic compensation network, including an EfficientNet-B3-based encoder, a Transformer-based decoder, a cross-layer semantic compensation module, and an auxiliary boundary monitoring module; Step 2: inputting the preprocessed original remote sensing image into the boundary-guided semantic compensation network for feature extraction; Step 3: performing semantic compensation and feature enhancement using a cross-layer semantic compensation module; Step 4 of performing global-local context decoding; Step 5: performing boundary guidance and auxiliary monitoring using an auxiliary boundary monitoring module; Step 6: Generate a segmentation result and perform joint optimization by combining the main segmentation loss and the auxiliary boundary loss; and (7) performing semantic segmentation using the optimized boundary-guided semantic compensation network.

2. The execution process of step 2 is as follows: Step 2.1: inputting the original RGB remote sensing image and preprocessing it; and (2) performing multi-scale feature extraction using an EfficientNet-B3 based encoder.

3. In the execution process of step 3, a cross-layer semantic compensation module is used to dynamically inject high-level semantic information into low-level features, and the execution process includes: Step 3.1 of matching high-level semantic features; Step 3.2: generating dual attention weights that integrate the spatial attention gating unit and the channel compression and excitation unit; Step 3.3 of optimizing the low-level feature space; 3.

4. The method for semantic segmentation of remote sensing images with boundary guidance and semantic compensation mechanism fusion according to claim 2, further comprising: step 3.

5.

4. The execution process of step 4 is as follows: Step 4.1: serializing features and processing them with a Transformer; 4.

2. The method for semantic segmentation of remote sensing images that combines boundary guidance and semantic compensation mechanisms, as described in claim 3, further comprising: step 4.

3. Reconstructing semantic mapping by performing upsampling and feature fusion in stages.

5. The execution process of step 5 is as follows: Step 5.1 of fusing multi-scale decoded features; Step 5.2 of edge-sensitive feature enhancement; 5. The method for semantic segmentation of remote sensing images with boundary guidance and semantic compensation mechanism fused according to claim 4, further comprising: step 5.3 of performing auxiliary boundary prediction.

Images