Semantic segmentation method for remote sensing images based on feature interaction enhancement

Abstract

This invention discloses a semantic segmentation method for remote sensing images based on feature interaction enhancement. The steps include acquiring remote sensing images and extracting features to obtain preliminary features; performing feature interaction enhancement based on the preliminary features to obtain enhanced features; and performing pixel-level segmentation based on the enhanced features. This invention combines standard convolution, depthwise separable convolution, and dilated convolution to interactively enhance the preliminary features, thereby improving the ability to capture multi-scale features in remote sensing images, achieving collaborative extraction of global and local features, and improving the accuracy and robustness of semantic segmentation.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically to a remote sensing image semantic segmentation method based on feature interaction enhancement. Background Technology

[0002] With the rapid development of remote sensing technology, high-resolution remote sensing imagery has been widely used in fields such as environmental monitoring, disaster assessment, and crop yield assessment. Semantic segmentation, as one of the key tasks in computer vision, can classify each pixel in an image.

[0003] However, high-resolution remote sensing images are characterized by complex backgrounds and large differences in target scale. In traditional semantic segmentation, feature extraction methods mainly rely on spectral and spatial features, but these methods often perform poorly when dealing with complex scenes.

[0004] In recent years, deep learning technology has made significant progress in the field of semantic segmentation, especially the application of methods such as convolutional neural networks (CNNs) and Transformers, including attention mechanisms and multi-scale feature fusion. Furthermore, Transformer-based semantic segmentation methods have also achieved remarkable progress, with Swin-Transformer demonstrating outstanding performance in object detection and segmentation tasks.

[0005] While the methods described above significantly improve segmentation accuracy and robustness, existing deep learning techniques (such as attention mechanisms and multi-scale feature fusion) still have some shortcomings and drawbacks in terms of global and local feature extraction. For example, attention mechanisms can effectively enhance the feature representation of important regions, but their computational complexity is high, especially when processing large-scale remote sensing images, which may lead to slower inference speeds. Furthermore, over-focusing on local features may lead to the neglect of global background information, thereby affecting the understanding of the overall semantics of the target.

[0006] Regarding segmentation accuracy, although multi-scale feature fusion can effectively extract image information at different scales, in complex scenes, targets with large scale differences may still be affected by inconsistencies in information between features at different scales, leading to reduced segmentation accuracy. Furthermore, these methods often rely on large amounts of training data and computational resources, limiting performance improvement when data is scarce or computational power is limited.

[0007] In summary, although deep learning technology has demonstrated excellent performance in semantic segmentation of remote sensing images, there are still areas for improvement when processing images with complex backgrounds and targets of different scales, such as computational efficiency and balanced processing of features at different scales. Summary of the Invention

[0008] In view of this, in order to at least partially solve the above-mentioned technical problems, the present invention provides a remote sensing image semantic segmentation method based on feature interaction enhancement. It aims to combine operations such as standard convolution, depthwise separable convolution and dilated convolution to enhance the ability to capture multi-scale features in remote sensing images, realize the collaborative extraction of global and local features, and thus improve the accuracy and robustness of semantic segmentation.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A semantic segmentation method for remote sensing images based on feature interaction enhancement includes:

[0011] Acquire remote sensing images and extract features to obtain preliminary features;

[0012] Enhanced features are obtained by performing feature interaction enhancement based on preliminary features.

[0013] And pixel-level segmentation based on enhanced features;

[0014] Feature interaction enhancement includes the following steps:

[0015] The initial features are input into the upper and middle branch networks respectively for local feature extraction. The results are then multiplied element-wise and then passed through an activation function to obtain the attention weight map.

[0016] The initial features are input into the lower branch network to extract global features, and the global features are multiplied with the attention weight map and then made into a skip connection with the initial features to obtain the initial enhanced features;

[0017] The initial enhancement features are batch normalized and then input into the spatial channel reconstruction convolutional branch and the depthwise separable convolutional branch, respectively. The results are multiplied together and then skipped to the initial enhancement features to obtain the enhancement features.

[0018] As a preferred embodiment, the upper branch network includes a first convolution and a full-dimensional dynamic convolution;

[0019] The middle branch network includes a second convolution and a first depthwise separable convolution;

[0020] The lower branch network includes dilated convolutions and a second depthwise separable convolution.

[0021] As a preferred embodiment, the spatial channel reconstruction convolution branch includes a third convolution and a spatial channel reconstruction convolution;

[0022] The depthwise separable convolution branch includes a fourth convolution, a third depthwise separable convolution, and an activation function.

[0023] As a preferred approach, the steps for feature extraction to obtain preliminary features include:

[0024] Extract local features from remote sensing images and reduce their dimensionality;

[0025] High-level semantic features are extracted stepwise from the dimensionality-reduced features using the ResNet backbone network;

[0026] Capture target details at different scales in high-level semantic features based on attention mechanisms;

[0027] As a preferred approach, dimensionality reduction includes sequentially performing batch normalization, ReLU activation function, and max pooling layer on local features.

[0028] As a preferred approach, attention mechanisms are used to capture target details at different scales within high-level semantic features, including:

[0029] Generate multi-scale features based on advanced semantic features;

[0030] Attention weights for features at each scale are determined using an attention network;

[0031] Multi-scale features are fused using an interactive weighting method.

[0032] Preferably, pixel-level segmentation is performed based on enhanced features, including:

[0033] By using multiple dilated convolutional layers, features of different scales are generated based on the enhanced features to improve the receptive field of the enhanced features;

[0034] Fuse features at different scales;

[0035] Upsampling of the fused features

[0036] The upsampled features are classified using convolutional layers to obtain the category label for each pixel.

[0037] As a preferred approach, after passing through the dilated convolutional layer, activation functions and batch normalization are added sequentially.

[0038] As a preferred method, the Softmax function is used to normalize each pixel to generate the probability distribution of each category, thus obtaining the segmentation result.

[0039] Compared with existing technologies,

[0040] 1) By constructing a feature interaction enhancement module, this invention can effectively capture multi-scale features in remote sensing images and achieve collaborative extraction of global and local features. Compared with traditional methods, it significantly improves the accuracy and robustness of semantic segmentation, making the model more adaptable and accurate when dealing with complex scenes and targets of different scales.

[0041] 2) This invention employs techniques such as batch normalization and skip links in the feature extraction process, which not only stabilizes network training and accelerates convergence, but also preserves the detailed information in the original features. At the same time, it integrates high-level features into the output features, forming enhanced features containing rich contextual information and detailed features. This provides higher-quality input for the subsequent segmentation network, further improves the segmentation effect, reduces the gradient vanishing or gradient exploding problem caused by inconsistent input feature distribution, and optimizes the training process and performance of the model. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the remote sensing image semantic segmentation method based on feature interaction enhancement according to the present invention;

[0044] Figure 2 This is a flowchart illustrating the enhanced interactive features of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0046] This invention discloses a remote sensing image semantic segmentation method based on feature interaction enhancement. By constructing a hybrid feature interaction module and combining operations such as standard convolution, depthwise separable convolution, and dilated convolution, the method enhances the ability to capture multi-scale features in remote sensing images, and achieves collaborative extraction of global and local features. This improves the accuracy and robustness of semantic segmentation, providing richer and more expressive input data for subsequent semantic segmentation networks, thus better serving the task of accurate segmentation of targets such as ships in remote sensing images.

[0047] In one embodiment, the remote sensing image semantic segmentation method based on feature interaction enhancement includes the following three steps:

[0048] Step 1: Acquire remote sensing images and extract features to obtain preliminary features;

[0049] Step 2: Perform feature interaction enhancement based on the preliminary features to obtain enhanced features;

[0050] Step 3: Perform pixel-level segmentation based on the enhanced features.

[0051] This invention not only enhances the ability to capture multi-scale features in remote sensing images and achieves collaborative extraction of global and local features, but also improves the accuracy and robustness of semantic segmentation.

[0052] In one embodiment, reference is made to... Figure 1 The execution process for each step is as follows:

[0053] Step 1: Extract features from the remote sensing image;

[0054] 1.1 Extract local features from remote sensing images and reduce dimensionality; dimensionality reduction includes batch normalization, ReLU activation function and max pooling layer for local features in sequence.

[0055] In this embodiment, the input remote sensing image is a 3-channel RGB image with dimensions of H×W×3, where H and W represent the height and width of the image, respectively, and 3 represents the number of color channels. The remote sensing image carries the original visual information of the target, including the texture, color, and other features of the target and background elements.

[0056] This application extracts local features from remote sensing images using convolutional layers. In one embodiment, the convolutional kernel is 7×7, the stride is set to 2, and the padding is set to 3. The size of the convolutional kernel determines the range of local regions that the model perceives in the image. In this embodiment, using a 7x7 convolutional kernel can capture a larger range of local features, such as edges and textures in the image. The stride of 2 and the padding of 3 ensure that the convolution operation does not lose important spatial information.

[0057] At the same time, the image was downsampled, reducing the image resolution and making subsequent feature extraction more efficient.

[0058] Through initial convolution operations, the model can capture low-level features in the image (such as edges, corners, and textures) while reducing image resolution and computational cost. The image size is reduced from H×W×3 to approximately... C1 represents the number of channels after the convolution operation (usually 64 or 128).

[0059] Furthermore, dimensionality reduction is performed on local features;

[0060] In this embodiment, after performing the convolution operation, a batch normalization layer is first applied. The function of this layer is to normalize the output of each layer to the same scale, thereby reducing the gradient vanishing or gradient exploding problem, helping to accelerate training and improve stability.

[0061] Secondly, the ReLU activation function (Rectified LinearUnit) is applied, which converts all negative values ​​to zero, introducing non-linearity and enabling the model to learn more complex feature representations. The purpose of this step is to enhance the model's expressive power, allowing it to extract more meaningful low-level features.

[0062] Next, max pooling is applied. This embodiment uses a 3×3 pooling window with a stride of 2. The purpose of this step is to further reduce the spatial resolution of the image, preserve the most important feature information, and enhance the spatial invariance of the model (i.e., robustness to image displacement). The pooling operation reduces the image size by selecting the maximum value in a local region while maintaining the salient features of the image. The size of the output feature map of the pooling layer will become... This further compresses the amount of information in the image.

[0063] 1.2 Utilize the ResNet backbone network to progressively extract high-level semantic features from the dimensionality reduction features;

[0064] This application introduces ResNet (Residual Network) as the backbone network. ResNet extracts high-level semantic features of images step by step through multiple stages. Each stage contains several convolutional layers and uses residual connections to avoid the gradient vanishing problem in deep networks, enabling the network to be trained deeper and extract more complex and abstract features.

[0065] Each residual module enhances the network's expressive power and effectively prevents performance degradation during deep network training by ensuring that information can flow through the network by directly adding the input to the output.

[0066] As the number of network layers increases, ResNet gradually extracts high-level semantic information from images. For example, the initial convolutional layers mainly capture local edge and texture features, while as the network deepens, ResNet begins to understand more complex patterns, such as the shape of objects, textures, and background distinctions. The output feature map of the ResNet network has a size of 4H×4W×C, where C is the final number of channels, typically 2048 or more. At this point, the feature map already contains rich semantic information, including the shape of objects and the texture of the background.

[0067] 1.3 Capturing target details at different scales in high-level semantic features based on attention mechanisms;

[0068] High-level semantic features extracted by the ResNet backbone have been generated, with feature maps of size 4H×4W×C, where H and W are the height and width of the image, and C is the final number of channels (typically 2048 or more). These feature maps contain semantic information about ship targets in remote sensing images, but may not fully capture target details at different scales.

[0069] Therefore, this application introduces the MAMBA (Multi-scale Attention Module with BilinearAttention) module, whose core is a multi-scale attention mechanism that enhances the expressive power of images by fusing features at different scales.

[0070] The MAMBA module mainly consists of the following two parts:

[0071] 1) Multi-scale attention mechanism: Assign different weights to feature maps of different scales to enhance the fusion ability of multi-scale information;

[0072] 2) Bilinear attention mechanism: Enhances the interaction between each scale, captures local and global information at different scales in detail, and avoids information loss.

[0073] In one embodiment, the capture step includes:

[0074] 1.3.1 Generating feature maps at multiple scales based on high-level semantic features;

[0075] This embodiment employs multiple convolutional or pooling layers to generate feature maps of multiple scales from the original feature map. For example, feature maps of multiple scales can be generated using convolutional kernels of different sizes (such as 3x3, 5x5, etc.) or different pooling operations (such as max pooling, average pooling, etc.). These feature maps of different scales can capture details of targets of different sizes.

[0076] 1.3.2 Determine the attention weights of features at each scale using an attention network;

[0077] Attention networks consist of convolutional or fully connected layers, and the resulting weights reflect the importance of features at different scales. In remote sensing images, the scale of targets varies considerably, and MAMBA dynamically adjusts the weights for each scale based on the content of the feature map.

[0078] 1.3.3 Multi-scale features are fused using an interactive weighting method.

[0079] In this embodiment, a bilinear attention mechanism is used to effectively fuse feature maps of different scales through an interactive weighting method, thereby enhancing the model's expressive power at different scales.

[0080] Furthermore, the resulting fused feature maps are further processed through an activation function (such as ReLU) to output feature maps for subsequent networks. These feature maps are richer and can capture details and background information of ship targets at different scales.

[0081] The output feature maps from the MAMBA module still have a size of 4H×4W×C, where C is the number of channels after processing by the MAMBA module (usually unchanged, but can be adjusted as needed). These output feature maps are more refined than those without the MAMBA module and contain multi-scale information, thus better supporting subsequent segmentation tasks.

[0082] The feature map output by the MAMBA module will be used as input for subsequent feature interaction enhancement, and then passed to the decoder for further upsampling and segmentation operations.

[0083] Through multi-scale feature enhancement using the MAMBA module, the network can better recover detailed information about the target and avoid losing important target information during the upsampling process.

[0084] This application extracts low-level to high-level semantic features step by step by applying convolutional layers, batch normalization, ReLU activation function, max pooling, and using a ResNet backbone network, which can efficiently capture the morphology and background information of ship targets.

[0085] Step 2: Enhance features through interactive mapping based on the preliminary feature map;

[0086] In one embodiment, this application constructs a feature interaction enhancement module, with the structure as follows: Figure 2 As shown, this module enhances the ability to capture multi-scale features in remote sensing images by combining standard convolution, depthwise separable convolution, and dilated convolution, achieving collaborative extraction of global and local features and improving the accuracy and robustness of semantic segmentation.

[0087] Reference Figure 2 The interaction enhancement steps are as follows:

[0088] 2.1 The preliminary feature maps are input into the upper branch and middle branch networks respectively for local feature extraction. The results are then multiplied element-wise and then passed through an activation function to obtain the attention weight map.

[0089] In a preferred embodiment, the preliminary feature map is first... Batch normalization is performed to obtain To stabilize network training and accelerate convergence.

[0090] Furthermore, The inputs are fed into the upper branch and middle branch networks respectively to extract local features and reduce computational complexity. The upper branch network includes a first convolution and a full-dimensional dynamic convolution; the middle branch network includes a second convolution and a first depthwise separable convolution.

[0091] Then, the features extracted from the two branches are multiplied element-wise, and the attention weight map is obtained by applying the Softmax activation function. The expression is:

[0092]

[0093] Wherein, DWConv represents depthwise separable convolution, which extracts local features; Conv represents convolution; and OCOnv represents full-dimensional dynamic convolution operation. σ represents the multiplication operation, and σ represents the Softmax activation function.

[0094] 2.2 To compensate for the insufficient attention paid to local features by the upper and middle branches, the feature interaction enhancement module of this application designs a lower branch to capture global context information; that is, the preliminary feature map is input into the lower branch network to extract global features. The lower branch network includes dilated convolution and a second depthwise separable convolution.

[0095] In one embodiment, for A 7×7 dilated convolution operation is applied, which expands the receptive field of the convolution kernel without increasing computation, thus effectively capturing a wider range of global features. Next, depthwise separable convolution is used to extract local features, yielding the next branch output.

[0096]

[0097] Here, DConv represents dilated convolution. By combining dilated convolution and depthwise separable convolution, long-range dependencies in remote sensing images are effectively captured, improving the ability to recognize complex scenes.

[0098] Furthermore, the global feature is multiplied by the attention weight map and then combined with the preliminary feature map. Perform skip connections to obtain preliminary enhanced feature maps. Preferably, before performing the skip connections, the multiplication result is applied through a convolutional layer, as shown in the expression:

[0099]

[0100] in, This indicates an addition operation. By using skip links, we can preserve the detailed information in the original features while fusing the high-level features extracted by the feature enhancement module into the output features, thus forming an enhanced feature map containing rich contextual information and detailed features.

[0101] 2.3 Preliminary Enhanced Feature Map Batch normalization is performed to accelerate model convergence and reduce gradient vanishing or exploding problems caused by inconsistent input feature distributions during training; the batch-normalized feature map is denoted as... By adjusting the mean and variance, it makes the numerical range of each feature channel more consistent, which helps the network learn more expressive features.

[0102] The inputs are then fed into the spatial channel reconstruction convolution branch and the depthwise separable convolution branch, respectively. The spatial channel reconstruction convolution branch includes a third convolution and a spatial channel reconstruction convolution. The spatial channel reconstruction convolution reduces redundant information by reconstructing the spatial and channel dimensions. The depthwise separable convolution branch includes a fourth convolution, a third depthwise separable convolution, and an activation function.

[0103] Multiplying the results further, we get

[0104]

[0105] Here, SCConv represents spatial channel reconstruction convolution, and represents pointwise multiplication.

[0106] Then, to Perform convolutional layer operations to extract higher-level features; and combine with initial feature enhancement. Adding them together yields the enhanced features.

[0107] This additive operation not only preserves the information of the original features but also integrates high-level features processed by multiple convolutions and depthwise separable convolutions, thereby enhancing the model's ability to understand complex scenes. Through this operation, the expressive power of the feature maps is significantly enhanced, enabling them to better serve the semantic segmentation task of remote sensing images.

[0108] Finally, the resulting enhanced feature map It combines the original feature maps with new features obtained through complex convolution operations. A feature map of size H×W×C contains rich semantic information about the target in the image. H×W represents the spatial resolution of the image, while C represents the number of channels, typically 2048 or more, depending on the configuration of the ResNet backbone network. This enhances the feature map. This provides richer input data for subsequent semantic segmentation networks.

[0109] Step 3: Perform pixel-level segmentation based on the enhanced features.

[0110] In one embodiment, the enhanced features The feature maps are fed into subsequent convolutional networks to generate the results of the segmentation task. The high-level semantic information has already been extracted through the previous steps, so it can be used as input to the segmentation network for further processing and to generate a segmentation map.

[0111] In this embodiment, dilated convolution is applied to capture a wider range of contextual information and further enhance the model's resolution and detail. Dilated convolution expands the receptive field of the convolution kernel by inserting gaps between the elements of the kernel without increasing computation or pooling operations. Dilated convolution can effectively preserve details in feature maps while capturing greater contextual information in the image, enabling the model to more comprehensively understand the structure in the image. By applying dilated convolutions at different scales, the model can simultaneously acquire local details and global semantic information, which is crucial for accurate object segmentation.

[0112] In one exemplary embodiment, the segmentation step includes:

[0113] 3.1 By using multiple dilated convolutional layers, feature maps of different scales are generated based on the enhanced features to gradually improve the receptive field of the enhanced features and capture a wider range of contextual information; preferably, after each convolution operation, an activation function (such as ReLU) and batch normalization are added to ensure stable network training.

[0114] 3.2 Feature maps of different scales are fused to make full use of contextual information at different scales.

[0115] In this application, the feature fusion method includes feature concatenation and feature summation. Through multi-scale feature fusion, the model can extract information from different scales, improving the accuracy of the segmentation results. Especially when facing small targets or complex backgrounds, it can effectively improve the performance of detection and segmentation.

[0116] Furthermore, this application processes the fused feature map through convolutional layers to further compress the size of the feature map while retaining important multi-scale information.

[0117] 3.3 Upsampling of fused features,

[0118] After dilated convolution and feature fusion, the resolution of the feature map may be low. To generate the final segmentation result, it is necessary to upsample the feature map during the decoding process to restore its resolution to the same size as the input image, i.e., H×W×C. out Cout This is the number of channels in the segmentation result, which is usually equal to the number of categories.

[0119] In this application, upsampling can be achieved through deconvolution, upsampling layers, or bilinear interpolation. Upsampling restores the spatial resolution of the feature map for pixel-level segmentation. During this process, the model gradually restores image details, generating a segmentation map with the same size as the input image.

[0120] 3.4 The upsampled features are classified using a 1x1 convolutional layer to obtain the category label for each pixel.

[0121] This step maps the features of each pixel to the final class space, obtaining a class prediction for each pixel. A 1x1 convolution can independently classify each pixel, outputting the probability or label of the corresponding class. Because the convolution kernel size is 1x1, it does not affect the spatial resolution of the image, but it can assign a classification label to each pixel based on the previously extracted features.

[0122] A 1x1 convolutional layer is used to generate a class label for each pixel, resulting in the final segmentation map. Then, the Softmax function is used to normalize each pixel, generating the probability distribution for each class, and finally obtaining a classification result for each pixel.

[0123] Compared with the prior art, the present invention has the following beneficial effects:

[0124] This invention not only improves segmentation accuracy but also reduces the resource requirements for model training and deployment through efficient parameter fine-tuning and the use of adapters, making it suitable for remote sensing image processing tasks in various computing environments. This method exhibits greater adaptability and accuracy when handling complex scenes and targets of different scales, significantly enhancing the performance of semantic segmentation in remote sensing images.

[0125] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0126] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semantic segmentation method for remote sensing images based on feature interaction enhancement, characterized in that, Acquire remote sensing images and extract features to obtain preliminary features; Enhanced features are obtained by performing feature interaction enhancement based on preliminary features. And pixel-level segmentation based on enhanced features; Feature interaction enhancement includes the following steps: The initial features are input into the upper and middle branch networks respectively for local feature extraction. The results are then multiplied element-wise and then passed through an activation function to obtain the attention weight map. The initial features are input into the lower branch network to extract global features, and the global features are multiplied with the attention weight map and then made into a skip connection with the initial features to obtain the initial enhanced features; The upper branch network includes the first convolution and the full-dimensional dynamic convolution; The middle branch network includes a second convolution and a first depthwise separable convolution; The lower branch network includes dilated convolutions and a second depthwise separable convolution; The initial enhancement features are batch normalized and then input into the spatial channel reconstruction convolutional branch and the depthwise separable convolutional branch respectively. The results are multiplied together and then skipped to the initial enhancement features to obtain the enhancement features. The spatial channel reconstruction convolution branch includes the third convolution and the spatial channel reconstruction convolution; The depthwise separable convolution branch includes a fourth convolution, a third depthwise separable convolution, and an activation function.

2. The remote sensing image semantic segmentation method according to claim 1, characterized in that, The steps for feature extraction to obtain preliminary features include: Extract local features from remote sensing images and reduce their dimensionality; High-level semantic features are extracted stepwise from the dimensionality-reduced features using the ResNet backbone network; Capture target details at different scales in high-level semantic features based on attention mechanisms.

3. The remote sensing image semantic segmentation method according to claim 2, characterized in that, Dimensionality reduction involves sequentially performing batch normalization, ReLU activation function, and max pooling layer on local features.

4. The remote sensing image semantic segmentation method according to claim 2, characterized in that, This involves capturing target details at different scales from high-level semantic features using an attention mechanism, including: Generate multi-scale features based on advanced semantic features; Attention weights for features at each scale are determined using an attention network; Multi-scale features are fused using an interactive weighting method.

5. The remote sensing image semantic segmentation method according to claim 1, characterized in that, Pixel-level segmentation based on enhanced features includes: By using multiple dilated convolutional layers, features of different scales are generated based on the enhanced features to improve the receptive field of the enhanced features; Fuse features at different scales; Upsampling of the fused features The upsampled features are classified using convolutional layers to obtain the category label for each pixel.

6. The remote sensing image semantic segmentation method according to claim 5, characterized in that, After passing through the dilated convolutional layer, activation functions and batch normalization are added sequentially.

7. The remote sensing image semantic segmentation method according to claim 5, characterized in that, The Softmax function is used to normalize each pixel to generate the probability distribution for each category, thus obtaining the segmentation result.

Images